Interfacing Geostatistics and GIS

Jürgen Pilz (Ed.)

Interfacing Geostatistics
and GIS

123
Editor
Prof. Dr. Jürgen Pilz
Universität Klagenfurt
Institut für Statistik
Universitätsstr. 65-67
9020 Klagenfurt
Austria
juergen.pilz@uni-klu.ac.at

ISBN: 978-3-540-33235-0 e-ISBN: 978-3-540-33236-7

DOI 10.1007/978-3-540-33236-7

Library of Congress Control Number: 2008938163


c Springer-Verlag Berlin Heidelberg 2009

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.

The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.

Cover design: deblik, Berlin

Printed on acid-free paper

9 8 7 6 5 4 3 2 1

springer.com
Preface

Most of the papers contained in this volume grew out of presentations given
at the International Workshop StatGIS03 – Interfacing Geostatistics, GIS and
Spatial Data Bases, which was held in Pörtschach, Austria, Sept. 29–Oct. 1,
2003, and ensuing discussions, afterwards. Some of the papers are new and
have not been given at the conference. Therefore, most of the papers should
not be considered as conference proceedings in its original sense but rather
more as self-contained and actual contributions to the theme of the conference,
the interfacing between geostatistics, geoinformation systems and spatial data
base management.
Although some progress has been made toward interfacing, we still feel
that there is only little overlap between the different communities. The present
volume is intended to provide a bridge between specialists working in different
areas. According to the topics of the above mentioned workshop, this volume
has been divided into three parts:
Part I starts with general aspects of geostatistical model building
(Pebesma) and then new methodological developments in geostatics are pre-
sented, in particular this pertains to neural networks (Parkin and Kanevski),
Gibbs fields as used in statistical physics (Hristopulos). Furthermore, new de-
velopments in Bayesian spatial interpolation with skewed heavy-tailed data
and new classification methods based on wavelets (Hofer et al.) and support
vector machines (Chaouch et al.) are presented.
Part II contains applications of geostatistics to such diverse areas as
geodetic network modelling (Čepek and Pytel), land use policy (Müller and
Munroe), precipitation fields modelling (Ahrens), air pollution monitoring
(Shibli and Dubois), soil characterization (Sunila and Horttanainen) and soil
contamination modelling (Palaseanu-Lovejoy et al.). But also new application
areas such as traffic modelling (Braxmeier et al.) and spatial modelling of
entrepreneurship data (Breitenecker et al.) are touched.
Part III is devoted to the issues of the integration of different types of
information systems. The paper by Krivoruchko and Bivand deals with the
problems of interfacing GIS and spatial statistics software systems, from the
VI Preface

perspecitves of users and developers. An application of GIS in connection with

spatial analysis of remotely sensed agricultural data is reported by Sambrakos
and Tsiligiridis.
The advance of the integration efforts with regard to epidemiological in-
formation systems is documented in the paper by Gómez-Rubio et al., similar
issues arising in biostatistical applications such as acute coronary heart di-
sease are considered by Niyonsenga et al. Non-standard developments and
applications of temporal GIS are reported by G. and N. Andrienko, whereas
A. Gebhardt reports on severeal possibilities for combining open-source (spa-
tial) databases and GIS.
Finally we would like to thank the Springer Publishing House for offering
us the opportunity to publish the present material on the important aspects
of integration and combination of spatial modelling branches which previously
had developed in a more or less isolated manner. We are looking forward to
reporting on further progress in this direction very soon.

Klagenfurt, Spring 2008 Jürgen Pilz

Contents

Part I Geostatistical Modeling Aspects and New (Geo)statistical

Tools

How We Build Geostatistical Models and Deal with Their

Output
Edzer J. Pebesma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Spartan Random Fields: Smoothness Properties of Gaussian

Densities and Definition of Certain Non-Gaussian Models
Dionissios T. Hristopulos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Bayesian Trans-Gaussian Kriging with Log-Log Transformed

Skew Data
Gunter Spöck, Hannes Kazianka, and Jürgen Pilz . . . . . . . . . . . . . . . . . . . . 29

Kriging and Splines: Theoretical Approach to Linking Spatial

Prediction Methods
Philipp Pluch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

ANNEX Model: Artificial Neural Networks with External

Drift Environmental Data Mapping
R. Parkin and M. Kanevski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Regional Classification of Indoor Radon Data with Support

Vector Machines and Geostatistical Tools
A. Chaouch, M. Kanevski, M. Maignan, A. Pozdnoukhov, J. Rodriguez,
and G. Piller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Daubechies Wavelets for Identification of Rock Variants from

IR Spectra
Vera Hofer, Jürgen Pilz, and Thorgeir S. Helgason . . . . . . . . . . . . . . . . . . . 79
VIII Contents

Part II Geostatistical Applications

Simulating the Effects of Rural Development Policies on

Land Use: Evidence from Spatially Explicit Modeling in the
Central Highlands of Vietnam
Daniel Müller and Darla K. Munroe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Kriged Road-Traffic Maps

Hans Braxmeier, Volker Schmidt, and Evgeny Spodarev . . . . . . . . . . . . . . . 105

On Evaluation of Precipitation Fields with Rain Station Data

Bodo Ahrens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Robust Spatial Correlation Analysis of the ETEX-1 Tracer

Data
Syed Shibli and Gregoire Dubois . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

Fuzzy Model of Soil Polygons for Managing the Imprecision

Rangsima Sunila and Pekka Horttanainen . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

Mapping the Contaminant Legacy of a Coking Plant, The

Avenue, Chesterfield, UK
Monica Palaseanu-Lovejoy, Ian Douglas, and Robert Barr . . . . . . . . . . . . . 161

A Note on Numerical Solutions of Least Squares Adjustment

in GNU Project Gama
Aleš Čepek and Jan Pytel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

Presentation of Entrepreneurship Data and Aspects of Spatial

Modeling
Robert J. Breitenecker, Jürgen Pilz, and Erich J. Schwarz . . . . . . . . . . . . . 189

Part III Integrated Information Systems: Combining

(Geo)Statistics, GIS and RDBMS

GIS, Users, Developers, and Spatial Statistics: On Monarchs

and Their Clothing
Konstantin Krivoruchko and Roger Bivand . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Reassignment of the Farm Structure Statistical Data Using

GIS and Spatialisation of the Results Based on Remotely
Sensed Data
M. Sambrakos and T. Tsiligiridis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Epidemiological Information Systems

V. Gómez-Rubio, J. Ferrándiz-Ferragud, and A. López-Quílez . . . . . . . . . . 235
 Contents IX

Geomatics, Epidemiology and BioStatistics: An Application

to Acute Coronary Syndrome
Théophile Niyonsenga, Josiane Courteau, Charmaine Dean, Abbas
Hemiari, Goze Bénié, and Alain Vanasse . . . . . . . . . . . . . . . . . . . . . . . . . . . 249

Interactive Cumulative Curves for Exploratory Classification

Maps
Gennady Andrienko and Natalia Andrienko . . . . . . . . . . . . . . . . . . . . . . . . . 261

Combining REmbeddedPostgres and PostGIS

Albrecht Gebhardt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
List of Contributors

Bodo Ahrens
Institut für Meteorologie und Geophysik, Universität Wien, Wien, Austria
Bodo.Ahrens@univie.ac.at

Gennady Andrienko
Fraunhofer Institute AIS Schloss Birlinghoven, Sankt Augustin, Germany
gennady.andrienko@ais.fraunhofer.de

Natalia Andrienko
Fraunhofer Institute AIS Schloss Birlinghoven, Sankt Augustin, Germany

Robert Barr
School of Geography, The University of Manchester, Manchester, UK

Goze Bénié
Geography and Remote-Sensing Department, Université de Sherbrooke,
Sherbrooke, QC, Canada
Goze.Bertin.Benie@USherbrooke.ca

Hans Braxmeier
Department of Applied Information Processing, University of Ulm, Ulm,
Germany
hans.braxmeier@uni-ulm.de

Robert J. Breitenecker
Department of Innovation Management and Entrepreneurship, University of
Klagenfurt, Klagenfurt, Austria
robert.breitenecker@uni-klu.ac.at

Roger Bivand
Norges Handelshøyskole, Bergen, Norway
Roger.Bivand@nhh.no
XII List of Contributors

Aleš Cepek
Faculty of Civil Engineering, CTU Prague, Prague, Czech Republic
cepek@fsv.cvut.cz

A. Chaouch
Institute of Mineralogy and Geochemistry, University of Lausanne, Lausanne,
Switzerland
aziz.chaouch@etu.unil.ch

Josiane Courteau
PRIMUS group, Clinical Research Center, Centre Hospitalier Universitaire de
Sherbrooke, Sherbrooke, QC, Canada
josiane.courteau@usherbrooke.ca

Charmaine Dean
Statistics and Actuarial Science, Simon-Fraser University, Vancouver, BC,
Canada
dean@stat.sfu.ca

Ian Douglas
School of Geography, The University of Manchester, Manchester, UK

Gregoire Dubois
Radioactivity Environmental Monitoring, Institute for the Environment and
Sustainability, Joint Research Centre, European Commission, Ispra, Italy
gregoire.dubois@jrc.it

J. Ferrándiz
Dpto. Estadística e Investigación Operativa, Universitat de València, València,
Spain
Juan.Ferrandiz@uv.es

Albrecht Gebhardt
Departement of Statistics, University of Klagenfurt, Klagenfurt, Austria
agebhard@uni-klu.ac.at

V. Gómez-Rubio
Dpto. Matemáticas, Universidad de Castilla-La Mancha, Albacete, Spain
Virgilio.Gomez@uclm.es

Thorgeir S. Helgason
Petromodel Ltd, Reykjavik, Iceland
thorgeir@petromodel.is
 List of Contributors XIII

Abbas Hemiari
PRIMUS group, Clinical Research Center, Centre Hospitalier Universitaire de
Sherbrooke, Sherbrooke, QC, Canada

Dionissios T. Hristopulos
Department of Mineral Resources Engineering, Technical University of Crete,
Crete, Greece
dionisi@mred.tuc.gr

Vera Hofer
Department of Statistics and Operations Research, Karl-Franzens University
Graz, Graz, Austria
vera.hofer@uni-graz.at

Pekka Horttanainen
Department of Surveying, Institute of Cartography and Geoinformatics,
Helsinki University of Technology (HUT), Espoo, Finland

Mikhail Kanevski
Institute of Geomatics and Analysis of Risk, University of Lausanne,
Switzerland
Mikhail.Kanevski@unil.ch

Hannes Kazianka
Department of Statistics, University of Klagenfurt, Klagenfurt, Austria
hannes.kazianka@uni-klu.ac.at

Konstantin Krivoruchko
Environmental Systems Research Institute Redlands, Redlands, CA, USA
kkrivoruchko@esri.com

A. López
Dpto. Estadística e Investigación Operativa, Universitat de València, València,
Spain
Antonio.Lopez@uv.es

M. Maignan
Institute of Mineralogy and Geochemistry, University of Lausanne, Lausanne,
Switzerland

Darla K. Munroe
Department of Geography, The Ohio State University, Columbus, OH, USA
munroe.9@osu.edu
XIV List of Contributors

Daniel Müller
Leibniz Institute of Agricultural Development in Central and Eastern Europe,
Halle (Saale)
mueller@iamo.de

Théophile Niyonsenga
Epidemiology and Biostatistics, Robert Stempel School of Public Health,
Florida International University (FIU), Miami, FL (USA)
theophile.niyonsenga@fiu.edu

Monica Palaseanu-Lovejoy
School of Geography, The University of Manchester, Manchester, UK
monica.palaseanu-lovejoy@stud.man.ac.uc

R. Parkin
Institute of Nuclear Safety (IBRAE), Moscow, Russia
park@ibrae.ac.ru

Edzer J. Pebesma
Institute for Geoinformatics (ifgi), University of Münster, Münster, Germany
edzer.pebesma@uni-muenster.de

Jürgen Pilz
Department of Statistics, University of Klagenfurt, Klagenfurt, Austria
juergen.pilz@uni-klu.ac.at

G. Piller
Swiss Federal Office of Public Health (OFSP), Bern, Switzerland

Philipp Pluch
Energy and Petroleum Resources Services GmbH, Vienna, Austria
ppluch@menpet.at

Alexei Pozdnoukhov
Institute of Geomatics and Analysis of Risk, University of Lausanne,
Switzerland
Alexei.Pozdnoukhov@unil.ch

Jan Pytel
Faculty of Civil Engineering, CTU Prague, Prague, Czech Republic
pytel@fsv.cvut.cz

J. Rodriguez
Swiss Federal Office of Public Health (OFSP), Bern, Switzerland
 List of Contributors XV

M. Sambrakos
InfoLab, Agricultural University of Athens, Athens, Greece
marios@aua.gr

Volker Schmidt
Department of Stochastics, University of Ulm, Ulm, Germany
volker.schmidt@uni-ulm.de

Erich J. Schwarz
Department of Innovation Management and Entrepreneurship, University of
Klagenfurt, Klagenfurt, Austria
erich.schwarz@uni-klu.ac.at

Syed Shibli
Landmark Eame Ltd, Aberdeen, Scotland, UK
syed.shibli@googlemail.com

Evgeny Spodarev
Department of Stochastics, University of Ulm, Ulm, Germany
evgeny.spodarev@uni-ulm.de

Gunter Spöck
Department of Statistics, University of Klagenfurt, Klagenfurt, Austria
gunter.spoeck@uni-klu.ac.at

Rangsima Sunila
Department of Surveying, Institute of Cartography and Geoinformatics,
Helsinki University of Technology (HUT), Espoo, Finland
rangsima.sunila@hut.fi

T. Tsiligiridis
InfoLab, Agricultural University of Athens, Athens, Greece
tsili@aua.gr

Alain Vanasse
Family Medicine Department, Université de Sherbrooke, Sherbrooke (QC),
Canada
alain.vanasse@usherbrooke.ca
How We Build Geostatistical Models and Deal
with Their Output

Edzer J. Pebesma

Institute for Geoinformatics (ifgi), University of Münster, Münster, Germany

edzer.pebesma@uni-muenster.de

1 Introduction
Multivariable linear geostatistical models extend multivariable, multiple lin-
ear regression models for cases where observations are spatially correlated,
enabling the prediction of values at unobserved locations. In multiple linear
regression, the goal is to explain a large part of the observed variability by a set
of regressors and possibly their interactions. The more variability explained,
the better the prediction. Geostatistics extends this by looking at spatial cor-
relation in the residual variability: at a prediction location a nearby residual
may carry predictive value to the residual value at that location. However,
much of the geostatistical curriculum (literature and software) does not start
off by attempting to explain variability in the observed variables, but rather
starts at describing and modelling the observed variability after assuming the
trend is a spatially constant, thereby potentially ignoring available informative
predictors.
Extensions are universal kriging and external drift kriging [5]. In universal
kriging, only coordinates are used to explain variability. It is of no surprise
that this has not become popular, as coordinates hardly ever carry a phys-
ical relation to the observed variable, and may lead to extreme, unrealistic
extrapolations near the border of the domain. External drift kriging does ex-
tend kriging interpolation to the linear using a linear regression model with
an external variable for the trend, but it is most often explained as being the
case where only a single predictor (external drift variable) is present. In the
following, we will not distinguish between universal kriging and external drift
kriging, as the procedures are equivalent [7].
Multivariable prediction has been known for a long time, and has been
applied especially when using one or more secondary variables to predict a
primary variable. The general case where m variables are used to predict m
variables, m being larger than say 3, is found seldom in literature. The reasons
for this do not have a statistical ground, but rather stem from the fact that
4 E.J. Pebesma

it is considered a burden to do so. Although algorithms exist to automate the

modelling of many direct and cross variograms [10, 11, 24], easily accessible
software implementations have been lacking.
This paper discusses several limitations found in statistical software (mostly
R) and Geographic Information Systems (GIS; [4]) software, with respect
to flexible modelling of the multivariable linear geostatistical model. In a
case study we show how to apply this model, using the recently developed
gstat package for R and S-PLUS [20]. The discussion closes with a highly
personalized view on the practice, limitations and possibilities of applied
geostatistics.

2 Geostatistical Prediction
In geostatistics, the variability in an observed variable Z, taken at location
si , is assumed to be the sum of a fixed trend and a random residual: Z(si ) =
m(si ) + e(si ), and the trend is modelled as a linear combination of p unknown
coefficients and p known predictors Xj (s):


p
Z(s) = Xj (s)βp + e(s) = X(s)β + e(s), s ∈ {s1 , ...sn }
j=1

with X1 (s) ≡ 1 when β1 is the intercept, and X(s) the n × p matrix with
predictors. Given knowledge of the (spatial) covariance of e, V = Cov(e), and
knowledge of the covariance between e(s) and e(s0 ), v = (Cov(e(s1 ), e(s0 )), ...,
Cov(e(sn ), e(s0 ))) , the best linear unbiased (or kriging) predictor is
obtained by
Ẑ(s0 ) = x(s0 )β̂ + v  V −1 (Z(s) − X(s)β̂)
where x(s0 ) contains the known predictors location s0 , and with

β̂ = (β̂0 , ..., β̂1 ) = (X(s) V −1 X(s))−1 X(s) V −1 Z(s)

the generalized least squares estimate of β. The prediction variance of Ẑ(s0 ) is

σ 2 (s0 ) = σ02 − v  V −1 v + η  (X(s) V −1 X(s))−1 η

with σ02 = Var(e(s0 )) and η  = (x(s0 )−v  V −1 X(s)). These equations reduce to
traditional multiple regression prediction if v = 0 and V is diagonal (weighted
least squares) or V = σ02 I (ordinary least squares) [9], and they reduce to
ordinary kriging if the regression only contains an intercept (i.e., X(s) and
x(s0 ) only contain a single column of ones).
When multiple, spatially cross correlated variables are present, they may
be used in a multivariable prediction [23], not only to enhance the predictions
 How We Build Geostatistical Models 5

of each individual variable, but also to assess the prediction error covariances
for all pairs of variables.
In practice, the application of these equations is often restricted to the
data available in a local neighbourhood around s0 . The reasons for this may be
computational, to avoid solving kriging systems with a very large (n  1000)
covariance matrix, or statistical, to reduce the assumption of globally constant
regression coefficients to the more flexible assumption of locally constant re-
gression coefficients.
Another specialty on the geostatistics menu is called change of support:
rather than predicting values Z(s0 ) for point  locations s0 , we may want to
predict the integral (mean) of Z(B0 ) = |B10 | u∈B0 Z(u)du, with |B0 | the area
or volume of integration. Block average values can be obtained by averaging
point kriging values, but block average prediction errors can not; for this
we need block kriging [5, 13]. The reason for wanting block kriging is that
highly detailed spatial predictions may not be wanted, and that block kriging
prediction errors are always smaller then point kriging prediction errors.
In addition to prediction, it may be useful to simulate realisations of ran-
dom fields Z(s) that honour both the observed data, the regression relations,
and the spatial correlation [19]. Abrahamsen and Espen Benth [2] describe an
algorithm where the simulation equivalent of universal (external drift) kriging
is given.

3 Case Study: Sea Floor Sediment Pollution

This case study analyses spatio-temporal data on sea floor sediment pollution,
collected from 1986 to 2000 in the Dutch part of the North Sea [17]. The data
set was provided by the Dutch National institute for Coastal and Marine Man-
agement (RIKZ). The variables measured comprise heavy metals, as well as
organic compounds like furans and polychlorinated biphenyls (PCB’s). Here,
we will look into a single PCB, named PCB138. Table 1 summarizes some of
the characteristics of the data. The programme initially aimed at monitoring
approximately five-yearly (the “main” monitoring years, ’86, ’91, ’96, ’00), over
which the number of samples range from 31 to 49; other measurements result
from additional sampling programs.
The main PCB138 source is sediment originating from the river Rhine; this
sediment is carried North bound along the coast by the local North Sea flow
direction. One of the main questions is whether temporal trends are present in
the data, how large the trend is, and to which extent spatially differentiated
trends can be inferred.
6 E.J. Pebesma

Table 1. PCB138 (μg/kg dry matter) data summaries; years marked with a ∗ are
the regular monitoring years, other years result from additional sampling programs
year 1986∗ 1987 1989 1991∗ 1993 1996∗ 2000∗ All
mean 7.29 8.39 4.08 3.70 1.03 1.58 1.27 4.20
median 6.90 7.50 2.65 3.05 0.775 1.40 0.90 2.85
max 21.1 19.7 12.3 13.1 2.7 4.9 3.3 21.1
min 1.60 2.10 1.00 0.70 0.25 0.20 0.20 0.2
n 45 29 14 42 6 49 31 216

3.1 Exploratory Data Analysis

Figure 1 shows a bubble plot with the spatial locations of the measurement
sites, per year. Symbol size is proportional to log-concentration, which is the
natural scale to view such variables. The summary statistics of Table 1 already
reveal that PCB138 decreases over time. Figure 1 furthermore shows that
high concentrations appear close to the coast. Simply looking at how PCB138

1986 1987 1989 1991

y−coordinate

1993 1996 2000

0.2
1
2
5
10
20

x−coordinate

Fig. 1. Maps of measured PCB138 concentrations in Dutch North Sea sediment

samples, for each year sampled from 1986 to 2000. The area shown has UTM31
x-coordinates ranging from 464000 m. to 739000 m. and y-coordinates from 5696500
m. to 6131500 m.
 How We Build Geostatistical Models 7

concentrations decrease with time may not be appropriate because the spa-
tial locations of sampling vary from year to year, and the sampling pattern
is not random. The sampling pattern (Fig. 1) is directed towards transects
perpendicular to the Dutch coast (the direction of the main gradient), and
seems clustered; many short distances are present.

3.2 Trend

10 20 30 40 10 20 30 40
1986 1987 1989 1991
20
10
5

2
1
.5
PCB138

.2
1993 1996 2000
20
10
5

2
1
.5

.2
10 20 30 40 10 20 30 40
water depth

Fig. 2. PCB138 concentration as a function of sea water depth, for each of the
measured years

Figure 2 shows spatial trends of (log) PCB138 concentrations as a function

of water depth, for each monitoring year. The figure suggests that on the log-
scale, the decrease with depth is more or less constant, and that the origin
(intercept) of the varies per year; the fitted line reflects this model. In terms
of a regression model, we could express this as
Zt (s) = β0,t + β1 + X1 (s) + e(s) (1)
with Zt (s) the log-transformed PCB138 measurement at year t and location
s, β0,t the intercept for year t, with X1 (s) the depth value at location s and
8 E.J. Pebesma

e(s) the residual. The regression model explains 77% of the variability in log-
PCB138. Under the assumption of independent data, (i) all terms were highly
significant (p < .001), and (ii) an interaction between year and depth (i.e., a
year-dependent regression slope with depth) was not significant. Clearly, these
significance assertions are of little value, as the data vary spatially, and we
may assume that they are spatially correlated.

3.3 Residual Spatial Correlation and Temporal Cross Correlation

Each of the monitoring years years has too few measurements to model a
residual variogram (Table 1). For that reason, the residual information of
all years was merged. Simply merging all residuals leads to the variogram
in the first panel of Fig. 3. This would be a valid approach if the residual
spatial pattern were constant over time. Constructing a pooled variogram
by only considering point pairs with both measurements in the same year
(rest of Fig. 3) shows that the hypothesis of a temporary constant spatial
pattern is not valid: a much stronger spatial correlation is revealed under the
hypothesis that only the spatial variability (variogram) is persistent over time.
On single residual variogram model was fitted for all within-year residuals,
γ(h) = 0.08δ(h + 0.224(1 − exp(−h/17247)) with δ(h) = 0 if h = 0 and
δ(h) = 1 if h > 0 (last panel of Fig. 3).

0 50000 100000 150000

ignoring year pooled within−year
154
0.4

183
218
235 177
75 0.3
151
106 144
151264 180
261 116
0.2

177
0.1
semivariance

0.0
pooled, boundaries adjusted pooled, with fitted model
154
0.4

183
218
235 177
0.3 75
151
106 144
151264 180
261 116
0.2
65
55
21
0.1
36

0.0
0 50000 100000 150000
distance

Fig. 3. Different approaches to modelling the sampling variogram for the residu-
als of the linear regression lines in Fig. 2; top left: residual variogram; top right:
pooled, within-year residual variogram; bottom left: short-distance variogram values
are split into smaller distance intervals; bottom right: a model fitted to the bottom
left variogram. Numbers reflect the point pairs that contribute to sample variogram
estimates
 How We Build Geostatistical Models 9

Before looking at residual cross correlation and temporal change, we will

restrict ourselves to the four “main” monitoring years, 1986, 1991, 1996 and
2000, because the other years have too few measurements for analyzing cross
variograms. The direct variograms for years in Fig. 4 (labeled “1986”, “1991”,...)
show indeed that each of them carry insufficient information for fitting a
separate model, and the fitted model used is the same for each, i.e. that
of Fig. 3. The cross variograms for pairs of years (labeled “1986.1991”, etc.)
are even more noisy, and attempts to automatically fit a linear model of
coregionalization [11] failed.
In order to proceed, we need a coregionalization model that requires less
parameters, such as the intrinsic correlation model [10]. This model only re-
quires a correlation between two years. Finding these correlations from the
cross variograms seems not so easy, partly because pairs of years do not share
common sample locations (Fig. 1). Therefore, we used the following approach:
for a pair of years, say 1986 and 1991, for each of the measurements in 1986
the (spatially) nearest measurement of 1991 was found, and from the pairs

1986
0.0 0.10.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.2 0.4 0.6

1986.1991 1991
0.4
0.2
semivariance

0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0

1986.1996 1991.1996 1996

0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3

1986.2000 1991.2000 1996.2000 2000

0.0 0.2 0.4 0.6
0.4
0.2
0.0

0 40000 80000 0 40000 80000

distance

Fig. 4. Sample direct and cross variograms for the four main measurement years,
and fitted Intrinsic Correlation model. The direct variogram model is that of 3, each
cross variogram is scaled down by a factor equal to the pointwise correlations of the
pair of years; pointwise correlations are approximated by joining spatially nearest
neighbours to form data pairs
10 E.J. Pebesma

thus found, the correlation coefficient was calculated. Next, the two years were
reversed, and a second correlation coefficient was calculated. The average of
these two correlations was used to model the cross variograms of Fig. 4. Be-
cause spatially nearest neighbours of year y were used to approximate the
measured value at a certain location in year x, the estimated correlations
must underestimate the true correlations.

3.4 Spatio-Temporal Prediction

Spatio-temporal prediction under model (1), given the data for each of the
four “main” years and given the direct variograms and the cross variograms
of Fig. 4 is simply a matter of universal cokriging. Universal cokriging yields
spatial predictions for each of the four years, shown in Fig. 5, and yields
in addition spatial prediction error variances for each of the four years, and
spatial prediction error covariances for prediction errors of all pairs of years.
Spatially differentiated estimates of trends can be assessed by combining the
yearly predictions and prediction error (co)variances.

−20
1986.pred 1991.pred 1996.pred 2000.pred
−10
−5

−2
y

−1

−0.5

−0.2
−0.1
x

Fig. 5. Cokriging predictions for the four main measurement years

3.5 Contrasts and Trends

Cokriging basically yields for each location s0 a vector with predictions, which
in our case could be called y(s0 ) = (y86 (s0 ), y91 (s0 ), y96 (s0 ), y00 (s0 )) , along
with the prediction error covariance matrix Cov(y(s0 )). Given this vector we
can calculate for each location s0 a contrast

C(s0 ) = λ y(s0 )

which has prediction error variance λ Cov(y(s0 ))λ. Examples of potentially

interesting contrasts are
• prediction for a single year, e.g. 1991: λ = (0, 1, 0, 0)
• prediction of the four-year mean: λ = ( 14 , 14 , 14 , 14 )
 How We Build Geostatistical Models 11

• prediction of the difference between the means of 1986 and 1991 versus
the mean of 1996 and 2000: λ = (− 12 , − 12 , 12 , 12 )
• prediction of the average yearly increase: λ = (−0.065, −0.02, 0.025, 0.061)
The weights of the latter contrast, which is obviously of major interest when
we want to assess spatially differentiated trends, are obtained as follows. Trend
estimation uses linear regression for predicting concentrations from years by
y(s0 ) = β0 (s0 )+β1 (s0 )t+e = Xβ(s0 )+e. The ordinary least squares estimate
of β is (X  X)−1 X  y. The contrast coefficients that estimate β1 (s0 ) are in the
second row of (X  X)−1 X  , with
⎡ ⎤
1 1986
⎢1 1991⎥
X=⎢ ⎥
⎣1 1996⎦ .
1 2000

Figure 6 shows the predicted trends, as well as the trend predictions di-
vided by their own prediction standard error. Clearly, the majority of the area

log PCB138: slope estimate

−−0.05

−−0.1
y

−−0.15

−−0.2

−−0.25

Fig. 6. Predicted trends (in ppm/year) for each point location (left); and relative
predicted trends, expressed as fraction of their own prediction standard error (right).
On the right, under the assumed model, relative predicted trends smaller than -2
tentatively indicate trends that cannot be attributed to pure chance (i.e. that are
significant)
12 E.J. Pebesma

shows a decrease in PCB138 concentration of a magnitude larger than twice

its standard error, tentatively indicating significant trends. Of course, these
assertions are only valid under the assumptions that were made along the way,
including (i) the model for the trend, shown in Fig. 2, (ii) the modeled vari-
ograms and cross variograms (Figs. 3 and 4), and (iii) that on the log-scale, a
second order stationary residual for PCB138 is a reasonable model. Stronger
positive correlations between the years result in smaller standard errors for
trends, so the underestimation of true between-year (spatial) correlations that
we mentioned in Sect. 3.3 results in conservative assessment of “significance”
of trends in the right part of Fig. 6.

4 Shortcomings
This case study shows some of the capabilities of the gstat package for R
[12], or for S-PLUS, which extensively uses the graphics capabilities of the
Trellis/lattice graphics package [6]. The gstat program [18] or R package
[20] offers flexibility with respect to trend modelling, multivariate variogram
modelling, multivariate prediction and simulation, change of support and pre-
diction in a local neighbourhood. Additional features that it does not address
are e.g. flexible three-dimensional anisotropic variogram modelling, Bayesian
handling of uncertainty in variogram model coefficients [22], and multivari-
able space time modelling in continuous time (i.e., where time is a dimension
rather than a discrete variable as in the case study of this paper). These fea-
tures are available in either other R packages or other environments, where
they are potentially hard to combine with the features offered by gstat.

5 Discussion
There may be various historical reasons for not starting off with a linear re-
gression model for the modelling the trend of spatial data. First, geostatistics
was developed by mining engineers, who usually did not have useful pre-
dictor variables available, other than spatial coordinates of observations and
prediction locations. Second, the sample variogram from estimated residuals
is biased because of the estimation of the trend, which would need the true
variogram—a chicken-and-egg problem raised by Armstrong [1] but settled by
Kitanidis [15]. Third, leading authors have suggested that predictors are not
needed [14], but that observations themselves carry enough information. This
is indeed the case when observations are abundant and not too noisy—the
case where even (geo)statistics could be ignored altogether and any contour-
ing algorithm would suffice. All these factors lead to situation where much of
the available geostatistical software packages (GSLIB, [8]; GsTL, [21]; ArcGIS
Geostatistical Analyst, [16]; Isatis, http://www.geovariances.fr) have lit-
tle flexibility with respect to modelling external drifts with multiple linear
regression models.
 How We Build Geostatistical Models 13

In the authors opinion, R (or S-PLUS) provides a very rich environment

for building geostatistical models, mainly because all basic building blocks are
there: linear and non-linear regression functions, regression diagnostic plotting
functions, and the Trellis graphics [6] tools—which we consider indispensable
for multivariable geostatistical exploration, analysis, modelling and predic-
tion. In addition, functions are present that allow straightforward calculation
of contrasts, or of aggregating quantities over large sets of Monte Carlo simula-
tions; these cases are usually much harder to deal with in a GIS environment.
A drawback or R (and S-PLUS) is that it still has no standard way of dealing
with spatial data, although several authors now work together in this direction
[3]. It will be a long way before R can touch upon commercial GIS’ capabil-
ities regarding visualisation of maps, but the basic building blocks are there.
In terms of analytical capabilities R has beaten commercial GIS by far.
References

1. Armstrong, M. (1984) Improving the estimation and modelling of the variogram.

In: G. Verly et al, (Eds.), Geostatistics for Natural Resource Characterization,
pp 1–20, Reidel, Dordrecht, Holland.
2. Abrahamsen, P. and F.E. Benth (2001) Kriging with inequality constraints.
Mathematical Geology 33 (6), 719–744.
3. Bivand, R.S. (2003) Approaches to classes for spatial data in R. In: K. Hornik
and F. Leisch (Eds.), Proceedings of the 3rd International Workshop on Dis-
tributed Statistical Computing (DSC 2003) March 20–22, Vienna, Austria.
http://www.ci.tuwien.ac.at/Conferences/DSC-2003/
4. Burrough, P.A. and R.A. McDonnell (1998) Principles of geographical informa-
tion systems. Oxford University Press, New York. 431 pp.
5. Chilès, J-P. and P. Delfiner (1999) Geostatistics, modeling spatial uncertainty.
Wiley, New York.
6. Cleveland, W.S. (1993) Visualizing data. Hobart Press, Summit, New Jersey.
7. Cressie, N.A.C. (1993) Statistics for Spatial Data, revised edition. Wiley,
New York.
8. Deutsch, C.V. and A.G. Journel (1990) GSLIB geostatistical software library
and user’s guide, second edition. Oxford University Press, New York.
9. Draper, N.R. and H. Smith (1981) Applied regression analysis, second edition.
Wiley, New York.
10. Goovaerts, P. (1997) Geostatistics for natural resources evaluation. Oxford Uni-
versity Press, New York.
11. Goulard, M. and M. Voltz (1992) Linear coregionalization model: tools for es-
timation and choice of cross-variogram matrix. Mathematical Geology 24 (3),
269–286.
12. Ihaka, R. and R. Gentleman (1996) R: a language for data analysis and graphics.
Journal of Computational and Graphical Statistics 5(3) 299–314.
13. Journel, A.G. and Ch.J. Huijbregts (1978) Mining geostatistics. Academic Press,
London, 600 pp.
14. Journel, A.G. and M.E. Rossi (1989) When do we need a trend model in kriging?
Mathematical Geology 21 (7), 715–739.
15. Kitanidis, P.K. (1993) Generalized covariance functions in estimation. Mathe-
matical Geology 25 (5), 525–540.
 How We Build Geostatistical Models 15

16. Krivoruchko, K. (2007)...., this volume.

17. Laane, R.W.P.M., H.L.A. Sonneveldt, A.J. Van der Weyden, J.P.G. Loch and
G. Groeneveld (1999) Trends in the spatial and temporal distribution of metals
(Cd, Cu, Zn and Pb) and organic compounds (PCBs and PAHs) in Dutch coastal
zone sediments from 1981 to 1996: a model case study for Cd and PCBs. Journal
of Sea Research 41, 1–17.
18. Pebesma, E.J. and C.G. Wesseling (1998) Gstat, a program for geostatistical
modelling, prediction and simulation. Computers & Geosciences, 24 (1), 17–31.
19. Pebesma, E.J. and G.B.M. Heuvelink (1999) Latin hypercube sampling of Gaus-
sian random fields. Technometrics 41 (4), 303–312.
20. Pebesma, E.J. (2004) Multivariable geostatistics in S: the gstat package. Com-
puters & Geosciences 30, 683–691.
21. Remy, N. (2001) GsTL: the geostatistical template library in C++. MSc thesis,
Dept. of Petroleum Engineering, Stanford University.
22. Stein, M.L. (1999) Interpolation of spatial data: some theory for kriging.
Springer, New York, 247 pp.
23. Ver Hoef, J.M. and N.A.C. Cressie (1993) Multivariable spatial prediction.
Mathematical Geology, 25 (2), pp. 219–240.
24. Yao, T. and A.G. Journel (1997) Automatic modeling of (cross) covariance tables
using fast fourier transform. Mathematical Geology 30, 589–616.
 Part I

Geostatistical Modeling Aspects and New

(Geo)statistical Tools
Spartan Random Fields: Smoothness Properties
of Gaussian Densities and Definition of Certain
Non-Gaussian Models

Dionissios T. Hristopulos

Department of Mineral Resources Engineering, Technical University of Crete,

Crete, Greece
dionisi@mred.tuc.gr

1 Introduction

Spartan spatial random fields (SSRFs) were introduced in [10]. Certain mathe-
matical properties of SSRFs were presented, inference of the model parameters
from synthetic samples was investigated [10], and methods for the uncondi-
tional simulation of SSRFs were developed [11]. This research has focused on
the fluctuation component of the spatial variability, which is assumed to be
statistically homogeneous (stationary) and normally distributed. The proba-
bility density function (pdf) of Spartan fields is determined from an energy
functional H[Xλ (s)], according to the familiar in statistical physics expression
for the Gibbs distribution

fx [Xλ (s)] = Z −1 exp {−H[Xλ (s)]} . (1)

The constant Z (called partition function) is the pdf normalization factor ob-
tained by integrating exp (−H) over all degrees of freedom (i.e. states of the
SSRF). The subscript λ denotes the fluctuation resolution scale. The energy
functional determines the spatial variability by means of interactions between
neighboring locations. One can express the multivariate Gaussian pdf, typi-
cally used in classical geostatistics, in terms of the following energy functional

H[Xλ (s)]= 2 ds ds Xλ (s)c−1

1  
X (s, s )Xλ (s ), (2)

where cX (s, s ) is the centered covariance function; the latter needs to be

determined from the data for all pairs of points s and s , or (assuming sta-
tistical homogeneity) for all distance vectors s − s . In contrast, the energy
functional in Spartan models is determined from physically motivated interac-
tions between neighbors. The name ‘Spartan’ emphasizes that the number Np
of model parameters to be determined from the data is small. For example,
18 D.T. Hristopulos

in the fluctuation – gradient – curvature (FGC) model, the pdf involves three
main parameters: the scale factor η0 , the covariance shape parameter η1 , and
the correlation length ξ . Another factor that adds flexibility to the model is
the coarse-graining kernel that determines the fluctuation resolution λ [10].
As we show below, the resolution is directly related to smoothness properties
of the SSRF. In previous work [10, 11], we have used a kernel with a boxcar
spectral density that imposes a sharp cutoff in frequency (wavevector) space
at kc ∝ λ−1 . We have treated the cutoff frequency as a constant, but it is
also possible to consider it as an additional model parameter, in which case
Np = 4.
A practical implication of an interaction-based energy functional is that
the parameters of the model follow from simple sample constraints that do not
require the full calculation of two-point functions (e.g., correlation function,
variogram). This feature permits fast computation of the model parameters.
In addition, for general spatial distributions (e.g., irregular distribution of
sampling points, anisotropic spatial dependence with unknown a priori prin-
cipal directions), the parameter inference does not require various empirical
assumptions such as choice of lag classes, number of pairs per class, lag and
angle tolerance, etc. [7] used in the calculation of two-point functions. In the
case of SSRFs that model data distributed on irregular supports, the definition
of the interaction between ‘near neighbors’ is not uniquely defined. Determin-
ing the neighbor structure for irregular supports increases the computational
effort [10], but the model inference process is still quite fast. Methods for the
non-constrained simulation of SSRFs with Gaussian probability densities on
the square lattice (by filtering Gaussian random variables in Fourier space and
reconstructing the state in real space with the inverse FFT) and for irregular
supports (based on a random phase superposition of cosine modes with fre-
quency distribution modeled on the covariance spectral density), have been
presented in [11].

2 FGC Energy Functional

The energy functional involves the SSRF states (configurations) Xλ (s). For
notational simplicity, we will not use different symbols for the random field and
its states in the following. As hinted above, the energy functional is properly
defined for SSRFs Xλ (s) with an inherent scale parameter ‘λ’ that denotes
the spatial resolution of the fluctuations. At lower scales, the fluctuations are
coarse-grained. The fluctuation resolution scale is physically meaningful, since
it would be unreasonable to expect a model of fluctuations to be valid for all
length scales. In contrast with classical random field representations, which
do not have a built-in scale for a fluctuation cutoff, SSRFs provide an explicit
‘handle’ for this meaningful parameter. In practical situations, the fluctuation
resolution scale is linked to the measurement support scale and the sampling
density. In the case of numerical simulations, the lattice spacing provides
 Spartan Random Fields 19

a lower bound for λ. The fluctuation resolution can also exceed the lattice
spacing, to allow for smoother variations of the field. The general probability
density function of continuum FGC Spartan random fields (FGC-SSRF) in
IRd is determined from the following functional

1
Hfgc [Xλ ] = ds hfgc [Xλ (s); η1 , ξ] , (3)
2η0 ξ d
where η0 is a scale factor with dimensions [X]2 that determines the magnitude
of the overall variability of the SSRF, η1 is a covariance shape parameter
(dimensionless), ξ is the correlation length, and hfgc is the normalized (to
η0 = 1) local energy at the point s. In the case of a Gaussian FGC random
field with mean (not necessarily stationary) mX;λ (s) = E [Xλ (s)] and isotropic
spatial dependence of the fluctuations, the functional hfgc [Xλ (s); η1 , ξ] is given
by the following
2 2  2
hfgc [Xλ (s); η1 , ξ] = [χλ (s)] + η1 ξ 2 [∇χλ (s)] + ξ 4 ∇2 χλ (s) , (4)

where χλ (s) is the local fluctuation field. The functional (4) is permissible
if Bochner’s theorem [3] for the covariance function is satisfied. As shown in
[10], permissibility requires η1 > −2. The covariance spectral density follows
from the equation 2

Q̃λ (k) η0 ξ d
G̃x;λ (k) = (5)
1 + η1 (k ξ)2 + (k ξ)4
where Q̃λ (k) is the Fourier transform of the smoothing kernel. If the latter
is the boxcar filter with cutoff at kc , (5) leads to a band-limited spectral
density G̃x;λ (k). For negative values of η1 the spectral density develops a
sharp peak, and as η1 approaches the permissibility boundary value equal to
−2, the spectral density tends to become singular. For negative values of η1
the structure of the spectral density leads to a negative hole in the covariance
function in real space. If Q̃λ (k) has no directional dependence, the spectral
density depends on the magnitude but not the direction of the frequency
vector k. Thus, the covariance is an isotropic function of distance in this case.
On regular lattices, the FGC spectral density is obtained by replac-
ing the operators ∇ and ∇2 in the energy functional with the correspond-
ing finite differences. Then, the local energy becomes hfgc [Xλ (s); η1 , ξ] =
hfgc [χλ {U (s); η1 , ξ}], where U (s) = s ∪ nnb(s) is the local neighborhood set
that contains the point s and its nearest lattice neighbors, χλ {U (s)} is the
set of the SSRF values at the points in U (s), and hfgc [·] is a quadratic func-
tional of the SSRF states that defines interactions between the fluctuation
values χλ {U (s)}. For irregular spatial distributions, there are more than one
possibilities for modeling the interactions. One approach, explored in [10], is
to define a background lattice that covers the area of interest and to construct
interactions between the cells of the background lattice. If CB (s) denotes the
cell of the background lattice that includes the point s and nnb {CB (s)} is
20 D.T. Hristopulos

the set of nearest neighbors of the cell CB (s), the local neighborhood set in-
volves the sampled points that belong to the cell CB (s) and its neighbors, i.e.
U (s) = s ∈ CB (s) ∪ nnb {CB (s)}.

3 Model Inference

The problem of model inference from available data is a typical inverse prob-
lem. In order to determine the model parameters experimental constraints
need to be defined that capture the main features of the spatial variability
in the data. These constraints should then be related to the interactions in
the SSRF energy functional. The experimental constraints used in [10] for the
square lattice are motivated by the local ‘fluctuation energy measures’ S0 (s) =
d 2 d (i) (j)
χ2λ (s), S1 (s) = i=1 [∇i χλ (s)] , and S2 (s) = i,j=1 Δ2 [χλ (s)] Δ2 [χλ (s)],
(i)
where Δ2 denotes the centered second-order difference operator. The respec-
tive experimental constraints are then given by S0 (s) (sample variance), S1 (s)
(average square gradient) and S2 (s), where the bar denotes the sample aver-
age. The respective stochastic constraints are E [Sm (s)], m = 0, 1, 2 and they
can be expressed in terms of the covariance function. For the isotropic FGC
model, calculation of the stochastic constraints involves a one-dimensional
numerical integration over the magnitude of the frequency. Matching of the
stochastic and experimental constraints is formulated as an optimization prob-
lem in terms of a functional that measures the distance between the two sets
[10] of constraints. Minimization of the distance functional leads to a set of
optimal values η0∗ , η1∗ , ξ ∗ for the model parameters. Use of kc as a fourth param-
eter needs further investigation. It should be noted that constraint matching
is based on the ergodic assumption, and thus a working approximation of
ergodicity should be established for the fluctuation field.

4 Smoothness of FGC Spartan Random Fields

The probability density of the FGC-SSRF involves the first- and second-order
derivatives of the field’s states. This requires defining the energy functional in
a manner consistent with the existence of the derivatives. In general, for Gaus-
sian random fields [1, 15], the nth-order derivative ∂ n Xλ (s)/∂sn1 1 ...∂snd d exists
in the mean square sense if (i) the mean function mX;λ (s) is differentiable,
and (ii) the following derivative of the covariance function exists [1, 15]

∂ 2n Gx;λ (s, p)
n1 nd n1

nd , n = n1 + ... + nd . (6)
∂s1 ...∂sd ∂p1 ...∂pd s=p

Since the FGC covariance function is statistically homogeneous and isotropic,

the above condition simply requires the existence of the isotropic derivative
 Spartan Random Fields 21

of order 2n at zero pair separation distance, i.e. the existence of the following
quantity  2n 
(2n) d Gx;λ (r)
Gx;λ (0) = (−1)n (7)
dr2n r=0
Equation (7) is equivalent to the existence of the corresponding integral of
the covariance spectral density
2
 2n 
∞
d Gx;λ (r) Q̃λ (k) k d+2n−1
d
= η0 ξ Sd dk (8)
dr2n r=0 1 + η1 (k ξ)2 + (k ξ)4
0
 d/2

where Sd = d k̂ = 2π Γ (d/2) denotes the surface of the unit sphere in d
2

dimensions. Note that if Q̃λ (k) = 1 , i.e. in the absence of smoothing, the
above integral does not exist unless d + 2n < 4, which can be attained only for
d = 1 and n = 1. If the smoothing kernel has a sharp cutoff kc (band-limited
spectrum), the 2n-th order derivative is expressed in terms of the following
integral

kc ξ
d2n Gx;λ (r) −2n κd+2n−1
= η 0 ξ S d dκ . (9)
dr2n r=0 1 + η1 κ2 + κ4
0
The integral in 9 exists for all d and n. However, if the correlation length ξ
exceeds significantly the resolution scale, i.e. ξ >> λ and kc ξ >> 1, for κ >>
(2n)
1 the integrand behaves as κd+2n−5 . Then, it follows Gx;λ (0) = regular +
−2n d+2n−4
αd ξ (kc ξ) , where ‘regular’ represents the bounded contribution of
the integral, while for fixed ξ the remaining term increases fast with kc ξ. The
constant αd depends on the dimensionality of space. Hence, for d ≥ 2 the
(2n)
singular term in Gx;λ (0) leads to large values of the covariance derivatives
for n ≥ 1. In [10] we focused on the case kc ξ >> 1, which leads to ‘rough’
Spartan fields. Based on the above, the Gaussian FGC-SSRF can, at least
in principle, interpolate between very smooth Gaussian random fields (e.g.,
Gaussian covariance function) and non-differentiable ones (e.g., exponential,
spherical covariance functions). The ‘degree’ of smoothness depends on the
value of the combined parameter kc ξ. Hence, the FGC-SSRF in effect has four
parameters, η0 , η1 , kc , ξ , and the value of kc ξ, which controls the smoothness
of the model. This property of smoothness control is also shared by random
fields with Matérn class covariance functions [14].

5 Non-Gaussian Probability Densities

An issue of significant practical importance is the ability of geostatistical
models to capture fluctuations with non-Gaussian distributions. Such distri-
butions can be developed in the Spartan-Gibbs framework by adding suitable
22 D.T. Hristopulos

(higher than second order) interaction terms in the energy functional. An ex-
ample is the energy functional of the Landau model e.g. [10], which includes
non-Gaussian terms and exhibits a transition between exponential and power-
law spatial dependence of the covariance function. Geostatistical probability
density models provide sufficient flexibility for fitting various types of non-
Gaussian data. The approaches typically used in geostatistics for modeling
asymmetric distributions with higher-than-normal weight in their tails em-
ploy the logarithmic and the Box-Cox transforms. In the former approach,
the initial distribution is assumed to be approximately lognormal. The log-
arithmic mean mY (s) = E [log Xλ (s)] is first estimated. Then, the fluctu-
ations yλ (s) = log [Xλ (s)] − mY (s) follow the Gaussian distribution, and
they can be modeled by means of the FGC-SSRF normalized energy den-
sity hfgc [yλ (s); η1 , ξ]. If the logarithm of the random field deviates from the
Gaussian distribution, it is possible to modify the energy functional by adding
a non-Gaussian term as follows

Hng [yλ (s); η0 , η1 , ξ, q] = Hfgc [yλ (s); η0 , η1 , ξ] + δH [yλ (s); q] , (10)

where δH is the non-Gaussian term that involves a parameter vector q. For

simplicity, below we are going to express (10) as H = HG + δH, where H is
the entire energy functional, and HG = Hf gc is the Gaussian FGC contribu-
tion. Now one has to determine the entire set of model parameters η0 , η1 , ξ, q
(and possibly kc ) simultaneously from the sample. The deviation of the dis-
tribution from the Gaussian dependence is captured by means of additional
constraints, e.g. based on the local terms S3 (s) = yλ3 (s) and S4 (s) = yλ4 (s).
The corresponding distance functional then becomes
 2  2
0] S2 E[S0 ]
Φs [Xλ (s)] = 1 − SS1 E[S + 1 − + (11)
0 E[S1 ] S0 E[S2 ]
 2  2
3/2 2
1 − S3 E[S0 ] + 1 − S4 E[S0 ]
3/2
S0 E[S3 ] 2
S0 E[S3 ]

 3/2  2
The ratio S3 S0 represents the sample skewness coefficient, while S4 S0
the sample kurtosis coefficient. In the case of the Gaussian FGC-SSRF model,
the stochastic moments E [Sm ] , m = 0, 1, 2 (which are used in determining
the model parameters) are expressed exactly in terms of the two-point covari-
ance function. The covariance spectral density also follows directly from the
energy functional. Such explicit expressions are not available for non-Gaussian
energy functionals. The moments must be calculated either by numerical in-
tegration (e.g., Monte Carlo methods) for each set of parameters visited by
the optimization method or by approximate, explicit methods that have been
developed in the framework of many-body theories, e.g. [5, 8, 9, 13].
In statistical physics, e.g. [4, 5, 6] there is a long literature on approxi-
mate but explicit methods (variational approximations, Feynman diagrams,
renormalization group, replicas) that address calculations with non-Gaussian
 Spartan Random Fields 23

probability densities. Preliminary efforts to apply these methods in geosta-

tistical research [9, 12], and references therein] should be followed by further
research on closed-form expressions for non-Gaussian Spartan densities and
the accuracy of such approximations in various areas of the parameter space.
In the variational approach [2, 5, 8], the non-Gaussian probability density is
expanded around an ‘optimal’ Gaussian. The variational Gaussian can then
be used as the zero-point approximation for low-order or diagrammatic per-
turbation expansions of the moments [9, 13]. Below, we outline the application
of the variational method [2, 4, pp. 198–200, 5, pp. 71–77].

5.1 The Variational Method

We present the formalism of the variational method assuming that the SSRF
is defined in a discretized space (e.g. on a lattice). The fluctuation random
field and its states are denoted by the vector y. The characteristic function
Z[J] corresponding to the energy functional H is defined as

Z[J] = Tr [exp(−H + J · y)] . (12)

The symbol ‘Tr’ denotes the trace over all the field variables in H. For a lattice
field the trace is obtained by integrating over the fluctuations at every point
of the lattice. The cumulant generating functional (CGF) is defined by

F [J] = − log Z[J]. (13)

The cumulants of the distribution are obtained from the derivatives of the
CGF with respect to J. For example, the mean is given by

∂F [J]
E [y(si )] = − , (14)
∂Ji J=0

and the covariance function by

∂ 2 F [J]
Gy;λ (s1 , s2 ) = . (15)
∂J1 ∂J2 J=0

Higher-order cumulants are given by higher order derivatives of the CGF. The
CGF of the Gaussian part H0 −J·y is denoted as F0 [J]. Let us now consider a
variational Gaussian energy functional H0 , which is in general different than
the Gaussian component HG of H. The average of an operator A with respect
to the pdf with energy H0 , is obtained by means of

Tr A e−H0
A
0 = . (16)
Tr e−H0
The following inequality [5] is valid for all H0

F [J] ≤ F0 [J] + H − H0
0 . (17)
24 D.T. Hristopulos

The optimal Ĥ0 that gives the best approximation of F [J], is obtained by min-
imizing the variational bound F0 + H − H0
0 with respect to the parameters
of H0 . The optimal Gaussian pdf has energy Ĥ0 and provides approximate
estimates of the non-Gaussian covariance function.
It is possible to improve on the variational approximation by expressing
the energy functional H as follows

H = Ĥ0 + (H − Ĥ0 ) = Ĥ0 + (HG − Ĥ0 + δH), (18)

and treating the component Hpert = HG − Ĥ0 +δH of the energy functional as
a perturbation around the optimal Gaussian Ĥ0 . Corrections of the stochastic
moments can then be obtained either by means of simple (low-order) per-
turbation expansions, or by means of diagrammatic perturbation methods.
However, there is no a priori guarantee that such corrections will lead to more
accurate estimates, and such approximation must be investigated for each
energy functional.

5.2 Example of Variational Calculation

Here we present a simple example for a univariate non-Gaussian pdf, which il-
lustrates the application of the variational method. Consider the non-Gaussian
energy functional
H(y) = a2 y 2 + β 4 y 4 , (19)
where y is a fluctuation with variance E[y 2 ], and the average is over the pdf
p(y) = Z −1 exp(−H). The following Gaussian variational expression is used
as an approximation of the non-Gaussian pdf
√ −1   
p0 (y) = 2 πσ exp −y 2 2σ 2 . (20)

Hence, the variational energy functional is H0 = y 2 2σ 2 and σ is the vari-
√
ational parameter. It follows that F0 = − log( 2 πσ) and H − H0
0 =
a2 σ 2 + 3 β 4 σ 4 − 1/2. The variational bound given by (17) is a convex upward
function of σ, as shown in Fig. 1. The bound is minimized for the following
value of σ

α   4−1
 1/2 α−1   4−1
 1/2
σ̂ = 3 1 + 12 ρ = 3 1 + 12 ρ . (21)
6β 2 6ρ2
β
In the above, ρ = α is the dimensionless ratio of the quartic over the quadratic
pdf parameters that measures the deviation of the energy functional from the
Gaussian form. The value of σ̂ 2 is the variational estimate of the variance.
The exact variance, calculated by numerical integration, and the variational
approximation for various values of the dimensionless coefficient ratio ρ = β/α
 Spartan Random Fields 25

2
β/α = 0.2
1.5 β/α = 0.6
β/α = 0.8
β/α = 1.0

Variational Bound
1

0.5

−0.5

−1

−1.5
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
σ

Fig. 1. Plots of the variational bound as a function of σ for four different values of
the ratio β/α

are plotted in Fig. 2, which shows that the variational estimate is an excellent
approximation of the exact result even for large values of the ratio ρ. Esti-
mates based on first-order and cumulant perturbation expansions around the
optimal Gaussian (these will be presented in detail elsewhere) are also shown
in Fig. 2. The additional corrections do not significantly alter the outcome
of the variational approximation for the variance, since all three plots almost
coincide. However, such corrections will be necessary for calculating higher
moments of non-Gaussian distributions. For example, the kurtosis of the

0.7
Exact
Variational
0.6 Variational & 1st Order
Variational & Cumulant
0.5
Variance

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6
Ratio β/α

Fig. 2. Plots of the exact variance (numerical) and approximate estimates based
on the variational approach as well as combinations of variational and perturbation
methods (first order and cumulant expansion)
26 D.T. Hristopulos

variational Gaussian is equal to 3, and thus it is not an accurate approxi-

mation of the kurtosis of the non-Gaussian distribution except for very small
values of β.

6 Discussion

Spartan random fields provide an alternative to classical geostatistics for mod-

eling the local variability of spatial processes. Spartan models are computa-
tionally efficient for large samples. In addition, they allow quantifying the
variability of spatially sparse data sets, since the model parameters can be
determined from a small number of measurements, in contrast with models
based on variograms. The SSRFs also include a resolution scale that controls
the smoothness of the field.
For SSRFs the structure of the energy functional, which may involve only
short-range interactions, also determines the spatial dependence at large dis-
tances. In principle, the impact of this property on geostatistical modeling
is mixed: On one hand, it does not allow estimating long-range dependence
directly from the data. On the other hand, the estimation of the variogram
at large distance often suffers from significant uncertainty due to insufficient
number of pairs. Hence, the ability of SSRFs to model the long-range behavior
of spatial processes needs to be investigated. It should also be mentioned that
it is possible to modify the energy functional of the SSRFs by adding explicit
long-range interactions.
Non-Gaussian distributions can be handled by means of the standard loga-
rithmic transform. It is also possible to define interactions in the energy func-
tional that lead to specific non-Gaussian probability densities. The complexity
of the inference problem in this case increases compared to the Gaussian case.
Certain methods that may be helpful for calculations with non-Gaussian den-
sities were suggested in this paper, and the variational method was presented
in more detail with the help of a specific univariate example.
Certain other methodological and numerical issues of SSRFs require further
investigation. The methodological issues include estimation at unsampled
points, Monte Carlo simulation, application to real data sets, formulation
of estimation uncertainty, stability of model parameters to uncorrelated
noise, modelling of spatial processes with multiple scales of variability and
anisotropic structures. Estimation has been briefly discussed in [10], and un-
conditional simulation in [11]. Numerical issues involve efficient algorithms
for optimization (model inference process), simulation, and the processing of
spatial information in problems with irregular supports.
References

1. Adler RJ (1981) The Geometry of Random Fields. Wiley, New York

2. Barthelemy M, Orland H, Zerah G (1995) Propagation in random media: cal-
culation of the effective dispersive permittivity by use of the replica method.
Phys Rev E 52(1):1123–1127
3. Bochner S (1959) Lectures on Fourier Integrals. Princeton University Press,
Princeton, NJ
4. Chaikin PM, Lubensky TC (1995) Principles of Condensed Matter Physics.
Cambridge University Press, UK
5. Feynman RP (1982) Statistical Mechanics. Benjamin and Cummings, Reading,
MA
6. Goldenfeld N (1993) Lectures on Phase Transitions and the Renormalization
Group. Addison-Wesley, New York
7. Goovaerts P (1997) Geostatistics for Natural Resources Evaluation. Oxford
University Press, NY
8. Hristopulos DT, Christakos G (1997) A variational calculation of the effective
fluid permeability of heterogeneous media. Phys Rev E 55(6): 7288–7298
9. Hristopulos DT, Christakos G (2001) Practical calculation of non-Gaussian
multivariate moments in spatiotemporal Bayesian maximum entropy analysis.
Math Geol 33(5): 543–568
10. Hristopulos DT (2003) Spartan Gibbs random field models for geostatistical
applications. SIAM J Sci Comp 24: 2125–2162
11. Hristopulos DT (2003) Simulations of spartan random fields. In: Simos TE
(ed) Proceedings of the international conference of computational methods in
sciences and engineering 2003: 242–247. World Scientific, London, UK
12. Hristopulos DT (2003) Renormalization group methods in subsurface hydrol-
ogy: Overview and applications in hydraulic conductivity upscaling. Adv Water
Resour 26(12): 1279–1308
13. Meurice Y (2002) Simple method to make asymptotic series of Feynman dia-
grams converge. Phys Rev Lett 88(14):1601–1604
14. Pilz J (2003) Bayesian spatial prediction using the Matern class of covariance
models. In: Dubois G, Malczewski J, De Cort M (eds) Mapping radioactivity
in the environment: Spatial interpolation comparison 1997: 238–252. Office for
Official Publications of the European Communities, Luxembourg
15. Yaglom M (1987) Correlation Theory of Stationary and Related Random Func-
tions I. Springer, New York
Bayesian Trans-Gaussian Kriging with Log-Log
Transformed Skew Data

Gunter Spöck, Hannes Kazianka, and Jürgen Pilz

Department of Statistics, University of Klagenfurt, Klagenfurt, Austria

gunter.spoeck@uni-klu.ac.at
hannes.kazianka@uni-klu.ac.at
juergen.pilz@uni-klu.ac.at

1 Introduction

Besides the assumption of stationarity it is conventional among geostatistical

practitioners to make the assumption of Gaussianity when applying the linear
kriging methodology. The standard procedure in geostatistics is to
• estimate an empirical variogram or covariance function,
• fit a theoretical variogram or covariance model to the empirical estimate
by means of least squares,
• apply linear kriging to the data by plugging in the estimated positive
semi-definite theoretical covariance model,
• report uncertainties of the predictions by means of the plug-in kriging
variance and Gaussian-based confidence intervals.
The above approach has certain deficiencies:
• The empirical variogram or covariance estimates at the different lags are
highly correlated and, therefore, can be very misleading.
• Once the empirical variogram is badly estimated, the same is also true for
the fitted theoretical model.
• Plugging the estimated theoretical covariance model in the kriging pre-
dictor and neglecting the uncertainty of the covariance estimate has the
following consequences:
– the plug-in kriging predictor is neither a linear nor best predictor.
– the kriging variance that is based on a plug-in estimate is not the true
variance of this highly non-linear plug-in kriging predictor and actually
underestimates the true unknown variance [5].
• In hardly any application the assumption of Gaussianity is fulfilled. Most
environmental processes like ore grades, soil and air measurements show
highly right-skewed marginal distributions.
30 G. Spöck et al.

All these negative facts were the motivation for us to look for more ad-
vanced kriging methodologies that relax the Gaussian assumption and the
disadvantage of not taking into account the uncertainty of the covariance
function.
One of the first papers that addressed the above deficiencies and influenced
our work was De Oliveira [15]. He developed a Bayesian trans-Gaussian krig-
ing where uncertainties could be specified on the trend, the covariance func-
tion and the parameter of the transformation function. The transformation
he used to make a skew random field Gaussian was the Box-Cox transforma-
tion. Motivated by the conjugacy of the normal-inverse-gamma family to the
normal sampling distribution he exactly used such kind of prior for the sill,
the range and the trend parameters of his model. His approach distinguishes
from our approach by the fact that his prior is informative. Berger et al. [3]
were the first to investigate also non-informative reference priors for correlated
stochastic processes. The disadvantage of their approach is that the assumed
random field must be Gaussian. Some time later the paper [16] appeared
where also a non-informative prior for the transformed Gaussian model was
proposed, but with fixed range parameter while the sill and the nugget are
variable.
Our approach seems to be the first completely non-informative approach
to trans-Gaussian Bayesian kriging. We avoid the specification of a prior dis-
tribution by means of a parametric bootstrap, where the sampling distribution
of maximum likelihood estimates is taken as the posterior for unknown
parameters.

2 Bayesian Trans-Gaussian Kriging Model

Throughout the paper we apply the convention that the underlying random
field is given by {Z(x) : x ∈ D} whereas the transformed Gaussian random
field is denoted by Y (x) = g (Z(x)). One transformation that is also conven-
tional in regression analysis is the Box-Cox transformation defined for positive
data z and given by
 λ
z −1
: λ = 0
gλ (z) = λ ,
log(z) : λ = 0
This transformation was used by De Oliveira [15] and Pilz and Spoeck [19] in
their approach to trans-Gaussian Bayesian kriging. Log-normal kriging forms
a special case, where the log-transform of the underlying random field {Z(x) :
x ∈ D} is assumed to be Gaussian: Y (x) = logZ(x) ∼ N (m(x), σ 2 ) where
σ 2 = var{Y (x)} is the (constant) variance of the log-transformed random field.
The Box-Cox transformation is able to model moderately skewed marginal
distribution.
The innovation of this paper is that we apply a new kind of transformation
that we call the log-log transformation and the fact that the work of Pilz et al.
 Bayesian Trans-Gaussian Kriging 31

[19] is extended to allow also for uncertainty in geometric anisotropy param-

eters. See Appendix for details. The log-log transformation is only defined for
positive data z and can be written as

gλ (z) = log (log (z) + λ) . (1)

The restriction log (z) > −λ for all z out of the support of the random process
must be fulfilled. This restriction is not relevant in applications since we only
have a finite set of observations and thus minz log (z) > −∞. The main ad-
vantage of the log-log transformation is that because of the double logarithm
highly skewed data can be potentially transformed to a normal distribution.
For the trend and error model of the transformed Gaussian random field
we assume the conventional geostatistical model:

E{Y (x)} = μ,

where μ is a constant trend. For the Gaussian error model we assume a co-
variance function Cθ,σ2 of the form

Cθ,σ2 (x1 − x2 ) = σ 2 kθ (x1 − x2 ) (2)

where σ 2 = var{Y (x)} denotes the variance (overall sill) of the random
field and kθ (·) the correlation function (normalized covariance function);
θ ∈ Θ ⊂ Rm stands for a parameter vector whose components describe
the range and shape of the positive definite correlation function. Under these
assumptions the probability density of the observed data takes the form

f (Z; μ, θ, σ 2 , λ) = Jλ (Z) ∗ ((2π)n det(σ 2 Kθ ))− 2

1

1
∗ exp{− (gλ (Z) − 1μ)T (σ 2 Kθ )−1 (gλ (Z) − 1μ)},
2
where Jλ (Z) is the determinant of the Jacobian of the specific transformation
used and σ 2 Kθ is the covariance matrix of

Y = gλ (Z) = (gλ (Z(x1 )), gλ (Z(x2 )), . . . , gλ (Z(xn )))T .

We consider the transformation parameter λ to be unknown and to be es-

timated. Christensen et al. [6] point out that the interpretation of the μ-
parameter changes with the value of λ, and the same applies to the covari-
ance parameters σ 2 and θ. Estimation of the parameter λ can be performed
by a profile likelihood approach. For fixed values of λ and θ, the maximum
likelihood estimates for μ and σ 2 are given by
−1 −1 T
μ̂OK T
λ,θ = (1 Kθ 1) 1 Kθ −1 gλ (Z), (3)

and
1 −1
2
σ̂λ,θ = (gλ (Z) − 1μ̂OK T
λ,θ ) Kθ (gλ (Z) − 1μ̂OK
λ,θ ),
n−1
32 G. Spöck et al.

respectively, where Kθ is the correlation matrix. The estimates for λ and θ

have no closed form expression and have to be found numerically by plugging
μ̂OK 2
λ,θ and σ̂λ,θ in the above likelihood function for numerical maximization.
The prediction problem can be faced by the help of the conditional mean.
The conditional mean – or ordinary kriging predictor – E{Z(x0 )|Z}, is optimal
with respect to the squared error loss. Moreover, it may be shown that the
conditional distribution, [Y (x0 )|Y], is normal with mean and variance equal
to this ordinary kriging predictor and kriging variance on the transformed
scale.
Since we want to apply the Bayesian paradigm we have to introduce ap-
propriate prior distributions for the mean parameter μ and the covariance
parameters θ and σ 2 . The first who made use of the Bayesian approach in
kriging were Kitanidis [13], Omre [17] and Omre and Halvorsen [18]. Con-
trary to us, they assumed a Gaussian random function model for Z(·) with
fixed, exactly known covariance function and only incorporated probabilis-
tic prior information for the trend function. The Bayesian ordinary kriging
predictor may be written as
2 2 2
θ,σ −1
ẐBK (x0 ) = μ̂θ,σ T θ,σ
BK + cθ Kθ (Z − 1μ̂BK ), (4)
2
μ̂θ,σ
BK = (1 T
K−1
θ 1
−1 −1
+Φ ) T
(1 K−1
θ Z
−1
+Φ μ0 ), (5)

where E (μ) = μ0 is the fixed a-priori mean and var (μ) = σ 2 Φ is the fixed
a-priori variance for μ. The vector cθ contains the correlations between the
point to be predicted and the observations at the n locations. It can be shown
that the total mean-squared error (TMSEP) of this predictor
2
θ,σ
E{ZBK (x0 ) − Z(x0 )}2 =
 
σ 2 1 − cTθ K−1θ c θ + ||1 − 1T
K −1
θ c ||2
θ (1T K 1)−1 ,
−1
θ

where ||a||2A is a short-hand for the quadratic form aT Aa, is always smaller
than the mean-squared-error of prediction (MSEP) of the ordinary kriging
predictor. Thus, by accepting a small bias in the Bayes kriging predictor and
using prior knowledge E (μ) = μ0 and var(μ) = σ 2 Φ one gets better predictions
than with ordinary kriging. We refer to Spöck [20], where these results are
investigated in more detail.
An obvious advantage of the Bayesian approach, besides its ability to deal
with the uncertainty of the model parameters, is the compensation for the lack
of information in case of only few measurements. This has been demonstrated
impressively by Omre [17], Omre and Halvorsen [18] and Abrahamsen [1].
Bayesian linear kriging is not fully Bayesian, since it makes no a-priori
distributional assumptions on the parameters of the covariance function. The
first to take also account of the uncertainty with respect to these parameters,
using a Bayesian setup, were Kitanidis [13] and Handcock and Stein [11]. A
prior different from the one of Handcock and Stein was used by Gaudard et al.
 Bayesian Trans-Gaussian Kriging 33

[10]. Further references to Bayesian spatial prediction approaches are Le and

Zidek [14], Handcock and Wallis [12], Cui et al. [7], Ecker and Gelfand [9] and
Banerjee et al. [2].
We now come to discuss the Bayesian approach to trans-Gaussian kriging.
Whereas, in conventional trans-Gaussian kriging [6] the uncertainty of the
transformation to Gaussianity and the uncertainty of the covariance function
was not considered, De Oliveira et al. [15] have proposed a Bayesian trans-
Gaussian kriging method which takes full account of these uncertainties. A
prior p(μ, θ, σ 2 , λ) is specified for all unknown parameters. “But the choice
of the prior distribution requires some care, because the interpretation of μ,
σ 2 and θ depends on the realized value of λ. Each transformation ( i.e. each
λ) will change the location and scale of the transformed data, as well as
the correlation structure, so assuming them to be independent a priori of λ
would give nonsensical results”, (citation from De Oliveira et al. [15]). Defining
τ = σ12 , their full prior specification is based on a proposal of Box and Cox
[4] and is given by the improper density
p(θ)p(λ)
p(μ, θ, τ, λ) = 1/n
.
τ Jλ (Z)
Observe, this prior is dependent on the data Z(xi ), i = 1, 2, . . . , n. De Oliveira
et al. [15] used their method for the spatial prediction of weekly rainfall
amounts. “It performed adequately and slightly better than several kriging
variants, especially regarding the empirical probability of coverage of the nom-
inal 95% prediction intervals”, (citation from De Oliveira et al. [15]).

3 The Bootstrap Approach

Our empirical Bayesian approach to trans-Gaussian kriging has delicate dif-
ferences to the approach taken by De Oliveira et al. [15]. They have to specify
priors p (θ) and p (λ) for covariance and transformation parameters. Up to
date there exists only very preliminary work on non-informative priors for
these parameters (Berger et al. [3]). So, nobody actually knows what the in-
formation content of informative priors is. Our approach is a kind of empirical
Bayes and non-informative step to reasoning using some Bayesian  paradigm.
For the mean parameter μ we use a normal prior μ ∼ N μ0 , σ 2 Φ but instead
of specifying prior distributions for θ, σ 2 and λ we implicitly use only informa-
tion from the data by means of parametrically bootstrapping ML-estimators
of these parameters. We then treat this bootstrap distribution as posterior
and apply the Bayesian predictive approach.
Generally, the parametric bootstrap of these ML-estimators works as
follows:
1. From the original data Z (x1 ) , . . . , Z (xn ) the desired covariance and
transformation parameters are estimated by means of ML to get θ̂0 , σ̂02
and λ̂0 .
34 G. Spöck et al.

2. The estimated parameters are subsequently used for simulating trans-

formed random fields as follows: With the estimated covariance parame-
ters a Gaussian random field with mean equal to the initial generalized
0 0 θ̂ ,σ̂ 2
least squares estimate μ̂BK is generated on the locations x1 , . . . , xn and
the estimated transformation parameter, λ̂0 , is then used to transform
the random field to a trans-Gaussian one. Arbitrary strictly monotone
transformations gλ to Gaussianity and back-transformations may be used.
Special cases would be the Box-Cox or the log-log transformation.
3. The second step is repeated for a large number N of simulations and
the covariance and transformation parameters are re-estimated from the
simulated data by means of ML.
4. 
The result  of this parametric bootstrap is a set of ML estimates
σ̂i2 , θ̂i , λ̂i .
i=1,...,N

The bootstrap distribution reflects the uncertainty of the covariance func-

tion and the correct transformation to Gaussianity. Treating this bootstrap
distribution as posterior for the unknown parameters in the Bayesian pre-
dictive approach thus takes account of all mentioned uncertainties. Since we
have simulations from our posterior distribution we can proceed by means
of a Monte-Carlo approximation to the predictive distribution at unknown
locations z0 :

1 
N
p(z0 |Z)  p(gλ̂i (z0 )|λ̂i , θ̂i , σ̂i2 , Z) ∗ Jλ̂i (z0 ) (6)
N i=1

Here p(gλ̂i (z0 )|λ̂i , θ̂i , σ̂i2 , Z) is the conditional predictive density,

λ̂ ,θ̂i ,σ̂i2
Y (x0 )|λ̂i , θ̂i , σ̂i2 , Z ∼ N (ŶBK
i
(x0 ), T M SEPλ̂i ,θ̂i ,σ̂2 ),
i

λ̂ ,θ̂ ,σ̂ 2
i i
where ŶBK i
(x0 ) is the Bayes kriging predictor applied to the transformed
data Y = gλ̂i (Z) for fixed (λ̂i , θ̂i , σ̂i2 ), and T M SEPλ̂i ,θ̂i ,σ̂2 is the correspond-
i
ing Bayes kriging variance. From this predictive distribution quantiles, the
median, the mean and probabilities above certain thresholds can easily be
calculated.

4 Application to the SIC2004 Joker Data Set

One of our aims is to have a methodology that is intrinsic Bayesian and can
be applied also to highly skewed data sets that often occur in applications.
In 2004 one such data set was investigated in detail during the spatial inter-
polation contest (SIC2004) [8]. Ten training data sets on radioactivity levels
were given to the participants of the contest in a certain period of time to
 Bayesian Trans-Gaussian Kriging 35

train their automatic interpolation algorithms. After training another data set,
called “Joker”, was given to the probands, which had one completely different
property than the training data. “A small corner located SW of the monitored
area was chosen and a dispersion process was modelled in order to obtain a
few values on the order of 10 times more than the overall background levels
reported for the first data set”, according to [8]. The automatic interpolation
routines applied to this data ranged from ordinary kriging, splines, support
vector machines to neural networks. The performance of the different algo-
rithms could later on be compared to the true values. Performance measures
such as mean absolute error (MAE) and root mean squared error (RMSE)
have been reported and published in [8]. Because we already know the true
data when we have looked at the performance of our Bayesian trans-Gaussian
kriging algorithm, the calculations we give here are outside the context of a
competition. But our results are in performance comparable to the best real
geostatistical algorithm from this contest. The winner was a neural network
algorithm.
Figure 1(a) gives the data locations of the joker data set (blue circles)
together with the locations where the prediction should take place (red stars).
For an exploratory data analysis we refer to Dubois [8]. A histogram of the
Joker data set is shown in Fig. 1(b). The histogram shows that the background
level is quite symmetric, however, there exist also some very large values that
can be interpreted as an accidental release of radioactivity.
The methodology we apply to this data set is Bayesian trans-Gaussian
kriging with the log-log transformation. Since we are not sure whether the
variogram model is linear or parabolic in the origin we have used a convex
combination of a Gaussian and an exponential variogram model. The advan-
tage of our method is that according to the data the bootstrap methodology
then takes account also of this uncertainty of the variogram model in the
origin. The convex combination parameter as well as Gaussian and expo-
nential range parameters, the overall sill and transformation parameter are
part of the bootstrap. The anisotropy is respected as well by including a
transformation matrix for the coordinates in the maximum likelihood boot-
strap. As already mentioned the main advantage of our approach is that all
uncertainties are taken into account and no prior specification is necessary.
Figure 2(a)–(d) show the bootstraped transformation parameters, covariance
parameters and variogram functions from the posterior. Although estimation
of geometric anisotropy was performed, it turned out that the boostrapped
semivariogram realizations show no anisotropy. Because we calculate posterior
predictive distributions (see Fig. 3) at all locations where prediction should
take place by means of Monte Carlo averaging with the samples from the
bootstrap, graphics like quantile maps, Fig. 4 (a)–(d) and Fig. 5(b), posterior
mean map, Fig. 5(a), and maps of the probability above thresholds, Fig. 6,
are available. To make our results comparable to the SIC2004 contest we
calculated the MAE=16.19 and RMSE=77.64. In terms of MSE this would
have been the second best result in the SIC2004 contest.
36 G. Spöck et al.
x 105
7

Northing
3

–1
–1 0 1 2 3
Easting
x 105

(a)
25

20

15

10

0
0 500 1000 1500
(b)

Fig. 1. (a) The data locations (blue circles) and the locations where prediction
takes places (red stars). (b) The histogram of the Joker data set

Because in most practical applications true data at the locations where

prediction is desired are not available, we have tested the performance of our
algorithm also in the context of cross-validation. Results are shown in Fig. 7.
Furthermore, we must mark that the Bayesian posterior mean is conditionally
biased as was expected. Small values are predicted quite well but the extreme
values from the dispersion process are underestimated. This has been expected
since only two very large values are included in the training set. Validation
results with the true data values are similar.
 Bayesian Trans-Gaussian Kriging 37

posterior of lambda posterior of nugget

10
20
18 9

16 8

14 7

12 6

10 5

8 4

6 3
4 2
2 1
0 0
–4 –3.9 –3.8 –3.7 –3.6 –3.5 –3.4 –3.3 –3.2 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09

(a) (b)
posterior of sill posterior of semivariograms
25 0.35

0.3
20
0.25
15 0.2

10 0.15

0.1
5
0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 1000 2000 3000 4000 5000
(c) (d)

Fig. 2. Results of parameteric boostrap taken as posterior. (a) posterior distri-

bution of transformation parameter, (b) posterior of nugget, (c) posterior of sill,
(d) semivariogram functions. Red stars and lines indicate the initial ML estimates
on which the parametric bootstrap is based

(x,y) = (44840,572913) x 10−3 (x,y) = (50787,195043)

0.06
9

0.05 8
7
0.04 6
5
0.03
4
0.02 3
2
0.01
1
0 0
20 40 60 80 100 120 140 160 180 200 100 200 300 400 500 600 700

(a) (b)

Fig. 3. (a) Posterior predictive distribution at a “normal” background location. (b)

posterior predictive distribution at a “hotspot”
38 G. Spöck et al.
q0.05 q0.25

105
140
600 600
100
130
500 95 500
120
90
400 400
110
85
300 300
80 100

200 200
75 90

100 70 100
80

65
0 0
70
60
0 100 200 0 100 200

(a) (b)
q0.75 q0.95

500
260
600 600
240 450

500
220
500
400

400 200 400 350

180
300 300 300
160
250
200 200
140

200
100 120 100

100
150
0 0
80 100

0 100 200 0 100 200

(c) (d)

Fig. 4. Maps of the quantiles of the posterior predictive distribution. (a) 5% quan-
tile, (b) 25% quantile, (c) 75% quantile, (d) 95% quantile

mean median

600 220 600 180

200
500 500 160

180
400 400
140
160
300 300

140 120
200 200
120
100
100 100
100

0 0 80
80

0 100 200 0 100 200

(a) (b)

Fig. 5. (a) posterior predictive mean. (b) posterior predictive median

 Bayesian Trans-Gaussian Kriging 39
threshold 90 threshold 110

0.8
600 0.8 600

0.7
0.7
500 500
0.6
0.6
400 400
0.5
0.5
300 300
0.4
0.4

200 200
0.3
0.3

100 100 0.2

0.2

0 0.1 0 0.1

0 100 200 0 100 200

(a) (b)
threshold 130 threshold 170
0.6

600 0.7 600

0.5
500 0.6 500

0.4
400 0.5 400

300 0.4 300 0.3

200 0.3 200

0.2

100 0.2 100

0.1
0 0.1 0

0 100 200 0 100 200

(c) (d)

Fig. 6. Maps of the probabilities above certain thresholds. (a) threshold 90, (b)
threshold 110, (c) threshold 130, (d) threshold 170

5 Conclusion

Skewed, non-Gaussian data are common in environmental and geostatistical

applications. Conventional geostatistical methods do not take into account
these facts. The present paper investigates a method to deal with this sort of
data. Because conventional kriging methodology does not respect the fact that
the used covariance function is an estimate and therefore always is uncertain to
some degree, our aim was to propose an approach that takes the uncertainty of
the covariance estimates into account. The suggested algorithms are Bayesian
in spirit but avoid the tedious specification of prior distributions for unknown
parameters for which up to date no, in the non-Gaussian context, practically
useful work exists. By applying a bootstrapping procedure and interpreting
the bootstrap distribution as posterior distribution we result in an empirical
Bayes method. A data example taken from the SIC2004 contest demonstrates
the usefulness of our methodology.
40 G. Spöck et al.
crossvalidation crossvalidation
100 100

expected percent data above threshold

90 90

percent data below quantile

80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 20 40 60 80 100 0 20 40 60 80 100
percent actual data above threshold quantile

(a) (b)
crossvalidation

220
200
180
posterior mean

160
140
120
100
80
60
40
200 400 600 800 1000 1200 1400
data

(c)

Fig. 7. Crossvalidation results. (a) percent of actual data vs. expected percent of
data above the thresholds 10, 20, . . . , 170, 1000, 1100, . . . , 1500. (b) posterior predic-
tive quantiles vs. percent of data below quantiles. (c) data vs. posterior predictive
mean

Appendix: Details on ML Covariance Estimation

For ML estimation of the parameters θ, σ 2 and λ we have to maximize the
likelihood function given by

L(μ, λ, σ 2 , θ; Z) = p(gλ (Z)|μ, σ 2 , λ, θ) ∗ Jλ (Z),

where the Jacobian is given by

∂
Jλ (Z) = | det( gλ (Z))|.
∂λ
and the density of the transformed data is

p(gλ (Z)|μ, σ 2 , λ, θ) = N {1μ, σ 2 Kθ ; gλ (Z)}.

Because we have prior knowledge about β in the form E(μ) = μ0 , var(μ) = σ 2 Φ

it makes sense to use the profile-likelihood instead:
2
L(λ, σ 2 , θ; Z) = p(gλ (Z)|μ̂θ,σ
BK , σ , λ, θ) ∗ Jλ (Z),
2

where
2 2
p(gλ (Z)|μ̂θ,σ θ,σ
BK , σ , λ, θ) = N {1μ̂BK , σ Kθ ; gλ (Z)}.
2 2
 Bayesian Trans-Gaussian Kriging 41

Numerical routines like Gauss-Newton but also Line-Search are sensitive to

the starting values (λ0 , σ02 , θ0 ) and to maximizing the likelihood for (λ̂, σ̂ 2 , θ̂)
at once. We have therefore adopted a profile-likelihood approach and used the
standard Matlab routines for maximization.
1. First estimate λ̂0 for λ: Least squares fitting of a Gaussian density to a
kernel density estimate of the transformed data.
2. First estimate for the covariance function:
(a) Transform the data with λ̂0 .
(b) Calculation of an anisotropic empirical variogram estimate from the
transformed data.
(c) Fitting of a theoretical anisotropic variogram model to the empirical
variogram to get estimates σ̂02 , θ̂0 and linear anisotropy transformation
matrix Â0 .
3. Repeat the following steps for i = 0, 1, 2, . . . until convergence of the
estimates:
(a) Fix σ 2 = σ̂i2 , θ = θ̂i , A = Âi and maximize the profile-loglikelihood
l(λ, σ 2 , θ, A; Z) for λ to get λ̂i+1 .
(b) Fix λ = λ̂i+1 , A = Âi and maximize the profile-loglikelihood l(λ, σ 2 ,
θ, A; Z) for σ 2 , θ to get σ̂i+1
2
, θ̂i+1 .
2 2
(c) Fix λ = λ̂i+1 , σ = σ̂i+1 , θ = θ̂i+1 and maximize the profile-
loglikelihood l(λ, σ 2 , θ, A; Z) for the anisotropy transformation matrix
A to get Âi+1 .
References

1. Abrahamsen, P. (1992), “Bayesian kriging for seismic depth conversion of a mul-

tilayer reservoir”, in Geostatistics Troia ’92, A. Soares ed., Kluwer, Dordrecht,
358–398
2. Banjeree, S., Carlin, B. & Gelfand, A. (2004), “Hierarchical modelling and anal-
ysis for spatial data”, Chapman and Hall/CRC, Boca Raton, Florida
3. Berger, J., De Oliveira, V. & Sanson, B. (2001), “Objective Bayesian analysis
of spatially correlated data”, Journal of the American Statistical Association,
96, 456, 1361–1374
4. Box, G. & Cox, D. (1964), “An analysis of transformations”, Journal of the
Royal Statistical Society, B, 26, 211–252
5. Christensen, R. (1991), “Linear models for multivariate time series and spatial
data”, Springer, Berlin
6. Christensen, O., Diggle, P. & Ribeiro, P. (2001), “Analysing positive-valued spa-
tial data: The transformed Gaussian model”, in GeoENV III – Geostatistics for
Environmental Applications, eds. P. Monestiez, D. Allard and R. Froidevaux,
Kluwer, Dordrecht, 287–298
7. Cui, H., Stein, A. & Myers, D. (1995), “Extension of sptial information,
Bayesian kriging and updating of prior variogram parameters”, Environmetrics,
6, 373–384
8. Dubois, G. & Galmarini, S. (2005), “Spatial interpolation comparison (SIC)
2004: Introduction to the exercise and overview of results”, in Automatic Map-
ping Algorithms for Routine and Emergency Monitoring Data, ed. G. Dubois,
Belgium
9. Ecker, M. & Gelfand, A. (1997), “Bayesian variogram modelling for an isotropic
spatial process”, Technical Report 97-01, Department of Statistics, University
of Connecticut
10. Goudard, M., Karson, M., Linder, E. & Sinha, D. (1999), “Bayesian spatial
prediction”, Environmental and Ecological Statistics, 6, 147–171
11. Handcock, M. & Stein, M. (1993), “A Bayesian analysis of kriging”, Technomet-
rics, 35, 4, 403–410
12. Handcock, M. & Wallis, J. (1994), “An approach to statistical spatio-temporal
modelling of meteorological fields”, Journal of the American Statistical Associ-
ation, 89, 368–378
 Bayesian Trans-Gaussian Kriging 43

13. Kitanidis, P. (1986), “Parameter uncertainty in estimation of spatial functions:

Bayesian analysis”, Water Resources Research, 22, 499–507
14. Le, N. & Zidek, J. (1992), “Interpolation with uncertain spatial covariance: A
Bayesian alternative to kriging”, Journal of Multivariate Analysis, 43, 351–374
15. De Oliveira, V., Kedem, B. & Short, D. (1997), “Bayesian prediction of trans-
formed Gaussian random fields”, Journal of the American Statistical Associa-
tion, 92, 440, 1422–1433
16. De Oliveira, V. (2007), “Objective Bayesian analysis of spatial data with mea-
surement error”, Canadian Journal of Statistics, 35, 283–301
17. Omre, H. (1987), “Bayesian kriging-merging observations and qualified guess in
kriging”, Mathematical Geology, 19, 25–39
18. Omre, H. & Halvorsen, K. (1989), “The Bayesian bridge between simple and
universal kriging”, Mathematical Geology, 21, 7, 767–786
19. Pilz, J. & Spoeck, G. (2007), “Why do we need and how should we imple-
ment Bayesian kriging methods”, Stochastic Environmental Research and Risk
Assessment, DOI: 10.1007/s00477-007-0165-7, Springer
20. Spöck, G. (1997), “Die geostatistische Berücksichtigung von a-priori Kenntnis-
sen über die Trendfunktion und die Kovarianzfunktion aus Bayesscher, Minimax
und Spektraler Sicht”, Diploma Thesis, University of Klagenfurt
Kriging and Splines: Theoretical Approach
to Linking Spatial Prediction Methods

Philipp Pluch

Energy and Petroleum Resources Services GmbH, Vienna, Austria

ppluch@menpet.at

1 Introduction
Spatial Statistics refers to a class of models and methods for spatial data that
aim at providing quantitative descriptions of natural variables distributed in
space or space and time (see Chiles and Delfiner [2], Cressie [3]). Examples
for such variables are ore grades collected in a mineral field,density of trees
of a certain species in a forest or CD (critical dimension) measurements in
semiconductor productions. A typical problem in spatial statistics is to pre-
dict values of measurements at places where they were not observed, or if
measured with error, to estimate a smooth spatial surface from the data. (Es-
timation of a regionalized variable.) A family of techniques, stochastic and non
stochastic ones were developed in geostatistics for that interpolation problem.
The general approach is to consider a class of unbiased estimators, usually
linear in the observations and to find the one with minimum uncertainty, as
measured by the error variance. A group of techniques, known loosely as krig-
ing, is a popular method among different interpolation techniques developed
in geostatistics by Krige [10], Matheron [11] and Journel and Huijbregts [8].
An interesting comparison of ten classes of interpolation techniques with char-
acteristics can be found in Burrough and McDonnell [1] and in a lot of papers
published recently a comparison of several interpolation techniques was made.
The goal of kriging like that of nonparametric regression is that the under-
standing of spatial estimation is enriched by the interpretation as smoothing
estimates. On the other hand random process models are also valuable in
setting uncertainty estimates for function estimates, specially in low noise sit-
uations. There are close connections between different mathematical subjects
such as kriging, radial basis functions (RBF) interpolations, spline interpola-
tions, reproducing hilbert space kernels (rhsk), PDE, Markov Random Fields
(MRF) etc. A short discussion of these links is given in Horiwitz et al. [7],
see Fig. 7. Splines link different fields of mathematics and statistics – and are
used in statistics for spatial modelling (see more in Wahba [15]).
46 P. Pluch

2 Interpolation Techniques
Interpolation techniques can be divided into techniques based on deterministic
and stochastic models. The stochastic approach regards the data {yi }ni=1 as a
realisation of a random field on Rd at ti = {xi1 , ..., xid }, i = 1, ..., n and sets
g(t) to be the best unbiased linear predictor of the random field at site t given
the measurements.
We assume here an intrinsic random field (second-order stationary). Splines
are smooth real valued functions g(t). Define a roughness penalty based on the
sum of integrated squared partial derivatives of a given order n. The choice
of g(t), which interpolates the data and minimises the roughness penalty, is
known as the smoothing thin plate spline introduced by Reinsch [13].

2.1 Univariate Spline Function

A theoretical definition and historical motivation can be found in Haemmer-

lin and Hoffmann [5]. The natural spline S(x) = Sn is a real valued function
S : [a, b] → R with n knots −∞ ≤ a < x1 < x2 < ... < xn ≤ ∞ with following
properties

S ∈ Π m−1 for x ∈ [a, x1 ] and x ∈ [xn , b]

S ∈ Π 2m−1 for x ∈ [xi , xi+1 ], i = 1, ..., n − 1
S ∈ C 2m−2 for x ∈ (−∞, ∞)
f (xi ) = fi for i = 1, ..., n

Π q is the class of polynomials with degree q and C p the class of con-

tinuous functions of order p. The historical problem is to find a function
f in a function space with continuos derivatives of order (m − 1) with

Fig. 1. Where splines link

 Kriging and Splines 47
b
minimum a (f (m) (x))2 dx under all functions with the property f (xi ) = fi
for i = 1, ..., n. Schoenberg shows that the solution is the natural spline.
The statistical approach turns attention to smooth the data and not to
interpolate them. The first access to this problem was done by Reinsch [13]
by means of finding a function g that minimises

xn  n
 g(xi ) − yi
,, 2
g (x) dx + ρ ( )2 + z 2 − S (1)
x0 i=0
δy i

where z and S are auxiliary variables and ρ is a Lagrange parameter. The

solution for (1) is the cubic spline:

g(x) = ai + bi (x − xi ) + ci (x − xi )2 + di (x − xi )3 for xi ≤ x ≤ xi+1 (2)

2.2 Multivariate Spline

In analogy with the interpolation problem in one dimension we now handle a

data set recording three dimensions with a grid of points (xi , yi ), i = 1, ..., n.
For polynomial interpolation we can use a polynomial with low degree r of
the form
r
Pr (x, y) = apq xp y q (3)
p+q=0

On an arbitrary grid it is in general not possible to construct a unique solution.

Here we assume a rectangular region on which we build a rectangular (n +
1)(k + 1) grid of the form

a = x0 < x1 < ... < xn = b

c = y0 < y1 < ... < yn = d,

where in x direction and in y direction a spline will be constructed. As an

analogue to the univariate B-Spline we can give the following spline base
(see [5]) ⎧ (xν+2 −x)(yκ+2 −y)
⎪
⎪
⎪ (xν+2 −xν+1 )(yκ+2 −yκ+1 ) for (x, y) ∈ I1
⎪
⎨ (x−xν )(yκ+2 −y)
−xν )(yκ+2 −yκ+1 ) for(x, y) ∈ I2
B1νκ = (xν+1 (4)
⎪
⎪
(x−xν )(y−yκ )
for (x, y) ∈ I3
⎪
⎪ (xν+1 −xν )(yκ+1 −yκ )
⎩ (xν+2 −x)(y−yκ )
(xν+2 −xν+1 )(yκ+1 −yκ ) for (x, y) ∈ I4

2.3 Additive Model

Now we are going to handle a (d + 1)-dimensional data record (xi , Yi ), with

xi = (xi1 , xi2 , ..., xid )T ; x1 , x2 , ..., xn are independent realisations of the ran-
dom vector X = (X1 , X2 , ..., Xd ), where xi and Yi fulfil

Yi = g(xi ) + εi (5)
48 P. Pluch

with 1 ≤ i ≤ n, where g is an unknown smoothing function mapping from Rd

to R, and ε1 , ε2 , ..., εn are independent error terms. Our aim is to find a valid
additive approximation of g by


d
g(X) ≈ g0 + gi (Xi ) (6)
i=1

The additive model is defined by


d
Y =α+ gi (Xi ) + ε (7)
i=1

Like in the multilinear regression case the error term is independent of Xi ,

E(ε) = 0 and var(ε) = σ 2 . From (7) we can conclude that for any predictor
and for all dimensions there exists at least one function g. More can be found
in Hastie [6].

3 Smoothing Spline
In this paragraph we consider the following problem: We want to find a func-
tion g(x) under all two times continuosly differentiable functions that min-
imises (1).

3.1 Univariate Approach

First we have to define:

hi = xi+1 − xi for i = 1, ..., (n − 1)

⎧
⎪ 1
⎨ Δi,i = hi
Δ = Δi,i+1 = −( hi + hi+1 ) a(n − 2) × n matrix
1 1
⎪
⎩ Δi,i+2 = 1
hi+1
⎧ hi
⎨ Ci−1,i = 6
C = Ci,i = 16 (hi + hi+1 ) a(n − 2) × (n − 2)matrix
⎩
Ci,i−1 = h6i

With this representation the minimisation problem in (1) is equivalent to

 y − g2 + ρg T Kg (8)

where K is a quadratic penalty matrix,

K = ΔT C −1 Δ.

Calculation of the inverse of C is possible because it is strictly diagonal dom-

inant (see [14]). The solution to (8) is given by
 Kriging and Splines 49

ĝ = Sy (9)

where S is the smoothing Matrix of the form

−1
S = (I + ρK) . (10)

Equation (10) we will later find in the solution of the kriging prediction prob-
lem and can be found also in ridge regression with natural basis and Demmler-
Reinsch basis (see more in Nychka [12])

3.2 Multivariate Approach

When we use this approach with the additive model we are able to find (with
help of (8)) the following result
⎛ ⎞2

d 
d 
n
g(X) ≈ ρi giT Ki gi + ⎝Yi − g0 − gj (xij )⎠ (11)
i=1 i=1 j=1

Minimisation leads to
⎛ ⎞

d
ĝl = Sl ⎝y − g0 − ĝj ⎠ (12)
j=1; j=l

with Sl = (I + ρl Kl )−1 for l = 1, ..., d (13)

Sl can be seen as smoothing matrix and Kl as penalty matrix. The following

matrix equation system must be solved:
⎛ ⎞⎛ ⎞ ⎛ ⎞
I S1 S1 . . . S1 g1 S1 y
⎜ S2 I S2 . . . S2 ⎟ ⎜ g2 ⎟ ⎜ S2 y ⎟
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎜ .. .. .. . . . ⎟⎜ . ⎟ = ⎜ . ⎟ (14)
⎝ . . . . .. ⎠ ⎝ .. ⎠ ⎝ .. ⎠
Sd Sd Sd ... I gd Sd y

shortly we can write P g = Qy where P is a (nd) × (nd) and Q is a (nd) ×

(nd) block matrix. A solution of that problem can be found by Backfitting
Algorithm (see [4]). One solution to (14) is given by


d
ĝ(x) = Ŷ = g0 + Oj−1 RjT gj (15)
j=1

where Rj is a reduction matrix and O an order matrix.

50 P. Pluch

3.3 Abstract Minimisation

In Wahba [15] a more general approach to the problem of spline smoothing

is given. She discusses the problem of finding a function f in the Sobolev
function space of the form

Wm : Wm [0, 1] = {f, f , , f ,, , ..., f (m−1) absolutelycontinuous,f (m) ∈ L2 },

The solution can be found by abstract analysis and abstract optimisation.

Wahba gives the following equivalent smoothing problem via r.h.k.s.

1
n
(yi − ηi , f
)2 + ρ  P1 f 2R → min (16)
n i=1 f ∈Wm

with the solution given by


M 
n
fρ = d ν φν + ci ξi
ν=1 i=1
ξi = P1 ηi , i = 1, ..., n
d = (d1 , d2 , ...dM )T = ((T T M −1 T )−1 T T M −1 )y
c = (c1 , ..., cn )T = M −1 (I − T (T T M −1 T )−1 T T M −1 )y
M = Σ + nρI
Σ = ( ξi , ξj
), i = 1, ...n and j = 1, ...M

A more general way to describe the roughness penalty function is in the form

+ ,2
d (r + 1)! ∂ r+1 g(t)
Jr+1 (g) = dt (17)
m!(r + 1 − m)! Rd ∂t1 · · · ∂tm
m1
d
d
|m|=r+1

When a particular penalty is chosen then the result is invariant under rotations
and translations of t.

4 Kriging

Let {Y (t), t ∈ Rd } be an intrinsic or stationary random field; we additionally

demand a polynomial drift of order n > 0. In that case the drift is a linear
combination of tm for | m |< n with unknown coefficients. The subspace of
the polynomial drift has an order less than n with dimension
+ ,
d+n
ν=
d
 Kriging and Splines 51

Let
+ ,
f (t0 )T
Vr =
F
F = (f (t1 ), ..., f (tn ))T
EY (t) = β T f (t),

Vr a (n + 1) × ν matrix of the drift at locations t0 , ..., tn , F a (n × ν) matrix

that describes the drift at locations t1 , ..., tn and ur,0 a vector of lentgh ν with
the elements tm 0 , | m |< n. The covariance function is defined as σ(ti − tj ) for
i, j = 1, ..., n and ϕi,j = σ(ti − tj ) for i, j = 0, ..., n.
+ 2 ,
σ σ0T
Φ=
σ0 Σ

where σ 2 = σ(0) and σ0 is a (n×1) vector with elements σ(t0 −ti ), i = 1, ..., n.
Σ has the elements σi,j .
If {Y (t)} is a stationary random field with a polynomial drift, then kriging
involves predicting Y (t0 ) by a linear combination Ŷ (t0 ) = αT y. The goal is
to minimise the prediction mean squared error subject to an unbiasedness
constraint.

var(Y (t0 ) − Ŷ (t0 )) = var(Y (t0 ) − αT y) = var(β T z) = (cT Φc) (18)

under the unbiasedness constraint EcT z = cT Vr β = 0. It’s straightforward to

minimise (18) by Lagrange multipliers to give

α = AFr,0 + Bσ0 (19)

where

A = Σ −1 F (F T Σ −1 F )−1 (20)
B = Σ −1 − Σ −1 F (F T Σ −1 F )−1 F T Σ −1 (21)

for more details see Kent and Mardia [9] who also give a solution for intrinsic
random fields where σ(h) is a polynomial in h with degree 2p. A and B can
be found with the use of Moore-Penrose generalised inverse.

B = [(I − Ur (UrT Ur )−1 UrT )Σ(I − Ur (UrT Ur )−1 UrT )]− (22)
A = (I − BΣ)Ur (UrT UR )−1 (23)

5 Link Between Kriging and Thin Plate Splining

Following theorem, which identifies the kriging solution with a thin-plate
spline is one of the main results in the literature (see [9]) for the discussion
“kriging vs. splines”.
52 P. Pluch

Theorem 1. Let r + 1 > 12 d (d is the dimension, r order of the polynomial

drift) and set α = r + 1 − d2 > 0. Then the problem of interpolating data
(yi , ti ), i = 1, ..., n subject to minimising the roughness penalty (17) has a
solution g ∗ (t0 ) given by
g ∗ (t0 ) = y T Aur,0 + y T Bσ0 , (24)
where A, B, ur,0 and σ0 are determined as before.
The link between kriging and thin-plate splines holds for the smoothing prob-
lem as well as for the interpolation problem shown by the last theorem. In the
thin-plate spline approach a smoothing function g(t) with square-integrable
(r + 1)th order derivatives has to be found such that

F (g, ρ) = | yi − g(ti ) |2 +ρJr+1
d
(g) (25)
is minimised. The solution to this problem can be found in Kent and Mardia
[9] or in an equivalent formulation in Nychka [12].
The optimal choice of g is given by
g(t0 ) = y T (I + κB)−1 Au0 + y T (I + κB)−1 Bσ0 , (26)
which is the same as the kriging predictor with a nugget effect.

6 Example for Kriging and Splining

Now we want to compare these two methods with the help of a modification
of Wendelberger’s test function (see Fig. 2), this function will be disturbed by
some noise (∼ N (0, 1)), see Fig. 3. This data set x, y and z will be smoothed

Fig. 2. Wendelberger’s test function

 Kriging and Splines 53

Fig. 3. Wendelberger’s test function with noise

Fig. 4. GCV fitted thin plate spline

54 P. Pluch

Fig. 5. Oversmoothed thin plate spline

by thin plate splines and with the help of GCV introduced in Wahba [15], see
Fig. 4 and Fig. 5.
The following Fig. 6 summarises the results of GCV.

Fig. 6. Summary of kriging with matern covariance function

 Kriging and Splines 55

Finally, Fig. 7 gives a contour plot of the surface, resulting from kriging
with a Matern covariance function.

Fig. 7. Contourplot of kriging prediction

References

1. Burrough, P.A. and McDonnell, R.A. (1998) Principles of Geographical Infor-

mation Systems, Oxford University Press
2. Chiles, J.P., Delfiner, P. (1999) Geostatistics, Modeling Spatial Uncertainty,
Wiley, New York
3. Cressie, N.A.C. (1991) Statistics for Spatial Data, Wiley, New York
4. Green and Yandell (1985) Semi-parametric generalised linear models
5. Haemmerlin, G., Hoffmann, K.H. (1992) Numerische Mathematik, Springer
Verlag, Berlin
6. Hastie, T.J., Tibshirani R.J. (1990) Generalised Additive Models, Chapman
and Hall, London
7. Horiwitz et al. (1996) Fast multidimensional Interpolation. 26th Proceedings of
the Applications of Computers and Operations Research in Mining Industry.
8. Journal, A.G. and Huijbregts, Ch. J. (1978) Mining Geostatistics, Academic
Press
9. Kent, J.T., Mardia K.V. (1994) The Link between Kriging and Thin-plateSpline
10. Krige, D.G. (1951) A statistical approach to some basic mine valuation prob-
lems on the Witwatersrand. Jornal of the Chemical, Metallurgical and Mining
Society of South Africa, 52, 119–139.
11. Matheron, G. (1965) Les Variables Regionalisées et leur estimation, Mason,
Paris
12. Nychka, D. (2000) Spatial Process Estimates as Smoothers. Smoothing and
Regression, Schimeck, M. (ed.), Wiley, New York
13. Reinsch, C. (1967) and (1970) Smoothing by spline functions I, II. Numerical
Mathematics.
14. Stoer, J. (1983) Einfuehrung in die Numerische Mathematik, Springer Verlag,
Berlin
15. Wahba G. (1990) Spline Models for Observational Data, Society for Industrial
and Applied Mathematics, Philadelphia, PA
ANNEX Model: Artificial Neural Networks with
External Drift Environmental Data Mapping

R. Parkin1 and M. Kanevski2

1
Institute of Nuclear Safety (IBRAE), Moscow, Russia
park@ibrae.ac.ru
2
Institute of Geomatics and Analysis of Risk, University of Lausanne, Switzerland
Mikhail.Kanevski@unil.ch

1 Introduction

The present research deals with the novel development in the field of envi-
ronmental spatial data modelling with the help of Artificial Neural Networks
(ANN). The following spatial prediction is considered: given the measurements
of some physical quantity at finite (and relatively small) number of points, the
objective is to make predictions over the considered region either on a regu-
lar dense grid (traditional mapping problem) or on irregular decision-oriented
grid. In many cases in addition to the available measurements of the main vari-
able there can be additional information: secondary variables, remote sensing
images, physical model of the phenomena, soft qualitative information, etc. In
the present paper a problem of spatial predictions of the primary variable us-
ing additional comprehensive information on secondary variable is considered.
If there is a relationship between variables (e.g. linear correlation) the second
one can be considered as an external drift. In order to solve this problem an
ANNEX model (ANN + EXternal drift) is proposed. The family of ANNEX
models developed for the spatial mapping problems is based on the idea of
incorporation of additional spatially distributed information into the ANN as
additional input(s). This secondary information is assumed to be related to
the primary variable. This approach considers that additional information is
available both at the training points and at all the points of the prediction
grid. The similar idea traditionally is used in geostatistical “kriging with ex-
ternal drift” model [2]. In general, the ANNEX approach can be considered
as a nonlinear modelling on a hypersurface described by input variables. In
the present work the application of the ANNEX model is applied to the real
case study dealing with the average long-term temperatures of air in June in
Kazakh Priaralie. Additional information that will be used is the elevation at
the measurement and prediction locations above the sea level. ANNEX model
results are compared with the ones of both standard MLP (without extra
58 R. Parkin and M. Kanevski

input) and linear geostatistical estimators: kriging, co-kriging, collocated co-

kriging, kriging with external drift [2].

2 ANNEX Model
The problem of spatial mapping of environmental data is rather traditional
and there exist a wide variety of different prediction models to solve it. In
most cases it is necessary to predict values of a spatial function (precipita-
tion, temperatures, contamination et al.) at the unsampled points, in partic-
ular on a regular grid. Geostatistics is the well-elaborated approach to solve
such problems. All geostatistical models rely on modelling of spatial correla-
tion structures (variography) and are mainly based on a linearity hypothesis.
Thus, geostatistics is a model-dependent approach: solutions highly depend
on a developed model of spatial correlation. Another data driven approach is
based on application of artificial neural networks [1, 3]. Neural networks are
robust, nonlinear and highly flexible tools for data modelling. It was shown
that ANN can be efficiently applied to spatial data modelling, especially in
combination with geostatistical tools [4]. Data analysis with ANN includes
several important steps: data selection and pre-processing, selection of archi-
tecture, training, testing, validation. In the present study multilayer percep-
tron (MLP), which is a workhorse of ANN data modelling is applied for spatial
prediction of temperature. MLP being very powerful modelling tools are able
to incorporate in a nonlinear manner different kinds of information and data
during modelling procedure. Usually in spatial data modelling input space of
ANN (independent variables) are described by geographical coordinates (e.g.,
x, y). Output unit (F) of ANN is a modelling function in case of univariate
prediction or a vector in case of multivariate predictions. The idea of ANNEX
model is as follows: if there is an additional information available at training
and prediction points and related to the primary one, we can try to use it as
additional inputs to the standard ANN.
Consider the examples of external information suitable for ANNEX type
of modelling:
1. Availability of “cheap” information on the secondary variable(s). Consider
that we are interested in a prediction of some physical quantity (primary
variable) whose measurements are rather complicated and/or expensive.
If there are other variables available or easily measured at all points (both
measurement and prediction grids) we can try to check and to use this
information in order to improve the quality of primary variable prediction.
2. Physical model of the phenomena. Consider that we are given the physical
model that describes phenomena under study. To include this model into
the data-driven ANNEX approach the output of the physical model at all
the prediction points and at measurements locations are used as an extra
input(s) for ANN. In general, secondary ANN model can be developed to
model (learn) physical phenomena.
 ANNEX Model 59

3. Physical description of the prediction region. Sometimes the additional

information on the prediction region is given: altitude map, soil map, etc.
Then this information can be extracted and used as an extra input to
ANN. The specific case is when the remotely sensed image with useful
information about the considered region is provided along with the mea-
surements of primary variable.
The primary variable and secondary (external) information are often cor-
related. If the correlation is linear the methods from multivariate geostatistics
can be used, e.g. cokriging. Linearity of correlation is not required by ANNEX
models.
Let us remind, that the objective of advanced modelling is quite clear: by
using ANNEX models we want to improve accuracy of the prediction and to
reduce corresponding uncertainties, for example measured by variances.
The general question for all kind of ANNEX models is: what relationship
(linear, nonlinear, stochastic or their combination) between the primary and
external secondary information should be in order to make use of ANNEX
efficient and how to measure these relationship? This is not an easy question in
case of nonlinear and stochastic relationships between variables. The problem
deals with a question of how much new information and/or new noise are
introduced with ANNEX approach. An external information can dramatically
change the solution in comparison with a standard model. In the present
study some of these tasks are considered in an empirical way using real case
study.

3 Case Study

3.1 Data Description

This case study deals with the prediction of air temperature in Kazakh Pri-
aralie. The selected region is covering 1, 400, 000 km2 with 400 monitoring
stations. The primary variable is average long-term temperatures of air in
June. Additional information that will be used as an extra ANN input is the
elevation of the locations above the sea level. This information is available on
a dense grid from Digital Elevation Model.
The correlation between air temperature and altitude is linear and is equal
to 0.9 (Fig. 1). The linearity of correlation allowed us to use traditional geosta-
tistical model (e.g., kriging with external drift) for modelling and comparing
the results with the one obtained by ANNEX model. The similar work on mod-
elling of air temperature applying kriging with external drift can be found in
Wackernagel [7].
Following the general methodology of ANN data modelling original data
were split into training and testing data sets. The spatial locations of train and
test data points are presented in Fig. 2. An important and difficult problem
60 R. Parkin and M. Kanevski

Fig. 1. The scatterplot between altitude and air temperature in June

deals with the criteria of data splitting. In most cases data are split ran-
domly. But in case of spatial and clustered data such approach can be not
adequate. In the present study the similarity of data sets was controlled by
comparing summary statistics, histograms and spatial correlation structures
(variograms). Since it is difficult to control both testing and training datasets,
more attention was paid to the similarity of the training data set to the initial
data structures of all data. Similarity of spatial structures of obtained datasets
with the initial data is even more important than statistical factors. Compar-
ison of the spatial structures was carried out with the help of variogram roses,
which model anisotropic spatial correlation structures (see Fig. 3). Such com-
parison provided grounds that split with 168 training and 67 testing points is
quite suitable for the following modelling. More advanced splitting methods
can use statistical tests.

Fig. 2. The spatial location of train (circles) and test (cross) data points
 ANNEX Model 61

Fig. 3. Variogram roses: raw (a), train (b) and test (c) datasets

3.2 Geostatistical Modelling

Firstly, traditional geostatistical models were used in order to compare ob-

tained results with ANNEX modelling results. The comparison allows under-
standing whether ANNEX model improves the efficiency of the prediction
in this case study. The following geostatistical models were used: kriging,
cokriging, collocated cokriging and kriging with external drift [2]. Results
of geostatistical modelling are presented below as errors on the test data
(Table 1). It can be seen that the best results among geostatistical methods
on a test dataset are obtained using kriging with external drift. Cokriging
results are worse than kriging ones because of screening effect [7]. Kriging
and collocated kriging demonstrate analogous patterns while kriging with ex-
ternal drift keeps not only the large-scale structure but also small-scale vari-
ability effects ignored by kriging and cokriging models. It should be noted
that results of mapping in some case should postprocessed using physical
scale of the phenomena variability and in order to avoid spurious small scale
structures.

Table 1. The air temperature test results for geostatistical models

model correlation RMSE MAE MRE
kriging 0.874 3.13 2.04 −0.06
cokriging 0.796 3.97 2.45 −0.11
collocated
cokriging 0.881 3.05 1.95 −0.07
kriging with
external drift 0.984 1.19 0.91 −0.03

3.3 ANNEX Modelling

In the present study, MLP models (as ANN) with the following parame-
ters were used: two (traditional ANN) or three (ANNEX) input neurons,
describing spatial co-ordinates (X, Y) and altitude, one hidden layer and
one output neuron describing air temperature. Backpropagation training with
Levenberg-Marquardt followed by conjugate gradient algorithm was used in
62 R. Parkin and M. Kanevski

order to avoid local minima [6]. ANN and ANNEX modelling results are pre-
sented below as errors on the test dataset (Table 2). MLP with structure
2-7-5-1 (7 and 5 neurones in the two hidden layers) showed the best result
among MLPs with 2 inputs, while ANNEX model with structure 3-8-1 (8
neurones in hidden layer) gave the best result among all considered models.
It is worth to mention that we used several MLP structures for the ANNEX
model and found the optimum model (see Table 2). Mapping on a grid with
the help of ANNEX model features similar pattern as kriging with external
drift.

Table 2. The air temperature test results for ANN and ANNEX models
model correlation RMSE MAE MRE
2-7-5-1 0.917 2.57 1.96 −0.02
3-3-1 0.989 0.96 0.73 −0.01
3-5-1 0.99 0.9 0.7 −0.007
3-7-1 0.991 0.85 0.66 −0.004
3-8-1 0.991 0.84 0.68 −0.001
3-9-1 0.991 0.88 0.69 −0.01
3-10-1 0.99 0.92 0.74 −0.01

3.4 Noise Effect

As it was mentioned above, an important problem concerns the question of

the quality of additional data: there is a dilemma between introducing new
information and/or new noise. In order to test the influence of noise on mod-
els (1) we have generated noise which was used as an additional input; (2)
external information (elevation) was contaminated by noise. In fact, the ob-
jective was to check the robustness and the stability of solution obtained with
different models. Firstly, we have considered 100% noise injection. This model
is similar to the models of noise injection into hidden layer. The noise was
modelled in the following way: the altitude values are remained the same
(distribution is the same), while the spatial coordinates interchanged ran-
domly (spatial structure is reduced to a nugget model). Such procedure is
well known in time series modelling: distribution of data is preserved while
correlation in time is destroyed. In the same manner way we changed only
10% of points and examined the obtained results (see Table 3). It can be
seen that presented ANNEX model is unsusceptible to noise in external
information.
 ANNEX Model 63

Table 3. The air temperature noise test results

model correlation RMSE MAE MRE
kriging 0.874 3.13 2.04 −0.06
kriging
external drift 0.984 1.19 0.91 −0.03
3-8-1 0.991 0.84 0.68 −0.001
3-8-1
(1000.839 3.54 2.37 −0.13
3-8-1
(100.939 2.32 -1.49 −0.003
kriging – ext. drift
(100.941 2.23 1.54 −0.06
3-8-1
(100.899 2.81 1.52 −0.08
kriging – ext. drift
(100.903 2.81 1.59 −0.103

4 Conclusions
An Artificial Neural Networks with External drift (ANNEX) model for the
analysis and mapping of spatially distributed data was applied to the real
data. It was shown that additional spatially distributed information can be
efficiently used by ANNEX and gives rise to better analysis and modelling of
environmental data. Promising results presented are based on the real case
study of air temperature mapping. Other kinds of Machine Leaning mod-
els (besides ANN) can be used with possible modifications in the proposed
framework. The advantage of the ANNEX model is its ability to model any
nonlinear relationships between variables. An interesting feature found in the
study is robustness and stability of the ANNEX solution versus noise. This
problem should be studied in more detail. ANNEX model performed better
even in the case of linear correlation between primary and secondary informa-
tion that is favourable to kriging with external drift. An even more interesting
study should consider nonlinear relationships between data and external in-
formation.

Acknowledgements

The analysis and presentation of the results as well as MLP and geostatistical
modelling were performed with the help of GEOSTAT OFFICE software ([5],
http://www.ibrae.ac.ru/ mkanev). The work was supported by INTAS Aral
Sea project #72. Authors thank S. Chernov and V. Timonin for programming
GEOSTAT OFFICE software, which was extensively used in the research.
References

1. Bishop CM (1995) Neural Networks for Pattern Recognition, Clarendon Press,

Oxford
2. Deutsch CV, Journel AG (1998) GSLIB Geostatistical Software Library and
User’s Guide, Oxford University Press, New York, Oxford
3. Haykin S (1999) Neural Networks, A Comprehensive Foundation, Second Edi-
tion, Prentice Hall International, Inc., Prentice Hall
4. Kanevski M, Arutyunyan R, Bolshov L, Demyanov V, Maignan M (1996) Artifi-
cial neural networks and spatial estimations of Chernobyl fallout. Geoinformat-
ics, vol. 7, pp. 5–11
5. Kanevski M, Demyanov V, Chernov S, Savelieva E, Serov A, Timonin V (1999)
Geostat Office for Environmental and Pollution Spatial Data Analysis. Mathe-
matische Geologie band 3 April 73–83, CPress Publishing House, Dresden
6. Masters T (1995) Advanced Algorithms for Neural Networks. A C++ Source-
book, John Wiley & Sons, Inc., New York
7. Wackernagel H (1995) Multivariate Geostatistics, Springler-Verlag, Berlin
Regional Classification of Indoor Radon
Data with Support Vector Machines
and Geostatistical Tools

A. Chaouch1 , M. Kanevski1 , M. Maignan1 , A. Pozdnoukhov2 , J. Rodriguez3 ,

and G. Piller3
1
Institute of Mineralogy and Geochemistry, University of Lausanne, Lausanne,
Switzerland aziz.chaouch@etu.unil.ch
2
Institute of Geomatics and Analysis of Risk, University of Lausanne,
Switzerland Alexei.Pozdnoukhov@unil.ch
3
Swiss Federal Office of Public Health (OFSP), Bern, Switzerland

1 Introduction
For the last 20 years, the Swiss Federal Office of Public Health (OFSP) has
performed more than 65,000 indoor radon measurements throughout whole
Switzerland. Swiss Indoor radon data are noisy and poorly spatially corre-
lated. They feature large low-scale variability and a strong spatial cluster-
ing. Univariate distribution is positively skewed and heavy-tailed. Thus one
possible way to deal with these data is to transform them into indicators
relative to a decision threshold and apply spatial statistics to these indica-
tors. Indeed when considering decision making, the task is often to classify
indoor radon data into low or high concentration level. This kind of two-class
classification task is commonly solved by geostatistical interpolations of in-
dicators using kriging and/or conditional simulations. However, geostatistical
approaches depend on several assumptions about the data (i.e. stationarity)
and require modelling of the variogram (see Chiles and Delfiner [1]), a task
that is, while sometimes possible, often very difficult and time consuming with
indoor radon data. In consequence, data-driven approaches such as support
vector machines (SVM) are considered as an alternative to geostatistical ap-
proaches. In this paper, their performance in application to indoor radon data
classification is assessed in comparison the one of indicator kriging (IK) and
sequential indicator simulations (SIS).

2 Data Pre-processing
This study will focus on a small square of 25 × 25 km located at the north-
eastern end of Switzerland, near the “Bodensee”. Over this region, stationarity
66 A. Chaouch et al.

might be assumed and the reasonable amount of data allows us to perform

a comprehensive spatial analysis with both geostatistical methods and SVM.
Selection of continuous radon levels recorded in ground floor of inhabited
houses was performed.

2.1 Indicator Transform

Continuous indoor radon levels are transformed into discrete binary indicators
[0;1] for geostatistical methods or [−1; +1] for SVM relative to a user-defined
threshold. In the present study, the threshold is set at 45 Bq/m3 , close to
the median of the regional distribution of indoor radon levels. That level was
chosen for the present methodological study and does not reflect decision level
defined by Swiss federal law [6]. Indicator I at location u is built by comparing
the local indoor radon level Z(u) to the decision threshold Z as follow:
3
I(u; Z) = 1 (geostat) or − 1(SV M ) if Z(u) < Z; Z = 45 Bq/m
I(u; Z) = 0 (geostat) or + 1(SV M ) otherwise

After binary coding of the data, the dataset contains 658 indicators. To
assess classification abilities of methods, a subset of 158 data is kept for vali-
dation purposes only, leaving 500 source data available for the spatial analysis,
see Fig. 1.

Fig. 1. Plot of indicator 1 (white box), indicator 0 (black box) and validation data
(marks)

2.2 Indicator Declustering

Obvious spatial clustering of source indicators induces a bias on univariate

statistics that are no more representative for the entire studied area. Morishita
 Regional Classification of Indoor Radon Data 67

diagram [4], measuring degree of spatial clustering at different scales, shows

that at any scale indicators 1 are slightly more clustered than indicators 0.
This preferential clustering tends to overestimate Indicator mean value. In
order to retrieve representative univariate statistics from the entire area, Cell
declustering and Voronoi polygon declustering [3] were used to compute the
weighted statistics of data, see Table 1.

Table 1. Univariate statistics of raw and declustered indicators

raw indicators cell-declustering polygon declustering
mean 0.508 0.462 0.477
stDev 0.500 0.499 0.499
nb of data 500 500 500

Average value of declustered mean (0.470) obtained with both declustering

techniques is assumed to be the representative indicator mean value over the
entire area of investigation.

3 Geostatistical Tools
When dealing with indicator transformed data [0;1], geostatistical tools such
as kriging and simulations provide an estimate of the probability that indicator
1 prevails at any unknown location u. This estimation requires definition and
modelling of the variogram of indicators.

3.1 Exploratory Variography of Indicator Data

The interpolation is based on the spatial structure of indicator data I which
is described and modelled by the semi-variogram (1) or variogram function γ
where h is the lag of the variogram (i.e. the Euclidean distance between pairs
of N points considered for the calculation).

1 
N
γ1 (h) = (I(ui ) − I(ui + h))2 (1)
2N i=1

Spatial structure of indicator data for threshold 45 Bq/m3 is weak but

existent. For improved visibility and easier modelling, common lag tolerance
of half the lag (traditional variogram) is here increased to three times the
lag. Resulting variogram shape is smooth and easier to model (regularised
variogram).
Correlation distance of indicator data is close to 2 km with a weak
anisotropy being present along NNW-SSE direction, see Fig. 3. High vari-
ability of indicator is obvious with nugget accounting up to 70% for the total
variability that is a reason of low classification efficiency.
68 A. Chaouch et al.

Fig. 2. Traditional and regularized omnidirectional variogram

Fig. 3. Variogram rose

3.2 Indicator (Simple) Kriging

Consider the problem of estimating the indicator value I(u; Z) with known
constant mean m at any location u using N available hard indicators Ik defined
at the threshold Z. The indicator kriging estimate is a linear combination of
available indicators.
- .

N 
N
I(u; Z) = λk · Ik (xk ; Z) + 1 − λk · m (2)
k=1 k=1

Weights λk assigned to available indicators Ik aimed at minimizing the

error variance and thus are provided by the simple kriging system:

N
λk · C1 (xj , ck ) = C1 (xj , u); j = 1, . . . , N (3)
k=1

where Ci (a, b) is the covariance function of indicators between locations a and

b. In case of second order stationarity the covariance function is linked to the
 Regional Classification of Indoor Radon Data 69

variogram function (1) of indicators γI (h) where h is the distance between

locations a and b and σI2 is the expected variance of indicators.

γI (h) = σI2 − CI (a, b) (4)

When available indicators are coded either as 0 or 1 relative to threshold

Z, the estimated indicator value can be seen as the probability that indicator
1 prevails at location u. It is then possible to decide on which probability
level the distinction between class 0 and class 1 is achieved. In this study,
declustered mean value was used as simple kriging mean m and the distinction
between classes is performed as follow:

I(u; Z) = 1 if I (u; Z) > 0.5

I(u; Z) = 0 otherwise

3.3 Sequential Indicator Simulations

Smoothing effect of kriging induced by the error minimization is well docu-

mented [4]. Comprehensive analysis of indicator response requires full repro-
duction of variability and thus sequential simulations are to be considered.
The key problem is here to rebuild both the distribution (i.e. proportions)
and the spatial structure (i.e. variogram) of available indicators still by con-
ditioning them using kriging. The task is performed as follow:
• choose a random path visiting each node ni of the interpolation grid
• perform simple kriging interpolation of available indicators at first node
n1 of the path. As defined in indicator simple kriging, this gives the prob-
ability that I(n1 ; Z) = 1.
• draw a random indicator [0;1] from probability function I(n1 ; Z)
• add the simulated indicator to the dataset and consider it as a new data
• proceed to next node until the interpolation grid is filled.
• restart the whole process many times using different random paths to build
many simulated images
Simulated images are all equally-probable and each one provide a different
possible reality of indicator configuration that honours both raw data and
their spatial structure, see Fig. 4.

Fig. 4. Some of equally-probable single realizations of the simulation process

70 A. Chaouch et al.

From the number of different simulated images, it is possible to infer the

probability to have indicator 1 at any location. As for indicator kriging, final
classification map is produced by choosing probability 0.5 as the decision level:

I(u; Z) = 1 if Z(u) < Z

I(u; Z) = 0 otherwise

4 Support Vector Machines (SVM) for Classification

SVM is a machine learning classification algorithm able to classify data into a
binary format [−1; 1] relative to a user-defined threshold by building a maxi-
mum margin hyperplane separating the two classes, see Kanevski et al. [5].

4.1 Statistical Learning Theory

Support Vector Machines is based on statistical learning theory defined by

Vapnik-Chervonenkis [7]. Statistical learning theory studies the process of
inferring based only on a finite amount of data. The fundamental problem
is thus called generalization. Best generalization abilities are retrieved with a
model of limited complexity. Indeed, simplistic models are not able to extract
enough information out of the data while very complex models tend to overfit
data and thus provide poor generalization performances. Therefore, SVM is
based on the structural risk minimisation principle, aiming to minimize both
the empirical risk (testing error) and the complexity of the model.

4.2 Linear Classification

Consider the classification problem of linearly separable training data labelled

−1; 1 relative to the decision threshold by a hyperplane H : w · z + b = 0 where
z is the input of training data (namely coordinates). Label yi of any point i
is defined by looking on which side of H point i is lying. Formally, this yields
to evaluation of (5).
yi = sign(w · zi + b) (5)
Several hyperplanes are able to separate these data but the unique optimal
solution considers a maximum margin hyperplane H where the distance δ
between all points z and H is maximum. It can be shown that maximizing this
distance yields to minimizing the norm of the weight vector w. The maximum
margin hyperplane should satisfy the condition that training data are correctly
classified:
yi · (w · zi + b) ≥ 1∀i (6)
This leads to an optimisation problem under constraints that can be solved
by introducing positive Lagrange multipliers α.
 Regional Classification of Indoor Radon Data 71

1  N
L(w, z, α) =  w 2 − αi (yi · (w · zi + b) − 1) (7)
2 i=1

L(w, z, α) must be minimized with respect to w and b. Applying the opti-

mality conditions yields to 2 solutions:

∂L  N
=⇒ w = αi · yi · zi (8)
∂w i=1

∂L  N
=0⇒ αi · yi = 0 (9)
∂b i=1

Equation (6) shows that the weight vector w and thus the hyperplane H
is defined only in terms of data with associated α > 0. These data are called
support vectors. Attribute b of the hyperplane is chosen so that it maximizes
the margin.
Classification of any new point is then performed by substituting (8) in
(5), what gives the decision function:
/#SV 0

ynew = sign αi yi ziSV znew +b (10)
i=1

In the linear case, the classification is thus achieved by solving the dot
product between support vector inputs and inputs of the new point to be
classified. Inputs are commonly spatial coordinates of points but they may
also contain additional information.

4.3 Non Linear Classification

Following Cover’s theorem, non linear classification of data can be achieved by

mapping the data into a higher dimensional feature space where data becomes
linearly separable [7]. As the dimension of the feature space where linear
separation is possible is unknown, kernel functions Script K are introduced.
They are able to solve the non linear classification problem in the input space
without explicitly going in the feature space where data would become linearly
separable. This is the so-called “Kernel trick”.
Eligible kernels should satisfy Mercer theorem [7, p. 423] and can take var-
ious forms. In this study, the Radial Basis Function Kernel (11) with standard
deviation σ (kernel width) has been used for SVM classification.
+ ,
 zi − z j 
K(zi , zj ) = exp − (11)
2σ 2
72 A. Chaouch et al.

4.4 Soft Margin Classifier

When dealing with noisy data such as indoor radon levels, it is not always ad-
visable to correctly classify all training data. Indeed, data are likely to contain
errors or incoherent values. The classifier is then built to allow misclassification
of points that have a too strong impact on the boundary definition in order to
improve generalisation abilities. The resulting boundary has a so-called “soft
margin”. Implementation of this additional constraint on the optimisation
problem is done by introducing a new parameter called C-value in addition
to slack variables ξi . Basically misclassified points i are on the wrong side of
their margin by an amount C · ξi [2]. The Lagrangian can then be rewritten
as follow:

N 
N 
N
L= 1
2  w 2 +C · ξi − α[yi · (w · zi + b) − (1 − ξi )] − μi ξi
i=1 i=1 i=1 (12)
ξ ≥ 0, i = 1, . . . , N
0 ≤ αi ≤ C

C value is hence a trade-off between margin maximisation and classification

error.

4.5 SVM Parameters Tuning

SVM parameters such as kernel width and C value are unknown and must be
tuned. Therefore, source dataset is split into training and testing data. Model
is built with training data and parameters are tuned according to testing data
using different combinations of kernel widths and C values. Optimality of SVM
parameters is reached following the structural risk minimization principle [7].
Testing error and complexity of the model that is defined in term of number
of support vectors are both considered. The more support vectors, the more
complex the model.
Ten different random splits of 400 training and 100 testing data were per-
formed to tune SVM parameters. Indeed optimal parameters may vary for
different random splits due to noisy data and/or possible algorithmic instabil-
ities. Average values of both testing error and normalized number of support
vectors for the ten random splits are presented on a map.
As suggested on Fig. 5, there is no clear unique solution. Minimal test
error lies on a straight line pointing out a fairly linear dependence between
optimal kernel width and logarithm of C value. However, two white patches
of low testing error can be seen on this line, one for a kernel width of 1000 m
and the other at approximately 3 km. The solution with kernel width of 3 km
and log(C) of 3 can be built by less support vectors (Fig. 6) and is therefore
preferred over the other. Final classification with optimal parameters is ap-
plied to training and testing data altogether. Under such configuration, the
final classification is performed using only 341 support vectors out of the 500
original data.
 Regional Classification of Indoor Radon Data 73

Fig. 5. Average classification error on testing data (map of empirical risk)

Fig. 6. Average normalized number of support vectors (map of model complexity)

5 Classification Results
Classification maps produced using both geostatistical methods and support
vector machines are presented in this section. Interpolation grid has square
cell-size of 200 m and present postplot of source (marks) and validation data
(circles).
Classifications obtained with geostatistical methods are very similar as
they use the same model of spatial structure: the variogram. However, SIS

Fig. 7. IK classification
74 A. Chaouch et al.

Fig. 8. IK standard deviation

classification is noisier as its purpose is to rebuild variability of data while

IK aim at minimizing error variance thus smoothing out resulting figures.
SVM classification produced with optimal parameters show large scale smooth
delimiting contours that feature similar patterns to the one obtained with
SIS and IK, see Figs. 7–9. However, it is obvious that SVM classification
provides less 0 indicators than other approaches. Unmatched classifications
are mainly located in poorly sampled regions where geostatistical estimates
are less accurate like in the limits of the area (Fig. 10).

Fig. 9. SIS classification

Fig. 10. SVM classification

 Regional Classification of Indoor Radon Data 75

Table 2. Validation of classification

classification method % of misclassified validation data
IK 30.38
SIS 29.75
SVM 32.91

Fig. 11. Omnidirectional regularized variograms of validation residuals

Accuracy of classification results is roughly assessed using validation data

(Table 2). Misclassification of validation data is rather similar for all reviewed
methods and stands around 30%. This poor predictability is explained by
the high variability of indicators as suggested by the elevated nugget seen on
the indicator variogram (Fig. 2). Apparently SVM cannot perform better than
geostatistical methods under such configuration. An explanation for this might
be found in the design of the validation experiment itself. As SVM are able
to misclassify some data according to their direct influence on the boundary,
validation data that contain as much noise and/or errors than source data
shouldn’t be necessarily correctly classified. Another validation experiment
consists then to apply exploratory variography on validation residuals, see
Fig. 11. Resulting variograms are fluctuating around sill and suggests that all
reviewed methods despite their different nature are able to extract the weak
spatial information out of the data.

6 Conclusions
Classification abilities of geostatistical approaches and SVM for a median
3
value test decision threshold (45 Bq/m ) are presented in this study. Despite
their poor classification abilities that are easily explained by the high variabil-
ity of indoor radon indicator data, all reviewed methods are able to efficiently
extract spatial information out of the data. In particular SVM are promising
for indoor radon binary classification as they don’t require any prior assump-
tion on data such as stationarity. They may therefore be applied to classifica-
tion problem over large regions if not over the whole country. A specific feature
76 A. Chaouch et al.

of SVM is its ability to consider data as noise while geostatistical approaches

force conditioning of data by kriging. SVM were then able to achieve same
classification abilities as geostatistical approaches with potentially less data
(support vectors). Despite the requirement of many different random splits of
training and testing data to assess convergence of optimal solution, the main
advantage of SVM is to avoid the difficult and time-consuming task of vari-
ogram definition and modelling which is known to be very tricky in case of in-
door radon data. Moreover SVM may incorporate additional data/knowledge
relevant to radon data classification (geology, soil permeability, etc.) as added
inputs to spatial coordinates.

Acknowledgments
The work was supported in part by the INTAS Aral sea grant #72.
References

1. Chilés J.P and Delfines P (1999) Geostatistics, Modelling Spatial Uncertainty.

Wiley-Interscience Publication, New York
2. Hastie T, Tibshirani R and Friedman J (2001) The elements of Statistical Learn-
ing. Data mining, inference and prediction. Springer-Verlag, Berlin
3. Isaaks E. H and Srivastava R. M (1989) An Introduction to Applied Geostatis-
tics. Oxford University Press, New York
4. Kanevski M and Maignan M (2004) Analysis and modelling of spatial environ-
mental data. Presses polytechniques et universitaires romandes (ISBN 2-88074-
507-1)
5. Kanevski M, Pozdnukhov A, Canu S, Maignan M, Wong P and Shibli S
(2001) Support vector machines for classification and mapping of reservoir data.
Soft Computing for Reservoir Characterization and Modeling. Springer-Verlag,
Berlin, pp. 531–558
6. OFSP (2002) Information on radon http://www.bag.admin.ch/strahlen/
ionisant/radon/f/
7. Vapnik V.N (1998) Statistical Learning Theory. Wiley-Interscience Publication,
New York
Daubechies Wavelets for Identification of Rock
Variants from IR Spectra

Vera Hofer1 , Jürgen Pilz2 , and Thorgeir S. Helgason3

1
Department of Statistics and Operations Research, Karl-Franzens University
Graz, Graz, Austria
vera.hofer@uni-graz.at
2
Department of Statistics, University of Klagenfurt, Klagenfurt, Austria
juergen.pilz@uni-klu.ac.at
3
Petromodel Ltd, Reykjavik, Iceland
thorgeir.helgason@petromodel.is

1 Introduction to the Data and the Problem

Aggregates, i.e. sand, gravel and crushed rock, are the most frequently used
construction materials worldwide, i.g. in concrete, cement, asphalt etc. Among
igneous rocks, granite and basalt are the most important. The properties of
all construction materials need to be appropriate for their intended purpose.
Some applications leave room for choice among many different rock variants,
others require a thorough inspection of the particular features of the rocks.
The question of whether the rock will resist physical and chemical loads, is of
particular importance. Specific imperfections in granite result from the trans-
formation of feldspar to kaolinite, or the decay of biotite and may lead to
reduced strength [6]. As rocks are increasingly being used up to the limits of
their mechanical strength, material tolerances are decreasing. This leads to
the demand for ever more careful assessment of rocks. In order to decrease
the costs of damages arising from improper use of aggregates, and to substan-
tially reduce production costs, the aggregates industry is interested in effective
quality control.
This calls for an efficient and fast method for classification of aggregates.
Automatic means for identifying suitable rock characteristics and their varia-
tion so have to be devised. It is well known that matter treated with light of
different wavelengths shows characteristic features that are suitable for quali-
tative and quantitative analysis and therefore for identification of substances
(e.g. [9]). The optical characteristics of the material investigated is expressed
in a spectrum, i.e. a plot of the absorption, transmission, reflection, or emission
intensity as a function of wavelength, frequency, or energy [2].
Theoretical studies show that different substances have characteristic
spectra in certain wavelength regions, even though the appearance of these
80 V. Hofer et al.

spectra may vary considerably, depending on the parameters used. This raises
the question, whether a statistical method for classification of aggregates is
possible.
The aim of the EUREKA project, PETROSCOPE, is to develop an auto-
matic testing instrument for process and quality control in the construction
aggregates industry [1]. In this paper, stemming from the PETROSCOPE
project, the identification of two types or variants of granite, called Granite
1 and Granite 2, by means of their reflectivity in mid-infrared light is investi-
gated. Ten samples, i.e. particles of size 16–32 mm for each of the two types of
Finnish granite, supplied by Lohja Rudus Oy from Hiiskula gravel pit, were
collected by the Geological Survey of Finland. Then the samples were irradi-
ated with infrared light at equidistant wavenumbers from 560 to 4000 cm−1
and from three positions for each particle. These measurements, performed at
VTT Electronics in Finland, resulted in three curves per particle or sample
and therefore 60 curves all together [15]. Figure 1 shows the spectral lines of
the samples for each of the two granite classes or variants.

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

1000 1500 2000 2500 3000 3500 4000 1000 1500 2000 2500 3000 3500 4000

Fig. 1. Reflectivity of two types of granite aggregates in infrared between wavenum-

ber 560 nm bis 4000 nm, first measurement or position for each sample

In general, the spectral lines of the curves seem to be very similar, although
there are also some specific characteristics. This impression is mirrored by the
mean curves of the classes in Fig. 2, which raises the statistical question,
whether the differences in the shape of the curves are systematic or just ran-
dom.
The data observed are continous curves not single observations of scalars
[7, 8], even though the curves were measured at discrete knots and therefore
represented by data vectors xi = (xi1 , . . . , xin ), where n indicates the dimen-
sion of the vectors, i.e. the number of observation knots. In statistical problems
dealing with spectra, the high dimensionality of the data causes problems in
applying the common techniques of multivariate statistics because the number
of samples compared to the number of observation knots is very small. In this
study the proportion of samples to predictors is about 1:8. Even partial least
 Daubechies Wavelets for Identification of Rock Variants 81

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

1000 1500 2000 2500 3000 3500 4000 1000 1500 2000 2500 3000 3500 4000

Fig. 2. Mean curves from each of three measurements of reflectivity for ten particles
or samples in both classes of granite

squares (PLS) does not lead to reasonable classification error rates because of
the similarity of the curves, as discussed in Sect. 2.
Several techniques exist in order to overcome the problem of high dimen-
sionality (multicollinearity). In the following section feature reduction by a
basis expansion is described. We use the fact that the curves observed are in
L2 (IR), which is a Hilbert space. The idea is to choose a proper basis and
then consider a subspace that contains the essential information of the signal.
This information is received by an orthogonal projection of the signal onto
this subspace.
In the current research we use a wavelet basis. Wavelets have adequate local
properties and have turned out to be appropriate for statistical modelling of
high dimensional data, as the characteristic features are summarized by a few
basis coefficients.

1.1 Wavelets

Wavelets are functions that are received from a mother wavelet ψ ∈ L1 (IR) ∩
L2 (IR), which is ([13]):
∞
(i) “waving” above and below the abscissa, i.e. −∞ ψ(t) dt = 0. This means
that ψ has mean zero.
(ii) well localized.
By translation and dilation of the mother wavelet ψ we obtain the family
j
ψj k (x) = 2 2 ψ(2j x − k) , (1)

where the integers j and k are the spatial parameters of the discret wavelet
transform. Two technical terms are usual in signal processing: The factor of
stretch or compression is called scale. The inverse of scale is the resolution.
The higher the resolution the better the approximation and the lower the
scale. The relation among level j, resolution and scale is shown in Table 1.
(In this paper Mallat’s indexing is used [13].)
82 V. Hofer et al.

Table 1. Relation between level, scale and resolution

level −2 −1 0 1 2
1 1
scale 4 2 1 2 4

1 1
resolution 4 2
1 2 4

Under mild conditions on ψ the family (1) constitutes an orthonormal

basis of the L2 (IR) so that each function f ∈ L2 (IR) can be expressed as

f= dj k ψj k .
j, k ∈ Z

To get a decomposition of a signal from low to high resolution, we define a

scaling function for a fixed level J0 , the father wavelet φ, so that the functions
J0
φJ0 k = 2 2 φ(2J0 x − k) k ∈ Z

constitute an orthonormal basis of

VJ0 = span ({φJ0 k | k ∈ Z}) = span ({ψj k | j < J0 ∧ k ∈ Z}) .

So we get a new orthonormal basis of the L2 (IR)

{φJ0 k , ψj k | j ≥ J0 ∧ k ∈ Z} .

For a fixed level j define

Wj = span ({ψj k | k ∈ Z}) .

Then it can be seen easily that the space Vj+1 is decomposed into two or-
thonormal subspaces
Vj+1 = Vj ⊕ Wj .
This concept leads to a multiresolution of L2 (IR):

L2 (IR) = VJ ⊕ WJ ⊕ WJ+1 ⊕ WJ+2 ⊕ · · ·

and to a multiresolution, i.e. a decomposition, of the function f ∈ L2 (IR) into

orthogonal components

f (t) = AJ (t) + Dj (t) ,
j≥J

where 
Dj (t) = dj k ψj k (t)
k∈Z
 Daubechies Wavelets for Identification of Rock Variants 83

is the detail on level j, i.e. on scale 2−j and


AJ (t) = Dj (t)
j<J

is the approximation of f in VJ . Figure 3 provides an overview of the decom-

positon of a signal according to the multiresolution analysis.

f(t)

VJ WJ WJ+1 WJ+2 ··· L 2( )

···
AJ DJ DJ+1 DJ+2

Fig. 3. Decomposition of a function into orthonormal components

Wavelets have several advantages that make them so popular [14]:

(i) they constitute an orthonormal basis,
(ii) they are local in time via translation, and in space via dilation, which,
for example, is not true for Fourier transform,
(iii) the coefficients are a measure of the local behaviour, depending on the
spatial parameters j and k,
(iv) it is very easy to apply them to functions of more than one variable.

1.2 The Final Model Under the Wavelet Approach

On using a wavelet basis, the following model results

fi (t) = gi (t) + εi (t) for i ∈ {1, . . . , n}


gi (t) ∈ VN i.e. gi (t) = AJ (t) + Dj (t) (2)
J≤j<N

εi (t) ∈ VN⊥ i.e. εi (t) = Dj (t) ,
j≥N

where gi (t) represents the systematic part in the curves observed, i.e. the
characteristic feature that should be investigated, and εi (t) stands for the
random error. A projection of the curves observed onto this subspace VN
separates the systematic components of the feature observed from the random
ones.
84 V. Hofer et al.

The question remaining is, which wavelet basis should be used in (2) and
which scale (resolution) should be chosen? Contrary to Fourier transform the
wavelet transform is not unique. There is no rule of thumb that prescribes the
proper basis for the current problem.
In many practial applications Daubechies wavelets are applied [3, 5]. They
have compact support, which is related to computational efficiency [10]. Be-
sides this, they also have some vanishing moments, which improves computa-
tional efficiency, since the higher the number of vanishing moments, the more
information will be concentrated in a smaller number of wavelet coefficients,
and the fine scale wavelet coefficients will be essentially zero where the func-
tion is smooth. However, this increases the support of the wavelets, so that a
trade-off is necessary [10, 12].
This study uses Daubechies wavelets with two vanishing moments. Differ-
ent levels of J were also investigated, and it turned out that J = 5 led to the
lowest classification error in the wavelet model with PCA.
The number of observation knots turns out to play an important role when
working with discretized signals ([3, 11] and the references there). Multireso-
lution analysis requires that the sample size is 2n for some integer n. In cases
where this condition is not satisfied, the problem is often solved by padding
the signal with zeros. This procedure usually introduces unnecessary edge ef-
fects because of the resulting discontinuity of the signal at the borders. These
edge effects are difficult to compensate for. For the data underlying this study
zero padding leads to very low classification error rates.

2 Classification Method and Results

The classification of the data was carried out according to the minimum of
Mahalanobis distance. As preparation for this, some basic transformations
were conducted, such as the log-transformation of the observations and the
standardisation of the observation range to an interval starting at zero, this
interval being devided into subintervals of length one. As each sample was
measured from three positions, the mean of these measurements was calcu-
lated and a baseline correction was carried out. Petrological examination of
the samples should ensure that the samples consisted only of one type of
granite. The signal resulting was projected onto V

P : L2 (IR) → V = span({φ−5 k , ψ−5 k | k ∈ Z}),

so that fi = gi + εi and gi = Ai5 + D5i . The other details were ignored

because different calculations have shown that the classification results im-
proved, when details were not used. This model leads to a reduction in the
number of variables by about 95%.
The classification was carried out by use of the coefficients of the princi-
pal components, calculated by Principle component analysis (PCA) or PLS
 Daubechies Wavelets for Identification of Rock Variants 85

estimation. In PCA and PLS, respectively, the number of features that enter
the classification is reduced to a further extent. Although this reduction is not
overwhelming – 16 scores remain in PLS and in PCA analysis (reduction of
about 1.5%!) – it has to be stated that the scores corresponding to low eigen-
values are important in cases when the direction of separation is orthogonal
to the first PCs (confer [4]).
Using this wavelet approach, complete classification of the mean measure-
ments was attained. The results of assigning single measurements in terms
of the leave-one-out model are summarized in the Tables 2 and 3 below and
depend on the additional method of dimension reduction.

Table 2. Results of statistical classification, based on measurement M for the ten

samples or particles of each of Granite 1 and Granite 2. Wavelet model with PCA
applied to the log-spectra by use of Daubechies wavelets of order two on level five.
M0 denotes the mean measurements, M1 to M3 the single measurements, and G1
and G2 stand for the two types of granite

assignment of

M0 M1 M2 M3

to G1 G2 G1 G2 G1 G2 G1 G2
from

G1 10 0 9 1 10 0 10 0

G2 0 10 1 9 1 9 0 10

The three measurements for each sample can be found in the columns M1
to M3 of Table 2, showing some misclassifications, whereas the classification
based on mean values gives correct assignments. This raises the question,
whether the assignment, based on the single measurements M1 to M3 , and
showing misclassification, belongs to the same particle or different particles;
unfortunately these are two different particles. But as can be seen in Table 2
3
the classification error rate for the single measurements is only 60 = 0.05,
which is very low for curves that are extremely similar as in the case here.
As the dimensionality of the data was only slightly reduced by PCA one
could ask, whether it was necessary to use PCA and whether PLS would
improve the results. The calculations showed that the error rate was unac-
ceptable in the case when classification was carried out only by use of wavelet
coefficients. Therefore, a further data reduction method was applied. After
calculation of the PLS estimation of the scores, instead of a PCA, the results
improved, but the number of scores was the same as in PCA, i.e. 16 scores
86 V. Hofer et al.

Table 3. Results of statistical classification, based on measurement M for the ten

samples or particles of each of Granite 1 and Granite 2. Wavelet model with PLS
applied to the log-spectra by use of Daubechies wavelets of order two on level five.
M0 are the mean measurements, M1 to M3 the single measurements, and G1 and
G2 stand for the two types of granite

assignment of

M0 M1 M2 M3

to G1 G2 G1 G2 G1 G2 G1 G2
from

G1 10 0 10 0 10 0 10 0

G2 0 10 1 9 1 9 0 10

were used in further classification. Table 3 shows that only two single curves
were misclassified after application of PLS for further dimension reduction.
These two curves belong to different samples.
PLS is often applied in chemometrics to reduce the dimension of spec-
tral data for classification. The following Table 4 gives an impression of
the classification error, which arises when PLS is used to select the original
log-transformed spectral lines. This means that PLS is used for identifying
the relevant observation knots. As can be seen from Table 4, there is even

Table 4. Results of statistical classification, based on measurement M for the ten

samples or particles of each of Granite 1 and Granite 2. PLS applied to the log-
spectra. M0 denotes the mean measurements, M1 to M3 the single measurements,
and G1 and G2 stand for the two types of granite

assignment of

M0 M1 M2 M3

to G1 G2 G1 G2 G1 G2 G1 G2
from

G1 9 1 7 3 9 1 9 1

G2 1 9 1 9 0 10 2 8
 Daubechies Wavelets for Identification of Rock Variants 87

classification error in the assignment of the mean measurements. As three

measurements of each sample were classified, the question on the number of
misclassified samples turns up again. In the case of the single measurements,
8 curves but only 5 samples were incorrectly assigned.

3 Summary

After a proper use of statistical spectra approximation it is possible to separate

variants of granite, even if they are very similar in terms of their reflectivity
of mid-infrared light. As reflectivity can vary due to the measurement posi-
tions, classification accuracy is improved by use of several measurements from
different positions of the sample. For such mean measurements, a reliable pre-
diction of the class membership can easily be derived by use of wavelets. A
further reduction of the number of classification features using PCA or PLS is
necessary. The use of PLS for estimation of the scores appeared to be superior
to the traditional PCA in the sense that the variability in measurements have
led to less of a deterioration in the classification results. The wavelet model
is superior to the PLS based reduction in the dimensionality of the original
data.
References

1. EUREKA (2004), PETROSCOPE – An Optical Analyser for Construction Ag-

gregates and Rocks, Project no. 2569, Announced 28 June 2001, EUREKA,
Brussels
2. Hunt, G.R. (1987) Spectroscopic properties of rocks and minerals, in:
Carmichael, R. S., (ed.), Handbook of Physical Properties of Rocks, CRC Press,
Boca Raton, 295–385
3. Jetter, K., Depczynski, U., Molt, K. and Niemöller, A. (2000) Principles and
applications of wavelet transformations to chemometrics, Analytica Chemica
Acta 420, 169–180
4. Jolliffe, I.T. (1986) Principle Component Analysis, Springer Verlag, Berlin
5. Leung, A.K., Chau, F. and Gao, J. and (1998) A review of wavelet transform
techniques in chemical analysis: 1989–1997, Chemometrics and Intelligent Lab-
oratory Systems, 43, 165–184
6. Müller, F. (1987) Gesteinskunde, Ebner Verlag, Ulm
7. Ramsay, J.O. and Silverman, B.W. (1997) Functional Data Analysis, Springer
Series in Statistics, Springer, New York
8. Ramsay, J.O., Silverman, B.W. (2001) Applied Functional Data Analysis,
Springer, New York
9. Schmidt, W. (1994) Optische Spektroskopie, Eine Einführung für Naturwis-
senschaftler und Techniker, VCH, Weinheim
10. Teppola, P. and Minkkinen, P. (2000) Wavelet-PLS regression models for both
exploratory data analysis and process monitoring, Journal of Chemometrics, 14,
383–399
11. Trygg, J. and Wold, S. (1998) PLS regression on wavelet compressed NIR spec-
tra, Chemometrics and Intelligent Laboratory Systems, 42, 209–220
12. Vannucci, M., Brown, P. J. and Fearn, T. (2003) A decision theoretical approach
to wavelet regression on curves with a high number of regressors, Journal of
Statistical Planning and Inference 112, 195–212
13. Vidakovic, B. (1999) Statistical Modeling by Wavelets, Wiley, New York
14. Vidakovic, B. and Müller, P. (1999) An Introduction to Wavelets, in Müller, P.,
Vidakovic, B., (eds.), Bayesian Inference in Wavelet-Based Models, Springer,
New York
15. VTT Electronics (2001) Petroscope Prestudy. Confidential Report to Petro-
model Ltd, VTT, Oulu, Finland
 Part II

Geostatistical Applications
Simulating the Effects of Rural Development
Policies on Land Use: Evidence from Spatially
Explicit Modeling in the Central Highlands
of Vietnam

Daniel Müller1 and Darla K. Munroe2

1
Leibniz Institute of Agricultural Development in Central and Eastern Europe,
Halle (Saale), Germany
mueller@iamo.de
2
Department of Geography, The Ohio State University, Columbus, OH, USA
munroe.9@osu.edu

1 Introduction

Land cover, the spectral characteristics of the earth’ surface, and land use,
the operational employment on that land, are closely related. However, there
is also a clear distinction between land use and land cover. While land
cover refers to the biophysical earth surface, land use is shaped by human,
socioeconomic and political influences on the land [7]. In essence, ‘land use
links land cover to the human activities that transform the landscape’ [15]. In
most practical applications the analysis of satellite images are used to infer
land use from land cover.
Land use is a common phenomenon associated with population growth,
market development, technical and institutional innovation, and related ru-
ral development policy. This paper attempts to assess the impact of policy,
technology, socioeconomic, and geophysical conditions on land use in the last
decade and combines data from a village-level survey with remote sensing
data derived from Landsat images. Our objective is to analyze the influence
of these explanatory variables on land use using a reduced-form, spatially ex-
plicit multinomial logit model. Simulations are then carried out to assess the
effects of three policy scenarios of rural development on land use. An empirical
application is presented for two districts of Dak Lak province in the Central
Highlands of Vietnam. Dak Lak exhibits an interesting case in the study of
land use dynamics with its abundant forest resources, ethnic diversity, high
immigration rates and dynamic agricultural and socioeconomic development.
In particular, the last decade was characterized by rapid, labor- and capital-
intensive growth in the agricultural sector.
92 D. Müller and D.K. Munroe

2 Methodology and Data

Land use is approximated by visual interpretation of a Landsat Enhanced

Thematic Mapper (ETM) image from 2000. The resulting land-use map is
categorized into three classes: (1) mixed agricultural land (cash and food
crops) including cultivated upland plots; (2) paddy fields, both with one and
two crops a year; (3) all non-agricultural land comprising forests of different
quality and small areas of mixed grassland as well as fallowed upland plots.
Results of earlier studies showed only modest land-use changes in the last
decade in absolute and relative terms. Forest cover increased slightly mainly
due to agricultural intensification, better market access and government poli-
cies on forest protection [10].3
We explain land use as a function of four types of regressors. First, we use
geophysical indicators such as the level and variance of rainfall derived from
interpolations of point measurements, surfaces of soil suitability for mixed
agriculture and paddy rice, the altitude and the slope of land. Second, we
employ historical variables that describe socioeconomic characteristics of the
villages. Here, we capture endogenous, and therefore lagged, population as
a time-variant variable to indicate the state at the beginning of the period.
A continuous population surface was generated from village recall data for
1990 by interpolating the point coverages of village locations with inverse
distance weighting [5]. Cultural influences on land use are captured by a
dummy for the ethnic composition of the villages. Third, exogenous policy
variables to describe government investments and macro-level policies are in-
cluded in the model. These include a quantification for the spatial placement
of policy-induced investments in road and market infrastructures and the in-
troduction of agricultural technologies, proxied by the increase in irrigation
per village and the year of introduction of a compound fertilizer contain-
ing nitrogen, phosphorus and potassium (NPK) as a yield-increasing input.
To capture market access we used a road network layer passable during the
whole year and existing since the time of French colonial rule in the first
part of the 19th century. Therefore, we assume the road layer to be exoge-
nous to present-day land use. In addition, areas delineated as New Economic
Zones (NEZ)4 were incorporated as a measure of the length since the inaugu-
ration of government-controlled immigration and associated investments. To
incorporate the patchiness of land use we include one indicator for landscape
fragmentation, measured by the ratio of perimeter to area of a particular
landscape patch. The lagged value of this variable is a proxy the spatial struc-
ture of individual landscape elements. Due to possible scale economies in a
3
Additional analysis and discussion of these issues are to be found in Müller [8]
and Müller and Munroe [9].
4
A resettlement scheme undertaken by the government after the end of the war
in 1975, where people from densely populated areas such as the Red River Delta
and the Mekong Delta were moved to less-densely settled areas for agricultural
production.
 Simulating the Effects of Rural Development Policies 93

situation of agricultural intensification and increasing market orientation, we

expect that pixels under intense agricultural production will tend to be more
spatially concentrated in areas close to market centers and in favorable natu-
ral conditions. Agricultural land uses are, therefore, anticipated to show more
homogeneous patterns.
We will explore the causal relationships between these variables in a spa-
tially explicit framework, quantify their respective direction and magnitude,
and simulate the effect of rural development policies. Detailed descriptive
statistics of the independent variables are reported in Table 1.

Table 1. Descriptive statistics for independent variables

variable mean std.dev. min. max. N

mixed agriculture 0.24 0.43 0 1 22,321
paddy 0.08 0.27 0 1 22,321
non-agricultural land 0.68 0.47 0 1 22,321

slope (degrees) 10.18 9.02 0 42 22,321

spatial lag slope (degrees) 21.01 16.74 0 77.9 22,321
elevation (100 m) 5.88 1.67 4.2 13.1 22,321
soil suitability for paddy (0/1) 0.14 0.35 0 1 22,321
soil suitability for mixed agriculture (0/1) 0.04 0.19 0 1 22,321
rainfall sum (100 mm), 1992–1999 16.41 1.78 9.9 18.6 22,307
rainfall variance (100 mm), 1992–1999 27.62 7.07 16.7 41.5 22,307

population, 1990 65.09 26.67 0.2 286 22,294

ethnic minority village (0/1) 0.62 0.49 0 1 22,321

distance to all-year road (km) 5.86 5.36 0 23.1 22,321

years since establishment of NEZ 9.26 8.55 0 24 22,321
years since introduction of NPK 4.26 3.7 0 20 22,321
increase in irrigated area (ha) 115.32 202.88 0 938 22,321
protected area Nam Ca (0/1) 0.08 0.27 0 1 22,321

landscape fragmentation, 1992 843.66 280.3 16.6 1,102.6 22,304

Source: Primary data on village level collected in village survey; secondary data
on geophysical and agroecological variables were provided by the Mekong River
Commission (Digital Elevation Model) and the Department for Agriculture
and Rural Development, Dak Lak (Digital Soil Map and protected areas);
rainfall data from own interpolation of data from nine meteorological stations,
classification of soil suitability dummies from expert opinion.

3 Spatial Sampling
Ideally, to integrate spatially explicit data derived from geographical informa-
tion systems (GIS) and remote sensing (RS) techniques with village survey
data, the scale of the analysis should match the agricultural plots as the
unit of decision-making. Yet, in Vietnam as in most developing countries,
94 D. Müller and D.K. Munroe

plot maps and village boundaries are not available [12]. This renders spatial
modeling a time-consuming and costly task due to the necessary delineation
of the spatial extent of plots or villages, e.g. using Global Positioning Systems
(GPS). To demarcate the spatial base unit for the integration of socioeco-
nomic variables, the geographic positions of all villages were recorded using
GPS and point coverages created. Village boundaries for all surveyed villages
were then approximated by applying a cost-distance algorithm to delineate a
set of explicitly defined ‘accessibility catchments’, generated around each vil-
lage location and based on estimated transport costs [4]. Spatial accessibility
is similar to Euclidean distance functions, but instead of calculating the actual
distance from one point to another, the shortest cost distance (or accumulated
transportation cost) from each cell to the nearest source cells is determined
(Fig. 1).5 The resulting catchment polygons were then used as a base unit
for village-level data in sub-sequent analysis. Hence, survey data – apart from
population – takes the value of the interviewed village for each point, which
has lower transportation costs to the geographic location of that village than
to any other village location.
The units of analysis are square pixels of 50 by 50 m, i.e. 0.25 ha. To focus
on changes at the forest margins influenced by human interventions, we restrict
the analysis to those pixels that have a cost of access below the mean for the

Fig. 1. Transportation cost surface with spatial sample and approximated village
borders [9]
5
For an example of spatial data integration using purely Euclidean distance mea-
sures, see [10].
 Simulating the Effects of Rural Development Policies 95

transportation cost surface (Fig. 1). In that way, we include nearly all the
agricultural area in 2000 and eliminate remote and high mountainous areas
covered mostly with thick primary forest, which are outside measurable human
influence.
At present, there are no models and test statistics available to account for
substantive spatial interaction in a qualitative dependent variable framework
[2]. To compensate for potential spatial dependence in the dependent vari-
ables, Besag’s coding scheme was used [3], also employed by [14] and [11] in
similar studies. The regular spatial sample was drawn by selecting every 5th
cell in the X and Y directions so that no selected cells are physical neighbors.
The sampling procedures allow us to apply standard estimation techniques [1]
and resulted in a dataset of 22,300 observations used for subsequent econo-
metric modeling. In addition, we include slope as a spatially lagged variable
[11, 13, 14]. These techniques help to reduce spatial autocorrelation although
they may not totally eliminate it [6].

4 Models and Results

4.1 Econometric Estimation

To explore relationships between exogenous and predetermined variables and

the land cover categories as left-hand side variables, a multinomial logit speci-
fication (MNL) was applied. MNL models estimate the direction and intensity
of the explanatory variables on the categorical dependent variable by predict-
ing a probability outcome associated with each category of the dependent vari-
able. MNL is based on the assumption that the probabilities are independent
of other outcomes.
Assuming that pixels are independent across villages, but not necessarily
within, we cluster all pixels based on approximated village areas. This affects
the estimated standard errors and the variance-covariance matrix of the esti-
mators, but not the estimated coefficients [16]. Further, we employ the Huber
and White sandwich estimator to obtain robust variance estimates.
The sampling procedure outlined in Sect. 3 resulted in 22,300 observations,
which we use to estimate the coefficients of the MNL. These coefficients are
then used to predict outcomes for the entire dataset (558,000 observations) in
order to generate continuous prediction and simulation maps.

4.2 Empirical Results

The MNL has three land cover classes as categorical, unordered dependent
variables. To control for potential endogeneity problems, only lagged values
for time-variant independent variables such as population growth and road
access are considered in the empirical applications. In addition, all variables
were tested for multicollinearity. We assess the assumption of independence
96 D. Müller and D.K. Munroe

Table 2. Multinomial logit results (non-agricultural land use as base category)

Mixed agriculture paddy

slope –0.088 –0.207
(6.97)∗∗∗ (4.28)∗∗∗
spatial lag slope –0.046 –0.072
(7.51)∗∗∗ (4.74)∗∗∗
elevation –2.621 –6.695
(4.56)∗∗∗ (6.73)∗∗∗
soil suitability for paddy 1.306 2.543
(5.52)∗∗∗ (8.97)∗∗∗
soil suitability for mixed agriculture 1.5 1.047
(4.51)∗∗∗ (2.74)∗∗∗
rainfall sum, 1992–1999 0.504 1.293
(1.90)∗ (1.97)∗∗
rainfall variance, 1992–1999 –0.116 –0.265
(2.86)∗∗∗ (3.85)∗∗∗
population, 1990 0.007 0.013
–0.82 –0.97
ethnic minority village –0.123 –0.433
–0.32 –1.13
distance to all-year road –0.001 –0.004
–0.02 –0.04
years since establishment of NEZ 0.041 –0.022
–1.6 –0.63
years since introduction of NPK 0.062 0.128
(1.65)∗ (2.95)∗∗∗
increase in irrigated area 0 0.001
–0.78 (2.25)∗∗
protected area Nam Ca –2.708 –2.74
(7.71)∗∗∗ (4.85)∗∗∗
landscape fragmentation, 1992 –0.017 –0.02
(4.07)∗∗∗ (3.02)∗∗∗
constant 6.488 11.168
–1.61 –0.95
observations 22,255 22,255
Source: own calculations.

of irrelevant alternatives using both the Hausman and the Small-Hsiao test
and can accept the null hypothesis that outcomes are independent of other
alternatives. Model results are reported as raw coefficients in Table 2 for non-
agricultural land as the comparison group. Overall predictive power is 88%,
measured as the locations predicted correctly. Equivalent to [13] we found
that wrong predictions frequently lie on the border between land use classes,
which is likely to be related to spatial errors in the source data and artifacts
inherent in our technique of data integration.
Coefficients for the geophysical variables are mostly significant at the 1%
level and show the expected signs with high predictive power. Agriculture
 Simulating the Effects of Rural Development Policies 97

is more likely at lower altitudes, flatter land, and more suitable soils. More
amount and less variance of rainfall increases the likelihood of paddy com-
pared to the other two categories. Surprisingly, access to all-year roads did
not have a significant effect on the probability of a certain land use class. We
assume that this is due to the relatively large areas under agricultural uses
far away from the all-year road network. Earlier introduction of mineral fer-
tilizer and more irrigated area increases the likelihood that pixels are under
paddy production. Lagged population does not seem to have an influence on
the amount of area cultivated as it was probably outweighed by effects from
agricultural intensification. In addition, the majority of migrants settled in
areas with high proportions of land suitable for paddy cultivation. Therefore,
lagged population has little influence on the amount of land used for cultiva-
tion. The dummy on ethnic composition is significant at the 5% level and has
a strong negative effect on the probability of paddy land. A more fragmented
landscape in an earlier period decreases the likelihood of present agricultural
uses and a significant amount of fragemented agricultural plots regenerated
into non-agricultural uses. Finally, forest protection has a strong effect on the
likelihood to observe both mixed agriculture and paddy land.

4.3 Policy Simulations

Changes in socioeconomic variables can be simulated by changing the val-
ues of the respective explanatory variables. Applying the estimated coeffi-
cients from the base estimation for the entire data set, new predictions and
probabilities for land use can then be generated with simulated values of ex-
planatory right-hand-side variables. Comparing simulated predictions of land
use to the base predictions yields an approximation of the effects of varying
levels of explanatory variables on land use [9, 13]. One main advantage of
the spatially explicit estimation is the possibility to assess locational changes
and to identify potential hot spots of land-use change following certain policy
interventions.
We distinguish three policy scenarios for rural development interventions
(see Table 3). The first scenario assumes an earlier introduction of NPK
fertilizers. Governments can influence such technology adoption by increasing

Table 3. Simulated policy scenarios

description proxy
1. Earlier introduction of fertilizer → introduction of NPK 5 years earlier
2. Forest protection → (a) protection of existing primary forest
→ (b) protection on slopes > 15 degrees
3. Earlier introduction of fertilizer → scenarios 2 and 3 combined
and forest protection

Source: authors.
98 D. Müller and D.K. Munroe

extensions services and by supporting fertilizer application through price in-

centives. The ‘forest protection scenario’ follows a government guideline that
discourages agricultural production above slopes of 15 degrees. We assume
here that this sloping land and, in addition, all primary forest with closed
crown cover are added to already existing protected areas. The last policy
option combines ‘earlier introduction of NPK’ with the ‘forest protection sce-
nario’ to assess the effect of this predominant policy strategy in Vietnam.
Because of the insignificant coefficients, we did not simulate the upgrading
of seasonal roads to year-round access as it would not result in considerable
land-use changes.
A comparison of simulated changes with the baseline predictions is ex-
pected to shed light on the impact of the rural development policy scenar-
ios on land use. The spatially explicit framework allows for an assessment
of both the magnitude as well as the location and spatial arrangement of
the simulated changes. This will enable the identification of areas, which
are likely to change into other forms of land use under a certain policy
setting [9].

5 Simulation Results

The scenario of ‘earlier introduction of fertilizer’ results in a reduction in non-

agricultural land uses by 14.4 km2 and an intensification of mixed agricultural
land into paddy of 22 km2. The observation of the resulting prediction map
reveals that most of the land switching from non-agriculture was bare soil
and grass land. However, the spatial fragmentation of the simulation map as
measured by the number of patches increases by 6%. The scenario of ‘forest
protection’ decreases agricultural land use by 4.8 km2. An increasing number
of patches leads to more scattered spatial arrangements, which is less desir-
able from an ecological viewpoint. The result suggests that forest protection
strategies ought to be combined with ecological valuations that explicitly take
into account the value of contiguous protection areas conserving precious bio-
diversity.
Due to space limitations we concentrate the following discussion on the
results of the third policy scenario. Significant changes in the spatial pat-
terns of land use result from a combination of policies to boost the in-
troduction of yield-increasing technologies combined with forest protection
(Table 4 and Fig. 2). It induces a reduction in area under mixed agricul-
ture by 26.9 km2. The intensification effect of an earlier introduction of NPK
fertilizer increases the probability for paddy cultivation while mixed agricul-
tural land retreats on marginal land following the simulated increase of forest
protection.
Figure 2 shows the spatial changes resulting from the policy interventions
(see also [8]). On flat land and in suitable areas agriculture is intensified as
can be observed in the northwestern part of the analyzed area. Agricultural
 Simulating the Effects of Rural Development Policies 99

Table 4. Earlier introduction of NPK combined with forest protection

simulated prediction

non-
mixed
paddy agricultural
agriculture
base prediction land total
mixed agriculture 347.5 21.9 5.0 374.3
paddy 0.0 88.0 0.0 88.0
non-agricultural 12.7 0.4 940.6 953.8
land
total 360.2 110.3 945.6 1,416.1
2
Source: own calculations; numbers reported in km .

Fig. 2. Prediction maps of policy scenarios on land use compared to base prediction

production is left abandoned in sloping areas close to the lowlands as in the

south and northeast. The identification of such potential hot-spots calls for
more in-depth field visits to monitor resulting land-use outcomes.

6 Discussion and Policy Implications

Agricultural land uses are concentrated on lower altitudes, flatter land, and
on better soils. Forest protection has a strong effect on the likelihood to ob-
serve both mixed agriculture and paddy land. An earlier introduction of yield-
100 D. Müller and D.K. Munroe

increasing fertilizers increases and higher investments in irrigation increase the

probability of paddy land, while in ethnic villages paddy land is less likely.
The changes in land use over the last decade show that agricultural intensifi-
cation and improved market access made farming more profitable and labor-
intensive. This led to higher outputs that compensated for the decline in the
availability of land per capita given the high population growth rates. Shifting
cultivation as the traditional land-intensive farming system practiced by the
indigenous population in the research area almost entirely disappeared during
the last decade [8, 10].
The spatially explicit simulations demonstrate possible land-use changes
resulting from rural policy interventions. Government investment in agricul-
tural intensification, proxied by an earlier introduction of NPK fertilizers,
has a strong positive influence on the area under paddy rice, by far the
most important food crop for farmers in the research area. The forest pro-
tection scenario induced higher pressure on agricultural land uses within
the simulated additional areas under forest protection and resulted in the
abandonment of more marginal lands. The main advantage we see in this
methodology is the facilitation of a spatial assessment of land-use changes.
In that way, the identification of hot spots, possibly requiring additional con-
servation efforts, and areas with untapped agricultural potentials becomes
possible.
The two districts in our study were purposively selected and can, there-
fore, not be generalized for the whole of Dak Lak province. Nonetheless, the
policy implications of this study call for a renewed emphasis on rural and
agricultural development that can address the subsistence and income needs
through technological progress and agricultural intensification. If combined
with forest protection, especially in locations with crucial watershed and bio-
diversity functions, agricultural intensification might safeguard locally and
globally valuable environmental services. With respect to implications for fur-
ther research, we conclude that problems for spatially explicit modeling and
spatial statistics are found more frequently in the combination of data on
natural resources with socioeconomic information at an acceptable scale and
under reasonable assumptions and simplifications. Our attempt to combine
land use data with accessibility catchments at the village level addresses this
issue. However, this aggregation masks decision-making processes at the farm
and plot levels. More disaggregated data is needed to consider the effects of
rural development policies on cropping patterns and micro-level changes in a
spatially explicit framework.
To obtain better spatial information for socioeconomic indicators, censuses
would yield more variation in explanatory indicators, improve the strength
of the estimations, and facilitate more sophisticated spatially explicit pol-
icy simulations. Combined with landscape metrics, more disaggregated data
can yield additional insights into the spatial composition and arrangement
of potential land-use changes resulting from investments in rural develop-
ment. In that way, spatially explicit econometric modeling can enhance the
 Simulating the Effects of Rural Development Policies 101

geographical targeting of rural development interventions and facilitate the as-

sessment of the magnitude and location of their economic and environmental
consequences.

Acknowledgments

We are grateful to Manfred Zeller and Regina Birner for useful comments and
suggestions on previous versions of this paper. The research was funded by
the German Ministry of Economic Development and Cooperation (BMZ) un-
der the Tropical Ecology Support Programme (TÖB) of the German Agency
for Technical Cooperation (GTZ). The writing of this chapter was sup-
ported by the German Research Foundation (DFG) under the Emmy Noether-
Programme.
References

1. Luc Anselin (2001) Spatial econometrics. In Badi Baltagi, editor, Companion

to Econometrics, pages 310–330. Basil Blackwell, Oxford.
2. Luc Anselin (2001) Spatial effects in econometric practice in environmen-
tal and resource economics. American Journal of Agricultural Economics,
83(3):705–710.
3. Julian Besag (1974) Spatial interaction and the statistical analysis of lattice
systems. Journal of the Royal Statistical Society, Series B 36:192–236.
4. David Bigman and Hippolyte Fofack (2000) Geographical targeting for poverty
alleviation: Methodology and applications. World Bank Regional and Sectoral
Studies Series. World Bank, Washington, DC.
5. Ian Bracken (1994) A surface model approach to the representation of
population-related social indicators. In Stewart A. Fotheringham and Peter
Rogerson, editors, Spatial Analysis and GIS, pages 247–259. Taylor & Francis,
London.
6. Alessandro De Pinto and Gerald C. Nelson (2002) Correcting for Spatial Ef-
fects in Limited Dependent Variable Regression: Assessing the Value of “Ad-
Hoc” Techniques. Paper prepared for the American Agricultural Economics
Association annual meeting, Long Beach, California.
7. Helmut J. Geist and Eric F. Lambin (2002) Proximate causes and underlying
driving forces of tropical deforestation. BioScience, 52(2):143–150.
8. Daniel Müller (2003) Land-use Change in the Central Highlands of Vietnam: A
spatial econometric model combining satellite imagery and village survey data.
Doctoral dissertation, Georg-August University Göttingen, Göttingen.
9. Daniel Müller and Darla K. Munroe (2005) Tradeoffs between rural develop-
ment policies and forest protection: Spatially explicit modeling in the Central
Highlands of Vietnam. Land Economics, 81(3):412–425.
10. Daniel Müller and Manfred Zeller (2002) Land use dynamics in the central high-
lands of vietnam: A spatial model combining village survey data and satellite
imagery interpretation. Agricultural Economics, 27(3):333–354.
11. Darla K. Munroe, Jane Southworth, and Catherine M. Tucker (2002) The dy-
namics of land-cover change in western honduras: Exploring spatial and tem-
poral complexity. Agricultural Economics, 27(3):355–369.
 Simulating the Effects of Rural Development Policies 103

12. Gerald C. Nelson and Jacqueline Geoghegan (2002) Deforestation and land use
change: sparse data environments. Agricultural Economics, 27(3):201–216.
13. Gerald C. Nelson, Virginia Harris, and Steven W. Stone (2002) Deforestation,
land use, and property rights: Empirical evidence from darien, panama. Land
Economics, 77(2):187–205.
14. Gerald C. Nelson and Daniel Hellerstein (1997) Do roads cause deforestation?
using satellite images in econometric analysis of land use. American Journal of
Agricultural Economics, 79:80–88.
15. Policy Division Committee on Global Change Research NRC (National Re-
search Council), Board on Sustainable Development (1999) Global Environ-
mental Change: Research Pathways for the Next Decade. National Academy
Press, Washington, DC.
16. StataCorp (2003) Stata Statistical Software: Release 8.0. Stata Corporation,
College Station, Texas.
Kriged Road-Traffic Maps

Hans Braxmeier1 , Volker Schmidt2 , and Evgeny Spodarev3

1
Department of Applied Information Processing, University of Ulm, Ulm,
Germany
hans.braxmeier@uni-ulm.de
2
Department of Stochastics, University of Ulm, Ulm, Germany
volker.schmidt@uni-ulm.de
3
Department of Stochastics, University of Ulm, Ulm, Germany
evgeny.spodarev@uni-ulm.de

1 Introduction
A common difficult problem of large cities with heavy traffic is the predic-
tion of traffic jams. In this paper, a first step towards mathematical traffic
forecasting, namely the spatial reconstruction of the present traffic state from
pointwise measurements is briefly described. For details, we refer to [1], where
models of stochastic geometry and geostatistics are used to spatially represent
the traffic state by means of velocity maps. A corresponding Java software that
implements efficient algorithms of spatial extrapolation is developed; see [5].
To illustrate our extrapolation method, we use real traffic data originating
from downtown Berlin. It was provided to us by the Institute of Transport
Research of the German Aerospace Center (DLR). Approximately 300 test
vehicles (taxis) equipped with GPS sensors transmit their geographic coor-
dinates and velocities to a central station within regular time intervals from
30 s up to 6 min; see Fig. 2. Thus, a large data base of more than 13 million
positions was formed since April 2001; see Fig. 1.

Fig. 1. Observed positions of test vehicles in downtown Berlin

106 H. Braxmeier et al.

In the first stage of our research, only a smaller data set (taxi positions on
all working days from 30.09.2001 till 19.02.2002, 5.00–5.30 pm, moving taxis
only) was considered. Furthermore, the observation window was reduced to
downtown Berlin to avoid inhomogeneities in the taxi positions.
The main idea of the extrapolation technique described in Sects. 2 and 3
below is to interpret the velocities of all vehicles at given time t as a realization
of a spatial random field V (t) = {V (t, u)} where V (t, u) is a traffic velocity
vector at location u ∈ R and time instant t ≥ 0. The goal is to analyze
the spatial structure of these random fields of velocities in order to describe
the geometry of traffic jams. Since V (t, u) can be measured just pointwise at
some observation points u1 , . . . , un , a spatial extrapolation of the observed
data is necessary. Notice that the velocities strongly depend on the location
and the direction of movement, e.g. the speed limits and consequently the
mean velocities are higher on highways than in downtown streets.

Fig. 2. Histogram of time intervals between consecutive GPS signals (in s)

The classical extrapolation methods of geostatistics (see e.g. [6]) either

make no use of additional directional information or provide measurements
V (t, u + ui ) and V (t, u − ui ) with equal weights. Both these features are not
adequate to the setting mentioned above. Thus, the standard extrapolation
methods had to be adapted to our specific problem. In Sect. 3, an ordinary
kriging with moving neighborhood is described that allows to extrapolate
directed velocity fields. First, the original data set should be split into four
subsets which are directionally homogeneous. A data unit (u, V (t, u)) belongs
to the data set i (i = 1, . . . , 4) if the polar angle of the vector V (t, u) lies
within the directional sector Si = {α ∈ [0, 2π) : (i − 1)π/2 ≤ α < iπ/2}.
By convention, the zero polar angle corresponds to the eastward direction on
the city map. The above data sets should be extrapolated separately for each
directional sector. This yields four velocity maps corresponding to the four
sectors of directions.
 Kriged Road-Traffic Maps 107

In what follows, the data from a given time interval, i.e. [5.00, 5.30] pm,
will be taken for extrapolation. Keeping this in mind, we shall omit the time
parameter t in further notation.
The extrapolation method described in Sects. 2 and 3 has been imple-
mented in Java, where a software library has been developed comprising the
estimation and fitting of variograms as well as the ordinary kriging with mov-
ing neighborhood; see [5]. As far as it is known to the authors, this is the
first complete implementation of such kriging methods in Java. Much atten-
tion was paid to the efficient implementation of fast algorithms. In contrast to
classical geostatistics operating with relatively small data sets, this efficiency
is of great importance for larger data sets with more than 10, 000 entries; see
[1] for details.
In Sect. 4, a numerical example is discussed which shows how the devel-
oped extrapolation technique can be applied to directional traffic data. Some
structural features of the resulting velocity maps (see Figs. 5 and 6) are also
discussed. In Sect. 5, this is combined with a statistical space-time analysis of
polygonal road-traffic trajectories which have been extracted from the original
traffic data. For example, it turns out that the distribution of the number of
segments in these traffic trajectories can be fitted quite well by a geometric dis-
tribution. The directional distribution of the segments reflects the anisotropy
of the street system of downtown Berlin, where the distribution of segment
lengths is demonstrably non-normal. Furthermore, the distributions of veloc-
ity residuals, i.e. the deviations from their means, show interesting skewness
properties which depend on the considered classes of low, medium, and high
mean velocities, respectively. A short outlook to simulation and prediction of
future traffic states is given in Sect. 6.

2 Random Fields
To model traffic maps, non-stationary random fields composed of a determi-
nistic drift and an intrinsically stationary random field of order two (residual)
are used. See e.g. the monographs [4] and [6] for details.
Let X = {X(u), u ∈ R2 } be a non-stationary random field with finite
second moment EX 2 (u) < ∞, u ∈ R2 . Then, X(u) can be decomposed into a
sum X(u) = m(u) + Y (u), where m(u) = EX(u) is the mean field (drift) and
Y (u) = X(u) − m(u) is the deviation field from the mean or residual. Assume
that {Y (u)} is intrinsically stationary of order two. Denote by
1
γ(h) = E[(Y (u) − Y (u + h))2 ] (1)
2
its variogram function. In practice, the field X can be observed in a compact
(mostly rectangular) window W ⊂ R2 . Let x(u1 ), . . . , x(un ) be a sample of
observed values of X, ui ∈ W for all i. The extrapolation method described
1
in Sect. 3 yields an “optimal” estimator X(u) of the value of X(u) for any
u ∈ W based on the sample variables X(u1 ), . . . , X(un ).
108 H. Braxmeier et al.

3 Kriging Based on Residuals

Among the variety of extrapolation techniques for non-stationary random

fields, our approach is similar to the so-called kriging based on residuals; see [4],
1
p. 190. First of all, an estimator m(u) for the drift m(u) has to be constructed.
Then, the deviation field Y ∗ = {Y ∗ (u), u ∈ R2 } defined by

Y ∗ (u) = X(u) − m(u)

1 (2)

is formed and its kriging estimator Y1 ∗ (u) is computed. Finally, the estimator
1
X(u) is given by
1
X(u) 1
= m(u) + Y1 ∗ (u) . (3)
1
If we suppose that the drift is known, i.e. m(u) = m(u) for all u, then we have
exact values of the deviation field Y (u1 ), . . . , Y (un ) since in this case

Y ∗ (u) = Y (u) = X(u) − m(u) .

Let {y(ui ) = x(ui ) − m(ui ), i = 1, . . . , n} be a realization of the sample

variables Y (u1 ), . . . , Y (un ). The extrapolation of Y (u) can be performed by
ordinary kriging making use of the variogram γ(h); see [4, 6].

3.1 The Kriging Estimator

A simpler version of the following ordinary kriging with moving neighborhood

can be found in [3], pp. 201–210 and [6], pp. 101–102. Consider the usual
indicator function 
1 if x ∈ B,
1{x ∈ B} =
0 otherwise.

Introduce the estimator Y1 (u) of Y (u) for u ∈ W as a linear combination of

the sample variables Y (ui ) with unknown weights λi = λi (u) by


n
Y1 (u) = λi Y (ui )1{ui ∈ A(u)} . (4)
i=1

The estimation involves only those sample random variables Y (ui ) that are
positioned in the “neighborhood” A(u) of u, i.e. if ui ∈ A(u). Being an arbi-
trary set, this moving neighborhood A(u) contains a priori information about
the geometric dependence structure of the random field Y . For instance, it
could be designed to model the formation of traffic jams; see Sect. 4.
Unbiasedness of the estimator introduced in (4) and minimizing its vari-
ance lead to the following conditions on the weights λi . For all i = 1, . . . , n
with ui ∈ A(u) it holds
 Kriged Road-Traffic Maps 109


n
λj γ(uj − ui )1{uj ∈ A(u)} + μ = γ(u − ui ) , (5)
j=1
n
λj 1{uj ∈ A(u)} = 1 .
j=1

To solve this system of equations, the knowledge of the variogram function

γ(h) is required. However, in most practical cases γ(h) is unknown and has
to be estimated from the data y(u1 ), . . . , y(un ).

3.2 Variograms

In this paper, the most simple and popular variogram estimator of Matheron
is used (cf. [3, 6]). It is defined by
1  2
1(h) =
γ (Y (ui ) − Y (uj )) (6)
2N (h)
i,j:ui −uj ≈h

where ui −uj ≈ h means that ui −uj belongs to a certain neighborhood U (h) of

vector h and N (h) denotes the number of such pairs (ui , uj ) for i, j = 1, . . . , n.
As shown in Fig. 4, the traffic data lead to empirical variograms that are
clearly zonally anisotropic. Below, we consider zonally anisotropic variogram
models constructed from isotropic ones (cf. [4, 6]). Put

γ(h) = γ1 (h) + γ2 (h) , (7)

where γ1 (h) is an exponential isotropic variogram model with nugget effect

a1 > 0, sill b1 and range c1 . The second term
 √


γ2 (h) = b2 1 − e− h Ch/c2 (8)

is a geometrically anisotropic exponential variogram model with sill b2 > 0

and a further parameter c2 > 0. For a vector h = (h1 , h2 ), the quadratic form
 
h Ch = λ2 h21 + λ1 h22 + (λ2 − λ1 ) (h22 − h21 ) cos2 α − h1 h2 sin(2α)

depends on two scale parameters λ1 , λ2 and a rotation parameter α ∈ [0, 2π).

Let γ1(h) be an empirical variogram estimated from the observed data
{y(ui )} for the field Y and γβ (h) the theoretical parametric variogram model
considered in (7) with parameter vector β = (a1 , b1 , c1 , b2 , λ1 /c22 , λ2 /c22 , α). In
practice, only a finite number m of values γ 1(hm ) can be computed.
1(h1 ), . . . , γ
In the case of traffic data, the classical least squares method is employed to
fit γβ to γ1. Since traffic data is substantially anisotropic, the variogram model
(7) has to be fitted to the data on the whole grid as well as in two directions
with polar angles α and α + π/2.
110 H. Braxmeier et al.

3.3 Drift Estimation

The mean field {m(u)} can be estimated from the data by various methods
ranging from radial extrapolation to smoothing techniques such as moving
average and edge preserving smoothing. In what follows, the moving average
is used because of its ease and computational efficiency for large data sets. By
moving average, the value m(u) is estimated as
1 
1
m(u) = X(ui ) (9)
Nu
ui ∈W (u)

where W (u) is the “moving” neighborhood of location u and Nu denotes the

number of measurement points ui ∈ W (u). For fast computation, we put
W (u) to be a square with side length τ centered in u.

3.4 Residuals formed with estimated drift

In the previous sections, we supposed that the drift m(u) is explicitly known.
However, if it has to be estimated from the data, the theoretical background
for the application of the kriging method breaks down (cf. [3], pp. 122–125,
[4] p. 72, [6], p. 214). Nevertheless, practitioners continue to use the ordinary
kriging of residuals with estimated drift based on the data y ∗ (ui ) = x(ui ) −
1 i ), i = 1, . . . , n legitimized by its ease and satisfactory results.
m(u

4 Extrapolation of the Velocity Field

In what follows, the extrapolation method of Sect. 3 is applied to real traffic
data of the directional sector S2 = {α : π/2 ≤ α < π}. This partial data set
contains 19699 entries of taxis moving northwest collected over 90 days.
In Fig. 3, the northwest movement direction of the taxis can be clearly
recognized in the mean velocity field {m(u)}. 1 Grey tones reflect speed vari-
ation. It clearly shows that the estimator m 1 preserves the spatial velocity
structure. To estimate the variogram γ ∗ of Y ∗ , the empirical variogram γ 1i∗
is computed
90for each day i = 1, . . . , 90 and then averaged over all days, i.e.
1∗ (h) =
γ 1i∗ (h) /90. The empirical variogram γ
i=1 γ 1∗ (h) with “maximum
range” in northwest direction and “minimum range” in orthogonal direction is
zonally anisotropic; see Fig. 4. The main directions of anisotropy are closely
connected to the road directions in downtown Berlin. See Sect. 5 and especially
Fig. 10(a) for details.
The zonally anisotropic variogram model (7) with two fixed parameters
α = 170◦ , λ1 /c22 = 1000 has been fitted to the empirical one. The classical
least squares fitting method applied to one-dimensional vertical slices of the
empirical variogram in orthogonal directions α = 80◦ and α = 170◦ yields
the remaining parameter values a1 = 31.772, b1 = 116.211, c1 = 245388.671,
 Kriged Road-Traffic Maps 111

1
Fig. 3. Mean field m(u) of data set 2

1∗ (h) and fitted variogram model γ ∗ (h) (level curves)

Fig. 4. Empirical variogram γ

b2 = 22.634, λ2 /c22 = 683964.794. Thus, the range values in directions 170◦

and 80◦ are r1 = 270 m and r2 = 162 m, respectively. It means that the
velocities of two vehicles separated by distances 3r1 = 810 m in horizontal
direction and 3r2 = 486 m in vertical direction are almost independent. These
range values are conform with the results stated in Sect. 5 for the typical
distance between two subsequent positions of the same test vehicle.
For extrapolation, the sample of velocities x(u1 ), . . . , x(un ) (n = 223) ob-
served on Monday, 18.02.2002 is used. Compared to the whole data set 2 rep-
resenting the “past”, it is interpreted as “actual” data. The random field Y ∗ of
deviations from mean velocities is extrapolated by kriging with moving neigh-
borhood (4) using the indicator function 1{ui ∈ A(u)} = 1{ϕ(ui − u) ∈ S2 }
where ϕ(ui − u) is the polar angle of the vector ui − u. This assumption is
rather intuitive since only those measurements at locations ui lying “ahead”
of the current position u can influence its velocity value.
The extrapolated residuals Y1 ∗ (u) and the resulting velocity map {X(u)} 1
are shown in Figs. 5 and 6, respectively. Due to the particular asymmetric
112 H. Braxmeier et al.

Fig. 5. Residual field Y1 ∗ (u)

1
Fig. 6. Velocity field X(u)

form of the indicators, the extrapolated field of residuals is strongly disconti-

1
nuous. Discontinuities of the realizations of {X(u)} caused by the kriging with
moving neighborhood are essential for precise localization of traffic-jam areas.
In Fig. 5, most of the deviation values are zero. The routes of taxis driving in
the streets are marked by peaks of the field Y1 ∗ with subsequent tails of non-
zero residual velocity values lying behind. Thus, one can distinguish separate
routes of different test vehicles. See also the extracted taxi routes in Fig. 8,
which are similar to those shown in Fig. 5.
 Kriged Road-Traffic Maps 113

1
Fig. 7. Traffic jams: X(u) ≤ 15 kph

Fig. 8. Taxi routes on 18.02.2002, from 5.15 to 5.30 pm

1
In Fig. 7, areas with velocities X(u) ≤ 15 kph are marked grey. Some
of these regions might be caused by traffic jams, others are regions with low
average velocities. Indeed, the most likely velocity value in downtown Berlin
is about 20 kph as it can be seen in Fig. 12.

5 Statistical Analysis of Traffic Data

In addition to the spatial statistical inference performed above, we now discuss

the histograms of velocity residuals and further traffic characteristics which
114 H. Braxmeier et al.

bring an extra insight into the structure of traffic data. In particular, they
help us to explain some features of anisotropy and spatial correlation which
we already mentioned in Sect. 4. For a more detailed treatment of the subject,
see [2].

5.1 Distributional Properties of Polygonal Taxi Routes

If we think about the way the traffic data are collected we understand that the
locations where the velocities are measured can not be deterministic. More-
over, they are stochastically dependent. In fact, each test vehicle follows a
route that consists of a random number of segments. Each segment connects
two locations where consecutive GPS signals were sent; see Fig. 8. The his-
togramm of the number of segments in the taxi routes is shown in Fig. 9.
It turns out that this histogram can be well approximated by a geometric
distribution with parameter p = 0.9365064 being the probability of enlarg-
ing a route by a new segment. Furthermore, the geometry of the taxi routes
explains the form of the variogram anisotropy mentioned in Sect. 4.
In particular, the distribution of the angles between the movement direc-
tion of a vehicle and the eastward direction in Fig. 10(a) reflects the distri-
bution of typical street directions with heavy traffic in downtown Berlin. The
majority of main roads goes east or west which corresponds to the angles of
0◦ , 180◦ , and 360◦ , respectively. This is certainly the reason for the charac-
ter of zonal anisotropy of the variograms in Fig. 4. Figure 10(b) shows that
the distribution of segment lengths is demonstrably non-normal. Furthermore,
with probability of ca. 0.9, the distances between two subsequent GPS signals
in the taxi routes do not exceed 1000 m. It is clear that the velocities at two
positions within this distance are correlated. The opposite statement is also
true. As it has been already mentioned in Sect. 4, the velocities of two cars

Fig. 9. Histogram of the number of segments in the taxi routes

 Kriged Road-Traffic Maps 115

(a) Directions of route segments (b) Lengths of route segments

Fig. 10. Histograms of segment directions (in degrees) and lengths (in m)

at a distance of more than 3 r12 + r22 ≈ 945 m from each other are almost
independent.

5.2 Distribution of Velocity Residuals

The histogram in Fig. 11 shows that the distribution of velocity residu als
can be well fitted by some normal distribution. Nevertheless, a more detailed
statistical inference shows that the distribution of velocity residuals depends
on the value of mean velocity. One reason for this is that the sum of the
residual and the mean has to be non-negative. Figures 13, 14 and 15 show
the histograms of the velocity residuals measured at locations with mean
velocities (in kph) belonging to three disjoint classes: [15, 20), [25, 30) and
[40, 45), respectively.

Fig. 11. Histogram of the velocity residuals (in kph)

116 H. Braxmeier et al.

(a) for the first route segments (b) for the remaining route segments

Fig. 12. Histogram of velocities (in kph)

Fig. 13. Histogram of residuals

1 ∈ [15, 20)
given that m

The right skewness of this his-

togram means that large positive
deviations from small mean values
are more likely than negative ones.

Fig. 14. Histogram of residuals

1 ∈ [25, 30)
given that m

This histogram is almost symmet-

ric. Thus, both positive and neg-
ative residuals of equal size occur
nearly with the same probability.
 Kriged Road-Traffic Maps 117
Fig. 15. Histogram of residuals
1 ∈ [40, 45)
given that m

It is clear from this histogram that

at positions with large mean values
large negative residuals are more
likely than positive ones.

If we add the mean velocity values to their residuals we see that most
velocities in downtown Berlin do not exceed 60 kph. The histogram of the
velocities themselves is given in Fig. 12 which shows that the most likely
velocity value in downtown Berlin (i.e., the modus of the empirical velocity
distribution) is about 20 kph. This explains the dominance of low velocity
values in the mean field m1 and the threshold maps in Figs. 3 and 7.

6 Outlook
The spatial extrapolation and statistical space-time analysis of traffic data
considered in the present paper is an important step towards stochastic mod-
elling, simulation and prediction of future road-traffic states. Our results can
be used to construct a Markov-type simulator by means of which future routes
of test vehicles can be generated, where the choice of the starting configura-
tion depends on the actually measured traffic situation. In particular, when
sampling the velocity residuals from histograms as given in Figs. 13, 14 and
15, the mean velocity field {m(u)}
1 will be actualized by the recently, say
at the given day, observed velocities. For example, suppose that significantly
larger velocities than usually have been observed at the considered day in a
certain neighborhood of location u. In this case, the velocity residual at loca-
tion u will be sampled from a histogram which corresponds to a larger class
1
of mean velocities than the “historical” value m(u). Further details concerning
our simulation algorithms can be found in [2].
Then, using the extrapolation technique described in Sects. 2 and 3, ve-
locity maps based on both the measured and simulated traffic data can be
computed. To evaluate the quality of these maps, they are compared with cor-
responding velocity maps computed exclusively from measured traffic data.
The comparison is based on morphological distance measures for digital image
data. These issues will be discussed in a forthcoming paper.
118 H. Braxmeier et al.

Acknowledgement

This research was supported by the German Aerospace Center (DLR) through
research grant 931/69175067. The authors are grateful to Reinhart Kühne, Pe-
ter Wagner and their co-workers from the DLR Institute of Transport Research
for suggesting the problem as well as for stimulating and fruitful discussions
on the subject.
References

1. Braxmeier H, Schmidt V, Spodarev E (2004) Spatial extrapolation of

anisotropic road traffic data. Image Analysis and Stereology 23, 185–198.
2. Braxmeier H, Spodarev E, Schmidt V (2005) Statistische Raum-Zeit-
Analyse und Simulation von Verkehrsströmen in Ballungsgebieten. Verkehrs-
forschung - Online 2, http://www.verkehrsforschung-online.de
3. Chilès JP, Delfiner P (1999) Geostatistics: modelling spatial uncertainty.
Wiley, New York
4. Cressie NAC (1993) Statistics for spatial data. Wiley, New York
5. GeoStoch (2008) Java library. University of Ulm, Department of Applied In-
formation Processing and Department of Stochastics, http://www.geostoch.de
6. Wackernagel H (1998) Multivariate geostatistics. 2nd ed., Springer, Berlin
On Evaluation of Precipitation Fields with Rain
Station Data

Bodo Ahrens

Institut für Meteorologie und Geophysik, Universität Wien, Wien, Austria

Bodo.Ahrens@univie.ac.at

1 Introduction
Nowadays, limited area numerical weather prediction models provide meteo-
rological forecasts with horizontal grid spacing of only a few kilometers and
grid spacing will decrease further in the coming years caused by progress in
high-performance computing [9, 26]. Precipitation forecasts are of primary in-
terest for both researchers and the public. For example, in flood forecasting
systems precipitation is the crucial input parameter, especially in mountain-
ous watersheds. Like the grid spacing of weather prediction models the grid
spacing of regional climate models is decreasing.
Precipitation forecasts have to be evaluated and errors have to be quanti-
fied. The most important evaluation method is comparison of meteorological
simulation results with meteorological observations. But, before errors can be
quantified two decisions have to be made. First, a set of useful statistics has
to be chosen. This shall not be the issue of this paper. The interested reader
is referred to, for example, Murphy and Winkler [23], Wilks [31], Wilson [32].
We apply for illustration a small set of simple continuous statistics.
Our focus is on the second problem: What is the observational reference?
Rain station data is commonly preferred to remote sensing data, in partic-
ular radar data, because of the relatively large measurement uncertainties
[e.g., 1, 14, 34]. Is it reasonable to compare precipitation forecasts valid for
grid boxes with several kilometers in diameter with sparsely distributed rain
station data valid for small areas of ∼ 1000cm2 ? This is often done in an op-
erational framework since it can be implemented by simple means. This area-
to-point evaluation is criticized and it is proposed to perform some upscaling
or regionalization of the station data up to forecast grid resolution [29, 12].
Regionalization can be done by some fitting approach yielding a precipita-
tion analysis. For example, a recent analysis of precipitation for the European
Alpes by Frei and Häller [17] has a time resolution of 24 h and a spatial grid
of about 25 km with regionally even lower effective resolution depending on
the available surface station network. This type of analysis is useful for model
122 B. Ahrens

validation at the 100 km-scale [see, e.g., 5, 16, 18], but not at 10 km-scale or
even less.
Analysis is a smoothing regionalization. This deteriorates application in
higher-moment statistics if the network is not dense enough. The statement
“dense enough” critically depends on the applied pixel support (is a pixel
value representative for boxes with diameter of ∼100, 10, or 1 km?) and the
analysis scheme. Another regionalization approach is stochastic simulation of
precipitation fields with conditioning on the available station data. The idea
of this is that the data is respected and the spatial variability is represented
more realistically than in the analysis. Then the forecast can be compared
with an ensemble of simulated fields. The ensemble mean field is an analysis
but the mean higher-moment statistics have not the same value than if the
forecast is just compared with the analysis alone.
This paper applies regionalization and performs area-to-point or area-to-
area comparison in evaluation of daily precipitation forecasts. The forecasts to
be evaluated by example are the forecasts of the NWP model ALADIN that is
operational at the Austrian national weather service with 10 km grid spacing.
ALADIN, the forecast days, and the available station data are introduced in
the next section. Section 3 discusses the applied evaluation approaches and
subsequent sections discuss the respective results. Finally, some concluding
remarks will be given.

2 Precipitation Events and Data

For illustrational purposes we investigate the August 2002 flood in Austria.

It was caused by two devastating, large scale rain events. The first rain event
(6–8 August) was dominated by an upper-air low over central Europe. This
flow pattern lead to torrential rainfall, especially on the windward-side of the
northern Alps and along the Austrian–Czech border. The second rain event
(11–13 August) was caused by strong cyclogenic activity over the Mediter-
ranean sea resulting in a south-easterly flow of warm and moist air into Cen-
tral Europe. Here, the daily precipitation fields over Austria for August the
6, 7, 11, and 12th are of interest. An investigated day starts at 06 UTC and
finishes at 06 UTC the next day.
Observational rain data is available from two sets of rain station data. The
first set is provided by the Hydrologische Zentralbüro with about 800 stations.
This set is named HZB in the following. The second data set is provided by the
Austrian national weather agency ZAMG with about 116 stations measuring
during the 4 days of evaluation. This second data set, named TAWES, is
independent from the HZB data set and generated by automatic weather
stations and available in near real time. Within this paper the daily time
scale is applied. Thus the TAWES data is accumulated to daily values.
This paper compares observational precipitation data with forecast pre-
cipitation fields from a numerical weather prediction (NWP) model. This
 On Evaluation of Precipitation Fields 123

comparison is a crucial step of model evaluation. Here, we apply forecast

fields simulated with the NWP model ALADIN (Aire Limitée Adaptation
Dynamique dévelopement InterNational, see, e.g., Bubnova et al. [11], Ahrens
et al. [4] and http://www.cnrm.meteo.fr/aladin/) in the setup operational in
the year 2002 by the Austrian national weather service with about 10 km
horizontal grid spacing. Thus, the precipitation forecast values are valid for
10 × 10 km2 blocks. Evaluated are sequences of 30-h forecasts initialized at
00 UTC discarding the leading 6 hours to account for model spin-up. Previ-
ous investigations have already shown that ALADIN quantitative precipita-
tion forecasts are useful [e.g., 2, 21, 22].

3 Evaluation Methods
In the following we will discuss the evaluation procedures applying a min-
imal set
Nof useful statistics. Most important is the mean distance bias =
1/Nx x=1 x
(mx − dx ) with the model forecast field mx , the observational field
dx , and with the space index x = 1, . . . , Nx . Additional statistics considered
are the coefficient of determination R2 (the square of the linear product-
moment correlation, possible values are between 0 and 1 with optimal value 1),
2
and the ratio of spatial variances SPREX = σm /σd2 (optimal value 1).
The applied evaluation methods are comparisons of model fields with (a)
station data (i.e., effectively with point data) and (b) with regionalized and
box averaged precipitation fields (i.e., with area data). Comparison with point
data is often done and, for example, standard in most European Meteorologi-
cal Services [see 10, 32]. It is simple in implementation. Two variants are com-
mon practice: direct comparison of the station data with the closest model grid
box values and thus performing an area-to-point comparison, or interpolation
of the model fields to the station locations and thus performing a point-to-
point comparison. In fact this interpolation smoothes the forecast field that is
eligible since single box values should not be interpreted [2, 20]. But, on the
other hand a simple interpolation like the often applied bi-linear interpolation
assumes that the precipitation field is continuous and introduces no additional
information. Consequently, interpretation of the interpolated values as point
data is delusive. The effective resolution of ALADIN is not the issue here,
and the raw forecasts of ALADIN with about 10 km horizontal resolution are
evaluated by example.
Comparison of model grid box output with regionalized rain fields with ap-
propriate pixel support is an area-to-area comparison and respects the scales.
The second potential advantage of regionalization is that station representa-
tivity problems (clustering of stations around larger cities or along valleys)
can principally be compensated. Here, we call regionalization by some opti-
mization involving data-fitting techniques (like regression, polynomial fitting,
spline functions, kriging, etc.) analysis and the estimated field is an analysis
field. A problem of analysis is that it is difficult to estimate analysis errors
124 B. Ahrens

(e.g., Kriging variances underestimate the analysis error in case of precipita-

tion since the Kriging assumptions are not fulfilled). A second problem is that
analysis fields are expected to underestimate the true field variance (e.g., the
smoothing relationship of Kriging states that the analysis variance at any lo-
cation is the data variance minus the kriging variance). These problems have
to be taken into account in an evaluation investigation.
Here, the details of the analysis scheme are of minor importance and
two schemes are applied. First, ordinary block Kriging with spherical vari-
ogram model and, second, inverse squared-distance weighting interpolation
with blocking is applied. Chosen blocks are 10 km in diameter and thus pixel
support of the analysis is 10×10 km2 like of the NWP model forecasts. In case
of HZB data analysis a neighborhood of 64 stations and in case of TAWES
data of 8 stations is considered. Kriging variants are often proposed and ap-
plied in precipitation analysis [7, 8, 15, 19]. Tests have shown that ordinary
Kriging with spherical variogram, whose parameters are estimated from the
actual data, performs slightly better than with other tested variograms or with
universal Kriging in the events investigated. Figure 1 shows the block kriging
results with pixel support of 10 km for the four days of comparison. Figure 2
shows the model forecasts. Obviously, the model performance is worst at day
one and best at day four. In the following we discuss the quantification of this
subjective conclusion.
Another regionalization approach is stochastic simulation. Here, the use-
fulness of precipitation field simulation conditioned on available data shall be
shown by application of Gaussian conditional simulation [e.g., 13, Chap. 7]
with 10 km blocks. The conditioning respects the station values and the esti-
mated variograms.

550 HZB analysis, block 10 km, 06.08.2002 550 HZB analysis, block 10 km, 07.08.2002
−120 −120
500
northing [km]

500 −100 −100

northing [km]

450 −80 450 −80

400 −60 400 −60
−40 −40
350 350
−20 −20
300 300
−0 −0
200 300 400 500 600 200 300 400 500 600
easting [km] easting [km]
550 HZB analysis, block 10 km, 11.08.2002 –120 550 HZB analysis, block 10 km, 12.08.2002
−140
–100 500
northing [km]

500 −120
northing [km]

–80 −100
450 450
–60 −80
400 400 −60
–40
350 350 −40
–20 −20
300 300
–0 −0
200 300 400 500 600 200 300 400 500 600
easting [km] easting [km]

Fig. 1. Daily analyses by ordinary block Kriging of HZB rain station data in Austria
for the four days investigated in this paper. The dots in the upper left panel indicate
the station locations. Units are mm/d
 On Evaluation of Precipitation Fields 125

550 ALADIN forecast, 06.08.2002

–160 550 ALADIN forecast, 07.08.2002 −120
500 –140 500 −100
northing [km]

northing [km]
–120
450 –100 450 −80

400 –80 400 −60

–60 −40
350 –40 350
–20 −20
300 300
–0 −0
200 300 400 500 600 200 300 400 500 600
easting [km] easting [km]
550 550 ALADIN forecast, 12.08.2002 −180
ALADIN forecast, 11.08.2002 −140 −160
500 −120 500

northing [km]
−140
northing [km]

−100 −120
450 450
−80 −100
400 −60
400 −80
−60
350 −40 350 −40
−20 −20
300 300
−0 −0
200 300 400 500 600 200 300 400 500 600
easting [km] easting [km]

Fig. 2. Operational NWP model forecasts for the days shown in Fig. 1. Units are
mm/d

Precipitation is a non-Gaussian, non-negative process and, therefore, the

chosen simulation method is far from optimal. There are unconditional simu-
lation methods described in the literature [e.g., 3, 30], but there is no appro-
priate conditional approach known to the author. Additionally, normal score
transformation of the precipitation data has not been proven superior to di-
rect simulation. Here, the advantages of regionalization by simulation shall
be discussed and the mentioned deficiencies are not crucial for the presented
conclusions. The made assumptions are the same as in Kriging. The statistics
calculated with the simulated fields are determined by the raw data, i.e. if
simulated with negative precipitation data. In Fig. 4 (negative values set to

−120
550 TAWES analysis, block 10 km, 12.08.2002

−100
500
northing [km]

−80
450
−60
400
−40
350
−20
300
−0
200 300 400 500 600
easting [km]

Fig. 3. Daily analysis by ordinary block Kriging of TAWES rain station data. Units
are mm/d
126 B. Ahrens

550 TAWES simulation, 12.08.2002

−120

500 −100
northing [km]

450 −80

−60
400
−40
350
−20
300
−0
200 300 400 500 600
easting [km]

Fig. 4. Realization of a conditional simulation to the daily accumulated TAWES

data applied in Fig. 3. Units are mm/d

zero) an example of an simulated precipitation field conditioned on TAWES

data is compared with the Kriging analysis of these data. The simulated field
is more heterogeneous with slightly larger maximum values. It is assumed that
a more realistic simulation approach would realize even more variable fields.
Kriging, inverse distance weighting interpolation and Gaussian simula-
tion are performed with the geostastistical software package gstat [24, and
www.gstat.org in the World Wide Web].

4 Evaluation Against Station Data

In a first evaluation step the comparison of model grid box data with rain
station data is discussed. The nearest model grid box is used to compare with
the point observations ignoring the corresponding error in location. Figure 5
and Table 1 present the comparison with the HZB and TAWES data set.
There is a large scatter in the results depending on the applied reference: the
TAWES or HZB data set. For example, the relative bias is +4% in comparison
with TAWES and −4% in comparison with HZB data at August 7th. At the
same day the forecast explains 39% of the TAWES data variability (R2 =
0.39) but only 20% of the HZB data variability. The model underestimates
the field variance at the 11th if compared with HZB data by 10% or rather
overestimates by 10% in comparison with TAWES data.
The impact of the station sample size is illustrated by the box plots in
Fig. 5. Twenty random sub-samples of 116 stations (the sample size of the
TAWES set) are drawn from the HZB set. These sub-samples are applied in
the evaluation process and the box plots show the quartiles of the twenty
evaluation results for each day and statistics. Twenty is a small number of
random sub-samples, but enough to illustrate the effects. The range of these
 On Evaluation of Precipitation Fields 127

100

3.0
60

2.5
80
40

2.0
20

SPREX [1]
60
R^2 [%]
bias [%]
0

1.5
40
–20

1.0
–40

20

0.5
–60

0
6. 7. 11. 12. 6. 7. 11. 12. 6. 7. 11. 12.
August 2002 August 2002 August 2002

Fig. 5. Comparison of NWP model forecasts against station data with the symbols
“+” indicating the evaluation statistics of forecast vs HZB station data for the four
days and with “×” indicating the comparison results against the TAWES data.
The box plots show the quartiles (the whiskers indicate the range) of results of
comparison against 20 random subsets with sample size 116 from the HZB data

Table 1. Comparison of forecasts by the NWP model against HZB or TAWES

station data by statistics. The given values are valid for August 6/7/11/12, respec-
tively

bias [%] R2 [%] SPREX [1]

NWP∼HZB 54/ − 4/ − 8/ − 9 2/16/20/44 1.4/0.5/1.3/0.9
NWP∼TAWES 65/4/ − 1/ − 13 5/19/39/52 1.6/0.6/1.6/1.1

results is substantial. For example, the relative bias range is about 20% for the
days with small bias and even 43% for August 6th. Interestingly, the TAWES
results are not within the interquartile range of the sub-sampling results most
of the time and the difference is systematic (besides the bias at August 12th).
The extremness can be explained by a more homogeneous distribution of the
TAWES stations (cf. Fig. 3) in comparison with the sub-sampled HZB stations
and more important by the different measurement system. The problem of
rain measurements can not be discussed further and the interested reader is
referred to [25, 33].
Instead of the next model box value often a bi-linearly interpolated value
is compared with station observations as discussed above. The bias results
are different to the results with next neighbor comparison, but within a scat-
ter range suggested by the box plots. The results of the pattern comparison
improves slightly but systematically (up to 5%). This is not surprising since
the model data is implicitly smoothed by the interpolation and thus the clas-
sic “double-penalty problem” [small location discrepancies of sharp peaks are
penalized twice, cf. 6] is reduced. Smoothing reduces the model variance and
thus the values of SPREX decrease (by 0.05 for the 7th to 0.2 for the 12th in
case of TAWES data).
128 B. Ahrens

5 Evaluation Against Analyses

With analyzed data it is possible to compare NWP forecast fields with analysis
fields at the same scale (pixel support, grid spacing, and coverage). Here, as
mentioned above, we apply ordinary block Kriging (OK) and inverse distance
weighing interpolation (IDW) with a block-size comparable to the box area
of the forecast model of 10 × 10 km2 .
Table 2 shows values of the statistics of comparison of forecast fields
(NWP) with analyses and intercomparisons of analyses. The scatter in the
statistics’ values due to analyzing the HZB or TAWES set in NWP model
evaluations decreases slightly in comparison with evaluation against station
data. This is supported by Fig. 6. The box-plots show the quartiles of values
from twenty comparisons of analyses by OK with the forecasts and the rela-
tive range in the statistics is smaller (at the 11th the absolute range increased
but the median approximately doubled).

Table 2. Comparison of NWP forecast fields with different analyses and of analyses
with analyses. Analyses considered are done by Kriging (OK) or inverse distance
weighting interpolation (IDW) based on different sets of rain station data (TAWES
or HZB). The last row shows the mean statistics of twenty comparisons of analyses
based on HZB sub-sets versus the total HZB data set
bias [%] R2 [%] SPREX [1]
NWP∼ OKHZB 73/ − 14/ − 0/ − 11 2/23/21/47 1.7/0.5/2.2/1.0
NWP∼ OKTAWES 70/ − 14/ − 1/ − 8 1/25/30/51 1.8/0.5/2.7/1.1
NWP∼ IDWHZB 73/ − 16/1/ − 12 2/23/23/50 2.1/0.5/3.1/1.2
NWP∼ IDWTAWES 71/ − 11/ − 2/ − 8 1/24/32/54 2.0/0.5/3.2/1.3
OKTAWES ∼ OKHZB 2/0/0/ − 4 84/93/62/87 0.9/1.0/0.8/0.9
IDWTAWES ∼ OKHZB 2/ − 4/2/ − 3 84/92/60/83 0.8/0.9/0.7/0.8
OKSS ∼ OKHZB  −2/2/ − 1/1 79/93/55/81 0.8/1.0/0.8/0.9

There are differences if NWP is compared against IDW or OK analyses.

The values for R2 and SPREX are larger with IDW since the IDW analyses
are smoother than the OK analyses as is quantified by SPREX in row 5
and 6 of Table 2. This is reasonable since the influence of distant data is
described in OK by the applied variogram that increases close to the origin
faster than quadratic which is assumed in IDW. The effective pixel support
is larger for IDW analysis than for Kriging analysis. Thus, even if formally
the pixel support is appropriate (both analyses are estimated for 10 × 10 km2
blocks) the second moment statistics are sensitive to the effective smoothness
(cf. analysis vs analysis SPREX values in Table 2 or the Fig. 3 in comparison
with the lower-right panel of Fig. 1). Keeping this in mind the area-to-point
comparison of the last section is not fair to the forecasts: R2 and SPREX are
too small in the mean.
Since the assumptions of Kriging are not fulfilled very well by precipitation
data it is not sure a-priori that Kriging is superior to IDW. Nevertheless, the
 On Evaluation of Precipitation Fields 129

impact of the analysis scheme is smaller than the impact of data sample size.
This is evident if, for example, the coefficient of determination is more closely
inspected for the third analysis day, August 11th. Both analyses based on
TAWES data explain only about 60% of the variance of the HZB analysis.
This day is less dominated by large-scale precipitation patterns and shows
the smallest field variance, but has the largest small-scale variability. Small
shifts in the analyzed small-scale pattern lead to more distinct double-penalty
effects than for the other days. And these shifts are less influenced by the
analysis method as the similar R2 values indicate than by the smaller data
sample size. This conclusion is supported by the last row in Table 2 where
the mean error of analyses based on random HZB subsets with sample size
116, as of the TAWES set, is shown. Nevertheless, a value of 60% is enough
if compared to the value of 21% explained by the NWP. What is missing is a
possibility to judge these 21% in comparison to the 23% for the 7th where the
precipitation field is far more easy to analyze as the more than 90% explained
variance indicates and thus should also be easier to forecast in the sense of
the applied statistics.
100

3.0
60

2.5
80
40

2.0
20

SPREX [1]
60
R^2 [%]
bias [%]
−20 0

1.5
40

1.0
−60 −40

20

0.5
0

6. 7. 11. 12. 6. 7. 11. 12. 6. 7. 11. 12.

August 2002 August 2002 August 2002

Fig. 6. Comparison of NWP model forecasts against analyses. The symbols “+”
indicate the evaluation statistics of forecast vs HZB based analysis for the four
days and “×” indicates the results vs TAWES based analysis. The box plots show
the quartiles (the whiskers indicate the range) of results of comparison against 20
analyses based on random HZB subsets with sample size 116

Table 3. Statistics’ mean of comparisons of NWP forecasts with twenty simula-

tions conditioned on TAWES data and of the simulations with the HZB or TAWES
analyses
bias [%] R2 [%] SPREX [1]

NWP ∼ SITAWES  69/ − 13/ − 2/ − 11 1/25/24/48 1.6/0.5/2.2/1.0

SITAWES ∼ OKHZB  2/1/1/ − 2 72/85/48/70 1.0/1.1/1.0/1.1
SITAWES ∼ OKTAWES  −0/0/1/2 84/91/77/90 1.2/1.1/1.2/1.2
130 B. Ahrens

6 Evaluation Against Simulations

Analysis underestimates spatial variability. The idea of stochastic simulation

is that the stochastic realizations show a good representation of the natural
field variability. Here, as mentioned above, Gaussian simulation with 10 ×
10 km blocks is applied. This is a oversimplified approach, but the principal
advantages can be discussed.
Figure 4 shows a realization of a simulation conditioned on TAWES data
that is obviously more variable than the corresponding analysis. Consequently,
R2 (due to the double-penalty effect) and SPREX of NWP in comparison with
simulated fields (Table 3) are smaller in the mean than in comparison with the
TAWES analysis (Table 2) and closer in the mean to the comparison with the
HZB analysis. This can also be seen in Fig. 7 through the systematic shifts of
the box plots from the TAWES to the HZB analysis values. In case of SPREX
the interpretation is that the natural variability is indeed more realistically
represented in the simulations than in the TAWES analysis. In case of R2 the
double penalty effect is larger in the realizations than in the TAWES analysis.
Since the simulation R2 s are rather close to the HZB R2 s we conclude that
the pattern at scales finer than the effective TAWES analysis resolution are
not forecast.
Also exciting are the comparisons of the simulated realizations against
the analyses in Table 3. In the mean the TAWES simulations explain only
about 50% of the HZB analysis variance for August 11th, but more than
70% for the other days. The 11th is also the day with smallest mean R2 if
the simulated fields are compared with the TAWES analysis. Therefore, the
available TAWES data are a relatively small constraint for the simulations for
this day, and the regionalizations, the simulations or the analysis, and thus the
evaluation results for this day are most uncertain. Additionally, small-scale
variability is important that is generally not very well forecasted by NWP
100

3.0
60

2.5
80
40
20

2.0
SPREX [1]
bias [%]

60
R^2 [%]
0

1.5
40
–60 –40 –20

1.0
20

0.5
0

6. 7. 11. 12. 6. 7. 11. 12. 6. 7. 11. 12.

August 2002 August 2002 August 2002

Fig. 7. Same as Fig. 6, but the box plots summarize the comparison of NWP
forecasts against precipitation simulations conditioned on TAWES data
 On Evaluation of Precipitation Fields 131

models due to predictability constraints. Thus, the forecast of August 11th is

more valuable than the forecast of the 7th. The simulations complement the
evaluation against regionalized precipitation fields.

7 Conclusions

Precipitation forecasts by numerical weather prediction or climate model pre-

cipitation forecasts have to be evaluated. Often this is done on the basis of
precipitation rain station data. This paper investigates the effect of different
scales and representativity of the compared data. As an evaluation test bed
the NWP model ALADIN with 10 km gridspacing is evaluated for four heavy
precipitation days in August 2002 by comparison with precipitation data mea-
sured with about 800 stations operated by the Austrian hydrological service
(HZB data set) and with a independent set of 116 stations operated in near
real-time by the Austrian weather service (TAWES data set). Thus, about
100 stations are available in Austria (total area: 84 000 km2 ) for a day-to-day
evaluation procedure. Somewhat less namely 72 quality controlled stations
are available in Austria for a climatic period of time (1948–2002, cf. Schöner
et al. [27]).
The experiments show that the uncertainty in bias evaluation is of the
order of the bias if the forecasts are compared with a station data set of
sample size 116. This scatter can be reduced by best estimate regionalization,
named analysis, of the data. A Kriging variant and inverse distance weighting
interpolation are applied. The dependence on the analysis scheme is smaller
than on data reduction by about a factor of eight.
It is shown that direct comparison of forecast fields with 10 × 10 km2 pixel
support to station data with about 0.1 × 0.1 m2 is inappropriate if second-
moment statistics shall be evaluated. This is exemplified by calculation of a
measure for pattern matching, R2 , and a ratio of variances SPREX. Compari-
son with analyses based on the reference station data set HZB and the smaller
data set TAWES proved the positive impact of regionalization by analysis with
10 × 10 km2 analysis blocks. Nevertheless, there is scatter in the results and,
especially, the values for SPREX show that analysis is a smoothing process
that may not provide a good representation of the variability, particularly in
regions with sparse observations coverage. Here, regionalization by stochastic
simulation conditioned on the station data complements the interpretation of
the evaluation results. The idea of stochastic simulations is that they represent
the natural field variability more realistically. This is confirmed by compar-
ison of TAWES analyses and simulations conditioned on TAWES data with
the HZB analyses which are the available reference.
Therefore, the following recipe for forecast evaluation with rain station
information is proposed: (1) compare with analyzed fields at the same scale,
and (2) apply stochastic simulation conditioned on the data. Step one is ap-
plied, for example, in Ahrens et al. [5], Ferretti et al. [16], Frei et al. [18] at
132 B. Ahrens

grid-scales of about 100 km in space and 24 h in space with analyses based

on denser data than assumed here to be available in near real-time. At these
relatively coarse scales the analyses represent the first and second moments
of the precipitation field with sufficient accuracy. At finer scales with even
less data like in the setup assumed here analyses are insufficient, but can be
complemented by stochastic simulations. Stochastic simulation is an efficient
method for detection and eventually avoidance of difficulties with analyses.
This paper discusses spatial representativity of rain station data in eval-
uation of NWP or regional climate model forecast. Here, only the horizontal
representativity issue is considered. In the mountainous areas the inhomoge-
neous distribution of stations in the vertical (most stations at valley floors)
can lead to systematic errors that are difficult to consider [e.g., 28]. A further
problem is systematic errors due to wind and evaporation loss of the rain
gauges [e.g., 25]. These are additional difficulties which should be considered,
but are often neglected in precipitation evaluation.

Acknowledgements

Precipitation data are provided by the Abteilung Wasserhaushalt im BML-

FUW, Hydrologische Zentralbüro, Vienna, and the Austrian weather service
ZAMG, Vienna. I thank the maintainers and developers of the statistical soft-
ware system R and especially E. Pebesma and R. Bivand for the development
of the software packages gstat and R/GRASS, respectively. The author is sup-
ported by the Austrian Academy of Sciences, Project HÖ 22.
References

1. R. F. Adler, C. Kidd, G. Petty, M. Morissey, and H. M. Goodman (2001)

Intercomparison of global precipitation products: The third precipitation inter-
comparison project (PIP-3). Bull. Amer. Meteor. Soc., 82(7):1377–1396.
2. B. Ahrens (2003) Evaluation of precipitation forecasting with the limited area
model ALADIN in an alpine watershed. Meteorol. Z., 12(5):245–255.
3. B. Ahrens (2003) Rainfall downscaling in an alpine watershed applying a mul-
tiresolution approach. J. Geophys. Res., 108(D8):8388–8399.
4. B. Ahrens, K. Jasper, and J. Gurtz (2003) On ALADIN precipitation modeling
and validation in an Alpine watershed. Ann. Geophysicae, 21:627–637.
5. B. Ahrens, U. Karstens, B. Rockel, and R. Stuhlmann (1998) On the vali-
dation of the atmospheric model REMO with ISCCP data and precipitation
measurements using simple statistics. Meteorol. Atmos. Phys., 68:127–142.
6. R. A. Anthes (1983) Regional models of the atmosphere in middle latitudes.
Monthly Wea. Rev., 111:1306–1335.
7. P. M. Atkinson and C. D. Lloyd (1998) Mapping precipitation in Switzerland
with ordinary and indicator kriging. J. Geogr. Inf. Dec. Anal., 2(2):65–76.
8. A. Beck and B. Ahrens (2004) Multiresolution evaluation of precipitation fore-
casts over the European Alps. Meteorol. Z., 13:55–62.
9. R. Benoit, C. Schär, P. Binder, S. Chamberland, H. C. Davies, M. Desgagné,
C. Girard, C. Keil, N. Kouwen, D. Lüthi, D. Maric, E. Müller, P. Pellerin,
J. Schmidli, F. Schubiger, C. Schwierz, M. Sprenger, A. Walser, S. Willemse,
W. Yu, and E. Zala (2002) The real-time ultrafinescale forecast support during
the special observing period of the MAP. Bull. Amer. Meteor. Soc., 83(1):
85–109.
10. P. Bougeault (2003) The WGNE survey of verification methods for
numerical prediction of weather elements and severe weather events.
WMO Report, available at http://www.bom.gov.au/bmrc/wefor/staff/eee/
verif/verif_web_page.html
11. R. Bubnova, G. Hello, P. Bernard, and J.-F. Geleyn (1995) Integration of
the fully elastic equations cast in the hydrostatic pressure terrain-following
coordinate in the framework of the ARPEGE/Aladin NWP system. Monthly
Wea. Rev., 123:515–535.
134 B. Ahrens

12. T. Cherubini, A. Ghelli, and L. Francois (2002) Verification of precipitation

forecasts over the alpine region using a high-density observing network. Weather
Forecast., 17(2):238–249.
13. J. Chilès (1999) Geostatistics: Modeling Spatial Uncertainty. John Wiley &
Sons, New York.
14. G. Ciach, M. L. Morrissey, and W. F. Krajewski (2000) Conditional bias in
radar rainfall estimation. J. Appl. Meteorol., 39(11):1941–1946, .
15. J. Creutin and C. Obled (1982) Objective analyses and mapping techniques for
rainfall fields: An objective comparison. Water Resour. Res., 18:413–431.
16. R. Ferretti, T. Paolucci, W. Zheng, G. Visconti, and P. Bonelli (2000) Anal-
yses of the precipitation pattern in the Alpine region using different cumulus
convection parameterizations. J. Appl. Meteorol., 39:182–200.
17. C. Frei and E. Häller (2001) Mesoscale precipitation analysis from MAP SOP
rain-gauge data. MAP Newsletter, 15:257–260.
18. Ch. Frei, J. H. Christensen, M. Déqué, D. Jacob, and P. L. Vidale (2003) Daily
precipitation statistics in regional climate models: Evaluation and intercompar-
ison for the European Alps. J. Geophys. Res., 108(D3):4124–4142.
19. P. Goovaerts (1999) Performance comparison of geostatistical algo-
rithms for incorporating elevation into the mapping of precipitation.
http://www.geocomputation.org/1999/023/gc_023.htm
20. L. Grasso (2000) The differentiation between grid spacing and resolution and
their application to numerical modeling. Bull. Amer. Meteor. Soc., 81(3):579–
580.
21. B. Ivancan-Picek, D. Glasnovic, and V. Jurcec (2003) Analysis and Aladin
prediction of a heavy precipitation event on the Eastern side of the Alps during
MAP IOP 5. Meteorol. Z., 12(2):103–112.
22. R. Mladek, J. Barckicke, P. Binder, P. Bougeault, N. Brzovic, C. Frei, J. F. Ge-
leyn, J. Hoffman, W. Ott, T. Paccagnella, P. Patruno, P. Pottier, and A. Rossa
(2000) Intercomparison and Evaluation of Precipitation Forecasts for MAP
Seasons 1995 and 1996. Meteorol. Atmos. Phys., 72:111–129.
23. A. H. Murphy and R. L. Winkler (1987) A general framework for forecast
verification. Monthly Wea. Rev., 115:1330–1338.
24. E. J. Pebesma and C. G. Wesseling (1998) Gstat: A program for geostatistical
modelling, prediction and simulation. Comp. Geosc., 24(1):17–31.
25. F. Rubel and M. Hantel (1999) Correction of daily rain gauge measurements
in the Baltic Sea drainage basin. Nordic Hydrol., 30:191–208.
26. C. Schär (2001) Alpine numerical weather prediction 2000–2020: A look back
to the future. MAP Newsletter, 14:6–12.
27. W. Schöner, I. Auer, R. Böhm, and S. Thaler (2003) Qualitätskontrolle und
statistische Eigenschaften ausgewählter Klimaparameter (Temperatur, Nieder-
schlag, Schneehöhe) auf Tageswertbasis im Hinblick auf Extremwertanalysen.
Report by ZAMG, Vienna. available from W. Schöner, ZAMG, Vienna, Austria.
28. B. Sevruk (1997) Regional dependency of precipitation–altitude relationship in
the Swiss Alps. Climatic Change, 36:355–369.
29. B. Tustison, D. Harris, and E. Foufoula-Georgiou (2001) Scale issues in verifi-
cation of precipitation forecasts. J. Geophys. Res., 106(D11):11775–11784.
30. E. Waymire, V. K. Gupta, and I. Rodriguez-Iturbe (1984) A spectral theory of
rainfall intensity at the meso-β scale. Water Resour. Res., 20:1453–1465.
 On Evaluation of Precipitation Fields 135

31. D. S. Wilks (1995) Statistical Methods in the Atmospheric Sciences, volume 58

of International Geophysics Series. Academic Press, San Diego.
32. C. Wilson (2001) Review of current methods and tools for verification
of numerical forecasts of precipitation. COST717–Working Group Report
WDF_02_200109_1, available at http://www.smhi.se/cost717/
33. WMO, Commission for Instruments and Methods of Observation (2001) Expert
meeting on rainfall intensity measurements. Technical report, World Metoro-
logical Organization.
34. C. B. Young, B. R. Nelson, A. A. Bradley, J. A. Smith, C. D. Peters-Lidard,
A. Kruger, and M. L. Baeck (1999) An evaluation of NEXRAD precipitation
estimates in complex terrain. J. Geophys. Res., 104(D16):19691–19703.
Robust Spatial Correlation Analysis
of the ETEX-1 Tracer Data

Syed Shibli1 and Gregoire Dubois2

1
Landmark Eame Ltd, Aberdeen, Scotland, UK
syed.shibli@googlemail.com
2
Radioactivity Environmental Monitoring, Institute for the Environment and
Sustainability, Joint Research Centre, European Commission, Ispra, Italy

1 Introduction
At 16:00 UTC on October 23, 1994, 340 kg of perfluoromethylcyclohexane
(PMCH) were released into the air from Monterfil in Brittany, France. Air
samples were collected at 168 stations in 17 European countries for a period
of 90 hours from the start of the release. The European Tracer Experiment
(ETEX) was initiated with the aim of collecting data for validating long range
transport and dispersion models used for emergency response applications
[4, 13]. Another release was made a month later under different meteorological
conditions.
Although the data have been used in numerous mechanistic atmospheric
dispersion studies, only previously was a geostatistical analysis performed in
order to provide some basis for a spatial interpolation [5]. Such an analysis
applied fractional Brownian motion models in order to summarise the spatial
correlation structure in terms of the power exponent of the variogram, which
is directly related to the fractal dimension.
Because of the distressing nature of the data set, which is highly skewed
and required a logarithmic transformation in the Dubois et al study, this
paper attempts to apply more robust variography on the raw data in order to
extract some order out of the chaos. We use the term “robust” in this paper
specifically to refer to the stability of variogram in the presence of a strong
direct proportional effect, which characterises the ETEX-1 dataset.

2 Analysis of Non-stationary Data

Non-stationary data, including skewed data showing a strong proportional
effect, poses significant challenges for spatial correlation analysis and subse-
quent interpolation. Often it is difficult to determine, from looking at the
sample variogram alone, whether what appears to be a power law variogram,
138 S. Shibli and G. Dubois

representing stationary data with an infinite capacity for dispersion, is not

just the manifestation of a quadratic growth due to a non-constant mean.
In theory if we assume that the regionalised variable is made up of two
components:
Z(s) = μ(s) + Y (s) (1)
where Y (s) is intrisically stationary with zero mean, and E{Z(s)} = μ(s),
i.e. the mean varies with location, then we can define the variogram of Z(s) as

2γZ = V ar{Z(s + h) − Z(s)} (2)

or, substituting (1) into (2)

2γZ (h) = V ar{[Y (s + h) − Y (s)] + [μ(s + h) − μ(s)]}

(3)
= V ar{Y (s + h) − Y (s)}

equation (3) shows that for Y (s) that is intrisically stationary, then the var-
iogram of Y (s) should be equivalent to the variogram of Z(s). However, in
practice, the sample variogram estimator is typically based on

2γZ (h) = E{[Z(s + h) − Z(s)]2 } (4)

which would only be valid for data with constant mean. Again, substituting
(1) into (4) would give the following expression for the variogram

2γZ (h) = E{(Y (s + h) − Y (s))2 } + {[(Y )(s + h) − Y (s))(μ(s + h) − μ(s))]}

+E{(μ(s + h) − μ(s))2 }
(5)
Since Y (s) is intrinsically stationary, the second term on the right hand
side of (5) should be zero, so we are left with the expression

2γZ (h) = 2γY (h) + [μ(s + h) − μ(s)]2 (6)

equation (6) shows that if the mean is constant everywhere, μ(s + h) = μ(s),
so the variogram for Z(s) should be equivalent to that for Y (s) and (4) would
thus apply. However, if the mean is not constant then we will derive a var-
iogram estimator that will exhibit a quadratic growth with h, which would
make the estimator invalid.
One method of incorporating a non-constant mean is to estimate a mean
surface and work with residuals assumed to be intrisically stationary, e.g. me-
dian polish kriging [1]. Nevertheless, in practice, data showing a strong pro-
portional effect (heteroscedasticity) might still require a rescaling of the var-
iogram. This is because the dispersion of the data now depends on its local
mean.
Since heteroscedasticity is commonly associated with highly skewed data,
one common approach is to transform the data to a logarithmic scale and per-
form variography and kriging in log space before back transforming, e.g. log-
normal kriging [2, 8]. Another approach would be to use alternative measures
 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 139

such as the relative variogram in order to account for the proportional effect
via some form of rescaling of the sample variogram values.
Both techniques of transformation and use of the relative variogram are
related, as demonstrated by Cressie [1]. If we assume that we have spatial
regions{Dj ; j = 1, . . . , n} within which the regionalised variable Y (s) is in-
trinsically stationary with mean μj and variogram 2γZj (h) for each region
j = 1, . . . , n, then using the δ method of Kendall and Stuart [9] we can apply
a transformation for the variable Z(s) as follows:

Y j (s) = g[Z j (s)]; s ∈ Dj (7)

where g(·) is sufficiently smooth to possess at least two continuous derivatives.

Kendall and Stuart applied a Taylor series expansion about E(Z(s)), i.e.

Y j (s) = g(μj ) + g  (μj )[Z j (s) − μj ] + g  (μj )[Z j (s) − μj ]2 /2! + . . . ; s ∈ Dj (8)

It follows from (7) that if we were to take the increments for Y j (s) then

Y j (s + h) − Y j (s) = g  (μj ) + g(μj )[Z j (s + h) − μj ] + . . . ; s ∈ Dj (9)

We can then define the variogram for Y j (s) by applying the variance op-
erator on both sides of (9) to derive

2γYj (h) = (g  (μj ))2 2γZj (h); j = 1, . . . , k (10)

(10) is similar to the general form of the local relative variogram, defined by
various authors [7, 8] in the following manner:

2γZj (h)
2γRY (h) = (11)
μnj

which is independent of j and where n is typically 2. Equating (10) and (11) we

note that when g(μj ) = 1/μj the process Y (s) is approximately intrinsically
stationary, i.e.
g(x) = log(x) (12)
This result shows that the relative variogram can be used as an alternative
to log transformation, in order reduce the influence of a proportional effect on
the traditional sample variogram. Srivastava and Parker [12] present another
deterministic form of the relative variogram, namely the pairwise relative var-
iogram described as follows:
 2
1  z(s + h) − z(s)
2γ̂P R (h) = (13)
Nh z(s + h) + z(s)

The above takes each squared difference between pairs of sample values
and divides it by the square of their average, hence the term “pairwise.”
140 S. Shibli and G. Dubois

Srivastava [11] also presents non-ergodic versions of the covariance and

correlogram functions to accomodate the proportional effect. Starting with
the definition of the covariance, i.e.

C(h) = E{Z(s)}E{Z(s + h)} − μ(s + h) (14)

Srivastava defines the following estimator for (14)

1 
Ĉ(h) = z(s + h)z(s) − z(s)z(s + h) (15)
N (h)

with the caveat that, although γ̂(h) = Ĉ(0) − Ĉ(h), the differences should
be sufficiently small that one can apply γ̂(h) in practice. Curriero et al. [3]
contends that the “head” and “tail” values in (15) are meaningless for the
omnidirectional direction since the common practice is to count the location
twice. Hence the re-scaling of the variance is implicitly taken into account by
incorporation of the mean for all the data contributing to the lag.
A common problem with using (15) to derive an approximate sample vari-
ogram is that the first few lags can result in negative values, since the number
of data used to derive the sample variance is typically greater than the num-
ber of data used to derive the sample covariance itself. Nevertheless, such a
difficulty does not exist if one scales each covariance value by the product of
the standard deviations for the head and tail values, thus giving the following
definition for the correlogram:

Ĉ(h)
ρ̂(h) = (16)
σ̂(s)σ̂(s + h)
By explicitly accounting for the possibility that some lags contain more
variable values than others, the correlogram is likely to suffer least from the
combination of heteroscedasticity and clustering [? ].
Yet another variation of (11) presented by Isaaks and Srivastava [7] is the
general relative variogram, defined as follows:
γ̂(h)
γ̂GR (h) = 1
 (17)
2Nh z(s) + z(s + h)

where γ̂(h) is the traditional method of moments variogram estimator. Equa-

tion (17) circumvents the potential problem of having too few pairs in the
local disjoint regions j = 1, . . . , n, which is required in order to estimate the
local relative variogram defined by (11).

3 Data Description
The ETEX-1 data consists of 155 raw concentration measurements of PMCH
3
(in units ng/m ) from the first ETEX release recorded at 26 different times,
 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 141

from 15 hours to 90 hours in increments of three hours, to give a total of

4,030 measurements. Thirteen stations out of the original list of 168 were
discarded a priori; nine due to the absence of any reliable measurements and
four being dismissed as outliers [5]. PMCH is an inert gas so is insensitive to
wet scavenging or uptake at the surface and thus is only affected by advection
and diffusion.
The time t = 15 hours was chosen as the start time since by this time all
of the tracer had been released to the atmosphere. A preliminary analysis of
the data indicated that a maximum of 95% of the stations did not record any
presence of the tracer at t = 15 hours and a minimum of 69% was found at
t = 45 hours. Due to the high positive skewness of the data, in the original
paper a logarithmic transformation was applied to minimise the effect of the
skewness on the variogram.

4 Univariate Statistical Analysis

Figure 1 shows the location map of the monitoring stations, which covers
part of Western Europe. The maximum diagonal distance for the analysed
area is some 3,093 km, and a quick analysis of the geographical locations
indicates that there is no preferential sampling strategy and that their spa-
tial distribuion appears to be random (Clark and Evans’s index=1.032). No
declustering was therefore performed on the data.
When plot on a time axis, we observe the mean of the concentrations to
increase between t = 15 hours and t = 45 hours before decreasing to zero
thereafter right up to t > 80 hours (Fig. 2) due to the large proportion of
zero measurements. The variance tracks a similar behaviour; the decrease in
the first and second order moments beyond t = 60 hours could possibly be
due to the cloud dispersing at a faster rate and splitting into more distinct
parts.
Skewness and kurtosis (Fig. 3) are quite constant in time from t = 15
hours to about t = 81 hours, indicating no general changes in shape of the
histograms (highly positively skewed to the right) for almost all time slices.
Figure 4 shows a direct proportional effect in time for the mean concen-
tration versus variance. The correlation coefficient between the mean and
variance is ca. 83%, which is high. Moving window statistics were also derived
for the data for t = 45 hours, and Fig. 5 shows that the linear relationship
between the mean concentration and variance is also present both in the east–
west and north–south directions.
Finally, Figs. 6–8 present the probability plots of the surface concentra-
tions for times from t = 15 hours to t = 75 hours. The skewness is highly
evident from these plots, with zero values making up a large proportion of
the recorded data. At times greater than 63 hours, the proportion of zeros
increases.
142 S. Shibli and G. Dubois

Fig. 1. Location of ETEX-1 monitoring stations

Fig. 2. Tracer concentration mean and variance over time

Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 143

Fig. 3. Skewness and kurtosis over time

Fig. 4. Tracer concentration mean versus variance for all times

Fig. 5. Moving window statistics at t = 45 hours

144 S. Shibli and G. Dubois

Fig. 6. Probability plot of tracer concentrations from t = 15 to t = 33 hours

 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 145

5 Spatial Correlation Analysis

An analysis of the spatial correlation structure of the surface concentra-
tion data is a prerequisite for spatial interpolation, whether for estimation
(e.g. kriging) or simulation (e.g. sequential Gaussian simulation). This phase
of the analysis quantifies, via a traditional measure such as the sample vari-
ogram, the spatial dependency between data separated some lag distance, h,
in space or time, described by (2) or (4).
In ordinary kriging, the estimated value ẑ(s) is given by

n
ẑ(s0 ) = λi z(si ) (18)
i=1

where λi are weights assigned to the observed data z(si ) that will determine
their role in defining the value taken by the variable at unsampled location s0 .
The main interest in applying geostatistical techniques is that these weights
are computed from a model of the spatial correlation of the analysed phe-
nomena. Hence, unlike other interpolators (for an overview of interpolation
methods refer to Lam [10]), geostatistics takes the spatial structure of the
variable explicitly into account.
A useful opening move in any geostatistical study is to derive the omnidi-
rectional semivariogram of the variable under study to determine the degree
of spatial correlation, often called a “structural analysis”. As noted earlier, the
heteroscedasticity observed in the data can makes the inference of a range and
sill very difficult using conventional measures such as the sample variogram.
Although by itself the shape of the semivariogram may not be affected by the
heteroscedasticity if the mean of the sample values is roughly the same for all
lags, the oridinary kriging variance, however, is dependent on the magnitude
of the variogram.
For the sake of brevity, the analyses will now be presented based on the
data for t = 45 hours. Figure 9 shows six different spatial correlation mea-
sures for this time slice, calculated using an angular tolerance of 90 degrees
(omnidirectional). Twelve lags were calculated at a lag increment of roughly
100 km, with a lag tolerance of about 50 km. This ensured that there were at
least 30 pairs for each lag distance.
The functions shown in Fig. 9 also include the sample general relative
variogram (17), pairwise relative variogram (13), non-ergodic covariance (14),
non-ergodic correlogram (16), and semimadogram (a measure of the average
mean absolute difference), defined by the following:
1 
2γ̂M (h) = |z(s + h) − z(s)| (19)
Nh
It can be noted that the both the madogram and variogram provide a poor
measure of the spatial correlation for this highly skewed data set, resulting in
an erratic sample variogram. Both measures rely only on the mean difference
146 S. Shibli and G. Dubois

Fig. 7. Probability plot of tracer concentrations from t = 36 to t = 54 hours

 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 147

(squared in the case of the variogram) between two data points located h m
apart, so no rescaling of the variogram is performed commensurate with the
proportional effect.
The madogram also forms the basis for other well known “robust” variog-
raphy techniques, e.g. that of Cressie and Hawkins [2] using the fourth power
of the square root of the madogram; and that of Genton [6] based on the k-th
quantile of the madogram. However these techniques rely on stability of the
variogram based on deviations from a primarily Gaussian distribution, and
are not expected to fare very well in the presence of a strong proportional
effect.
From Fig. 9, for the other four measures which scale as some function of
the mean value, some semblance of structure can be inferred, giving an omni-
directional range for the tracer concentration values at around 40,000 m. The
pairwise relative variogram results in a usable albeit smooth model because
values at each pair are rescaled by the mean of the values contributing to
that pair. The correlogram appears to be less erratic than the covariance, due
to the advantage of having the values standardised by the data variances.
Both the correlogram and the pairwise relative variogram result in smoother
spatial correlation structures compared to the general relative variogram and
covariance.
The above results confirm that some sort of spatial dependence exists
for the concentration cloud at t = 45 hours, which was not evident from
the traditional variogram measure. Figure 10 shows the directional pairwise
relative variogram for four principal directions 0 deg, 45 deg, 90 deg, and
135 deg based on an angular tolerance of 22.5 deg. Little or no anisotropy is
evident; for the direction 135 deg a somewhat shorter range of some 300 km
is observed, and this appears to be the direction of minimum continuity. The
range in the direction of maximum continuity appears to be between 400 and
500 km.
For convenience, and as a prerequisite for spatial interpolation, we will
model the spatial variability by assuming a power law model for the sample
variogram. This is performed by curve fitting a log–log plot of the pairwise
relative variogram values versus lag distance (refer Fig. 11) and inferring the
slope of the fit.
The variance of increments for random fractional Brownian motion (fBm)
models satisfying a distribution with fractal geometry can be written as:

2γ̂(h) = VH (h)2H (20)

where H is the intermittency or Hurst exponent and is related to the fractal

dimension D = 2−H. The intermittency exponent falls between zero and one;
the special case H = 0.5 characterising normal Brownian motion. For H < 0.5,
the phenomena is anti-persistent (more zero crossings) and less continuous; for
H > 0.5 the phenomena is persistent, or values tend to be clustered together
and high values are more likely to be followed by high values as well.
148 S. Shibli and G. Dubois

Fig. 8. Probability plot of tracer concentrations from t = 57 to t = 75 hours

 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 149
20 1.2
18 1
16
0.8
14
12 0.6
10 0.4
8
0.2
6
0
4
2 –0.2
0 –0.4
0 200000 400000 600000 800000 1000000 1200000 0 200000 400000 600000 800000 1000000 1200000
Lag distance [m]
Lag distance [m]

2.5 0.08

0.06
2
0.04
1.5 0.02

1 0

–0.02
0.5
–0.04

0 –0.06
0 200000 400000 600000 800000 1000000 1200000 0 200000 400000 600000 800000 1000000 1200000
Lag distance [m]
Lag distance [m]

0.2 2.5
0.18
0.16 2
0.14
0.12 1.5
0.1
0.08 1
0.06
0.04 0.5
0.02
0 0
0 200000 400000 600000 800000 1000000 1200000 0 200000 400000 600000 800000 1000000 1200000
Lag distance [m] Lag distance [m]

Fig. 9. Spatial correlation measures at t = 45 hours (omnidirectional)

Fig. 10. Directional pairwise relative variograms at t = 45 hours

150 S. Shibli and G. Dubois

Fig. 11. Log–log model fit for directional variograms

The four directional variograms in Fig. 10 give the following fractal

dimensions:

direction fractal dimension

0 deg ± 22.5 deg 1.87
45 deg ± 22.5 deg 1.93
90 deg ± 22.5 deg 1.90
135 deg ± 22.5 deg 1.91

The dimensions are characteristically high, although within the same range
of values reported by Dubois et al. [5], who based their results on log–log
plots of variograms based on log transformed variables. These high dimen-
sions translate to very low intermittency exponents, corresponding to anti-
persistence, or less continuous phenomena. Such behaviour is also reflected in
the shape of the relative variograms themselves.

6 Discussion

For the ETEX-1 data, the traditional variogram estimator performs poorly in
the presence of a strong proportional effect, showing erratic behaviour at all
lags. Use of the alternative estimators which re-scale the variogram value by
some function of the mean allows us to infer large scale structure from the
 Robust Spatial Correlation Analysis of the ETEX-1 Tracer Data 151

highly skewed data. This obviates the need to perform a logarithmic transform
in order to temper the impact of such skewness. The next step in the analysis
would be to perform a spatial interpolation of the concentrations based on
the inferred structural range and curvature. Use of power law (fBm) models
is one alternative.
References

1. Cressie, N. (1993) Statistics for Spatial Data, Revised edition. John Wiley &
Sons.
2. Cressie, N.A.C and Hawkins, D.M. (1980). Robust Estimation of the Variogram:
I. Mathematical Geology, 12(2):115–125.
3. Curriero, F., Hohn, M., Liebold, A., and Lele, S. (2002). A statistical evaluation
of non-ergodic variogram estimators. Environmental and Ecological Statistics,
9(1):89–110.
4. Girardi, F., Graziani, G., van Veltzen, D., Galmarini, S., Mosca, S., Bianconi, R.,
Bellasio, R. and Klug, W. (eds.) (1998). “The ETEX Project”. EUR Report
18143 EN. Office for Official Publications of the European Communities, Lux-
embourg.
5. Dubois, G., Galmarini, S., and Saisana, M. (2005). Geostatistical Investigation
of ETEX-1: Structural Analysis. Atmospheric Environment, 39: 1683–1693.
6. Genton, M. (1998). Highly Robust Variogram Estimation. Mathematical Geol-
ogy, 30(2):213–221.
7. Isaaks, E. and Srivastava, R.M. (1989). An Introduction to Applied Geostatis-
tics, Oxford University Press, Oxford.
8. Journel, A.G. and Huijbregts, C.J. (1978). Mining Geostatistics. Academic
Press, New York.
9. Kendall, M.G., and Stuart, A. (1969). The advanced theory of statistics, Vol. 1,
3rd ed. Griffin, London.
10. Lam, N.S. (1983). Spatial interpolation methods: a review. The American Car-
tographer, 10(2):129–149.
11. Srivastava, R.M. (1987). A Non-ergodic framework for variograms and covari-
ance functions. M.Sc. Thesis, Stanford University.
12. Srivastava, R.M. and Parker, H.M. (1988). Robust Measures of Spatial Conti-
nuity, In: Geostatistics Volume 1, Proceedings of the Third International Geo-
statistics Congress, Kluwer Academic Publishers.
13. van Dop, H. and Nodop, K. (eds). (1998). ETEX: A European Tracer Experi-
ment, Atmospheric Environment 32:4089–4378.
Fuzzy Model of Soil Polygons
for Managing the Imprecision

Rangsima Sunila and Pekka Horttanainen

Department of Surveying, Institute of Cartography and Geoinformatics, Helsinki

University of Technology (HUT), Espoo, Finland
rangsima.sunila@hut.fi

1 Introduction
It is impossible to create a perfect representation of the world in a GIS
database since all GIS data are subject to uncertainty [5]. There is no perfect
information that contains 100% accurate presentation in a GIS database. The
impression of certainty usually conveyed by GIS is at odds with the uncertain
nature of geographic information, a contradiction that has been acknowledged
as an important research topic for nearly two decades [4]. Incorrectness in
measurement or errors in observations due to rich information bring about
imperfection in geographic information. The information is taken and used as
if it were accurate or believed to be true. In fact, the reliability of the infor-
mation is not yet considered in terms of level of accuracy or uncertainty. If
the geographic information is looked over carefully, it contains vagueness, im-
precision and inaccuracy particularly when it presents an invisible object like
soil. Soil polygon boundaries are defined based on accurate field observation
and compared with human interpretation and soil types are classed according
to geologists’ prowess or expertise. For many reasons, one can say that soil
map is one of the most imprecise maps in the world.

1.1 The Background of the Research

In Finland, mapped areas are big and field observations are relatively sparse.
Manual interpretation is therefore the only feasible alternative in creating
soil maps. Samples of soil are taken randomly by soil mapping surveyors for
soil type classification and some of these samples may be taken back to the
laboratory for detailed tests in case of inadequacies (Fig. 1).
For defining soil polygon boundaries, geologists use aerial photos, geologic
maps, topographic maps in scale 1:20,000 together with knowledge based of
geomorphology. Nevertheless, there are no specific rules to define the impreci-
sion of these boundaries and neither imprecision in data nor expert knowledge
154 R. Sunila and P. Horttanainen

Fig. 1. Soil mapping fieldtrip in Porvoo, Finland

is recorded. As a result, imprecision in soil data is carried along the survey

process and it continues to the stage when the data is entered to soil database
then the final result comes out in the form of printed map that certainly
contains uncertainty information.
The problem to be solved is to try to collect and describe the knowledge
and experience of geologists. To cope with this and discussion about uncer-
tainty in geographic databases, fuzzy logic will be used as a tool to describe
and solve the problem of modeling imprecise objects like soil polygons.

1.2 The Goal

The core idea in this research is to apply fuzzy modeling to the management
of expert knowledge in soil mapping. Fuzzy soil maps are then used in map
overlay type analysis [11]. The goal is not to construct a fuzzy model of soil
layers but a fuzzy soil map presenting non-crisp soil polygon boundaries. The
map will be created in certain scale for a certain purpose and we believe
that fuzzy soil layer with imprecision is better input to the analysis than
artificial crisp polygon map with no information about the uncertainty of the
boundaries.
 Fuzzy Model of Soil Polygons 155

2 Literature Survey

Recently, there have been many researches that were related to fuzzy concepts.
Brown [2] carried out a research in classification and boundary vagueness in
mapping presettlement forest types. In his research, he conducted a model to
test the role of classification ambiguity in affecting boundary vagueness us-
ing fuzzy concepts and Kriging. He explained methods of determining species
memberships and interpolating membership values. Stefanakis et al. [10] con-
ducted a research on incorpolating fuzzy set methodologies in a Database
Management System (DBMS) repository for the application domain of GIS.
They considered that fuzzy set methodologies seemed to be instrumental in
the design of efficient tools to support the spatial decision-making process.
The results showed that Zadeh’s fuzzy concepts and fuzzy set theory [12]
might be adopted for the representation and analysis of geological data. Jiang
et al. [6] proposed the application of fuzzy measures and argued that the
standardized factors of multi-criteria evaluation belong to a general class of
fuzzy measures and the more specific instance of fuzzy membership. Develop-
ing of classification algorithms for using auxiliary information in fuzzification
and fuzzy set operations to reduce uncertainty in classification process was
researched in 2000 by Oberthür et al. [9]. The research was conducted in order
to study how to define fuzzy membership functions (FMF) and reduce classi-
fication uncertainty hedge operators. Zhu et al. [13] introduced soil mapping
using GIS, expert knowledge, and fuzzy logic. The scheme consisted of three
major components: a model employing a similarity representation of soils, a
set of inference techniques for deriving similarity representation and use of
the similarity repreentation. Basically they invented an automated soil infer-
ence under fuzzy logic. To produce a raster soil database for the study areas,
the knowledge base and the spatial data in the GIS database were combined
under the fuzzy inference engine. The output was the comparison of soil se-
ries referred from Soil Land Inference Model (SoLIM) and derived from the
soil map against the field observations for the study area and it showed that
SoLIM has higher correctness. The derivation of the fuzzy spatial extent was
developed by Cheng et al. [3]. Three fuzzy object models and the data ex-
tracted from field observation were introduced and modeled. Software such as
FUZZEKS [1], FuzME Version 3.0 [8] and ASIS [7] were programmed to deal
with fuzzy logic and data analysis.

3 Pilot Studies

At this stage of the project, expert knowledge is being collected. The re-
search team is currently trying to document a rule-based model of soil polygon
boundaries. As the geographic environment is varied depending on regions and
geomorphy, each region needs to be differentiated and taken into consideration
156 R. Sunila and P. Horttanainen

separately. To construct a fuzzy model, the values of membership functions

are required. The membership functions from different regions should present
different values, as the behavior of geographic data in each region is distinct.
This may require a lot of work in the beginning but once the system is set up,
everything will be processed smoothly. When fuzzy membership functions of
soil types are defined, fuzzy models of soil polygons will be constructed.

4 Methodology
From the dataset received from Geological Survey of Finland, the data will
be read in numerical format as in Fig. 2. These numbers represent different
types of soil and the connection with different numbers implies where the
polygon borders are. As it was mentioned earlier, to construct fuzzy models,
membership functions of classification are needed.

Fig. 2. Example of soil data in numerical format; soil type 1: bedrock, soil type 2:
sand, soil type 3: clay

Therefore, Table 1 is a fundamental design for geologists to provide their

expert opinions of the level of certainty in soil polygon boundaries. For exam-
ple, number 0.75 in the table is the value of the pixel’s membership function
adjacent to the boundary, which means that it is not crisp. The numbers on
the diagonal represent the value of the membership function within the soil
polygon.

Table 1. Level of certainty of polygon borders between different soil types ranged
from 0 to 1
eastern Finland bedrock sand clay
bedrock 0.95 0.92 0.88
sand 0.9 0.75
clay 0.85
 Fuzzy Model of Soil Polygons 157

From the numbers in the table, logical rules will be constructed in a fuzzy
tool. The expected result is the soil data layer that shows imprecision on the
soil polygon boundaries. Next phase, the layer of elevation could be added to
adjust the values of fuzziness.

5 Expectations
It takes time to start up the system and develop a good connection that will
lead to success in the future. Collecting geologists’ knowledge and trying to
construct the documentation of expert knowledge is the first priority. From
the documentation, a rule-based model is created which will lead to the con-
struction of fuzzy models. Currently, there are three expectations.
• Documentation of geological knowledge used in interpretation
• Development of a rule-based model of imprecise soil polygons which could
be used for GIS analysis. This is not to create a new soil model for geolo-
gists or even for soil mapping but to improve the represetation of soil data
layer that shows the imprecision of the classification especially around the
boundaries of soil polygons.
• Fuzzy modeling of soil maps to understand uncertainty in geographical
information for better uses in spatial analyses

6 Future Plan
For further research, there are still many possibilities to continue studying
imprecision in soil polygon boundaries, for instance, implementing Kriging to
test out the result of fuzziness. The question may arise here whether Kriging
can be used, as the values seem to be from a discrete function. Clearly, Kriging
is not going to be used for better classification of soil types, instead, it will be
applied together with a fuzzy model to verify the imprecision on soil polygon
boundaries to smooth out the result. Sample points could be taken from the
real site to study soil type misclassification and these numbers will be used
together with fuzzy membership functions for better results.

6.1 An Example
Figure 3 shows an example of fuzzy and Kriging application on soil polygon
boundaries.
158 R. Sunila and P. Horttanainen

2 2 2 2 3 3 3 3 3 .75.75
2 2 2 2 2 3 3 3 3 .75.75
2 2 2 2 2 2 3 3 3 .75.75
2 2 2 2 2 2 3 3 3 .75.75
Sand 2 2 2 2 2 2 3 3 3 .75.75
2 2 2 2 2 2 3 3 3 .75.75
2 2 2 2 2 3 3 3 3 .75.75
Clay 2 2 2 2 2 3 3 3 3 .75.75
2 2 2 2 3 3 3 3 3 .75.75
(a) (b) (c) (d)

2 3
2 2
2 2
2 2
2 2
2 2
2 3
3 3
3 3
(e)

(f) (g) (h)

Fig. 3. (a) Consider a map that contains only two soil types, sand and clay. (b) The
soil map data is transferred to a raster format. (c) From the raster format the data
is coded: sand = 2 and clay = 3. (d) Next step, membership functions represent the
level of classification certainty on soil polygon boundaries. The result is the data
layer that contains imprecise soil polygon boundaries. (e) Kriging is used to test
out the soil property in each class. In this example, misclassification is discovered
and it affects the boundary. Comparison: (f ) The original soil polygon boundary in
raster format. (g) The new soil polygon boundary resulted from Kriging. (h) The
highlighted area shows the error or fuzziness along the boundary. The highlighted
areas are the areas that should be taken into consideration in order to adjust the
values of membership functions

7 Difficulties in Using Kriging

Although Kriging is a well-known geostatistical tool, it is not being used in
Finnish soil mapping. Finland is big with its size of about 338,000 km2 so it is
difficult to have very high density of soil sampling in a large scale. Moreover,
much effort and skill are needed for setting up test areas in every region. This
will be too time consuming and expensive, which is an important matter for
concern. Besides, the results are believed to be different in each particular area
so it is hard to set up a standard value for error measurements. In conclusion,
this method has high requirements in order to conduct the research.
 Fuzzy Model of Soil Polygons 159

However, it is hoped that developing membership functions in fuzzy mod-

els together with Kriging would give better solutions in studying imprecise
information in soil databases especially in some specific areas that are worth
for studying in depth for better spatial analysis.

8 Conclusion

Soil polygon boundaries are not crisp in reality. Moreover, soil maps contain
a large amount of uncertainty and imprecision. Therefore a natural way to
model vagueness of soil polygons is to include imprecision in their boundaries.
One way to do this is to develop a fuzzy model for raster data using fuzzy
membership functions for each soil layer. This model will be created using
expert knowledge and fuzzy logic.
In Finland there are no sufficient metadata to assess the uncertainty of
soil polygon boundaries. Thus, the first step of the research is to collect and
document expert knowledge about soil mapping. Only then can the rule-based
fuzzy model be created. The information of imprecision provided by the fuzzy
model will eventually be used in GIS to give an estimation of uncertainty in
soil maps.

Acknowledgements
The project is funded by the Ministry of Forestry and Agriculture of Finland
and in cooperation with Geological Survey of Finland. This project is an off-
spring from the military terrain analysis project conducted by the Scientific
Board of the Finnish Defence Forces. The authors would like to thank Profes-
sor Kirsi Virrantaus for introducing the project and giving her comments and
suggestions, Jukka-Pekka Palmu and Maija Haavisto-Hyvärinen for support-
ing the information, Professor Vesa Niskanen for instruction of fuzzy mod-
eling, Olga Křemenová for her assistance and cooperation and StatGIS 2003
conference committee for giving a chance to present the project.
References

1. Bartels F (1997) FUZZEKS. The Fuzzy Evaluation and Kriging System.

http://www.fuzzeks.de
2. Brown D (1998) Classification and boundary vagueness in mapping presettle-
ment forest types. Int. J. Geogr. Inf. Sci. 12(2):105–129.
3. Cheng T, Molenaar M and Lin H (2001) Formalizing fuzzy objects from uncer-
tain classification results. Int. J. Geogr. Inf. Sci. 15(1):27–42
4. Duckham M (2002) Uncertainty and geographic information: computational
and critical convergence. In: Representation in a digital geography, Wiley: New
York.
5. Goodchild M (2003) Models for Uncertainty in Area-Class Map Presentation
lecture note, Uncertainty in Geographical Information course, Helsinki Univer-
sity of Technology, 2–3 June
6. Jiang H and Eastman J R (2003) Application of fuzzy measures in multi-criteria
evaluation in GIS. Int. J. Geogr. Inf. Sci. 14(2):173–184
7. Mazaheri S A, Koppi A J, Mcbratney A B and Constable B (1995) Australian
Soil Identification Spreadsheet (ASIS). A program for allocatiing soil profiles
to Australian Great Soil Groups (GSG)
8. Minasny B and McBratney AB (2002) FuzME version 3.0, Australian Centre
for Precision Agriculture, The University of Sydney, Australia
9. Oberthür T, Dobermann A and Aylhard M (2000) Using auxiliary information
to adjust fuzzy membership functions for improved mapping of soil qualities.
Int. J. Geo. Inf. Sci. 14(5):431–454
10. Stefanakis E, Vazirgiannis M and Sellis T (1999) Incorporating fuzzy set
methodology in a DBMS repository for the application domain of GIS. Int.
J. Geogr. Inf. Sci. 13(7):657–675
11. Virrantaus K (2003) Analysis of the Uncertainty and Imprecision of the Source
Data Sets for a Military Terrain Analysis Application, Proceedings of the 2nd
International Symposium of Spatial Data Quality, Hong Kong, 20–22 March
12. Zadeh L A (1965) Fuzzy sets. Information and Control 8:338–353
13. Zhu A, Hudson B, Burt J, Lubich K and Simonson D (2001) Soil Mapping using
GIS, Expert Knowledge, and Fuzzy Logic, Soil Sci. Soc. Am. J. 65:1463–1472
Mapping the Contaminant Legacy of a Coking
Plant, The Avenue, Chesterfield, UK

Monica Palaseanu-Lovejoy, Ian Douglas, and Robert Barr

School of Geography, The University of Manchester, Manchester, UK

monica.palaseanu-lovejoy@stud.man.ac.uc

1 Introduction
With 300,000 ha of contaminated land, 1.2% of the Britain land area [13, 14],
the UK has a major need for effective environmental risk assessment for land
remediation and reclamation [22]. Such a risk assessment is usually based on
the characterization of potential site contaminants and analysis of source –
pathway – target scenarios [8, 9, 11]. A risk-based contaminant description re-
quires a conceptual model of the site that includes qualitative and quantitative
analyses of pollution sources, contaminant pathways and pollutant receptors
[4, 22]. This characterization typically has to rely on limited and irregularly
distributed point data. In addition, soil and surface material heterogeneity,
as well as the quasi-random nature of contamination sources add to the com-
plexity of developing good spatial models of pollutants on old industrial sites
[13, 22]. Improved and appropriate geostatistical tools and GIS based anal-
ysis can help to overcome some of these problems. This paper tackles the
development of such a methodology for a former coking plant by examining
the sources and pathways of Polycyclic Aromatic Hydrocarbons (PAHs), as
part of an analysis of a wider range of contaminants at the Avenue Coking
Works, near Chesterfield, UK (Fig. 1). In the UK, coking works were estab-
lished alongside the iron and steel plants from the mid 18th century. By the
end of the 19th century surplus gas from coking works was sold as town gas,
and by 1912 coke ovens were being installed at town gas works [10]. Each
works occupied between 0.3 and 200 ha. By 1995, only four of the total of
400 such works were still operating [10]. Tar distillation took place on coal
and, or coke works sites, and was the primary source of organic chemicals
for different industries until petrochemical products took over in the 1960s
[10]. The contamination at former gas and coke works varies with the range of
products and by-products manufactured. On such sites, ground contamination
arises from by-products, waste products from landfills and lagoons, and ancil-
lary products such as ammoniacal liquor, coal tar, spent oxide and foul lime
[12]. The organic contaminants are derived from constituents of coal tar such
162 M. Palaseanu-Lovejoy et al.

as aromatic hydrocarbons, polycyclic aromatic hydrocarbons (PAHs), phenils

and phenols, nitrogen compounds, and organo – sulphur compounds, natural
gas processing compounds such as alcohol, glycols, resins, heavy oils, and or-
ganic fuels such as petroleum and naphthalene [12]. Several of these coking
works were developed on sites with a long industrial history, adding further
contamination to that derived from the original activity. This study sets out
to construct a conceptual model for the Avenue site that distinguishes lo-
cal point source PAH16 (Polycyclic Aromatic Hydrocarbon) pollution arising
from the coking plant activities, from the general historical diffuse pollution
caused by other older industrial operations on the same site. Legacies of later
phases of pollution that contribute to the same type of contamination can
hide the special pattern of diffusion of contamination due to the earlier phase
of industrial activity. One common problem in environmental risk assessment
is to determine the value of a continuous attribute at any particular unsam-
pled location; the uncertainty of any unsampled value; and the probability
that a regulatory threshold for soil pollution or a criterion for soil quality
is exceeded at any unsampled location, when only few sampled values are
known [6, 7, 15, 16, 17, 19]. Geostatistics provide the basis for analysing data
that vary continuously spatially and permit the inference values of the same
variable at unsampled locations through interpolation techniques. Two key
assumptions in geostatistical analysis are that (1) sample values are expected
to vary continuously from one location to another; (2) at any particular lo-
cation the value of the variable comprises a fixed component of the variation
trend, which is usually unknown, and a random variable following one specific
distribution [5, 15, 19] expressed by:

z(x) = a ⊕ Z(x) (1)

– z(x) = value of the variable z at location x;

– a = fixed unknown component of the variation trend;
– Z(x) = random variable described by:

Z(x) = m(x) + ε (x) + ε (2)

– m(x) = deterministic function that describes the structural component
with a constant mean or trend [3];
– ε (x) = a random but spatially correlated component, known as the varia-
tion of the regionalized variable, and it is the locally varying but spatially
dependent residual of m(x);
– ε = is residual, spatially independent noise, having a mean of zero and
a standard deviation or variance σ 2 .
If the assumption that an element varies continuously over a certain area is
true, then it is customary to assume that the value at any point will be influ-
enced much more by a closer known value of that element than by one farther
away. Interpolation techniques are based on a normal or Gaussian distribution
 Mapping the Contaminant Legacy of a Coking Plant 163

Fig. 1. Avenue zoning and soil sampling location

of data, and good spatial correlation. Problems arise when potentially contin-
uous processes have not yet led to a normal spatial distribution, yet overlie
an older continuous, but random process, which is spatially correlated.

Lark [21] models complex soil properties by assuming that the soil con-
tamination is formed by a continuous but random component combined with
a quasi point process. The quasi point process characterizes contamination
164 M. Palaseanu-Lovejoy et al.

(or any other process) in a small area of finite extent, which is represented
by only one (or very few) soil sample(s) and does not diffuse continuously
towards its neighbours. The continuous random processes are representative
for the native metal content of the soil parent material and diffuse sources of
pollution, while the quasi point process is defined by localized point sources
of pollution. This situation may describe the pollution of an industrial site. If
we consider that contamination with the same pollutant can result from both
diffuse and point sources, we can expect its measured values to show very little
spatial correlation, if any. These values, which have the point process values
embedded, are called outliers and are considered unusual in their spatial con-
text. The outliers do not belong to the continuous, but random, distribution
of the majority of data, and are not necessarily extreme low or high values
[2, 23]. In the case of pollution, if the outliers are not statistical anomalies
due to errors in measurement or recordings, they indicate different processes
superimposed on the same area and affecting the same variable [2, 18, 24].

2 Site Description

The colliery built in the 1880s at Avenue (Fig. 1), and the adjacent later lime
and iron works were dismantled by 1938 and the site reverted to agriculture
use. When the new, up-to-date 98 ha Avenue coking plant, built in the early
1950’s to supply the Sheffield steel industry, was working at full capacity, it
carbonised 2,175 tons of coal a day, producing 1,400 tons of smokeless fuel,
65 tons of 77% sulphuric acid, 35 tons of ammonium sulphate, 70,000 litres of
crude benzole, and 250 tons of tar. Operations ceased in 1992 and since 1999
environmental reclamation work has been going on under the supervision of
the Babtie Group [1].

3 Statistics

As part of the reclamation work, the site owners’ consultants drilled 108 bore-
holes (BH) and 266 trial pits (TP). Seven hundred and twenty nine soil sam-
ples from depths between 10 cm and 18 m below surface level were analysed
for PAH16. The PAH16 levels in parts of the site are two orders of magnitude
higher than the PAH16 environmental threshold of 1,000 ppm. Overall the
concentrations span from 0.05 ppm to over 20,000 ppm [1]. The soil samples
were divided into four categories according to the depth of sampling, 0–1,
1–2, 2–4 m, and more than 4 m. It was originally hypothesised that the 265
soil samples between 10 cm and 1 m below surface level would be spatially cor-
related. To test this hypothesis the empirical PAH16 semi-variogram (Fig. 2)
was built using the Geostatistical Analyst tools and ArcGIS 8.3, considering
that it is more likely to have similar measured values close to the estimated
 Mapping the Contaminant Legacy of a Coking Plant 165

Fig. 2. Empirical semi-variogram for PAH16

point, but different measured values at further away. In this case the assump-
tion is that the difference in values between two samples depends only on the
distance between them and their relative orientation [5, 16]. In this case the
variance, or standard deviation, of the sample value differences, varies only
with the distance and the direction h between samples and it is known as a
variogram.
1
n
2γ ∗ (h) = [z(x) − z(x + h)]2 , where : (3)
n 1
– n = number of data pairs within a given class of distance and direction;
– γ ∗ (h) = calculated semi-variance;
– z(x) = value of the sample at location x;
– z(x + h) = value of the sample at location x + h;
The results are plotted into a graph in which the horizontal axis represents
the distance h and the vertical axis the experimental semi-variance, respec-
tively. If two samples were picked from the same location, therefore h equals
0, we expect that the semi-variance value to be 0 for both calculated and mea-
sured semi-variance [5, 16]. The semi-variogram (Fig. 2) shows the presence of
both global and local outliers for PAH16 and no spatial correlation. The global
outliers are defined as very high or very low values relative with all the values
166 M. Palaseanu-Lovejoy et al.

in the dataset [2, 24], and in the semi-variogram they plot as distinct hori-
zontal groupings of points [20]. The local outliers are values which, although
not out of the dataset range, are abnormal relative to the surrounding values.
Consequently, the local outliers have high semi-variogram values for pairs of
points close to each other. These points plot close to the semi-variogram axis
γ [2, 20, 23]. The arrows in Fig. 2 show the links between the local outliers
in the empirical semi-variogram (pair of points very close in space with high
semi-variance) and actual data locations in space (Avenue map).
A Q–Q normal plot for this data set suggests at least two different pop-
ulations (Fig. 3a). This is interpreted as one population set modelling the
diffuse pollution process, and the other describing the point source pollution
process. The statistical outliers were identified through a box and whisker plot
(Fig. 3b). This graph is depicting the first quartile, median, and the third quar-
tile of a dataset. The box’s lowest and highest horizontal limits represent the
first and the third quartile positions on the y-axis, respectively. Fifty percent
of the data values are plotted inside this box. The "whiskers" represent the
percentile of the most extreme data-point, which is no more than 1.5 times the
interquartile range from the box [25]. The outliers are values larger or equal
to the sum of the third-quartile and the box interquartile range multiplied
by 1.5 [26]. In Fig. 3b, the identified outliers are above the 80th percentile
and range from 371 ppm to 12,340 ppm (49 samples out of 265, or 18.5%
of the data) and represent the point source pollution from the coking works,
while the remaining values represent the historical, diffuse contamination on
the site. A Q–Q normal plot for these 49 outlier untransformed point source
pollutant values suggests a single statistical population (Fig. 3c)
Spatially, the PAH16 outlier values cluster in three main areas, with a few
isolated outliers elsewhere on the site (Fig. 4). The three main clusters are
associated with (a) waste disposal and tar lagoon 4, or point source 1 (PS1),
(b) stoking area, or point source 2 (PS2), and (c) main plant area, or point
source 3 (PS3). The isolated outliers may be the result of individual spills or
leakage from underground tanks.
The Avenue site was divided into Thiessen polygons based on the PAH16
sampling points (Fig. 5). This procedure took into account each sample depth,

Fig. 3. a: normal Q–Q plot; b: box-whisker plot; c: normal Q–Q plot outliers
 Mapping the Contaminant Legacy of a Coking Plant 167

Fig. 4. Spatial distribution of PAH16

and it was assumed that each polygon has the characteristic of the sampled
data. A narrow uncontaminated area of 12–18 m depth separates PS1 and
PS2. Both PS2 and PS3 have small, unpolluted areas surrounded by high-
polluted areas. Pollution may be present below those uncontaminated areas
that were sampled only to a maximum depth of 2 m, since contaminated areas
around them have very high PAH16 values below this depth. This may imply
168 M. Palaseanu-Lovejoy et al.

Fig. 5. Identification of point source pollution

that uncontaminated material has been deposited on top of the contaminated

ground.
A 3D model of source point pollution plume (PS1) based on 6 BH and 3
TP data, totalling 35 soil samples, was generated using a Thiessen polygon
procedure (Fig. 6). Each sample represents the features of a 3D object with its
Thiessen polygon as the base. The height of the object is obtained by taking
the mid points between the sample depth and the sample depths above and

Fig. 6. 3D modeling of source point pollution plume PS1

 Mapping the Contaminant Legacy of a Coking Plant 169

below it. For the first and last sample depth, 20 cm were subtracted and added
respectively, in order to also place these samples inside 3D objects. The PAH16
concentration decreases from tens of thousands ppm to a few thousands ppm
to tens of ppm over just 2.5–5 m vertical distance. This demonstrates that not
only the pollution is extremely localized, but also that diffusion and dilution
of the point source pollution over 50 years has been relatively slight.

4 Conclusions

Superimposed processes over the same area contributing to variation in the

same phenomenon can mask the spatial correlation of continuous variables.
Identification of outliers helps to decide which data reflects the process that
failed to diffuse over the entire study area and thus obliterate the initial spatial
variability, but was significant enough to disturb the studied phenomenon’s
spatial correlation. The outlier dataset clusters spatially with the main clus-
ters being associated with the waste disposal tip and tar lagoon 4, stocking
area, and the main plant area. The remaining outliers probably represent
the outcome of individual spills or leaks from underground tanks. The three-
dimensional modelling of the pollution plume PS1 proves that the pollution
diffusion and dilution over 50 years has been relatively insignificant. Three-
dimensional modelling of pollution plumes combined with qualitative and
historical data most likely will provide valuable information for remediation
programmes, and increase our understanding of pollution processes.

Acknowledgements

I would like to thank all who have helped me in this work, particularly Babtie
and East Midlands Development Agency, UK, for access to the site; Babtie,
UK, for permission to use the site data; Nigel Lawson for help in gaining
access to the site; and my supervisors: Prof. Ian Douglas and Prof. Robert
Barr.
References

1. Babtie Group (2000) The Avenue Coking Works, Ground model & preliminary
contamination assessment. Babtie Group, Fairbairn House, Ashton Lane, Sale,
Manchester
2. Barnett V, Lewis T (1994) Outliers in Statistical data. Wiley series in proba-
bility and mathematical statistics, 3rd edn. John Wiley & Sons: New York
3. Burrough PA, McDonnell RA (1998) Principles of geographical information
systems. Oxford University Press: Oxford
4. Carlon C, Critto A, Marcomini A, Nathanail P (2001) Risk based characteriza-
tion of contaminated industrial site using multivariate and geostatistical tools.
Environmental Pollution 111, pp 417–427
5. Clark I, (1978) Practical geostatistics. Elsevier Applied Science: New York
6. Clark and Harper (2000) Practical geostatistics 2000. Geostokos (Ecosse) Lim-
ited: Scotland, UK
7. Clark I, Harper WV (2001) Practical geostatistics 2000. Ecosse North America
Llc.: Columbus, OH
8. D.E.T.R. (2000a) Contaminated land: Implementation of Part II A of the En-
vironmental Protection Act 1990. London: HMSO
9. D.E.T.R. (2000b) Guidelines for environmental risk assessment and manage-
ment. Revised Departmental Guidance. London: HMSO
10. D.o.E. (1987) Problems arising from the redevelopment of gas works and similar
sites. Environmental Resources Limited, 2nd edn. London: HMSO
11. D.o.E. (1995a) Gas works, coke works and other coal carbonisation plants,
Industry Profile. London: HMSO
12. D.o.E. (1995b) A guide to risk Assessment and risk management for environ-
mental protection, London: HMSO
13. DTLR (2002) Development on land affected by contamination. Consultation
paper on Draft planning technical advice, London: HMSO
14. Environment Agency (2002) Dealing with contaminated land in England,
Progress in 2002 with implementing the part IIA regime, Environment Agency,
Rio House, Bristol, UK
15. Goovaerts P (1997) Geostatistics for natural resources evaluation, Oxford
University Press
 Mapping the Contaminant Legacy of a Coking Plant 171

16. Goovaerts P (1999) Geostatistics in soil science: state-of-the-art and perspec-

tives, Geoderma 89, pp 1–45
17. Goovaerts P (2001) Geostatistical modelling of uncertainty in soil science, Geo-
derma 103, pp 3–26
18. Hawkins DM (1980) Identification of outliers, Monographs on Applied Proba-
bility and Statistics. Chapman and Hall
19. Issaks EH, Srivastava RM (1989) An introduction to applied geostatistics.
Oxford University Press: New York, Oxford
20. Johnston K, Ver Hoef JM, Krivoruchko K, Lucas N (2001) Using ArcGIS geo-
statistical analyst, ESRI
21. Lark RM (2002) Modelling complex soil properties as contaminated regionalized
variables, Geoderma 106, pp 173–190
22. Petts J, Cairney T, Smith M (1997) Risk-based contaminated land investigation
and assessment. John Wiley & Sons Ltd.
23. Rousseeuw PJ, Leroy AM (1987) Robust regression and outlier detection. John
Wiley & Sons, Ltd.
24. Shekhar SLC Zhang P (2003) A Unified approach to detecting spatial outliers,
GeoInformatica 7:2, 139–166
25. The R Development Core Team (2003) The R environment for statistical com-
puting and graphics, Reference Index, Version 1.8.0, http://www.r-project.org/,
accessed 8/10/03
26. Tukey JW (1977) Exploratory data analysis. Addison-Wesley Publishing
Company
A Note on Numerical Solutions of Least Squares
Adjustment in GNU Project Gama

Aleš Čepek and Jan Pytel

Faculty of Civil Engineering, CTU Prague, Prague, Czech Republic

cepek@fsv.cvut.cz
pytel@fsv.cvut.cz

1 Introduction
Project Gama for adjustment of geodetic networks was started at the de-
partment of mapping and cartography, Faculty of Civil Engineering, Czech
TU Prague, in 1998. Formerly it was planned to be only a local project with
main goal to demonstrate students the power of object programming and at
the same time to be a free independent tool for comparison of adjustment
results from other sources. The Gama project received the official status of
GNU software in 2001 and now contains a C++ library (including small C++
matrix/vector template library gmatvec) and two programs gama-local and
gama-g3, that correspond to two development branches of the project.
Stable branch of the Gama project is represented by command line pro-
gram gama-local for adjustment of three-dimensional geodetic networks in a
local coordinates system. New development branch of the project (gama-g3)
is aimed to the adjustment of geodetic networks in global geocentric system.
The stable branch (gama-local) enables common adjustment of possibly cor-
related horizontal directions and distances, horizontal angles, slope distances
and zenith angles, height differences, observed coordinates (used in sequen-
tial adjustment, etc.) and observed coordinate differences (vectors). Although
such an adjustment model is now obsoleted by global positioning systems, it
can still serve as an educational tool for demonstrating adjustment procedures
to students and as a starting platform for the development of new branch of
the project (gama-g3).
Numerical solution of least squares adjustment in geodesy is most com-
monly based on the solution of normal equations. As the Gama project
was also meant to be a comparison tool, it was desirable to use a differ-
ent method and Singular Value Decomposition (SVD) was implemented as
the main numerical algorithm. As an testing alternative Gama implements
another algorithm from the family of orthogonal decompositions based on
Gram–Schmidt orthogonalization (GSO). Practical experience with both al-
gorithms are discussed. In the Gama project geodetic input data are described
174 A. Čepek and J. Pytel

in Extensible Markup Language (XML). The primary motivation for usage of

XML was to define structured input data for adjustment of local geodetic
network. The most important feature of XML is probably the ease of defin-
ing a grammar for user data (a class of XML documents) that consequently
can be validated even independently of our applications. One of the goals
of the Gama project is to build a collection of model geodetic networks de-
scribed in XML. The lack of reliable testing data was one of major obstacles
when testing implementation of numerical solution of the geodetic network
adjustment.

2 Adjustment and Analysis of Observations

Geodesy as the scientific discipline is studying geometry of the Earth or, from
the practical point of view, positioning if objects located on the Earth surface
or in its relatively close boundaries. The input information is represented by
geodetic observations.
The spectrum of observation types dealt by geodesy is very wide and ranges
from classical astro-geodetic observations (astronomical longitude and lati-
tude, variations and position of the Earth pole), measurements of geophysical
quantities (gravity acceleration and its local anomalies), through traditional
geometric observables like directions, angles and distances to photogrammet-
ric measurements of historical monuments. But of the main importance in
geodesy today are satellite global positioning systems (first of all NAVSTAR
GPS and complementary other systems like DORIS or GLONASS).
The key role in processing of geodetic data belongs to the sphere of ap-
plied statistics in geodesy traditionally called adjustment of observations. The
processing of geodetic observations is determined by the choice of appropriate
mathematical model, which can be symbolically expressed as

f (c, x, l) = 0, (1)

where f is a vector of functions describing relations between constants c, un-

known parameters x and observed quantities l. Corresponding to the three
components of this model are three mathematical spaces: parameter, obser-
vation and model space [1].
Three basic components of the mathematical model (1) are depicted in
Fig. 1, where A, B, G and H are matrices of corresponding linearized relations
(values of constants c are not estimated in geodesy and we can consider them
to be a part of model space). Models can be direct, indirect or implicit; linear
or nonlinear; can occur individually or in combinations

model explicit in x : x = g(l), x = Gl + v

model explicit in l : l = h(x), l = Hx + v
implicit model : f (x, l) = 0, Ax + Bl + v = 0
 A Note on Numerical Solutions 175

Fig. 1. Linear relations between parameter, observation and model spaces

3 Least Squares and Singular Systems

On adjustment of geodetic observations we are relatively often faced with
models leading to singular sets of linear equations. Typically these are mod-
els without fixed points, ie. no points with fixed coordinates are given, or
the number of fixed points is not sufficient (free networks, see [2] for more
information).
Lets take as an example local network with observed directions and dis-
tances from Fig. 2. Relationship between unknown adjusted coordinates and
observations can be expressed after linearization as the project equations

Fig. 2. Example of local geodetic free network

176 A. Čepek and J. Pytel

Ax − l = v (2)

where A is design matrix, x vector of unknowns, l vector of reduced observa-

tions and v vector of residuals (misclosure vector).
In geodesy the number of observation is always higher then number of
unknowns. Project equations (2) thus represent overdetermined system and
matrix A has more rows then columns. Least Squares is the basic method
used in geodesy for observation adjustment, it gives us the unique solution x
of system (2) that minimizes Euclidean norm of residual vector
√
min v v. (3)

A method commonly used for solving projects equations (2) (model explicit
in observations) is based on normal equations

N = A A, n = A l,
x = N−1 n (4)

Apart from unknown vector x (and residuals v) we are always interested in

geodesy in estimates of precision of adjusted quantities, in geodetic practice
represented by variance-covariance matrix of adjusted unknowns Cxx and
adjusted observations Cll

Cxx = m0 N−1 (5)

Cll = ACxx A (6)

The geometric shape of our adjusted network is defined by observed direc-

tions (or angles) and directions. If we fixed coordinates of two or more points
the network shape would be necessarily distorted. Normal equations would
lead to an adjustment solution in which residuals would be dependent on the
coordinates of fixed points. This way we would degrade our observations in
cases when coordinates of network points are either unknown or known with
lower precision.
On the other hand if we consider coordinates of all points to be free,
the corresponding matrix N is inevitably singular; columns of matrix A are
linearly dependent (network can float freely in the coordinate system) and
normal equation matrix N is positive-semidefinite

p Np ≥ 0, p = 0.

To get a unique solution we have to define additional constraints regularizing

the system, preferably without deformation of the network shape. In geodetic
practice we most often meet the following approaches
• Singular system is regularized by introducing pseudo-observations, typi-
cally with huge weights, that play a similar role as a set of constraint
equations.
 A Note on Numerical Solutions 177

• Explicit system of constraint equations is defined to make the given system

regular
Cx = c. (7)
Normal equations then become
+ ,+ , + ,
N C x l
= , (8)
C 0 Λ c

where Λ is the vector of Lagrange multipliers. In this case the matrix C

is problem dependent and need to be known explicitly in advance.
• Euclidean norm of certain subset of unknown parameters vector x is min-
imized 
min x2i , i ∈ O. (9)
xi

The set of indices O can contain all elements, but more often only selected
elements of x.
In the case of plane geodetic free network we can geometrically interpret
the last constraint (9) as follows. By minimization of the Euclidean norm of
residual vector (3) the shape and scale (if at least one distance is available) of
the adjusted network together with covariances of adjusted observations are
uniquely defined. The second additional constraint (9) then defines localiza-
tion of the network in the coordinate system. Apart from the adjusted network
shape we define simultaneously its shift and rotation in the coordinate system.
Another equivalent interpretation is that constraint (9) defines the par-
ticular solution of (2) in which the trace of variance-covariance submatrix
corresponding to indices i ∈ O is minimal.

4 Normal Equations and Numerical Stability

Numerical solution of adjustment of observed quantities based on normal

equations can be numerical unstable and in certain case we should prefer
other numerical algorithms that solve directly the project equations (2). The
possible source of troubles are the normal equations itself, or more precisely
the condition number of normal equations. Let us restrict our discussion here
to the simple case when matrix A does not contain linearly dependent columns
and matrix N is positive-definite.
Condition number of matrix A is defined as
2
λ(A A)max
κ(A) = (10)
λ(A A)min

where λ(A A)∗ denotes maximal and minimal eigenvalue of matrix A A. If

we solve a linear set of equations then its condition number represents the
178 A. Čepek and J. Pytel

minimal upper estimate of the ratio of relative error of x and relative error of
right hand side l.
From the (10) directly follows that the condition number of normal equa-
tion matrix N is the square of condition number of the project equation
matrix A
2
κ (N) = (κ (A)) (11)
We can say that when solving poorly conditioned normal equations we are
loosing twice as much of correct decimal digits in a solution x compared with
any direct solution of project equations.
Probably the most important class of algorithms for direct solution of
project equations (2) is the family of orthogonal decomposition algorithms.
Apart from other goals, GNU project Gama has been planned to be a kind of
etalon, i.e. a tool for checking adjustment results from other software prod-
ucts. For this reason it was desirable to have adjustment based on a different
numerical method other then traditional solution of normal equations and
Singular Value Decomposition (SVD) was implemented as the main numeri-
cal algorithm. As an alternative another orthogonal decomposition adjustment
algorithm GSO (based on Gram–Schmidt orthogonalization) is also available.
We describe briefly both algorithm in the following section.

5 Gram–Schmidt Orthogonalization
Gram–Schmidt orthogonal decomposition is algorithm for computing factor-
ization
A = QR, Q Q = 1 (12)
where Q is orthogonal matrix and R is upper triangular matrix. Matrix R
here is identical to the upper triangular matrix of Cholesky decomposition of
normal equations
N = A A = R Q QR = R R. (13)
Gram–Schmidt orthogonalization is a very straightforward and relatively
simple algorithm that can be implemented in several variants differing in the
order in which vectors are orthogonalized. The following three algorithms are
adopted from [3, 300–301].
Algorithm 1.1 [Modified Gram–Schmidt (MGS) row version]
for k = 1, 2, . . . , n
(k)
q̂k := ak ; rkk := (q̂kT q̂k )1/2 ;
qk := q̂k /rkk ;
for i = k + 1, . . . , n
(k) (k+1) (k)
rki := qkT ai ; ai := ai − rki qk ;
end
end
 A Note on Numerical Solutions 179

Algorithm 1.2 [Modified Gram–Schmidt (MGS) column version]

for k = 1, 2, . . . , n
for i = 1, . . . , k − 1
(i) (i+1) (i)
rik := qiT ak ; ak := ak − rik qi ;
end
(k)
q̂k := ak ; rkk := (q̂kT q̂k )1/2 ;
qk := q̂k /rkk ;
end

Algorithms 1.3 [Classical Gram–Schmidt (CGS)]

for k = 1, 2, . . . , n
for i = 1, . . . , k − 1
rik := qiT ak ;
end

k−1
q̂k := ak − rik qi ;
i=1

rkk := (q̂kT q̂k )1/2 ; qk := q̂k /rkk ;

end

It must not be forgotten that the variant known as Classical Gram–Schmidt

has very poor numerical properties in that there is typically a severe loss of or-
thogonality among the computed qi . Rearrangement of the calculation, known
as Modified Gram–Schmidt, yields a much sounder computational procedure
[4, pp. 230–232].

5.1 Generalized Orthogonalization Algorithm (GSO)

Generalized orthogonalization algorithm (GSO), a method based on Gram–

Schmidt orthogonalization, for numerical solution of various adjustment mod-
els in geodesy was elaborated by František Charamza [5, 6]. GSO was imple-
mented in GNU Gama to conserve this rarely used but interesting method and
to represent an alternative numerical algorithm to SVD (which we expected
to give better numerical results for numerically unstable systems).
Algorithm GSO operates on a block matrix structure
+ , + ,
M1 M2 Q1 Q2
−→ (14)
M3 M4 Q3 Q4

where transition from M to Q is defined by equations

180 A. Čepek and J. Pytel

Q 1 Q1 = 1 (15)
M1 = Q1 R (16)
Q1 = M1 R−1 , Q2 = M2 − Q1 Q 1 M2 (17)
−1 
Q3 = M3 R , Q4 = M4 − Q3 Q 1 M2 (18)

and R is upper triangular matrix.

Algorithm MGS is applied to block matrix M so that column dot products
are computed only for submatrices (M1 , M2 ), projections rki qk are computed
for full columns of M and the whole process is terminated after processing of
all columns of submatrix (M1 , M3 ) . This step is called first orthogonalization
in algorithm GSO.
Lets take as an example system of project equations (2)

Ax − l = v.

and apply algorithm GSO to the block matrix

+ , + ,
A −l Q v
−→
1 0 R−1 x

The result is directly the vector of unknown parameters x and vector of residu-
als v. Cofactors (weight coefficients) of adjusted parameters qxi xj are available
as dot products of rows i and j of submatrix R−1 , cofactors of adjusted ob-
servations qlm ln are computed as dot products of rows m and n of submatrix
Q and mixed cofactors qxi ln similarly as dot products of i-th row of R−1 and
n-th row of matrix Q.

5.2 Algorithm GSO and Singular Systems

Let us suppose now that project equations matrix A contains r linearly inde-
pendent columns and remaining d linearly dependent columns. Without a loss
generality we can assume that linearly dependent columns are located in the
right part of matrix A. We denote linearly independent columns A1 , linearly
dependent columns A2 and the matrix of their linearly combinations α
+ ,
x1
A = (A1 , A2 ) , A2 = A1 α, x= (19)
x2

Now we can rewrite project equations as

v = A1 x1 + A2 x2 − l = A1 (x1 + αx2 ) − l = A1 x̃ − l (20)

As the matrix A1 does not contain linearly dependent columns, the unique
solution x̃ of (20) exists that minimize Euclidean norm of v.
 A Note on Numerical Solutions 181

If we know the matrix α and the vector x̃ then any solution x of

x̃ = x1 + αx2 = (1, α) x (21)

is at the same time the Least Square solution of (20) with the same vector of
residuals v.
If we apply algorithm GSO to the matrix
/ 0 ⎛A A2 −l
⎞
I I 1
M M
MI = 1 2
=⎝ 1 0 0 ⎠ (22)
MI3 MI4 0 1 0

we receive a block matrix

⎛ ⎞
Q1 0 v
QI = ⎝ R−1
1 −α x̃ ⎠ (23)
0 1 0

In the case of singular systems in GSO we have the first orthogonaliza-

tion, that defines particular solution in which the unknowns parameters cor-
responding to linearly dependent columns of A are set to zero. From CGS
directly comes out that matrix α is the matrix of linear combinations from
(19). Cofactors are computed the same way as in the case of regular systems.
On numerical computation of GSO we naturally do not obtain exactly
zero vectors on positions of (almost) linearly dependent columns. We declare
to be linearly dependent those columns of A whose norms drop under a given
tolerance. During the first orthogonalization we set to zero corresponding
subvectors in the area of A2 . These values can be considered just a random
noise that adds no information to the whole solution.
The result of first orthogonalization are first of all the vector of residuals
and cofactors of adjusted observations. Now remains to determine the vector
of unknown parameters x that satisfies condition (9) and its cofactors (weight
coefficients). This step of GSO is called second orthogonalization.
By solving the system of linear equations
+ , + ,
−α x̃
x2 = (24)
1 0

we get, according to (21), a vector x with minimal norm. If we select from

(24) only certain rows, we obtain the solution minimizing the corresponding
subvector. This system can naturally be solved using GSO.
If we need cofactors of adjusted unknowns, as is the standard case with
geodetic applications, we have to process during second orthogonalization the
whole lower submatrix that resulted from the first orthogonalization step.
+ ,
II
 II II
 −α R−1
1 x̃
M = M1 M2 = (25)
1 0 0
182 A. Čepek and J. Pytel

During the first orthogonalization linearly dependent columns in M1 are

identified and are explicitly zeroed. The results of first orthogonalization is
a particular solution in which unknowns corresponding to linearly dependent
columns are all set to zero. Naturally their cofactors are zero as well.
During the second orthogonalization step only the submatrix (QI3 , QI4 ) is
influenced and the orthogonalization process is carried as follows
• Gram–Schmidt orthogonalization runs only through columns correspond-
ing to linearly dependent columns of M1 as if they were numbered
1, 2, . . . , d, where d is the nullity of M1
• dot products are computed only for indexes i ∈ O from the regularization
condition 
min x2i , i∈O
xi
Linearly dependent columns are zeroed during second orthogonalization even
in the region of submatrix (Q3 , Q4 ). Cofactors after second orthogonalization
are computed the same way as in the case of regular systems.

6 Singular Value Decomposition (SVD)

For any real m×n matrix A, m ≥ n, there exists singular value decomposition
A = UWV (26)
U U = 1 VV = V V = 1
where U is m × n matrix with orthogonal columns, W is diagonal matrix
n × n with nonnegative elements and V is square orthogonal matrix n × n.
(this variant is referred to as the thin SVD [4]).
The matrix W is uniquely determined up to the permutation of its diag-
onal elements. Diagonal elements wi are called singular values of matrix A.
Their squares are eigenvalues of n × n matrix A A. Thus, the condition num-
ber of matrix A can be compute as ratio of maximal and minimal singular
value.
wmax
κ(A) = (27)
wmin
With singular decomposition we can directly express the vector of unknown
parameters x from project equations
Ax = l, x = VW−1 U l, W−1 = diag(1/wi ) (28)
If matrix A has more rows then columns (overdetermined system), then the
Euclidean norm of residual vector
v = Ax − l
is minimal and the vector x is the Least Squares solution to project equations
(2).
 A Note on Numerical Solutions 183

For a matrix A with linearly dependent columns d singular values are zero
(d is dimension of null space of A). Singular value decomposition explicitly
constructs orthonormal vector basis of the null space and the range of A.
Columns of the matrix U corresponding to nonzero singular values wi form
the orthonormal base of the range of A. Similarly columns of matrix V cor-
responding to nonzero singular values form the orthonormal basis of the null
space of A.

NA = {x | Ax = 0, x ∈ Rn }
RA = {y | y = Ax, x ∈ Rn }

In the case of rank deficient systems, we set into the diagonal of inverse
matrix W−1 zeros instead of reciprocals for elements corresponding to linearly
dependent columns A

−1 1/wi pro wi > 0
W = diag (29)
0 pro wi = 0
Resulting particular solution x minimizes both Euclidean norm of residuals
and at the same time the norm of unknown parameters x.
Rather surprising replacement of reciprocal 1/0 ≡ ∞ by zero can be ex-
plained as follows. Solution vector x of overdetermined system

Ax = l

can be expresses as the linear combination of columns of matrix V

n + ,
1
x= U(i) l V(i) (30)
i=1
wi

Coefficients in the parenthesis are dot products of columns U and right hand
site l multiplied by reciprocal value of the singular value. Zero singular values
correspond to linearly dependent columns of matrix A that add no other infor-
mation to the given system. Setting corresponding diagonal elements of matrix
W−1 to zeros is equivalent to elimination of linearly dependent columns from
the matrix A.
With matrix W−1 defined according to (29), cofactors are computed the
same way for regular and singular systems
−1 −1
Qxx = N−1 = (A A) = (VW U UWV ) = VW−1 W−T V (31)
Qll = AQxx A = (UWV )(VW−1 W−T V )(VW U ) = UU (32)
 −1 −T  −1 
Qlx = AQxx = UWV VW W V = UW V (33)

Cofactors (weight coefficients) for adjusted parameters, observations and

mixed cofactors are computed, similarly is in the case of GSO, as the dot
products of rows of matrices U and V; multiplied by diagonal elements of
W−1 in the case of cofactors of x.
184 A. Čepek and J. Pytel

6.1 Algorithm SVD and Singular Systems

What now remains is to show how to compute the particular solution that
minimizes only a given subset of subvector x according to the second regular-
ization condition (9). We compose overdetermined system of linear equations

ψc + x = x̂ (34)

where columns of matrix ψ are vectors of null space basis

 
ψ = V(i1 ) , V(i2 ) , . . . , V(id ) , , win = 0

and c is the vector of coefficients of linear combination of null space basis

vectors that, when added to vector x, minimizes selected subvector of unknown
parameters x̂ (here they act as residuals).
From comparing (34) with (24) and (25) it is obvious that for computing
x̂ we can use second orthogonalization of algorithm GSO. If the GSO second
orthogonalization is applied to the matrix V from singular decomposition
   
MII = MII 1 M2
II
= ψΨ , (35)
 
Ψ = V(j1 ) , V(j2 ) , . . . , V(jr ) , , wjk = 0

we obtain a matrix V̂. If we now replace singular value decomposition matrix

V by the matrix V̂, we can compute vector x̂ and all cofactors according to
the same formulas (30), (31), (32) and (33) as in the case of standard SVD
solution x.

7 Network Adjustment in GNU Gama

Gama was started in 1998 as a local educational project, mainly to demon-

strate our students the power and capability of object programming (the
project is written in C++) and at the same time to show some alternatives
to traditional approach of numerical solutions of Least Squares adjustments
based on normal equations. Project Gama was released under the terms of
GNU General Public license and in 2001 received the official status of GNU
software.
Numerical solution of geodetic network adjustment in Gama is based on
an abstract C++ class and currently two derived classes are available imple-
menting algorithms SVD and GSO. SVD is the primary algorithm used in
Gama (one of our long term goals is to add more numerical solutions, namely
solutions exploiting sparse structure of project equations). From this perspec-
tive algorithm GSO was implemented in Gama only as an testing alternative,
either for comparing numerical results and for verification of the adjustment
classes hierarchy in practice.
 A Note on Numerical Solutions 185

It is generally agreed that a bad implementation of GSO can produce dis-

astrous results. For example, during first orthogonalization step of GSO we set
to zeros unknown parameters corresponding to linearly dependent columns.
In the case of free geodetic network adjustment these are coordinates of some
points—the whole network is pinned on these points and clearly, if close points
are selected the regularization is unstable. The order of columns in orthogo-
nalization is important.
From practical experience we know that vector norms in GSO orthogo-
nalization process generally tend to decrease. As GSO is just an alternative
algorithm in Gama and its performance is not a crucial point, we implemented
it with full pivoting, ie in each orthogonalization cycle the vector with max-
imal norm is selected as a pivot (with this modification GSO is about twice
as slow compared to SVD for large networks).
Singular Value Decomposition is a very robust method for dealing with
systems that are either singular or numerically close to singular. Even with
full pivoting we expected GSO to prove to be inferior compared to SVD, at
least in cases with ill-conditioned matrices. Surprisingly, with all real geodetic
networks we have available this was not the case. Apart from real data we
used for testing series of random generated three-dimensional networks.
Our implementation of SVD is based on a classical algorithm published by
Golub and Reinsch [7] (the ALGOL procedure SVD). The decomposition is
constructed in two phases. It starts with Householder reduction to bidiagonal
form followed by diagonalization. Contrary to our expectations, SVD as used
in Gama has not proved to give numerically better results and in some cases
it even lost convergence in the diagonalization phase.
Simple and tempting explanation, that comes first to mind, would be that
SVD implementation in Gama is somehow wrong. After all testing and revi-
sions this does not seem to be the point. A possible explanation might give
us the following quotation from [4]
. . . Finally, we mention Jacobi’s method . . . for the SVD. This trans-
formation method repeatedly multiplies A on the right by elementary
orthogonal matrices (Jacobi rotations) until A converges to U Σ; the
product of the Jacobi rotations is V . Jacobi is slower than any of the
above transformation methods (it can be made to run within about
twice the time of QR . . . ) but has the useful property that for certain
A it can deliver the tiny singular values, and their singular vectors,
much more accurately than any of the above methods provided that
it is properly implemented . . .
Surely to have more numerical methods implemented in Gama would be help-
ful, for example the above mentioned Jacobi’s method for SVD.
A practical problem, during testing of the adjustment methods in Gama,
was relative shortage of reliable observation data and their adjustment results
for testing. To enable easy comparison with other softwares we defined descrip-
tion of geodetic networks in XML (we use DTD for the definition of the formal
186 A. Čepek and J. Pytel

syntax of our structured data). Conversion from a well defined data format
into XML is relatively simple task but processing of XML is not a trivial task
and cannot be done without a XML parser. In GNU Gama project we use
XML parser expat by James Clark, see http://expat.sourceforge.net/.
We believe that XML is the best data format for description and exchange of
structured data in Gama project. One of the goals of our project is to compile
a free collection of geodetic networks described in XML.
References

1. Petr Vaníček and Edward J. Krakiwsky (1986) Geodesy: The Concepts, 2nd ed.,
North-Holland, Amsterdam
2. Karl-Rudolf Koch (1999) Parameter Estimation and Hypothesis Testing in Lin-
ear Models, 2nd ed., Springer-Verlag, Berlin
3. Åke Björck (1994) Numerics of Gram–Schmidt Orthogonalization, Linear Alge-
bra and Its Applications 197, 198:297–316
4. Gene H. Golub and Charles F. Van Loan (1996) Matrix Computations, 3rd ed.,
The John Hopkins University Press, Baltimore
5. Charamza, F. (1979) GSO—An Algorithm for Solving Least-Squares Problems
with Possibly Rank Deficient Matrices, Optimization of Design and Computation
of Control Networks, Akadémiai Kiadó, Budapest
6. Charamza, F. (1978) An Algorithm for the Minimum-Length Least-Squares So-
lution of a Set of Observation Equations, Studia geoph. et geod., Vol 22, pp.
129–139
7. G. H. Golub and C. Reinsch (1971) Singular Value Decomposition and Least
Squares Solutions, Numer. Math. 14, 403–420 (1970), Handbook for Auto.
Comp., Vol II—Linear Algebra, 134–151.
Presentation of Entrepreneurship Data
and Aspects of Spatial Modeling

Robert J. Breitenecker1 , Jürgen Pilz2 , and Erich J. Schwarz3

1
Department of Innovation Management and Entrepreneurship, University
of Klagenfurt, Klagenfurt, Austria
robert.breitenecker@uni-klu.ac.at
2
Department of Statistics, University of Klagenfurt, Klagenfurt, Austria
juergen.pilz@uni-klu.ac.at
3
Department of Innovation Management and Entrepreneurship University
of Klagenfurt, Klagenfurt, Austria,
erich.schwarz@uni-klu.ac.at

1 Introduction
Positive effects of new firms on the job market, technology transfer, and contri-
butions to structural change has turned political attention to start-ups. Every
year the number of new firm creations increases, where on the other hand the
number of major enterprises decreases.4 More and more firms have no employ-
ees and a trend to small-scale self employment can be recognized.5 Political
and economic support programs try to revoke regional discrepancies of busi-
ness activity, firm development and foundation activity. These programs have
to be evaluated and improved continuously.
An meaningful curatorial foundation statistic for entrepreneurship research
and a statistic to assess the entrepreneurial activity in Austria for political
decision making does not exist until now.6 Solely the Federal Economic Cham-
ber of Austria (WKO) reports a statistic of foundation activity of commercial
firms every year.7 This statistic permits to observe a trend of firm foundations,

4
Cp. Wirtschaftskammer Österreich [16]: 23.
5
Cp. Schwarz and Grieshuber [11]: 103ff.
6
For the statistical situation in Germany see e.g. Fritsch et al. [3]: 2f. For the effort
to the development of the curatorial statistical system in Germany cp. Struck [14]:
41ff.
7
The Federal Economic Chamber of Austria is the legal representation of interests
of Austrian entrepreneurs. In its founding statistic the number of new start-ups
are calculated from new entrants into the membership database of the WKO. To
exclude pseudo foundations and multiple data set entries the database has been re-
vised. A detailed description of data revision can be found in Wirtschaftskammer
Österreich [16].
190 R.J. Breitenecker et al.

but does not map the overall Austrian foundation activity. Firms which are
not in the scope of the WKO are not included in this statistic.8 Few other
Austrian public services cumulate data from newly founded firms, but the
access to these data sources is limited and the data is not appropriate for
research.9
To ensure a continuous and complete evaluation of supporting programs
it would be wise to design a monitoring system for all Austrian enterprises
which includes all commercial and noncommercial firm foundations and clos-
ings. Such a system would be a valuable source for entrepreneurship research,
but suitable statistical measures and methods for presenting and modeling
non-normal spatial entrepreneurship data are needed to build such an overall
monitoring system.
This paper briefly reports measures and methods commonly used in de-
scriptive statistics for presenting entrepreneurship data in reference to its
spatial distribution. We illustrate these by an example of numbers of new
start-ups in Austria in the year 2001.10 By applying different measures we
show how sensible presenting regional differences in foundation activity can
be. Further we will give a brief introduction of spatial general linear mod-
els [5] and the hierarchical Bayesian models for count data [15] for modeling
non-normal spatial data.

2 Presenting and Mapping Entrepreneurship Data

Charts and tables are instruments for descriptive and explorative data anal-
ysis. They are helpful in visualizing data, building hypotheses and presenting
results from statistical computation with spatial reference. To compare re-
gional discrepancies, the right measure has to be specified to include different
area or population of regions in the calculation or presentation.
For example if the differences in firm foundation activity of Austrian
provinces are to be compared, the absolute numbers or the percentage of
counted foundation will not be practical. Regions with different size and pop-
ulation cannot be compared with non standardized measures. Although the
absolute values and the percentage are improper measures, both can be found
in regional comparisons.11
We give an example of how different measures of foundation activity can
influence the ranking of regions. Figure 1 shows the number of firm foundations
8
Foundations in the field of agriculture and forestry and freelancers which belong to
another chamber or not, are not registered by this and any other official statistic.
9
E.g. social insurance institution, finance office, commercial credit agencies.
10
Data from Wirtschaftskammer Österreich [16].
11
On the web site of the Lower Austria’s Business Portal the foundation statis-
tic of WKO can be found, but on the site only the absolute counts and the
percentage to all new Austrian firms for 2002 are presented (www.loweraustria.
biz/upload/downloads/Betriebsgruendungen%202002.pdf).
 Presentation of Entrepreneurship Data 191

Fig. 1. Percentage of new firm foundations in Austrian provinces 2001

of Austrian provinces in percent to all new start-ups in Austria in a map. In

the year 2001, Vienna is ranked first with 6,145 new firm foundations or a
percentage of 23.1%, before Lower Austria with 5,351 new firms or 20.6%.
It follow Styria, Upper Austria, Tyrol, Salzburg, Carinthia and Vorarlberg.
The province of Burgenland with 869 or 3.3% of new enterprises of Austria
is found at the lower end in this statistic.12 We will see how the ranking of
provinces (especially Upper Austria and Burgenland) concerning foundation
activity will change when applying two different standardized measures.
A standardized measure for presenting the number of firm foundations on
the level of provinces is the foundation intensity, presented in the foundation
statistic of the WKO.13 The intensity is the percentage of start-ups in relation
to all active members of the WKO in the according region, where active
members are all existing and active firms registered by the WKO. Figure 2
shows that Burgenland is now on the fourth place with a percentage of 9.2%
new firms above the average of Austria with 8.7%. Upper Austria is with 7.4%
below average only on the sixth place.14
Egeln et al. [2] calculate a measure for start-ups in relation to the potential
of foundations, where the potential founders are the employees of the observed
region and province, respectively. Start-ups per employees is the number of
firm foundations in relation to 1,000 employees in the region.15 Figure 3 shows
this statistic. Compared to Figs. 1 and 2, differences in ranking of the nine

12
Cp. Wirtschaftskammer Österreich [16]: 21.
13
Cp. Wirtschaftskammer Österreich [16]: 22; Fritsch and Niese [4]: 4. Fritsch and
Niese calculate the number of new start ups in relation to 100 existing firms of
the respective region.
14
Cp. Wirtschaftskammer Österreich [16]: 22.
15
Cp. Egeln et al. [2]; Fritsch and Niese [4]: 3f.
192 R.J. Breitenecker et al.

Fig. 2. Start-ups per 100 existing firms in Austria 2001

Austrian provinces can be recognized. Now Burgenland is ranked first, Upper

Austria is ranked at last.16
But also this measure is not eligible for regional comparison, because every
province has a different industrial history and therefore a different firm struc-
ture. Figure 4 shows the number of employees per working place in the nine
provinces, from which the firm structure in the provinces can be deduced.

Fig. 3. Start-ups per 1,000 employees in Austria 2001

16
Data from Wirtschaftskammer Österreich [16]: 21 and Statistik Austria [13] with
own calculations.
 Presentation of Entrepreneurship Data 193

Fig. 4. Employees per place of work in Austria 2001

Upper Austria with its iron, steel and chemical industry has many major
enterprises. Burgenland seams to have more small enterprises.17
This short example shows that choosing the right measure for such data is
very sensitive, particularly if political decisions have to be made on the basis
of such statistics. Area, population and firm structure should be considered
by comparing different countries or provinces.
To make statistical information from different EU countries and regions
comparable, EUROSTAT has established the Nomenclature des unites terri-
toriales statistiques (NUTS). Using NUTS classification ensure that regions
of comparable size all appear at the same level and making it possible to com-
pare. Each NUTS unit contains regions which are similar in terms of area,
population, economic weight or administrative power.18 In Table 1 eigth dif-
ferent levels for presentation and comparison of regions in Austria, including
the NUTS units, are presented.19 It shows the configuration and the number
of regions of these units.
With the exception of the post code, all lower levels can be aggregated to
a higher level. To assign new firm foundation activity to the levels of NUTS
3, political districts or communities, the addresses or the post codes of the
firms can be used. A main problem in that case is that classification with the
post codes to these levels is not unique. The range of post districts overlap

17
Data Source: Statistik Austria [13].
18
For more information about NUTS see EUROSTAT on the web
(http://europa.eu.int/comm/eurostat/ramon/nuts/splash_regions.html).
19
Data from http://www.statistik.at/fachbereich_topograph/tab2.shtml for politi-
cal units in Austria, http://www.statistik.at/verzeichnis/nuts.pdf for NUTS clas-
sification and http://www.statistik.at/verzeichnis/gemeindeverzeichnis.shtml to
count post regions.
194 R.J. Breitenecker et al.

Table 1. Regional units for presentation in Austrian maps

unit regions of unit no. of regions

NUTS 0 Country – Austria 1
NUTS 1 Groups of Provinces 3
NUTS 2 Austrian Provinces 9
NUTS 3 Groups of Political Districts 35
district Austrian Political Districts 99
post code Post Regions 2,073
community Austrian Communities 2,359
locality Austrian Localities 17,364

with administrative boundaries. To ensure a unique classification, the total

firm address, with the knowledge in which community it falls, has to be used.
An alternative to the regional levels is to assign the firm address to ge-
ographic coordinates (x, y). Different commercial providers offer the service
to assign addresses to geographic coordinates.20 A subsequent aggregation to
other levels can be done ex post.

3 Spatial Model Building for Entrepreneurship Data

By presenting count data in epidemiology it is common practice to smooth

over regions to avoid extreme values for regions with few observations or less
population. Such extreme values are hard to interpret and lead to misinter-
pretations. With kernel smoother geographical trends and implications will
be more trusty.21 Considering the example of new start ups the foundation
rate of region i will be calculated as weighted sum of the rates of neighbored
regions.22
There are several possibilities to calculate spatial dependences and neigh-
borhoods between coordinates and/or regions, but it should be decided
whether these methods make sense applied to entrepreneurship data. A basic
method to define the neighborhood between two coordinates is the euclidean
distance. For calculating this or any other metric between regions, the geo-
graphic centers, central places or the main cities of these regions have to be
defined. The method to define neighborhoods of regions over common borders
is well known in epidemiology. For epidemiological and environmental data
these methods seem to be acceptable but don’t seem to be for entrepreneur-
ship data. It makes no sense to define the distance between two regions or
enterprises by a straight line, when practically streets and/or mountains have
20
E.g. WIGeoGIS for Austria (http://www.wigeogis.com).
21
Cp. Koboltschnig [8]: 9.
22
A detailed introduction in kernel density estimation can be found in Silverman
[12].
 Presentation of Entrepreneurship Data 195

to be negotiated or two regions with common borders don’t have to share one
street. On this account an alternative defining neighborhoods between regions
could be common infrastructure. In this case two regions are neighbors, if they
have a street, a highway or a railroad line in common. For a detailed illustra-
tion of different neighborhood definitions and different spatial linear modeling
strategies in the field of entrepreneurship research see Breitenecker [1].

3.1 Generalized Linear Model for Spatial Prediction

Gotway and Stroup [5] introduce a spatial approach for analyzing non-normal
data. We give a brief introduction how in terms of Gotway and Stroup [5] the
theory of generalized linear models can be extended to include discrete and
categorical data for spatial prediction.
Let Z = (Z(s1 ), . . . , Z(sn )) be a vector of random variables, each hav-
ing a distribution in the exponential family, and z = (z(s1 ), . . . , z(sn ))
the corresponding vector of data values at observed spatial locations s =
(s1 , . . . , sn ) . Suppose we want to predict a vector of k random variables Z 0 =
(Z(s0,1 ), . . . , Z(s0,k )) at unobserved spatial locations s0 = (s0,1 , . . . , s0,k ) .
We assume that the mean function for Z and Z0 can be written as

E(Z) = μ(s)
E(Z 0 ) = μ(s0 ),

where μ(s) and μ(s0 ) are n × 1 and k × 1 dimensional mean vectors associated
with data locations s and prediction locations s0 , respectively.
We define the link function

η = g(μ(s)) = Xβ, (1)

where X is an n × p matrix of explanatory variables or a design matrix at

observed locations and β is a p × 1 vector of parameters. Alternatively the
mean function may be written as

μ = E(Z) = h(Xβ),

where h(·) = g −1 (·) is the inverse link function. In case of our example with
the number of new start-ups, the canonical link function is the log link η =
log(μ(s)) for Poisson distributed data.
Further we assume that
+ ,   
Z
var ≡ V = ZZ Z0 ,
Z0 0Z 00
  
where ZZ , Z0 , and 00 are known positive definite matrices of dimensions
n × n, n × k and k × k, respectively. In practice the general symmetric positive
definite variance-covariance matrix V can be calculated by
196 R.J. Breitenecker et al.

V = var(Z) = υμ1/2 R(α)υμ1/2 , (2)

where υμ = diag[υ(μi )], and υ(μi ) is the general form of the variance function.
The matrix R is the correlation matrix that describes the spatial dependence
among the observations, which in practice will be estimated by a semivari-
ogram model with parameter vector α, denoting the parameters nugget effect,
partial sill, range. In terms of mátern semivariogram model, introduced by
Handcock and Stein [6], α will be extended by the smoothness parameter. For
Possion data υ(μi ) = μi and the variance-covariance matrix V in (2) can be
written as
V = diag[μi ]1/2 R(α)diag[μi ]1/2 .
Gotway and Stroup [5] emphasize that estimation with the generalized lin-
ear model is a maximum likelihood procedure, but that the full log-likelihood
for estimating β is not needed. It is sufficient to describe the relationship be-
tween the mean and the model with the link function, the form of the variance
and the relationship between the variance and the mean. This is fulfilled by
the quasi-likelihood procedure (cp. [5,9: 323ff]).
Prediction with generalized linear models can be accomplished by obtain-
ing β̂ G as an iterative solution vector corresponding to equation

X  W Xβ = X  W z ∗ ,

where W = D V −1 D, D = diag[∂μi /∂ηi ] is an n × n matrix, and z ∗ =

η + D−1 (z − μ). Further estimating η as η̂ = X β̂ G , corresponding to the
data, and ηˆ0 = X0 β̂ G , corresponding to the variables to be predicted. Then
μ̂(s) = h(η̂) and μ̂(s0 ) = h(η̂ 0 ) can be used with the appropriate covariance
Matrix V in (2), to calculate Ẑ 0 from
−1
Ẑ 0 = μ̂(s0 ) + Σ0Z ΣZZ (Z − μ̂(s)),

Gotway and Stroup [5] emphasized that the estimated parameter vector
β̂ G will still be consistent for β even if the correlation matrix is not correctly
specified.
A further extension of general linear model theory introduced by Hastie
and Tibshirani [7] are the general additive models. The generalized additive
model
 differs from the generalized linear model in that an additive predictor
f
j j (X j ) replaces the linear predictor Xβ in (1). This theory should be
adapted to apply it with spatial data.

3.2 Spatial Bayesian Model with Count Data

Ver Hoef and Frost [15] develop a Bayesian hierarchical model for analyzing
trend, abundance, and effects of covariates for monitoring programs of multi-
ple sites and apply it to counts of harbor seals in Alaska. This approach also
 Presentation of Entrepreneurship Data 197

has broader application in other monitoring situations. We show in this sec-

tion, how we can apply this Bayesian hierarchical model for analyzing trend
and effects of covariates for monitoring the number of new start-ups.
Let Zij be a random variable of the numbers of new firm foundations in the
j-th year and in the i-th region. The Bayesian hierarchical model in Ver Hoef
and Frost [15] begins with Poisson regression for each observation. We write
z
f (zij ) = exp(−λij )λijij /zij !,

with
ln(λij ) = θij + x β + ij ,
where θij is an intercept, x = (x1ij , . . . , xpij ) is a p × 1 vector of observed
values of covariates in region i and year j, β = (β1i , . . . , βpi ) is a p × 1
vector of parameters, and ij is an overdispersion parameter. We assume that
conditional on the covariates, all observations are independent, then we can
write the joint density 3
f (z|θ, β) ≡ f (zij ).
Further a separate trend model for each region is developed. In Ver Hoef
and Frost [15] f (θij |τi , δ 2 ) = N (τi , δ 2 ) is a normal distribution and the joint
distribution can be written as
33
f (θ|τ , δ 2 ) = f (θij |τi , δ 2 ).
i j

In the next level of hierachy the region specific covariate parameters are
grouped. The joint distribution is given by
33
f (β|μ, σ) = f (βpi |μp , σp2 ),
p i

where in Ver Hoef and Frost [15] the region specific covariate parameters are
given a normal distribution with mean μp and variance σp2 .
Further in this level of hierachy the region specific covariate parameters
for the trend parameters are grouped. Ver Hoef and Frost give them a normal
distribution with mean η and varince γ 2 . The joint distribution is given by
3
f (τ |η, γ) = f (τi |η, γ 2 ).
i

The overdispersion parameters are grouped and the joint distribution is

33
f (|0, ξ) = f (ij |0, ξi2 ),
i j

and 3
f (ξ|νa , νb ) = f (ξi |νa , νb ).
i
198 R.J. Breitenecker et al.

In Ver Hoef and Frost [15] f (ij |0, ξi2 ) is a normal distribution with mean 0
and variance ξi2 and f (ξi |νa , νb ) is a gamma distribution with parameters νa
and νb .
In the fourth and final level of the hierarchy diffuse priors have to be given
to μp , σp2 , δ 2 , ηq , γq2 , νa and νb . Ver Hoef and Frost [15] give the mean pa-
rameters μp and ηq a normal distribution with mean 0 and variance 1,000,000
to constitute the uncertainty. A gamma distribution with parameter a and b
equal 0.001 is given to σp2 , δ 2 , γq2 , νa and νb .
With the Bayes theorem we can write the posterior distribution

f (θ, β, τ , , δ 2 , ξ, μ, σ, η, γ, νa , νb |z) ∝
f (z|β, θ)f (β|μ, σ)f (θ|τ , δ 2 )f (|0, ξ)f (τ |η, γ)f (ξ|νa , νb )
f (δ 2 )f (μ)f (σ)f (η)f (γ)f (νa )f (νb ).

Markov Chain Monte Carlo technique (MCMC) make it possible to ob-

tain samples from posterior distribution. From these samples functions and
summaries of the posteriori distribution can be computed. For an alternative
method to obtain the predictive density see Pilz and Spoeck [10].

4 Conclusion
In this paper we have briefly discussed some ideas for presenting entrepreneur-
ship data and have pointed out how sensible the right choice of a suitable
measure can be, to compare regions or political districts. Further we have
briefly presented two approaches for spatial modeling entrepreneurship data,
the generalized linear model from Gotway and Stroup [5] and the Bayesian
hierarchical model for count data from Ver Hoef and Frost [15]. In a future
project we will test how good these two approaches can be applied to model
firm foundations and firm survival in respect to their spatial location. A main
problem in the latter approach will be to find suitable prior distributions for
the parameters and to calculate the posterior distribution with Markov Chain
Monte Carlo techniques.
References

1. Breitenecker R (2007) Räumliche Lineare Modelle und Autokorrelationsstruk-

turen in der Gründungsstatistik - Methodische Analyse und empirische Tests.
PhD thesis, University of Klagenfurt
2. Egeln J, Gassler H, Schmidt P (1999) Regionale Aspekte von Unternehmen-
sneugründungen in Österreich. Nomos-Verlag, Baden-Baden
3. Fritsch M, Grotz R, Brixy U, Niese M, Otto A (2002) Gründungen in Deutsch-
land: Datenquellen, Niveau und räumlich-sektorale Struktur. In: Schmude J
und Leiner R (eds) Unternehmensgründungen: Interdiszipiläre Beiträge zum
Entrepreneurship Research. Physica-Verlag, Heidelberg: 1–31
4. Fritsch M, Niese M (1999) Betriebsgründungen in den westdeutschen Raumord-
nungsregionen von 1983–97. Freiberg Working Papers, No. 20
5. Gotway CA, Stroup WW (1997) A Generalized Linear Model Approach to
Spatial Data Analysis and Prediction. Journal of Agricultural, Biological, and
Environmental Statistics Vol. 2, No. 2: 157–178
6. Handcock MS, Stein ML (1993) A Bayesian Analysis of Kriging. Technomet-
rics, Vol. 35, No. 4: 403–410
7. Hastie TJ, Tibshirani RJ (1990) Generalized Additive Models. Chapman &
Hall, London
8. Koboltschnig RG (1998) Anwendungen Bayesscher Modelle in der Räumlichen
Epidemiologie am Beispiel von Lungen-Karzenomdaten in Westösterreich. PhD
thesis, University of Klagenfurt
9. Mc Cullagh P, Nelder JA (1989) Generalized Linear Models. Second Edition,
Chapman & Hall, London
10. Pilz J, Spoeck G (2008) Why do we need and how should we implement Baysian
Kriging methods?, Stochastic Environmental Research and Risk Assessment,
Vol. 22, No. 5: 621–632
11. Schwarz EJ, Grieshuber E (2003) Vom Gründungs- zum Jungunternehmen:
Eine explorative Analyse. Springer, Wien
12. Silverman BW (1986) Density Estimation for Statistics and Data Analysis.
Chapman & Hall, London
13. Statistik Austria (2003) Arbeitsstättenzählung 2001: vorläufige Ergebnisse.
http://www.statistik.at/wdbs/jsp/aztabellen.jsp
200 R.J. Breitenecker et al.

14. Struck J (1999) Quo Vadis Gründingsstatistik?. DtA, Wissenschaftliche Reihe,

Band 10, Berlin
15. Ver Hoef JM, Frost KJ (2003) A Bayesian hieracichal model for monitoring
harbor seal changes in Prince William Sound, Alaska. Environmental and Eco-
logical Statistics, Vol. 10: 201–219
16. Wirtschaftskammer Österreich (2003) Unternehmensgründungen in Österreich.
Wien
 Part III

Integrated Information Systems: Combining

(Geo)Statistics, GIS and RDBMS
GIS, Users, Developers, and Spatial Statistics:
On Monarchs and Their Clothing

Konstantin Krivoruchko1 and Roger Bivand2

1
Environmental Systems Research Institute Redlands, Redlands, CA, USA
kkrivoruchko@esri.com
2
Norges Handelshøyskole, Bergen, Norway
Roger.Bivand@nhh.no

1 Introduction and Motivation

The development and documentation of software for the analysis of geograph-

ical data is maturing, and the needs and desires of varying user communities
are becoming clearer3 . Certainly today there are more users in more commu-
nities, and in general much more data than before, even though data is more
accessible in some countries than in others. Many more users are now meeting
geographical data through geographical information systems software (GIS).
GIS are general-purpose environments for handling geographical data, and do
not assume that the user will need to make predictions or draw inferences from
the data, or error propagation in geographical data analysis. Indeed, much of
current progress in GIS is in making it easier for users to construct maps at
the front end and in providing open and consistent data base support at the
back end. Neither of these two areas lie close to the central concerns of statis-
tical data analysts, such as making predictions with associated uncertainties,
but can be of great value to them.
In meeting and undertaking dialogues with users and developers, it seems
both valid and important to attempt to explore some of the assumptions the
different communities hold themselves, have about each other, and the tasks
they undertake separately and jointly. Some of the points to be made will draw
on the ontology discourse in geographical information science (GIScience),
which may be helpful in throwing light on different assumptions made by
different communities, not just technical/motivational, but also related to the
sociology of organizations and of scientific disciplines.
The paper discusses these issues in general terms, but more specifically
touching on tools and methods that may propagate between communities of
users, and on difficulties associated with the use of inference in inappropriate
3
This paper represents the views of the authors as they were when it was written
in 2003.
204 K. Krivoruchko and R. Bivand

settings. In particular, we will present and discuss selected examples of an-

alytical practice that are common, although alternatives exist that resolve
or avoid some of the difficulties of these methods. In some cases, choices
were limited at the time the methods were proposed by lack of access to
computing power and capacity, in others by lack of access to appropriate
software. In further cases, choices of methods seem to reflect barriers to the
diffusion of practise across discipline boundaries, in particular from the sta-
tistical sciences, especially applied statistics, to other fields. This also re-
flects some of the organizational relationships, in that for example work in
one field is most frequently refereed within that field, so that duplications
may not be made plain for some time if at all. On the other hand, scien-
tific progress in ever-more specialized fields is difficult to track, so multiple
apparently original work is more the rule than the exception, and indeed
should be welcomed as providing replicated indications of the potential fruit-
fulness of an approach. It does however introduce the risk of “borrowing”
between different fields of applied science without sufficient reference being
made, and/or without an adequate understanding of the underlying assump-
tions. Analogies can be useful, but can also be misleading, because the his-
tory and rationale of the development of a method may be discarded when
it is “transplanted” into a new field. The transfer and rephrasing of ideas
about geostatistics from meteorology to geology is a classic example: compare
[14, 6, 7, 8, 20, 21].
We will also examine needs for documenting analyses leading to decisions,
because decision support involves not only giving policy advice, but also pro-
viding a clear statement of how the advice was reached. This means not only
documenting data sources and collection methods, but also how the data has
been handled and analyzed. Analysis is partly a matter of the formal defini-
tions of methods, but should specify the implementation or implementations
used to create derivative products used as a basis for policy advice. This means
that it should be possible, as in food and drugs appraisal procedures, to as-
sign responsibility for each step in data analysis, so that another researcher
could replicate it and confirm that the results achieved are in accord with the
data and the specified analytical scheme. This is analogous to the statement
of authority in metadata sources. For statistical analysis, this is known as
reproducible statistical research.
This leads on to a discussion of the benefits different software implementa-
tions can offer one another, be they closed or open source. While open source
software by its nature provides full insight into algorithm implementation, it
may be no better or worse than other implementations in providing facilities
for recording the steps taken in data analysis. It should also be acknowledged
that different user communities expect and require different levels of support
and documentation, and that developers should have some grasp of their needs
for and use of implemented methods. This may be balanced by developers of-
fering guidance, either narrowly prescriptive: do this with data of that class,
or broadly prescriptive: with data of that class, consider the assumptions you
 On Monarchs and Their Clothing 205

are making, and try at least some of the following alternative methods to
check the robustness of your conclusions.
Before proceeding, the reader deserves an explanation for the extension
of the title we have chosen. We found it helpful to stress that even when an
analyst feels “well-clothed” in relation to the expectations of his or her commu-
nity, it may well be that others will have different opinions. This is about as-
sumptions of what “well-clothed” means in different circumstances, and about
the sometimes self-reinforcing views expressed about this in inward-looking
communities of both users and developers. Different users and research com-
munities do things in various ways, often tradition-based, and it is a user’s
inherent right to choose software and tools to use in his/her research and
decision-making. But this right requires that the user (or developer making
tools available) accepts responsibility at least for documenting how the analy-
sis has been done. It is not enough to rely on the assurances of courtiers that
we are fashionably clothed, when our apparel is awry or absent in the view of
others.

2 Assumptions Held by User and Developer

Communities
The landscape in which we are living and working is a reality that is much
more complicated than the statistical framework in most studies. It is defi-
nitely not a flat surface, see Fig. 1. There is nothing to measure using simple
straight-line distances. There are many natural and artificial barriers between
spatial objects. Among these are disciplinary barriers, which make it difficult
to interact fruitfully with people from other disciplines and traditions. The
naïve GIS user is shocked at how different reality is from the models that
describe it. There is no such thing in Nature as the Gaussian distribution and
data homogeneity. There are no clear boundaries between polygonal objects.
Both GIS users and developers are users of reality, but they may approach its
conceptualization differently.
In the real world users of GIS and spatial statistical software vary in their
insight both into statistics and into their own discipline-based domains in
depth and breadth. Some users do not have either the programming expe-
rience or the motivation to modify and customize the software to suit the
needs of analysis, while others will want to do so. Kuan [18] terms this the
integration of the consumer into production, and this is applicable not only to
software development, but to many kinds of scientific endeavors. An immedi-
ate consequence is that if we really want to make Spatial Statistics available to
the average GIS user, which includes tightly integrated spatial statistics in a
GIS environment and “protecting” the user from themselves (that is, inappro-
priate use of methods), then an attractive approach is what is already done
by ESRI, building Geostatistical Analyst within ArcMap. There are some
problems with such an approach for users who feel themselves well educated
206 K. Krivoruchko and R. Bivand

Fig. 1. Typically the landscape in which we are living and working is much more
complicated than our models assume

in statistics/geostatistics and thus able and willing to use additional meth-

ods or modify existing ones; we will discuss one possibility to resolve existing
problems below.
The researcher as user and as developer is often stressed by conflicting
understandings of why analysis is being undertaken and what kinds of results
are acceptable, and to whom. In academic settings, results are scanned for
propriety by referees, using the standards of their communities. Curiously,
they quite often disagree, either on standards, or on how they “read” the
product with which they have been presented. In applied settings, reviewing
and evaluation is also practiced, but there are most often also instrumental
goals for the analysis being carried out. These typically do not adequately
acknowledge the fact that uncertainty is certain. This is a key point of dis-
cord between the use of spatial statistical methods, in which we need both
predictions and estimates of uncertainty, and the demands of practical users,
or rather the organizations for which they work, for certainty.
There are as many logics as we can imagine. This is because logic is based
on systems of axioms and rules for deriving logically true statements. For
example, the greatest mathematician of the 20th century, Andrei Kolmogorov,
formulated the following rule of human logic: Let [P ⇒ Q] and [Q is nice];
then [P ]. Example: If my parents have money, I’ll have a new bicycle [P ⇒ Q];
 On Monarchs and Their Clothing 207

It is nice to have a new bicycle [Q is nice]; Hence my parents have money [P ].

In fact, many users follow such logic, but there are other possibilities.
Chrisman [3] points to weaknesses in the use of ontology-derived terms,
such as bona fide and fiat objects, arguing that the distinction between them
is not very helpful in practice: “The notion of fiat objects does recognize the
human dimensions of practice, but some of these same issues recur in the
objects that are meant to be beyond human interference. Bona fide objects
are just as subject to conventions and standards developed from disciplinary
practice.” This observation has direct application to spatial statistics, in that
disciplinary practices mediate in the treatment of objects, especially when the
observed data objects have been used as a basis, with selected methods, for
modeling and prediction.
One specific difficulty is that spatial statistics as a field is broader than the
application of these methods to geographical data. Some methods, in particu-
lar point pattern analysis, but also others, have direct relevance at much larger
and smaller scales, such as in the analysis of patterns on microscope slides, or
in medical imaging. Because of this, the views of reality embodied in methods
of analysis may or may not be well suited to geographical data. Other views
of reality within geographical scales are difficult to represent using legacy GIS
data models, for example time and a third dimension.
It is arguably the case that even good statistical training will not help
someone who is lacking in domain knowledge terms, so that some “positive”
or “self-reinforcing” intersection of methods practice and domain knowledge
is desirable. If a researcher does not understand his own discipline, methods
or software will not help, but well-structured methods and software can “en-
able” scientists who are aware of the often difficult assumptions of their own
domain. Some of the methods may actually be simple, just good practice in
data analysis, like good laboratory practice. This is associated with learning,
engaging users by asking them questions to try to get them to grasp and op-
erationalise their research problem in a way that lets them both solve that
problem (or admit that it cannot be solved as posed), and learn something
more generic that is transferable to other situations encountered later.
In a similar way, user and developer communities functioning in relation
to spatial statistics software, especially but not only extensible software, allow
participants to become more familiar with each others’ assumptions. Learning
is here the key, conditioned by the willingness of participants to make their
positions explicit in understandable terms. Some software includes methods
and implementations that would not be proposed by statisticians, but are pro-
vided because the domain scientists expect them to be present, if just because
that were once considered appropriate, and were fashionable when the scien-
tists were trained. It is also important to acknowledge that statistical methods
are often seen by domain scientists as not very enjoyable, compared for exam-
ple with fieldwork. Promoting positive attitudes towards analysis ought to be
included in the development of such software, and is in many cases neglected,
because the developer does not share this dislike. If developers were forced
208 K. Krivoruchko and R. Bivand

to do fieldwork instead of coding, they might empathize better with users

of their software. The message needs to be sent clearly that the users’ data,
collected with considerable commitment and interest, and often expense, de-
serve the respect of the developer, embodied in the software – executables,
documentation, and training, and in the virtual community (online discus-
sions, meetings, conferences). Advances in computing hardware may lead to
another problem: the use of ever more popular MCMC methods risk “dumbing
down” science, making statistics mindless computing in cases where compute
power is used instead of analysis.

3 Selected Examples of Conceptual Discourses

and Discords
We will present one extended example of discord between practices in spatial
statistics, concerning indicator kriging usage. The presentation of probability
values for local indicators of spatial association is another topic that will be
mentioned more briefly.

3.1 Indicator Kriging Usage

Journel proposed indicator kriging [13] as an alternative to disjunctive and

multiGaussian (that is kriging after data transformation) kriging, because
they require good understanding of the assumptions involved and were con-
sidered difficult to use at that time. In indicator kriging the data are pre-
processed first. Indicator values are defined for each data location as the fol-
lowing: an indicator is set to zero if the data value at the location s is below
the threshold, and one otherwise:

0 Z(s) < threshold
I(s) = I(Z(s) < threshold) =
1 Z(s) > threshold
Indicator transformation for one-dimensional data is illustrated in Fig. 2.
Transformed input data inside the interval around threshold (for example,

Values above
threshold Threshold μ
become 1’s
1
Measured Value

Measured Value

ε(S)
Values below
threshold 0
become 0’s Z(S)

0 5 10 15 20 25 30 0 5 10 15 20 25 30
X-Coordinate X-Coordinate

Fig. 2. Illustration of the indicator transformation for one-dimensional data

 On Monarchs and Their Clothing 209

this can be measurement error interval) are displayed as red points in the
indicator transformation in the graph to the right. It is quite possible that
they can be exchanged if one more measurement will be taken.
Notice that after indicator transformation, input values near and far from
the threshold became either zeroes or ones. This means that we are loosing
information when transforming the data.
Then these indicator values are used as input to ordinary (sometimes to
simple or universal) kriging. Ordinary kriging produces continuous predic-
tion and we might expect that prediction at the unsampled locations will be
between zero and one (this is often not fulfilled in practice, however). The
prediction is interpreted as the probability that the threshold is exceeded at
location s. For instance, if the prediction equals 0.71, it is interpreted as a
71% chance that threshold was exceeded. Predictions made at each location
form a surface that can be interpreted as a probability map that the specified
threshold is exceeded. If a set of indicators is used as input to ordinary kriging
(for example, 10 quantiles of the input data distribution), the resulting set of
predictions at each location can be combined to give a cumulative probability
distribution from which a probability density distribution can be estimated
and the prediction mean and variance can be calculated.
Although indicator kriging became very popular immediately, a number
of problems have been found. If different semivariogram models are used for
different thresholds, then internally inconsistent results may be obtained. One
possible workaround for this problem is to use the median indicator variogram
for all indicators. However, this nullifies the potential advantage of the model
that the spatial structure of a variable depends on its value. For instance,
we might expect that range of correlation is smaller and variance is larger
for large values. Nowadays indicator kriging is mostly used to provide risk-
qualified predictions (probability that a specified threshold is exceeded) at the
unsampled locations, and not for prediction itself.
Consider this kriging model for the signal Y (s), (see [17]):

Y (s) = m(s) + S(s) + η(s)

where m(s) is a large scale variation (trend), known or estimated, S(s) is a

random process with zero mean and known covariance (small scale variation),
and micro-scale variation η(s) is the variation at a scale, too fine to be rec-
ognizable from the data. Measurement Zi in the location si is a sum of the
signal and independent random error with zero mean and known variance.

Zi = Y (si ) + εi , i = 1, n,

where n is a number of measurements. This allows for more than one mea-
surement at the same data location.
Geostatistical prediction and conditional simulations should not honor the
data if there is measurement error and all real data are not exact. But geosta-
tistical programs usually assume that data are perfect, that is εi = 0, which
210 K. Krivoruchko and R. Bivand

contradicts common sense. Probably this comes from a distrust of statistical

models. Users think that their data are all they really know, see discussion in
[17].
The idea behind indicator kriging is to estimate probability that the spec-
ified threshold T was exceeded:
P r(Y (s) ≥ T |Z) = E(I(s) ≥ T |Z)
assuming that there is no measurement error, that is Z(s) ≡ Y (s). A reason
for this assumption is that
E(I(Z(s0 )) ≥ T ) = P r(Z(s0 ) ≥ T )
= P r(Y (s0 ) + ε(s0 ) ≥ T )
= E(I(Y (s0 )) ≥ T ),
meaning that indicator kriging is a biased predictor of a signal and this bias
can be substantial if measurement error is large. The non-existence of the
filtered predictant is a serious disadvantage of the indicator kriging model. In
practice, predictions in the close vicinity of the data locations are usually not
close to 0 or 1 and predictions jump to 0 or 1 at the data locations. Such a
prediction surface for discrete input data must be questioned.
Kriging is the best linear predictor for Gaussian random variables, but
I(Z(si )) ≥ T ) are Bernoulli random variables and the indicator predictor may
be far from optimal [5]. A semivariogram might be inappropriate measure of
spatial continuity of discrete data, [25].
There is also a problem with block estimation of I(Z(B)) ≥ T ), where B
is an area, from point data Zi since
+
,

1 1
I(Z(B)) ≥ T ) = I Z(u)du ≥ T = I(Z(u) ≥ T )du
|B| B |B| B
see discussion in [11], meaning that an additional assumption concerning the
covariance between point and block indicators, cov(I(Z(si )) ≥ T ), I(Z(B)) ≥
T )), needs to be made.
A basic assumption behind any standard geostatistical model is an as-
sumption about data stationarity. In reality, data often more or less depart
from stationarity, and the solution is to use detrending and transformation
techniques to make data close to stationarity, see case study with comparison
of indicator, disjunctive, and other krigings performance in Krivoruchko [15].
However, indicator kriging uses original data and there is no possibility to
transform data to stationarity. Also, even if the original measurements are
stationary, there is no guarantee that the transformed indicator variable will
be stationary. For simple statistical models, departures from the stationarity
assumption are more serious in their consequences for the reliability of infer-
ence than violation of the distribution assumption for more complex models.
An important advantage of statistical models over deterministic ones is the
possibility to estimate prediction uncertainty. Without data pre-processing,
 On Monarchs and Their Clothing 211

Fig. 3. Comparison of standard error of indicators maps created using indica-

tor (left) and disjunctive kriging (right) after data transformation using threshold
15Ci/km2

the kriging standard error map does not depend on data values, only on mea-
surement density. If input data are transformed to an approximately Gaussian
distribution, prediction standard errors depend on data values. For example,
Fig. 3 taken from [15], compares standard error of indicators maps created
using indicator and disjunctive kriging with data transformation for radionu-
clide soil contamination interpolation in Southern Belarus. The probability
map created using disjunctive kriging is data dependent, and the largest un-
certainty corresponds to areas close to the selected threshold value. Without
reliable information on modeling uncertainty, decision-making may be mis-
leading.
This example shows how the practical use of methods can become encum-
bered with what we can call “encrustations”. A method became established,
that was introduced to address a pragmatic issue, or a group of issues, at
least partly because other methods, acknowledged to be more adequate, were
seen as practically or computationally infeasible, as well as poorly matched
to users’ possibilities. Over the intervening period, not only computational
resources, but also the research bases, have changed, but not least for prag-
matic reasons, analytical practice has not necessarily followed up. We could
have chosen to present other examples of areas where spatial statisticians dif-
fer sincerely in their approaches to analysis, and others have been noted above
in brief. This will for now have to be sufficient to indicate some of the features
of one of many debates.
Because of the problems described above regarding indicator kriging usage,
it is safe to use it as ESDA technique, but not as a prediction model for
decision-making.
It is not advisable to use the conditional indicator simulation model as well,
because of above-mentioned problems with indicator kriging and because of
some other problems, see Gotway and Rutherford [10].
212 K. Krivoruchko and R. Bivand

3.2 Local Indicators of Spatial Association

The main difference between geostatistical and polygonal data analysis is in

the order of specifying covariance/semivariogram matrix and weights of neigh-
bors involved in spatial prediction. In geostatistics, the correlation between
locations separated by a specified distance is modeled first. Then weights are
calculated automatically. In polygonal data analysis weights are defined first.
It is supposed that they reflect the statistical distance between polygons. Cliff
and Ord ([4], p. 11–13) provide the initial formalization of the relationships
as a generalized weighting matrix, most usually termed W. It is usual in the
literature to define the contiguity relation in terms of sets of neighbors of
zone or site i. These are coded in the form of a weights matrix W, with a zero
diagonal, and the off-diagonal non-zero elements often scaled to sum to unity
in each row, with typical elements:
cij
wij =

N
cij
j=1

where cij = 1 if i is linked to j and cij = 0 otherwise. This implies no use of

other information than that of neighborhood set membership. In practice set
membership is almost always defined arbitrary and the most popular way to
define it is on the basis of shared boundaries, centroids lying within distance
bands, and “rook” or “queen” rules, terms borrowed from chess. Often it is
unclear how rook or queen will behave near the boundary of the area under
investigation.
Spatial autocorrelation is the term given to a measure of the correlation
among neighboring values. There are many different ways to quantify spatial
autocorrelation, but the most common index for regional data is Moran’s I
[22, 4]:
N N
N i=j j=1 wij (ri − r̄)(rj − r̄)
Iglobal = N N N
i=1 (ri − r̄)
wij 2
i=j j=1

where r̄ is the global mean value, defined and calculated as a simple average
value based on all the data. The data could be counts or rates, although in our
opinion working with count data is misleading since the underlying population
also varies among the regions, see detailed discussion in [16].
A local version, called a Local Indicator of Spatial Association or LISA by
Anselin [1] is:
+ , N + ,
local ri − r̄  rj − r̄
Ii,std = wij
s j=1
s

where r̄ and s are the overall mean and standard deviation, respectively,
and the weights wij reflect the spatial proximity between regions i and j.
 On Monarchs and Their Clothing 213

This statistic provides a measure of local similarity (or dissimilarity) for each
region. There are problems with LISA. For instance, it is hard to understand
how to interpret a case where, when using adjacency weights, two adjacent
regions have very different statistics.
Getis and Ord [9] and Anselin [1] give expectations and variances for the
local indicators, using both assumptions of normality and randomization, fol-
lowing Cliff and Ord [4] for the global measures. The standard route to draw-
ing inferences has been to treat the square root of the difference between the
observed measure and its expectation divided by its variance, as a standard
normal deviate. The Gaussian distribution can be a good model for continuous
data, but count data are inherently discrete. In randomization, assuming the
observed values are exchangeable, the assumption of stationarity is actually
made and this is violated by counts and rates in any case, and also when sta-
tionarity is not present. Other indices that allow the mean and variance of the
data to vary with the population in each region, and are thus more suitable
for measuring clustering in regional populations are available, see discussion
in [16].
Moran’s I can be modified to relax the assumption of constant mean and
variance. One such statistic for rates, see Walter [26], is:
+ ,
N + ,
yi − r̄ni yj − r̄nj
IiWM = √ wij √
r̄ni j=1
r̄nj

assuming that the underlying risk r̄ to be constant over all regions and esti-
mated from the data over the entire region. This statistic is based on properties
of the Poisson distribution assuming that E(Y i) = r̄ni . The p-values can be
computed using Monte Carlo simulation as follows [16]:
1. Generate simulated values for each region, under the null (or default)
hypothesis of spatial independence; Here we assume the data follow a
Poisson distribution with mean E(Y i) = r̄ni ; these values are simulated
from the Poisson distribution and are not a permutation of the observed
counts.
2. Compute the statistic of interest, in this case U = IiWM for each simulated
data set.
3. Repeat M times. This gives U1 , U2 , . . . , UM .
4. Compare the observed statistic calculated from the available data, say
Uobs to the distribution of the simulated Uj and determine the proportion
of simulated Uj values that are greater than Uobs
The idea is to obtain the proportion of simulated values that are more
extreme that the value determined from the data.
It is natural and users regularly ask for probability values to be made
available for global and local indices of spatial association, but it is a rather
delicate procedure. Some software permutes all the data values across the set
of units, as is typically done for global measures. But this does not provide an
214 K. Krivoruchko and R. Bivand

adequate basis for inferring about the local neighborhood, in which the range
of values found may be much more restricted. One could attempt to simulate
for each neighborhood, but because the numbers of neighbors are small, very
few draws can be made before all possible combinations have been exhausted.
An underlying problem is that global autocorrelation, perhaps reflecting a
trend in the data, will yield apparently significant local measures, and will also
make the use of the whole pool of data values for simulation wrong, because
within a local neighborhood, the trend limits values to a narrow band. Users
are at risk of drawing conclusions from the output of local indicators that
are not robust, and it is not obvious how to indicate to them how dependent
these indicators, and derived measures, such as probability values, are on the
assumptions being made.
Often software users and developers assume that the data are independent
and follow a stationary Gaussian distribution, but this is an unreliable condi-
tions in practice, at least for aggregated data, such as cancer and crime rates.
In fact almost all polygonal data are not continuous and Moran’s I should
arguably be used only for pedagogical purposes. The best approach is to use
Monte Carlo testing. In this approach we generate realizations from a specified
univariate distribution that describes the data, calculate local index for each
polygon and then compute the p-value as in example using Walter’s modified
I above. This certainly is best approach in the case of global statistics, but
it still may be misleading for LISA because number of neighbors is usually
small, less than 10, and any statistics might be insufficient. One possible so-
lution to the problem is to use several different indices as in the case study
by Krivoruchko et al. [16]. If all or most of indices give similar results, we
can safely make conclusion about data clustering or cross-correlation. If not,
further research is required.

4 Reproducible Research
Leisch and Rossini [19] present arguments for making statistical research re-
producible, so that given the same data, another analyst will be able to re-
create the research outcome used in a paper or report. If software has been
used, the implementation of the method applied should also be documented. If
arguments used by the implemented functions can take different values, then
these also need documentation. An example is the way in which a geostatisti-
cal layer in ESRI ArcMap is defined [12]. Most ArcMap layer types store the
reference to the data source, the symbology for displaying the layer, and other
defining characteristics. A geostatistical layer stores the sources of the data
from which it was created (usually a point feature layers), the symbology, and
other defining characteristics, but it also stores the model parameters from the
interpolation, including type or model for data transformation, covariance and
cross-covariance models, estimated measurement error, trend surface, search-
ing neighborhood, and results of validation and cross-validation.
 On Monarchs and Their Clothing 215

Another example of model storage for reproduction and updating is the

geoprocessing environment, see [2]. Geoprocessing tools allow researchers to
combine and interpret data obtained from different sources. As such, they
form important components of an underlying model that takes input data
(coverages, shape files, raster grids) and assimilates them in a meaningful way
to produce output information that can provide a more suitable interpretation.
The ease with which geoprocessing can be implemented within a GIS
makes it easy to forget that such processes inherently alter the input data.
In most cases, the geometric properties of the features of the input data are
altered to form new features and functions of the input attribute values are
transferred to the new features. In many GIS applications, geoprocessing is
just a means to an end. In others, however, a more thorough understanding
of the model may be desirable. The overall goal of modeling may be to under-
stand how assumptions, parameters and variation associated with the input
data affect the resulting output data and the conclusions made from them. In
such cases, a probabilistic framework for model building with geoprocessing
tools may be desirable. Consider the following geoprocessing scenarios:
• A buffer function is used to create a zone of a specified distance around the
features in a layer. How do we know what distance to use? What happens
to our results and conclusions if we increase the distance slightly? Given a
choice of distance, how wrong can our conclusions be? This latter question
can have huge implications in environmental justice, for instance, where
we are trying to decide if under-privileged people are more likely to live
near toxic waste sites, landfills, or other environmental hazards.
• Data are usually available at different resolutions. Union and intersection
geoprocessing operations require that data be aggregated and disaggre-
gated. For example, a soil classification map with polygonal features is
often converted to raster for use in geoprocessing; DEM data exist in just
several resolutions and are often upscaled or downscaled to provide ele-
vation estimates needed for geological and hydrological applications; spa-
tial interpolation methods such as kriging and inverse distance-squared
are often used to provide maps of environmental variables whose values
are then aggregated to in order to link them to public health and dis-
ease information summarized for administrative regions. However, raster
conversion, interpolation, and aggregation and disaggregation procedures
provide only estimates for attributes associated with the newly created
features, not the true values. The user may need to know the accuracy of
the resulting estimates, and the impact of estimation error on additional
calculations.
• The accuracy of spatial data is a very important concern. Locational (po-
sitional) errors occur when the geographical coordinates of a point feature
are not known precisely. This can be due to measurement error, projection
distortion, or reporting errors. Even if locational error is relatively small,
the uncertainty arising from lack of precise geographical coordinates can
216 K. Krivoruchko and R. Bivand

have a substantial impact on spatial analysis. For example, spatial prox-

imity is usually calculated using distances between pairs of data locations
and uncertainty in location coordinates will influence the results of the
raster-based interpolation methods in geostatistical analyses. When these
output raster layers are used as input to other geoprocessing operations,
the errors propagate through the calculations and small errors can quickly
add up to large errors if many calculations are performed. If uncertain
locations are buffered, attribute values from another layer may be mis-
classified. GIS users may want to make sure that locational errors do not
greatly impact their results and conclusions, and if they do, they might
want to be able to track them or adjust their results for them.
In all of these cases it is important to maintain the lineage of the operations
being performed, so that a trail exists allowing either an audit to re-create the
research, or to permit additional examination of changes in conclusions when
data or function arguments are supplemented or modified. Introducing error
propagation in geoprocessing operations may lead to irreproducible results
(layers and maps), but nearly reproducible research: if result is a prediction
with associated uncertainty, it is a realization of the true process, which in-
herently unknown, hence, irreproducible.
An example of a setting in which the user may wish to supplement the
geoprocessing model is when the output of a function or procedure is in a
form that is harder to display in map or tabular form. Say that we have a
map layer with point locations of some events. They appear clustered in some
sense, and we can construct and plot maps of the density of the pattern using
different bandwidths. But we would like to test whether the spatial pattern of
points could have been generated by a cluster process, within a given study
area polygon. Typically, the output will be a plot or summary statistic for the
pattern as a whole, and may vary depending on arguments to the function.
Consider the example of a test of a point pattern representing the places of
residence of juvenile offenders in a part of Cardiff, Wales.
Point pattern analysis is concerned with the location of events, and with
answering questions about the distribution of those locations, specifically
whether they are clustered, randomly or regularly distributed. Point pattern
analysis is very sensitive to the definition of the study area, since a regularly
distributed pattern can be made to seem clustered by including large margins
within the study area. Measures are also subject to boundary corrections,
and most often study area boundaries have to be defined as convex polygons
over the study area, or in the simplest form as rectangles bounding the points
under analysis. The simplest way of exploring point pattern data is by ex-
amining a two-dimensional frequency distribution of counts within equal-area
units imposed on the study area, giving an impression of how the intensity of
the point process varies; this can be extended to kernel estimation. Nearest
neighbor distances are also used to analyze intensity of points, the mean num-
ber of events per unit area at point s. Spatial dependence is captured by the
 On Monarchs and Their Clothing 217

second order properties of a spatial point process, which involve the relation-
ship between numbers of events in pairs within the chosen study area. The K
function is a summary measure of second order effects, and is estimated for a
sequence of rings of distance h by:

R   Ih (dij )
K̂(h) =
n2 i=j wij

where R is the area of the study area polygon, n is the number of points,
Ih (dij ) is an indicator function which is 1 if dij < h and 0 otherwise, and wij
is a edge adjustment – the proportion of the circumference of a circle centered
on i and going through j that is within the study area. K̂(h) is often reported
as L̂(h), where: 2
K̂(h)
L̂(h) = −h
π
Observed values of K̂(h) for a given study area polygon boundary can be
compared with simulated values of the same measure for a given spatial point
process model. Most often the model chosen is that of complete spatial ran-
domness, which involved simulating n points within the study area polygon
for each simulated pattern following a homogeneous Poisson process. Results
are displayed by recording K̂(h) for each simulation, and plotting the largest
and smallest values for each h as a simulation envelope. If the observed K̂(h)
leaves the envelope, this may be taken to show that it is unlikely that – for
the chosen number of simulations – that the observed pattern could have been
generated by the process used in the simulation. When we wish to test whether
a pattern is clustered, it may be more natural to use a process model that
suits this hypothesis. The Poisson cluster process involves the inclusion of a
spatial clustering mechanism into the model, so that observed K̂(h) falling
within a Poisson cluster process simulation envelope show that the observed
pattern could have been generated by such a model.
The following code example run in the R statistical computing environ-
ment [23] will generate reproducible results, in this case the plot shown
in Fig. 4. The function being called to generate the simulation envelope is
pcp.sim(), contributed to the splancs package [24] by Giovanni Petris, and
using faster code changing the order of calls to the random number generator
contributed by Nicolas Picard, which can be turned off by setting argument
vectorise.loop=FALSE.
# Load the "splancs" package
library(splancs)
# Load the Cardiff juvenile offenders domiciles point
# data set and bounding polygon, assign the distance
# sequence and compute Khat
data(cardiff, package="splancs")
r <- seq(2, 30, by = 2)
K.hat <- khat(as.points(cardiff), cardiff$poly, r)
# Compute the fitted Poisson Clustering Process
pcp.fit <- pcp(as.points(cardiff), cardiff$poly, h0=30, n.int=30)
218 K. Krivoruchko and R. Bivand

m <- npts(as.points(cardiff))/(areapl(cardiff$poly)*pcp.fit$par[2])
# Set the random number generator seed and perform the simulation
#to find the simulation envelope bounds
RNGkind(kind="Mersenne-Twister", normal.kind="Inversion")
set.seed(123)
K.env <- Kenv.pcp(pcp.fit$par[2], m, pcp.fit$par[1], cardiff$poly,
nsim = 20, r = r, vectorise.loop=TRUE)
# Create a function to convert Khat values to Lhat
Lhat <- function(x, r) sqrt(x/pi) - r
# Apply the function to the simulation results
L.env <- lapply(K.env, Lhat, r)
# plot the observed Lhat values
limits <- range(unlist(L.env))
plot(r, Lhat(K.hat, r), ylim = limits,
main = "L function with simulation envelopes and average",
type = "l", xlab = "distance", ylab = "", lwd=3)
# Add the simulation average and envelope to the plot
lines(r, L.env$lower, lty = 5)
lines(r, L.env$upper, lty = 5)
lines(r, L.env$ave, lty = 6)
abline(h = 0)

L function with simulation envelopes and average

3
2
1
0
–1

5 10 15 20 25 30
distance

Fig. 4. Observed function for Cardiff juvenile offenders places of residence, with
Poisson cluster process simulation envelope using the “Mersenne-Twister” random
number generator
 On Monarchs and Their Clothing 219

Running the code calculates the K̂(h) function from the observed spa-
tial pattern for the chosen sequence of distances, using edge correction for a
bounding polygon, and plots its L̂(h) transformation. We test against a Pois-
son cluster process by simulating such a process within the bounding box,
here only 20 times, and plotting the maximum, mean, and minimum sim-
ulated L̂(h) values around the observed values (Fig. 4). It appears that the
observed data pattern could have been generated by a Poisson cluster process.
Having the code, the specified version of the splancs package and R, a
reviewer can re-investigate the impact on our conclusions of changing the
boundaries used for calculating edge effects, the number of simulations, the
distance sequence, the random number generator, and other parameters of
the model. Research should be documented not just for academic reasons,
and the provision of mechanisms for journaling methods used and thereby
securing the lineage of objects in documents within or derived from GIS is
necessary. Review and decision-making based on reproducible research using
well-documented closed (Geostatistical Analyst) or open (R) code software
is transparent and verifiable and does not require the participation of skilled
researchers. If the steps taken are documented and can be reproduced, the
results are available for checking in the future by the same or other users.
Should newer methods or fresh data become available, the documentation of
the lineage of results means that they can continue to be valuable for the
organizations that have invested in their collection and processing.

5 Concluding Remarks

In a perfect world, we should be able to combine the strengths of statistical

software and GIS, but we do not see it happening in practice except by simply
passing data sets back and forth between the two. Of course, there are open
source environments for both GIS and statistics, for instance, GRASS and R,
but it is very unlikely that this mixture will be used for decision-making by
very large community of commercial GIS software (millions of users), but only
by special interest and minority groups of academics, consultants, and others
needing the low-level flexibility this makes available.
Using the trivial example from the end of the previous section, we can
suggest that there need be no unhealthy competition between proprietary
software like ArcGIS and free software like R. Indeed, access is possible be-
tween ArcGIS and R under Windows using, for example, the R(D)COM Stat-
Connector. R is a good candidate for testing and prototyping new statistical
models for further implementation as commercial extensions for broader GIS
communities. This is because it provides support for documentation and re-
lease organization, guided by a well-regarded core team. More importantly,
commercial GIS software cannot be updated very often and the best way to
develop and test new models is to use more flexible environment, such as
growing R.
220 K. Krivoruchko and R. Bivand

R is not polished, for example lacking an integrated GUI, and it functions

more like a prototyping “kitchen”, where the ingredients can be tasted before
the meal is composed. The code above exemplifies this, as does access to
source code for the functions at least back to the operating system libraries,
if the need arises. This means that data can potentially be exchanged, in this
case a set of point locations, and one or more bounding polygons, for use in
R. The results could then be created as a document for display and inclusion
in further work, or to provoke changes in arguments passed to the underlying
R statistical compute engine.
Returning to our question about the way monarchs are clothed, we feel
that there are benefits to be drawn from raising questions about the ways
assumptions are handled in the statistical analysis of spatial data. In some
cases, users are neither able to make nor interested in the appropriate choice of
methods. In these cases, the developer should provide guidance, and document
the choices made and methods used in the resulting data objects. In other
cases, users are more like developers, working much more closely with the
software in writing scripts and macros, and in trying out new models. Here, the
accessibility of the input data objects to user-written functions is important,
and for some purposes, the linking of GIS software with external statistical
or modeling software may provide the level of customization some users need
for their research. In both scenarios – in fact on a continuum from black-box
to white-box – focus on the degree to which the assumptions of the applied
methods are met will let the user, or an auditor of the user’s work, find out
what has been done, and hopefully avoid unnecessary blunders.

Acknowledgements
This work was supported in part by EU Contract Number Q5RS-2000-30183.
The authors would like to thank Carol Gotway-Crawford for her helpful
discussions.
References

1. Anselin L (1995) Local indicators of spatial association – LISA. Geographical

Analysis 27:93–115
2. ArcNews (2003) http://www.esri.com/news/arcnews/summer03articles/
arcgis9.html
3. Chrisman N (2000) Building GIS without Foundations: Ontology from a So-
cial Practice Perspective, First International Conference on Geographic Infor-
mation Science, Savannah, Georgia, USA, October 28–31, 2000, http://www.
giscience.org/GIScience2000/invited/Chrisman.pdf
4. Cliff AD, Ord JK (1973) Spatial autocorrelation. Pion, London
5. Cressie N (1993) Aggregation in geostatistical problems. In Soares, A (ed) Geo-
statistics Troia 1992, vol 1, Kluwer Academic Publishers, Dordrecht, 25–36
6. Gandin LS (1959) The problem on optimal interpolation. Trudy GGO 99:67–75
(In Russian)
7. Gandin LS (1960) On optimal interpolation and extrapolation of meteorological
fields. Trudy GGO 114:75–89 (In Russian)
8. Gandin LS (1963) Objective Analysis of Meteorological Fields. Gidrometeoro-
logicheskoe Izdatel’stvo (GIMIZ), Leningrad (translated by Israel Program for
Scientific Translations, Jerusalem, 1965)
9. Getis A, Ord JK (1996) Local spatial statistics: an overview. In Longley, P,
Batty M (eds) Spatial analysis: modelling in a GIS environment Geoinformation
International, Cambridge, 261–277
10. Gotway CA, Rutherford BM (1994) Stochastic simulation for imaging spatial
uncertainty: Comparison and evaluation of available algorithms. In Armstrong
M, Dowd PA (eds.) Geostatistical Simulations, Kluwer Academic, Dordrecht,
1–16.
11. Gotway CA, Young LJ, (2002) Combining incompatible spatial data. Journal
of the American Statistical Association, 97:632–648
12. Johnston K, Ver Hoef J, Krivoruchko K, Lucas, N (2001) Using ArcGIS Geo-
statistical Analyst: GIS by ESRI, ESRI, Redlands CA
13. Journel AG, (1983) Nonparametric estimation of spatial distributions. Mathe-
matical Geology, 15:445–468
222 K. Krivoruchko and R. Bivand

14. Kolmogorov AN (1941) Interpolation and extrapolation of stationary ran-

dom sequences. Isvestiia Akademii Nauk SSSR, Seriia Matematicheskiia 5:3–14
(Translation, 1946, Memo RM–3090–PR, Rand Corp., Santa Monica, CA)
15. Krivoruchko K (2001) Using linear and non-linear kriging interpolators to pro-
duce probability maps. Available from ESRI online at http://www.esri.com/
software/arcgis/arcgisxtensions/geostatistical/research_papers.html
16. Krivoruchko K, Gotway C, Zhigimont A (2003) Statistical Tools for Re-
gional Data Analysis Using GIS. ACMGIS’03, November 7-8, 2003, New Or-
leans, Louisiana, USA. Available from ESRI online at http://www.esri.com/
software/arcgis/arcgisxtensions/geostatistical/research_papers.html
17. Krivoruchko K, Gribov A, and Ver Hoef J, 2000, A New Method for
Handling the Nugget Effect in Kriging. Available from ESRI online at
http://www.esri.com/software/arcgis/arcgisxtensions/geostatistical/
research_papers.html
18. Kuan J (2000) Open Source Software as Consumer Integration into Production,
unpublished working paper, University of California at Berkeley
19. Leisch F, Rossini A (2003) Reproducible statistical research. Chance 16(2):
46–50
20. Matheron G (1962) Traité de Géostatistique Appliquée, Tome I. Mémoires du
Bureau de Recherches Géologiques et Minières, No. 14, Editions Technip, Paris
21. Matheron G (1965) Les variables régionalisées et leur estmation. Une applica-
tion de la théorie des fonctions aléatoires aux Sciences de la Nature. Masson,
Paris
22. Moran PAP (1950) Notes on continuous stochastic phenomena. Biometrika
37:17–23
23. R Development Core Team (2003) R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria, http:
//www.R-project.org
24. Rowlingson B, Diggle P (1993) Splancs: spatial point pattern analysis code in
S-Plus. Computers and Geosciences, 19:627–655
25. Walder O, Stoyan D (1996) On variograms in point process statistics. Biomet-
rical Journal 38:895–905
26. Walter SD (1992) The analysis of regional patterns in health data I: dis-
tributional considerations. American Journal of Epidemiology, Supplement
132:S136–S143
Reassignment of the Farm Structure Statistical
Data Using GIS and Spatialisation of the
Results Based on Remotely Sensed Data

M. Sambrakos1 and T. Tsiligiridis2

1
InfoLab, Agricultural University of Athens, Athens, Greece
marios@aua.gr
2
InfoLab, Agricultural University of Athens, Athens, Greece
tsili@aua.gr

1 Introduction
From a rural land use perspective, an important development in Europe is that
agricultural activities are being combined with other activities such as envi-
ronmental care, maintaining the landscape, forestry, preserving recreational
and tourist areas, etc. As a result, there is a strong need for statistical data on
rural populations and particularly on landscapes and land use, which are by
their nature spatial in form. The management, the processing and the display
of such statistical data is therefore, to a large extend, a spatial process. In
this respect, GIS is considered necessary in the production of census maps,
for dealing with census logistics, for monitoring census activities, and for data
dissemination [2].
With the advent of GIS, a wide range of spatial analysis methods has
been developed for carrying out data transformations between different spa-
tial structures. These methods help to present the data in a more meaningful
and consistent manner and enable different data sets, based on different ge-
ographical units, to be brought together and overlaid. They also facilitate
the spatial analysis of statistical data required in the development and/or
calculation of more reliable indicators for the determination of the state and
quality of the environment, and the ability to measure the effect of the agri-
cultural economy, across regions and countries. Most policy makers concerned
with agri-environmental issues at the national level are confronted with frag-
mented information and it is accordingly difficult to use the information in a
way that effectively contributes to policy decision making.
A necessary step in the assessment of agricultural policies and of their
impact on the countryside and landscapes is the study of spatial units that
constitute the underlying structure of these areas. Most statistical data in the
European Union (EU), by means of the Farm Structure Survey (FSS) data,
224 M. Sambrakos and T. Tsiligiridis

is organized and presented on the basis of NUTS (Nomenclature des Unites

Territoriales Statistiques) system, to provide a single, uniform breakdown of
a country. Nevertheless, these units are geographical areas that may vary
substantially not only in their size and shape, but also over time.
This work presents an interface between the statistical and geographical
databases and provides a comparison between them by means of the FSS and
CORINE Land Cover (CLC). The geographical database can be used as a
means for the spatial disaggregation of FSS data into a more accurate geo-
graphical level and it is the first step towards a satisfactory spatial analysis.
FSS and CLC commonly describe land cover and land use. Definition of an
interface between their nomenclatures is a precondition for this spatial disag-
gregation. Notice that the comparison requires determining the aggregation
level of the classes for which the correspondence has already been set, as well
as, validation of the result by comparing the respective surface areas of the
related classes. After the reclassification of the above data, common classes
are created and presented on a map using an embedded GIS environment.
FSS data also require a comparison with other sources of information, as for
example topography, climatology of the different types of agricultural land, if
someone wants, for example, to evaluate the risks of erosion or of pollution of
watercourses by pesticides.
Knowing agricultural areas by type of crop within survey districts is in-
sufficient. It is necessary to localize this information more precisely. This will
allow the reallocation of data into suitable areas, such as drainage basins, while
limiting the loss of information. Notice that land use is difficult to define by
photo-interpretation as well as from a large distance (i.e. > 100 m). However,
this mode of observation is not preponderant concerning unused land, such
as, shrub land, forest, bare land, permanent grassland and water/wetland.
It concerns south Mediterranean areas where the land cover is very likely
to be shrub land or bare lands. On the ground, the high rate of shrub land
must reflect a difficulty for defining clearly the activity on such intermediate
biotope. These areas can also be considered as unused because of the climate
and the low density of population. As the forest unused areas, they might
be used as rough grazing areas. From a general point of view, unused areas
are much more located on homogenous land cover types (shrub land, land
without tree, permanent grassland without tree, forest, etc.). Unused areas
occupy a large part of Greece where the proportion rises to almost 40% of the
country.
To test the interface and provide the appropriate links between certain
classes of the two databases the region of the island of Crete has been chosen.
The statistical data used has been provided by the Basic FSS of 1999/2000
(Census of Agricultural for Livestock Breeding or simply Agricultural Census).
However, to achieve compatibility between census and photo-interpretation
data a recently developed, improved version of the CLC geographical database
has been used. The new geo-statistical database, which takes into account the
FSS nomenclature and definitions, provides a much better acquisition period
 Reassignment of the Farm Structure Statistical Data Using GIS 225

(Landsat-TM 1998–1999) which is the same as the census reference period

(1998–1999).
The structure of the paper is as follows: The next section describes briefly
the recently introduced geo-statistical database. It provides the main charac-
teristics of the classification scheme used and it resolves the problems encoun-
tered when linking the data sources (i.e., the FSS and the CLC databases).
Then, the section dealing with the software tool follows, which provides a
sufficient description of the development. It should be noted that the devel-
oped tool is quite general; however, for validation purposes a case study has
been conducted. This section is followed by the section of Data Analysis in
which the results from the comparison of the related nomenclatures are pre-
sented. Finally, the last section presents the conclusions and discusses further
developments of this work.

2 Material and Methods

2.1 The Hellenic Geo-Statistical Database

In the light of recent developments concerning land use statistics and in

order to produce more objective information on this sector an up-to-date
methodology is adopted using GIS techniques. The National Statistical Ser-
vice of Greece (NSSG) is testing a methodology to produce a detailed land
cover map for the Hellenic territory. The data sources of the land cover
map include aerial ortho-photographs, satellite images as well as agricul-
tural census (FSS), and it is based on the same with the CLC minimum
mapping unit. The new geo-statistical database aims to cover the needs of
land use/cover statistics as far as the distribution of the Hellenic total area
into basic categories of land use is concerned. The new database is properly
generalized as reference data and harmonized with the FSS nomenclature,
by means of characteristics and definitions. As a result, the distribution of
the main land uses in Hellas has been organized into 16 classes. For the
year 1990, the CLC1990 database is used and therefore a correspondence
between the two nomenclatures has been set. Interesting to note that us-
ing the 44 CLC classes one may capture the total land cover diversity, i.e.
that linked to urban and natural areas. Nevertheless, our interest is to shed
light on the relationship between agriculture and the landscape in rural areas
and therefore the pre-mentioned reduction in the number of CLC classes is
obvious.
Spatial analysis of the information to be recorded is realized by determin-
ing the area of the minimum recorded surface, which is taken according to the
proposed nomenclature of 16 classes, the methodology of use/cover definition,
the requirements of 1:100.000 scale and the user needs. The method, by which
the theme information is drawn up, is a comparative photo-interpretation of
226 M. Sambrakos and T. Tsiligiridis

new satellite data collected in 1998–1999 in relation to those used for the cre-
ation of the Hellenic geo-statistical database. The digital photo-interpretation
of the new satellite data is made using image processing software and other
data such as those from land recordings. The recording, planning and the
use of the data from the field work also define the reliability of the specific
photo-interpretation.
The new geographical database for the country’s area has numerous ad-
vantages, the most important of which are the following:
• It provides a land use/cover map covering all Hellenic territory using 16
classes.
• It takes into account the FSS nomenclature and definitions.
• It enables comparability between different time periods, using the same
source of information, namely census or photo-interpretation.
• It enables comparability between the two sources of information, namely
census versus photo-interpretation. In the case of Hellenic Republic, the
acquisition period of the data is spread over 2 years for both, the LTM
1998–1999 and the FSS 1999/2000 (reference year the 1998–1999 crop
year).
• It enables the integration of the chrono-geographical co-ordinates of the
satellite images sources of CLC. This will help in the identification of
districts for which image interpretation is one year apart (minus or plus)
from the census year (1990 or 2000, respectively). In addition, using the
intermediate FSS data that correspond closely to the date of the satellite
image, it will be possible to mitigate the effect of time.
As it appears, the new geo-statistical database is in principle more accurate
than CLC. It can be used to calibrate diversity measurements computed from
CLC, although there are some problems because the reference dates may not
coincide. The methodology has been tested in the region of Crete (NUTS II).
The Crete island is about 8, 267, 45 km2 , it is located in the most south part of
Hellenic Republic and it is divided into four administrative areas (NUTS III).

Differences in Data Models

To describe the methodology adopted in the problem we are studying, one has
to take into account the non-matching areal units and the problem Modifiable
Area Unit (MAUP) [5]. Note that, the temporal incompatibilities problem and
the procedure of matching the data points by non-matching due to collection
cycles is not considered here.
Starting with the non-matching areal unit problem, as this appears in the
pilot case, a new object, called interoperable geo-object is introduced. This
object includes all the required procedures in order to solve the following two
problems.
Reassignment of the Farm Structure Statistical Data Using GIS 227

Fig. 1. The conceptual model

228 M. Sambrakos and T. Tsiligiridis

• The different boundaries definition of the administrative units that have

been used during the collection of the FSS data (1991, 2000) in the pilot
regions (NUTS II and NUTS III).
• The geodetic datum used in order to represent jointly the statistical and
the ancillary geographical data on a map.

The first problem has been solved with the appropriate transformation
between different spatial structures. This transformation determines the pro-
cess of the aggregation and the disaggregation within nested, nonnested and
neighboring polygons. To overlay the data the conceptual model of Fig. 1
has been designed. This model contains and maintains all the polygons and
the related geometric data (lines, nodes etc), representing the areal units. To
link the descriptive with the spatial information, the data of the geographical
area has been divided into smaller parts in order to determine the field that
identifies the specific entity (PolyKey), which has been used as a reference
key to the GIS. Also, a set of spatial queries has been developed to carry out
the above transformation. The second problem has been solved through the
development of a common geodetic datum, which represents jointly the sta-
tistical and the ancillary geographical data on a map. Finally, an automated
procedure has been developed to convert the data from the original to the
target geodetic datum.
The MAUP problem has been faced by increasing the spatial detail, using
ancillary geographical data [4] such as contour lines, lines representing rivers,
or polygons representing lakes etc. This allows the synthesis of geographical
data along with the statistical data. Further, it allows the combination of
different scenarios to be considered in order to simulate the plotting of the
statistical data on a map. For validation and/or prediction purposes, the re-
sults have been compared visually with other spatial quantitative information
or sampling data presented on thematic maps.
To automate both the transformation between different definitions of ad-
ministrative units and to achieve the connection between a file containing
quantitative data (usually statistical) along with GIS data (ancillary and sta-
tistical), two-object classes have been developed namely, the class for data
manipulation and the class for GIS manipulation.

3 Software Development – A Case Study

As it has been pointed out, the linkage of the two nomenclatures, by means of
the structure survey and the geographical databases, require the development
of a software tool able to display maps and descriptive data in a tabular
form. This has been achieved by linking the geographical information with
the multi-dimensional tabular information of FSS. Thus, the user becomes
part of the GIS without the necessity of having specific skills and intimate
knowledge of the data used.
 Reassignment of the Farm Structure Statistical Data Using GIS 229

The application consists of four items, namely, the relational database, the
class of objects for data manipulation, the class of objects for GIS manipu-
lation and the main body of the application software containing the above
items along with the functions required by the end user.
To begin with, a step-by-step analysis of the software design is required.
The appropriate design steps are described below:
1. The ancillary geographical features, such as contour lines, roads, cities,
lakes and rivers are added on the geographical layer of the area of interest.
This will help to localize the geographical data.
2. From the FSS database only themes associated with agricultural products
have been selected. Note that the use of the geo-object offers the capability
to work at different levels of administrative units. However, in the pilot
case, the FSS data have been selected at prefecture level (NUTS III), in
thousands of hectares, as they are reported in the 2000 census.
3. We develop the entity relationship model as well as the relational database
of the software tool, based on the data provided by the FSS database and
geo-database.
4. The geographical data have been stored in some database tables of the
software tool, using some especially developed functions. Further, the
OLEServer method of the QuantitativeInput object has been used with
the appropriate DLLs, which have been provided by the FSS, in order to
transfer the FSS data into the database.
5. We define the appropriate functions and queries, and we developed object
classes in order to achieve uniformity at both the user and the developer
levels.
6. We developed an application in which the RDBMS, the GIS and the pre-
mentioned object class have been used. The basic capabilities offered by
this application are the following:
• Compose (aggregate) a new FSS theme by selecting one or more
classes, and vice versa.
• Decompose (disaggregate) an existing FSS theme to one or more
classes, and vice versa.
• Correspond (relate) the new FSS themes to classes.
• Classify (sort) the results by date, county (region), or by class.
• Observe the results plotted on a map and classify these by some ge-
ographical characteristics (e.g. allocation of the selected growth by
elevation).

4 The Relational Data Base

The GIS tool used for the geo-database construction is the ESRI ArcInfo
software. This tool stores a set of tables in DBF format, containing both
the spatial and descriptive information about map’s features, which are log-
ically organized into themes of information. Each theme consists of topolog-
ically linked polygons along with the associated descriptive data. Generally,
230 M. Sambrakos and T. Tsiligiridis

X-Base formats, cannot easily aggregate, desegregate, isolate, and combine

geographical data with other sources. Furthermore, due to severe limitations
associated with the temporal component of data in the GIS raster databases,
a comparison between geographical data obtained in the past is very difficult
to be achieved in practice [3].
To support the exchange of heterogeneous data into an integrated database
environment a conceptual model is required [6]. In the design of such a model
one has to take into consideration loading and refreshing the descriptive geo-
graphical data for each attribute of the GIS, at any time it will be required,
and then linking them with the information derived from other sources, such
as the FSS data. The conceptual model of the proposed database is described
in Fig. 1.

5 The Classes of Objects for Data and Geo-Database

Manipulation

The class of objects for data manipulation is based on a PowerBuilder object

called DataWindow [1]. This class provides a simple way of retrieving, display-
ing and updating data from a specified data source. Although the data source
is usually a database, it can also be a text file or other data structure. The
class of DataWindow object, named PBDWO, inherits the basic functionality
and encapsulates the ability to dynamically, at run time, bind and combine
data from different sources.
The geo-database contains only agricultural classes, which can be mapped
on one or more regions. To compare them with other ancillary geographical
data, for example road, lakes, contour lines, etc., and to process the geograph-
ical data, the development of a class of objects that inherit their properties,
methods and functions is required. This class of objects encapsulates more
functions and customized events to finally communicate with the database,
and vice versa. This class is called interoperable geo-object. For the develop-
ment of this class the ESRI MapObjects is used.

6 The Application Software

As it has been pointed out, this application is computer-based software able to

display maps and descriptive data in a tabular form. This has been achieved
using geographical information from CLC database linked with tabular infor-
mation of the multi-dimensional tables of the FSS. The user becomes part of
the GIS without the necessity of specific skills and intimate knowledge of the
data used.
To test the application software a preliminary study, using the 1991 Basic
FSS data of the island of Crete has been prepared. The island of Crete is a
region (NUTS II level) and consists of four districts (NUTS III level); Chania,
 Reassignment of the Farm Structure Statistical Data Using GIS 231

Rethimno, Iraklio and Lasithi. The geographical presentation is based on the

16 class geo-statistical database (1991 and 2000) and is constructed using the
Hellenic Geodetic Reference System 1987 (HGRS 87). Any additional geo-
data used such as roads, lakes, contour lines, etc. are constructed using World
Geodetic System 1984 (WGS 84). To solve the problem of geodetic datum
transformation without making changes in the application source code a map
layer object is added. This object has a property to specify the path of the
ASCII file, which contains the appropriate transformation parameters. Fur-
thermore, the basic geographic layer is constructed using detailed geographical
data, such as coastlines, contour lines, roads, airports etc.
The main application window includes the standard GUI controls (menu
and buttons) as well as the PB-DWO and the interoperable geo-object. The
PB-DWO contains the rows of the entity cordesc matching the selected area.
The interoperable geo-object displays the corresponding polygons of the above
entity. Complementary, details of the method followed may be found in Sam-
brakos et al. [7]. As it may be seen it becomes an easy task for the user to
incorporate the geo-statistical and/or the FSS data of any year (and thus for
the 2000 year) into the application.

7 Data Analysis
Although the new geo-statistical nomenclature has been harmonized with the
FSS nomenclature, there are still some problems related to the two different
methodologies. The analysis of the above problems has been carried out by a
comparison between the respective areas of the related classes. The available
data from the 2000 FSS is based at the Municipality/Commune level (NUTS
IV), whereas the data drawn from the new geostatistical nomenclature is at
the district level (NUTS III). The data of two databases have been compared
in a pilot study of four Hellenic regions at a district (NUTS III) and pre-
fecture (NUTS II) level. The comparison shows large difference between in
the agricultural areas. Generally, the examined agricultural areas in the geo-
statistical nomenclature are greater than the corresponding agricultural areas
in the 2000 FSS. The differences are because of the difficulties in correlating
the pastures areas between the two databases, whereas the differences in the
arable areas and the areas under permanent crops are related to the different
methodologies.
The results found so far are presented in Table 1. Table 1(a) presents the
differences (%) in arable areas, areas under permanent crops, and cultivated
areas, as they were recorded in the districts (NUTS III) of the examined
regions, between the two nomenclatures. Positive sign is in favor of the geo-
statistical nomenclature, whereas negative sign is in favor of the FSS nomen-
clature. Note that the actual differences in the above classes are not as high
as they are in the remaining classes, namely agricultural areas (Table 1(b)),
pastures and meadows (Table 1(c)), and heterogeneous areas (Table 1(d)). To
facilitate the comparison for the last cases the actual values are presented.
232 M. Sambrakos and T. Tsiligiridis

Table 1. Results showing the differences between classes, as they have been recorded
in the 2000 FSS and the geo-statistical databases
Table 1(a) districts (%) differences (2000 FSS – GeoStat)
(NUTS II) (NUTS III) arable areas permanent crops cultivated areas
crete IRAKLIO −71 4 −4
LASITHI 54 47 48
RETHIMNO −91 −7 −24
CHANIA −72 4 −4
total −66 6 −3
Table 1(b) Districts agricultural areas (ha)
(NUTS II) (NUTS III) 2000 FSS GeoStat Differences
crete IRAKLIO 221,982 139,733 82,249
LASITHI 127,252 37,864 89,388
RETHIMNO 115,842 101,182 14,660
CHANIA 116,472 109,191 7,281
total 581,548 387,970 193,578
Table 1(c) Districts pastures and meadows (ha)
(NUTS II) (NUTS III) 2000 FSS GeoStat Difference
crete IRAKLIO 36,412 69,070 32,658
LASITHI 16,817 61,631 44,814
RETHIMNO 62,470 53,241 −9,229
CHANIA 63,410 40,167 −23,243
total 179,109 224,109 45,000
Table 1(d) Districts heterogeneous areas (ha)
(NUTS II) (NUTS III) 2000 FSS GeoStat Differences
crete IRAKLIO 143 54,339 54,196
LASITHI 12 34,433 34,422
RETHIMNO 159 33,372 33,213
CHANIA 14 32,420 32,406
total 328 154,564 154,237

It has been observed that the above differences in the regions (NUTS II)
are generally smaller from the corresponding inter-regional ones (district level;
NUTS III). This is due to the fact that the mapping unit of 25 ha in the new
CLC is not able to identify parcels of smaller size. This is the case of Greece, in
which the average holding size is around 4,5 ha and the average parcel size is
around 0,7 ha. An additional reason is that in FSS all the holdings are recorded
at the place of residence of the holder (natural person) or headquarter (legal
person) of the holding.

8 Conclusions

The work presented so far is a pilot study merging with the use of a soft-
ware tool the statistical data, available at the administrative level, with the
 Reassignment of the Farm Structure Statistical Data Using GIS 233

geo-referenced land cover in order to identify and explain the most significant
differences encountered between the aggregates of agricultural land cover
classes. This has been achieved thanks to the creation of a new geo-statistical
database, which is based on both, the FSS and the CLC nomenclature.
The above geo-statistical database seems to provide a good mapping base
for Greece, which could be improved further by using suitable satellite images
that are able to produce scaled maps of at least 1:50,000. Note that the im-
posed minimum mapping unit of 25 ha results in an overall underestimation
of the diversity of landscapes something, which is particularly important in
the case of Greece for which the average size of the holdings is 4,5 ha. Ad-
ditional sources may be used providing detailed complementary information,
such as aerial ortho-photographs, the cadastral map of Greece, IACS (In-
tegrated Administrative Control System), MARS (Monitor Agriculture with
Remote Sensing), NATURA2000 database, or other ongoing analysis of the
European landscape.
The methodology of using the interoperable geo-object in conjunction with
RDBMS settings and the OOP logic means that many of the objects can
be used in similar GIS applications with a little effort of maintenance. The
application developed is an easy-to-use tool, ideal for comparison of descriptive
census results and interpreted geo-data, as well as, to conclude about the
correctness of these data. If the expert combines the ability of simultaneous
comparison and appearance of results of different years, the conclusions will
be more reasonable.
Future research is three fold. Firstly, it is to continue improving the idea
of interoperable geo-object by adding methods and properties for uncertainty
manipulation and to investigate requirements of GIS in a fuzzy object data
model. Our final objective is to provide the geo-object with the ability to
generate and visualize transitions from one state to another, using the rules
of an expert spatiotemporal system. Related work on this aspect is given in
[8]. Secondly, this study may be considered as a first step in the direction
of presenting geo-reference statistical and/or agricultural and environmental
data. As soon as this initiation will be completed, it will become possible to
redistribute quantitative data other than land use from the FSS by defining
some distribution rules using co-variables. Finally, this research will facilitate
the spatial analysis of statistical data required in the development and/or
calculation of more reliable indicators.
References

1. Candak R., Chandak P. (1999) Advanced in PowerBuilder 7 Techniques. Wiley,

New York, pp. 202–253.
2. Deichmann U. (1997) Geographical information systems in the census process
Technology options, costs and benefits. Workshop on Strategies for the 2000
Round of Population and Housing Censuses in the ESCWA Region, Cairo,
December 6–10.
3. Dragicevic V., Marceau D. (2000) An application of fuzzy logic reasoning for
GIS temporal modeling of dynamic processes. Fuzzy sets and Systems Vol 113:
69–80, Elsevier.
4. Flowerdew, R., Green M. (2001) Areal Interpolation and Types of Data. In
Stewart Fortheringham and Peter Rogerson (eds) Spatial Analysis and GIS.
London, Bristol: Taylor & Francis Ltd.
5. Openshaw S. (1984) The Modifiable Areal Unit Problem In Norwich: Geo
Books, Norwich.
6. Parent C., Spaccapietra S., Zimanyi E. (1999) Spatio-temporal conceptual mod-
els: Data structures + space + time. Proceedings of the 7th ACM Symposium
on Advances in Geographic Information Systems, GIS’99, Kansas City, USA.
7. Sambrakos M., Fillis I., Tsiligiridis T. (2001) Integration of Spatial Descriptive
Statistical Data and Geographic Information. Proceedings of 8th Pannellenic
Conference on Informatics, paper 91C, Nicosia, Cyprus, November 2001.
8. Sambrakos M., Yialouris C., Tsiligiridis T. (2002) Fuzzy Interoperable Geo-
graphical Object (FIGO). An approach of enhancing spatial objects with fuzzy
behavior. Proceedings of the 1st Conference of HAICTA (Hellenic Association
of Information and Communication Technology in Agriculture, Food and En-
vironment), Session 3B, pp 209–219, Athens University of Agriculture, Athens,
Greece, June 6–7, 2002.
9. Sambrakos M., Tsiligiridis T. (2003) A Comparative Landscape Pattern
Analysis Using Remotely-Sensed and Statistical Data to Evaluate Regional
Diversity. Proceedings of 1st International Congress of ITAFE 03 (Informa-
tion Technology in Agriculture, Food and Environment), O-GIS-33-0, Edge
University, Izmir, Turkey, October 7–10, 2003.
10. Yuan Y., Smith R., Limp W. (1997) Remodeling Census Population with Spa-
tial Information from Landsat TM Imagery. Computers, Environment and Ur-
ban Systems, 21:245–258.
Epidemiological Information Systems

V. Gómez-Rubio1 , J. Ferrándiz-Ferragud2 , and A. López-Quílez3

1
Dpto. Matemáticas, Universidad de Castilla-La Mancha, Albacete, Spain
Virgilio.Gomez@uclm.es
2
Dpto. Estadística e Investigación Operativa, Universitat de València, València,
Spain
Juan.Ferrandiz@uv.es
3
Dpto. Estadística e Investigación Operativa, Universitat de València, València,
Spain
Antonio.Lopez@uv.es

1 Introduction

1.1 GIS and Public Health

Public Health Authorities can take advantage of GIS in order to perform

studies in disease surveillance tasks, to know how diseases spread and to lo-
cate outbreaks. Facts involved in these studies come from a wide range of
sources: hospital registers, clinicians, environmental organisations, etc., so it
is important to collect and store all them for it being easier to access and
analyse.
The increasing interest in GIS in Public Health has also been reflected
in the literature. In last years, a number of books have appeared devoted
to GIS and Public Health affairs [4, 9, 15], statistical methods for Spatial
Epidemiology [19, 23, 24], and several journals, such as Statistics in Medicine
[11, 30] and Journal of Royal Statistical Society [32] have devoted special issues
to related subjects.

1.2 A GIS for Spatial Epidemiology

Since the range of applications of GIS in Public Health is nearly unlimited (like
in many other fields), we will focus on Spatial Epidemiology, which refers to
different topics about the spatial spread of diseases: disease mapping, detection
of clusters of disease, ecological analysis, etiology, etc.
In this paper we describe how a Geographic Information System for Spatial
Epidemiology can be developed and we briefly discuss the main points to which
attention should be paid.
236 V. Gómez-Rubio et al.

In Sect. 2 we describe the main issues concerning data. Section 3 covers

statistical methodology. Section 4 comprehends how the whole GIS can be de-
veloped, integrating and analysing data. What we have developed is explained
in Sect. 5.

2 Managing Data
Data needed are determined by the conclusions we want to draw from the
studies. Usually, the main concern is to explore the spatial distribution of
a group of diseases (mortality and/or morbidity), which is accomplished by
means of Disease Mapping (see Sect. 3.2). The second step is often the detec-
tion of those regions where there exists a higher risk of suffering from these
diseases, known as Risk Assessment (as explained in Sect. 3.3).
When a high risk has been detected an explanation is usually required.
Sometimes it is possible to look for relationships between risk and a number
of covariates. This is done via Ecological Analysis, as described in Sect. 3.4.
Studies are always restricted to a period of time and to a particular area.
Data available are aggregated on the basis of units used to measure space and
time. For this, year is often used, while there is no clear preference for the
spatial units, since it usually depends on administrative boundaries.
The level of aggregation is restricted due to confidentiality issues. Data
available in the studies are usually in a form that prevents from identifying
single individuals. This means that short periods of time or quite small areas
can’t be used.
Data quality uses to be quite good for mortality, but it doesn’t happen
to morbidity, excepting for a few set of ‘important’ diseases for which special
registers are drawn up (like, for example, children malformations, AIDS or
cancer).
Measuring exposure to a risk factor is always difficult and it is often im-
possible to take exact measurements for every person at risk. Residence is
often used as a proxy to the place of exposition, although it can be misleading
since people is also expected to spend quite time at work, for example.
Nevertheless, government agencies and other data providers usually link
their information to the appropriate administrative areas, be it quarters, elec-
toral districts, etc. When this doesn’t happen it is not difficult to refer the
actual location to the standard administrative areas [27].
Finally, it is important to update data on a regular basis, so that recent
problems can be investigated.

2.1 Population Data

Population at risk is a key issue, since it helps to measure the incidence or

prevalence of a disease. When calculating risk rates, a ratio between the num-
ber of affected people and population at risk is calculated (see Sect. 3.1).
 Epidemiological Information Systems 237

Inaccuracies in the denominator can lead to wrong estimations, especially in

low populated areas.
Measurement of population at risk is even harder than morbidity or mor-
tality since it may not be clear how to define at risk. The most common
approach is to consider residence as the main place where people stay, hence
where they are exposed. But it can be inaccurate since work is an important
place where people can be exposed. This is especially important when the
disease under study may be related to the kind of employment.
Migration is another source of bias. Information about residence is col-
lected on a regular basis every ten years by census offices. If population of any
year in between is needed an estimation is carried out [29]. Taking migration
into account is difficult since there are no clear clues about how to approach
it because the high number of economic, social and political factors that may
be influencing.

2.2 Health Data

Health data are collected by a number of health institutions, such as hospi-

tals, and gathered by the main Local Health Authority. Usually, mortality is
recorded with a high level of accuracy, while morbidity is more difficult to
obtain.
For those diseases of high interest, such as AIDS or Cancer, separated
records are maintained by a specialised institution (for example, Cancer Reg-
istries) and provided information is more accurate.

2.3 Geographic Data

The previously described information is always referred to an administrative

level and, if not, it is not difficult to make a conversion so they are. Since
administrative levels are usually structured into layers (from smaller to larger
areas), knowing how to aggregate (or disaggregate) data can help to conduct
studies at different levels.
Administrative boundaries can be used to plot maps that summarise re-
sults from analysis. Cloropleth maps are the most common representation, in
which the variable under study is categorised and every administrative region
is filled with a colour according to these categories.

2.4 Environmental Data

Sometimes, disease outbreaks are related to environmental conditions. Several

issues may increase the risk of suffering from certain diseases. For example,
exposure to pollutants of a nearby petrochemical complex may increase the
incidence of leukaemia. For this case, distance to the putative source can be
used as an approximation to the level of exposure [27].
238 V. Gómez-Rubio et al.

This kind of data usually comes from a wide range of sources and variables
of interest depend on the kind of study to be carried out. Some of them may
be available from government agencies but others must be collected ad hoc
for the studies.
Maps can be created to represent the spatial distribution of these vari-
ables (see Fig. 1). When a continuous representation is required, geostatistical
methods [8] can be used to provide estimations at those points where a sample
hasn’t been taken.

Fig. 1. Maps produced by the RIF (described in Sect. 5) in an investigation of

the relationship between hardness of drinking water and cerebrovascular mortality
in the province of Valencia (Spain). On the left, we have the spatial distribution of
calcium in drinking water, while on the right, SMRs (top) and SMRs standardised
also by calcium levels (bottom) are shown

2.5 Socio-Economic Data

Social and economical factors may also influence results. It has been proved
that way of life, diet, deprivation, etc. may have a strong influence over disease
risk. A deprivation index is created taking into account different variables and
it is often used in standardisation to filter socio-economic inequalities [12].
 Epidemiological Information Systems 239

2.6 The Right Data for the Right Analysis

Once all these data have been collected, the next step is to make them available
from a common source. This compilation implies a debugging of the data in
order to detect and correct (or remove) wrong entries. This point is crucial
since data quality is the key to what kind of studies can be carried. Imputation
methods must be used to fill in missing data.
The same database should be used to store all the information available
in order to integrate different types of data and a link among them must be
established. This is done by means of the spatial location of the data, so it
is necessary to unify how it is specified: coordinates, municipalities, counties,
etc. This is a challenging task since there can be a lot of different spatial units.
When data are measured at individual locations (i.e., different measures
at single, maybe different, points) a mechanism to associate these points to
administrative regions may be needed [27]. Point data can be analysed, but
following a different approach from the one employed for area based data.

3 Statistical Analysis

3.1 General Analysis

Although standard and simpler methods can be used to carry out a number
of analysis, we should take advantage of the spatial nature of our data and
the special characteristics of the problem we are facing [8, 19].
As a rule, cases (mortality and/or morbidity) and population are stratified
according to sex, age group and, if available, a measure of deprivation or
poverty. These factors have been proved to be important in the analysis, and
this approach helps to reduce bias in the estimations and to remove the effect
of this factors [22].
We will also find different measures for administrative regions and periods
of time, so that we have variables Oijk and Pijk , for the observed number
of cases and population, respectively, in region i, age-sex-deprivation strata
j and period of time k. When working with fully spatial models, the third
component will be missed and data will be aggregated based on the third
component.
When calculating incidence rates, a quotient between affected people (nu-
merator) and population at risk (denominator) is made [22]. When a compar-
ison between different studies at different locations is required, another region
is used as a reference to calculate standard rates for each age-sex-deprivation
stratum.
If we call rjk the reference (or comparison) rate for stratum j and period
k and supposing no region – age-sex-deprivation interaction [34], we can get
an estimation of the expected number of cases in region i and period of time
k by Eijk = rjk · Pijk . This process is called standardisation [22].
240 V. Gómez-Rubio et al.

Most models suppose Oijk drawn from a Poisson distribution whose mean
is θik Eijk , where θik is called the Relative Risk [34]. Its Maximum Likelihood
 
Estimator is θ1ik = j Oijk / j Eijk , called the Standardised Mortality Ra-
tio (S.M.R.). Usually, a confidence interval is calculated to test significant
departure from value 1, which marks the standard risk.
Another layer can be added to the model in order to explain how θik is
distributed (see Sect. 3.5). Usually, a Generalised Linear Model is constructed
[26], although Bayesian Hierarchical Models with spatial structure have been
used to smooth relative risks (as shown in Sect. 3.5), so that neighbouring
regions are also taken into account in the local estimations.

3.2 Disease Mapping

Disease Mapping is used to display the spatial distribution of a disease. Usu-

ally, a statistic associated to the disease is calculated and later split into dif-
ferent categories. Then, areas belonging to the same category are filled with
the same colour (see Fig. 1).
Although these maps can be really helpful as an early summary, they can
be misleading, since higher areas will attract more attention and low populated
areas tend to show more extreme values [6]. Some authors have also addressed
other problems related to the influence of scale, map projection and colours
used [9].
Common statistics to be mapped include p-values [5], S.M.R.s [34] and
residuals from a fitted regression model [8] (usually used to see whether any
spatial structure remains unexplained by covariates).

3.3 Clusters of Disease

By taking a look at a disease map, groups of areas with higher risk (clusters)
can be detected. Due to the problems mentioned in the previous subsection,
this method can be misleading and not accurate.
Since the detection of clusters of disease is one of the priorities for epi-
demiologists, a number of methods have been developed for this purpose, and
a few reviews have been made by authors in the last years [19, 25, 33].
In the investigation of clusters of disease we can mainly distinguish two
types: search for clusters in the study area [33] or investigating a known pu-
tative pollution source [10]. Clearly, statistical assumptions are quite different
depending on which one we are working on.

3.4 Ecological Regression Analysis

The relationship between disease and risk can be investigated through Ecolog-
ical Regression. Generalised Linear Models [26] are often used for this purpose,
although Generalised Additive Models has also been used [20]. An example of
 Epidemiological Information Systems 241

Ecological Regression using GLMs is the study about the relationship between
hardness of drinking water and cerebrovascular mortality [14].
When performing an Ecological Regression, it is important to pay atten-
tion to how risk exposure has been taken. If different levels of aggregation
are used it may happen that some measurements have been made at a broad
level, i.e., the same risk is associated to a wide range of population and a bias
may be introduced in the analysis. This problem is know as the Ecological
Fallacy [28].

3.5 Bayesian Hierarchical Models

The Bayesian paradigm has been successfully applied to all fields of Spa-
tial Epidemiology. It is based on Bayes’ Theorem, so that we calculate the
posterior distribution (after observing the data) as the product of likelihood
and priors of the random variables. When the posterior distributions can’t be
worked out, as it happens most times, MCMC techniques [16] are employed
to simulate them.
Probably, the first Bayesian Hierarchical Model applied to Spatial Epi-
demiology was the one proposed by Clayton and Kaldor [6], which proposes
a prior Gamma for all the relative risks and produces a globally smoothed
estimate of the relative risk θ1i = (Oi + ν)/(Ei + α), which is a compromise
between the S.M.R. of region i and the prior mean (ν/α), reducing extreme
values.
Other models, such as the one proposed by Besag et al. [2], also produce
smoothed estimates of relative risks. The logarithm of the relative risk is
expressed as the sum of the effect of neighbouring areas plus the effect of
the local area, which can be a linear function of covariates [13]. Then, the
estimations of the relative risks obtained have been smoothed by taking into
account the effect of neighbours.
Smoothed estimates of relative risks obtained in these models can be used
to produce cloropleth maps. Comparing these maps to those made from SMRs
will show how the effect of extreme values in low population areas is reduced.

4 Integrating Data and Statistics

4.1 Previous Consideration

First of all, it is necessary to know current and future needs before designing
the whole system. Perhaps, we only must care about the spatial spread of
diseases, without paying attention to possible causes. Or, on the contrary,
the main concern is to investigate sources of pollution to see how they affect
health. This is important because data and statistical methods will depend
on these needs.
242 V. Gómez-Rubio et al.

For basic models with few data, they could be imported into any statistical
software available, while for huge amounts of data, statistical methods will
require better integration with the database in order to make analysis possible
in a short time (or even possible!).
Time required for the analysis may also be important. Exploratory tools
may be used as a first, rapid look into a problem to decide whether a further
investigation should be carried out. If it is decided so, more time consuming
methods (such as, for example, Bayesian methods computed via MCMC) may
be required for a more accurate analysis.

4.2 Managing Data

Since data are collected from many different sources, it is crucial to integrate
them into a single database. Before doing this, data quality must be assured
and it should be checked for inadequacies, wrong and missing values, etc. Data
can be linked by referring to their spatial location, as explained in Sect. 2.
Health data can be stored as single events, but usually a minimal aggre-
gation is defined for space and time, and data will be aggregated on the fly
when performing a study. Depending on the amount of data available it may
be useful to create separated tables of aggregated data to speed up future
investigations.
It is important to provide a way to move up and down the different ad-
ministrative layers in order to be able to carry out studies at different levels.
This can be done by providing a conversion table from one level into another.

4.3 Computing Statistics

Basic statistics can be computed with any statistical software available. Al-
though some GIS software are incorporating statistical methods for spatial
statistics, they are mostly focused on geostatistics and it is difficult to find
methods used in Spatial Epidemiology.
Accessing data directly to the database can solve this problem but, as
commented before, when the amount of data is big, it can be very slow and
it should be better to implement statistical methods inside the database.
Unfortunately, as stated in [9], there is a lack of software in the field of
Spatial Epidemiology, and what can be found so far, are isolated programs to
compute a few methods.

4.4 Interfacing Data and Statistics

Although GIS, databases and statistical software can be used separately when
performing studies, it can be more helpful (and less time consuming) to de-
velop an unified tool.
 Epidemiological Information Systems 243

An interface can be created to define the investigation to be carried out,

query the database and perform the statistical computations. Most GIS pro-
vide a programming language which enables the development of internal rou-
tines. Sometimes, it is possible to link the GIS to statistical software and
the database, so that all results are returned to the GIS and, for example,
cloropleth maps can be developed.
Statistical outcomes can also be displayed in reports, and results are often
grouped by region and age-sex-deprivation stratum. This is useful when a
comparison among them is required, although a Comparison Test could be
used for this purpose [12].

5 Software Developed
5.1 Rapid Inquiry Facility

The Rapid Inquiry Facility (RIF) was initially developed at the Small Area
Health Statistics Unit [1], but it was rewritten within the framework of EU-
ROHEIS Project [7], funded by the European Commission. It was intended
to be a Health Information System for Disease Mapping and Risk Assessment
around putative pollution sources.
It is based on ArcView 3.2, Oracle Database 8i, and Oracle Forms and
Reports. A graphical user interface developed in Avenue (ArcView’s internal
programming language) allows the selection of study and comparison regions
together with the period of time and diseases to investigate. Two types of
studies can be done: Disease Mapping and Risk Assessment around putative
pollution sources.
The RIF was developed with a standard structure (see Fig. 2) in order
to allow all the partners of the project to customise it at their Local Health

ARC/VIEW 3.2 G.I.S.

DISPLAY RESULTS

DIS. MAP. DISPLAY

SELECT SELECT SELECT CREATE
DEFINE
TYPE OF RISK ASS. STUDY COMPARISON STUDY STATISTICAL
STUDY REGION REGION MAPS
COVARIATES RESULTS

ARC/VIEW ORACLE FORMS/RPTS.

AVENUE SCRIPTS

BASIC STATS. RESULTS

HEALTH POPULATION ENVIRONMENTAL GEOGRAPHIC FROM
SMOOTHING STUDY

DATA TABLES PL / SQL DATA TABLES

ORACLE DATABASE 8i

Fig. 2. RIF structure

244 V. Gómez-Rubio et al.

Authorities with their own data. As the Spanish Partner, we did it at the
Conselleria de Sanitat (Comunidad Valenciana, Spain).
As numerator tables, we have mortality, and hospital admissions (removing
repeated registers) are used as a proxy to morbidity. Population, as provided
by Census register, is used as denominator. All these data are available from
1989 to 1998 at the level of municipality, due to confidentiality.
A deprivation index [12] was also developed and incorporated into the
system to be used, together with sex and age group, in standardisation.
While mortality and morbidity are stored as single cases in the database
(and later aggregated in the studies) population is stored as habitants per
age-sex-deprivation stratum in each municipality.
Administrative boundaries at three levels (municipality, province and au-
tonomous community) are available in the system, and any of these levels can
be used when carrying out a study. They are stored as shapefiles, which are
used by ArcView to create maps. The code of the municipality is used to link
different data available.
For Disease Mapping, basic statistics (expected cases, observed cases, rel-
ative risk and its 95% confidence interval) are calculated for each region under
study and Poisson-Gamma Model [6] is used to provide smoothed estimations
of relative risks.
For Risk Assessment, one or several points regarding putative pollution
sources (nuclear plants, waste incinerators, tile industry, etc.) are selected,
and regions are grouped according to their distance to these points and the
same basic statistics as before are calculated.
Beside maps based on the results, a report with all the statistics is created
by Oracle Forms and Reports. Results are grouped by administrative region
and 6 groups depending on sex (males, females and both sexes) and type of
standardisation (whether deprivation index is taken into account). This way of
presenting data is useful to compare risk between different sex and deprivation
levels. All studies are stored in the database so they can be accessed later.

5.2 Enhancing the RIF

Although the RIF provides a rapid look into the data, we missed a few ca-
pabilities in the system. For example, it doesn’t perform any test to compare
results among different sex-deprivation strata, which is quite important. Fur-
thermore, covariates can’t be used in the studies and there is no possibility of
exporting results to be analysed with an external statistical software.
The use of covariates in the study was implemented inside the RIF using
the existing structure. Covariates are treated as sex, age group and deprivation
index in standardisation [12]. Each covariate is split into groups defined by
the user and rates are calculated with and without standardising by covariate
groups. If the covariate really has any relationship with the disease, we expect
to have different results.
 Epidemiological Information Systems 245

For further statistical analysis we decided to use the R Statistical Envi-

ronment [21] because it provides a wide range of options for spatial analysis
and it is freely available. We developed a couple of packages, RArcInfo [17]
and DCluster [18] to be used for Disease Mapping and detection of clusters of
disease, respectively.
In order to use Bayesian Hierarchical Models, WinBUGS [31] is a suit-
able software. Although MCMC calculations must be used and analysed with
care, a few models can be predefined so that they are automatically run after
plugging-in results from the RIF. This can be done within R, since a link
to WinBUGS has been developed and convergence of Markov Chains can be
checked in R by using package CODA [3].

Acknowledgements

This work has been partly funded by Conselleria de Sanitat, Comunidad Va-
lenciana (Spain) and EUROHEIS Project, code SI2.329122 (2001CVG2-604),
funded by European Commission.
References

1. P. Aylin, R. Maheswaran, J. Wakefield, S. Cockings, L. Jarup, R. Arnold,

G. Wheeler, and P. Elliot (1999) A national facility for small area disease
mapping and rapid initial assessment of apparent disease clusters around a
point source: The U.K. Small Area Health Statistics Unit. Journal of Public
Health Medicine, 21(3):289–298.
2. J. Besag, J. York, and A. Mollie (1991) Bayesian image restoration, with two
applications in spatial statistics. Annals of the Institute of Statistical Mathe-
matics, 43:1–59.
3. N. G. Best, M. K. Cowles, and S. K. Vines (1995) CODA Convergence Diagnosis
and Output Analysis software for Gibbs Sampler output: Version 0.3.
4. D. J. Briggs, P. Forer, L. Järup, and R. Stern, editors (2002) GIS for Emergency
Preparedness and Health Risk Reduction. Kluwer Academic Publishers.
5. M. Choynowski (1959) Map based on probabilities. Journal of the American
Statistical Society, 54(286):385–388.
6. D. Clayton and J. Kaldor (1987) Empirical bayes estimates of age-standardized
relative risks for use in disease mapping. Biometrics, 43:671–681.
7. S. Cockings and L. Järup (2002) A European Project and Environment In-
formation System for exposure and disease mapping and risk assessment (EU-
ROHEIS). In D. J. Briggs, P. Forer, L. Järup, and R. Stern, editors, GIS
for Emergency Preparedness and Health Risk Reduction, volume 11 of NATO
Sciences Series, chapter 11, pages 201–226. Kluwer Academic Publichers.
8. N. A. C. Cressie (1993) Statistics for Spatial Data. John Wiley & Sons, Inc.
9. E. K. Cromley and S. L. McLafferty (2002) GIS and Public Health. The Guilford
Press.
10. P. Diggle and P. Elliot (1995) Disease risk near point sources: Statistical issues
for analyses using individual or spatially aggregated data. Journal of Epidemi-
ology and Community Health, 49: S20–S27. (Suppl. 2).
11. K. H. Falter, D. R. Betts, D. B. Rolka, H. R. Rolka, and W. K. Sieber, editors
(1999) Statistics in Medicine, volume 18 (23). Wiley and Sons, Inc. Special
Issue: Symposium on statistical bases for Publich Health decision making: from
exploration to modelling.
12. J. Ferrándiz, J. J. Abellán, V. Gómez-Rubio, A. López-Quílez, P. Sanmartín,
C. Abellán, M. A. Martínez-Beneito, I. Melchor, H. Vanaclocha, O. Zurriaga,
 Epidemiological Information Systems 247

F. Ballester, J. M. Gil, S. Pérez-Hoyos, and R. Ocaña (2004) Spatial analysis of

the relationship between cardiovascular mortality and drinking water hardness.
Environmental Health Perspectives, 112(9):1037–1044.
13. J. Ferrándiz, J. J. Abellán, A. López, P. Sanmartín, H. Vanaclocha, O. Zurriaga,
M. A. Martínez-Beneito, I. Melchor, and J. Calabuig. Geographical distribution
of cardiovascular mortality in Comunidad Valenciana (Spain). In D. J. Briggs,
P. Forer, L. Järup, and R. Stern, editors (2002) GIS for Emergency Preparedness
and Health Risk Reduction, volume 11 of NATO Sciences Series, chapter 15,
pages 267–282. Kluwer Academic Publichers.
14. J. Ferrándiz, A. López, V. Gómez-Rubio, P. Sanmartín, M. A. Martínez-
Beneito, I. Melchor, H. Vanaclocha, O. Zuriaga, F. Ballester, J. M. Gil, S. Pérez-
Hoyos, and J. J. Abellán (2003) Statistical relationship between hardness of
drinking water and cerebrovascular mortality in Valencia: A comparison of spa-
tiotemporal models. Environmetrics, 14(5):491–510.
15. A. Gathrell and M. Löytönen, editors (1998) GIS and Health. Number 6 in
GISDATA. Taylor & Francis.
16. W. R. Gilks, S. Richardson, and D. J. Spiegelhalter, editors (1996) Markov
Chain Monte Carlo in Pratice. Chapman and Hall.
17. V. Gómez-Rubio and A. López-Quílez (2005) RArc Info: Using GIS data with
R. Computers & Geosciences 31:1000–1006.
18. V. Gómez-Rubio, J. Ferrándiz-Ferragud, and A. López-Quílez (2005) Detecting
clusters of disease with R. Journal of Geographical Systems 7(2):189–206.
19. R. Haining (2003) Spatial Data Analysis. Theory and Practice. Cambridge
University Press.
20. T. Hastie and R. Tibshirani (1986) Generalized additive models. Statistical
Science, 1:297–318.
21. R. Ihaka and R. Gentleman (1996) R: A language for data analysis and graphics.
Journal of Computational and Graphical Statistics, 5(3):299–314.
22. M. Jenicek and R. Cléroux (1982) Epidemiologie. Principes, Techniques, Ap-
plications. Edisem Inc., 2 edition.
23. A. Lawson, A. Biggeri, D. Böhning, E. Lesaffre, and J.-F. Viel, editors (1999)
Disease Mapping and Risk Assessment for Public Health. John Wiley & Sons
Ltd.
24. A. B. Lawson (2001) Statistical Methods in Spatial Epidemiology. Wiley.
25. R. J. Marshall (1991) A review of methods for the statistical analysis of spa-
tial patterns of disease. Journal of the Royal Statistical Society, Series A,
154(3):421–441.
26. P. McCullagh and J. Nelder (1989) Generalized Linear Models. Chapman and
Hall.
27. J. Pekkanan, E. Pukkala, M. Vahteristo, and T. Vatiainen (1995) Studying
cancer incidence around an oil refinery as an example of a small area study
based on map coordinates. Environmental Research, 71(2):128–134.
28. S. Piatandosi, D. P. Byar, and S. B. Green (1988) The ecological fallacy. Amer-
ican Journal of Epidemiology, 127(5):893–904.
29. P. Rees, P. Norman, and D. Brown (2004) A framework for progressively im-
proving small area population estimates. Journal Royal Statistical Society A,
167:5–36.
30. W. K. Sieber, T. A. Green, G. S. Haugh, M. jo Kresnow, E. T. Luman, and
H. G. Wilson, editors (2001). Statistics in Medicine, volume 20 (9–10). Wiley
248 V. Gómez-Rubio et al.

and Sons, Inc. Special Issue: Symposium on Emerging Statistical Issues in

Public Health for the 21st Century.
31. D. J. Spiegelhalter, A. Thomas, and N. G. Best (1999) WinBUGS Version 1.2
User Manual. MRC Biostatistics Unit.
32. J. Wakefield, M. Quinn, and G. Raab, editors (2001) Journal of the Royal
Statistical Society (Series A), volume 164 (1). Blackwell.
33. J. C. Wakefield, J. E. Kelsall, and S. E. Morris (2000) Clustering, cluster
detection and spatial variation in risk. In P. Elliot, J. C. Wakefield, N. G. Best,
and D. J. Briggs, editors, Spatial Epidemiology. Methods and Applications.,
chapter 8, pages 128–152. Oxford University Press.
34. J. C. Wakefiled, N. G. Best, and L. Waller (2000) Bayesian approaches to
disease mapping. In P. Elliot, J. C. Wakefield, N. G. Best, and D. J. Briggs,
editors, Spatial Epidemiology: Methods and Applications, pages 104–127. Oxford
University Press.
Geomatics, Epidemiology and BioStatistics:
An Application to Acute Coronary Syndrome

Théophile Niyonsenga1 , Josiane Courteau2 , Charmaine Dean3 , Abbas

Hemiari2 , Goze Bénié4 , and Alain Vanasse2,5
1
Robert Stempel School of Public Health, Florida International University (FIU),
Miami, FL (USA)
theophile.niyonsenga@fiu.edu
2
PRIMUS group, Clinical Research Center, CHUS, Sherbrooke (QC), Canada
3
Statistics and Actuarial Science, Simon-Fraser University, Vancouver, BC,
Canada
dean@stat.sfu.ca
4
Geography and Remote-Sensing Department, Université de Sherbrooke,
Sherbrooke, QC Canada
Goze.Bertin.Benie@USherbrooke.ca
5
Family Medicine Department, Université de Sherbrooke, Sherbrooke (QC),
Canada
alain.vanasse@usherbrooke.ca

1 Introduction

The emerging field of Geomatics has found useful application in several re-
search areas. A new phase of its development merges it into the realm of
epidemiology and public health to bring insight into the regional disparities
of disease incidence and for disease surveillance. New advances in biostatistics
include spatial statistics methods which aim to specifically understand and
model the spatial variability. Spatial statistics, when combined with geomat-
ics, constitutes an excellent and powerful analysis approach to handle and
better understand health issues, specifically in disease related prevention and
intervention studies.
The main goal of this paper is to explore the integration of geomatics and
spatial statistics with an application to a specific health issue. The outcome of
interest is acute coronary syndrome (ACS) incidence in the province of Que-
bec (Canada) between 1996 and 1998, and hospital readmission at one month
post-discharge. It is an established fact that mortality and hospital-acquired
infection rates are indicators of the quality of care [24] but more recently
some studies are turning their attention to the early hospital readmission
rate [2, 26]. Within this context, following specific questions are addressed: Is
there spatial heterogeneity and/or spatial aggregation in the ACS incidence
250 T. Niyonsenga et al.

and early readmission rates? Is there any geographical trend in the rates?
Is there an explanation for the spatial heterogeneity? By ACS, we mean the
occurrence of myocardial infarction (MI) or unstable angina. By early read-
mission, we mean readmission of a patient for coronary heart disease (ACS
and angina) 30-days post discharge. Although practice guidelines have been
in circulation to standardize the treatment and follow-up of acute myocardial
infarction [16, 22], regional variations are currently reported in the literature
[20, 21, 25]. The presence of a complex network of factors influencing care
quality [2], hospital readmission and the interaction between them have been
put forward as the potential explanation of the observed spatial variability in
hospital readmission rates. Most of these factors center on the patient while
the data available and/or the interests of public policy makers focus on rates
of local health units over a given time period. Over these administrative ge-
ographical units, interest centers on variables that could explain spatial vari-
ability, such as deprivation indices and other area specific characteristics, and
help to understand inequalities within health care services and accessibility.

2 Methods
2.1 Population
The studied population consists of all the patients living in the Québec
province in Canada, hospitalized for an ACS. The first registered hospital-
ization during the study period (1996–1998) will be considered as the “index
hospitalization”. The Québec register “Maintenance et Exploitation des Don-
nées pour l’Étude de la Clientèle Hospitalière (MED-ECHO)” made possible
the identification of the patients that fit inclusion criterion. This register lists
all summary administrative data collected when any patient is treated in an
acute care hospital in the Québec province. The validity of this data, concern-
ing MI, was studied and its results published [14, 19, 27, 28]. The inclusion
criterion is the inscription of the code 410 (MI) or 411 (unstable angina) of
the international classification of disease 9th revision (IDC-9) as the main
diagnosis for the hospitalization. In order to increase the study’s internal va-
lidity, we excluded patients with an error of code of residence and patients
that were younger than 25 years old, because they are more likely to have had
a MI caused by a different pathophysiological process. For the same reasons,
we wanted ‘new’ cases of ACS, so we excluded patients that were hospitalized
with an ACS in the year preceding the index hospitalization.

2.2 Geographical Unit

Spatial data sources used in this study are ESRI [7], DMTI Spatial [5] and
the Ministère de la santé et des services sociaux [3, 17]. The geographic co-
ordinate system used in mapping is GCS North American 1983. Spatial data
and cartographic representations were managed using ArcGIS [1].
 Geomatics, Epidemiology and BioStatistics 251

2.3 Variables

The hospitalization for ACS and early hospital readmission are our depen-
dent variables. The former variable is defined as the first occurrence of an
ACS hospitalization (main diagnosis IDC-9 codes 410 or 411) in the 3 years’
study (index hospitalization), and the latter is defined as a early readmission
for a coronary heart disease (main diagnosis IDC-9 codes 410 to 414) in the
30-days following the index hospitalization. The incidence and the readmis-
sion rates were calculated by Local Health Unit (LHU). For the readmission
rate, we excluded all in-hospital deaths and all deaths encountered within
30-days post discharge. Two deprivation indices were retained as potential
explanatory variables for the spatial heterogeneity in health outcomes. Pam-
palon et al. [18] calculated two sets of deprivation quintiles for approximately
9,000 enumeration areas (EA) using a principal component analysis (PCA).
The two principal components of this PCA reflected a material dimension
of deprivation (percent of people without a secondary certificate, the ratio
employment by population, the average income) and a social dimension of de-
privation (the percent of people divorced, separated or widowed, the percent
of single-parents, the percent of persons living alone). For each of these PCA
components, the EA were ordered by their factor score – from the least to
the most deprived – and then the population was fragmented in quintiles (the
fifth quintile being the most deprived). For each LHU, the population that
belongs to each quintile was calculated. Based on these quintiles, we defined
two variables, denoted by material deprivation index (MDI) and social depri-
vation index (SDI), by the percent of the LHU population that belongs to the
fifth quintile.

2.4 Analyses

The analyses were performed in two steps. We first focussed on the geographi-
cal distribution of rates (spatial heterogeneity and clusters). To detect clusters
and particular hot spots, we used SaTScan [12]. To see the general spatial
trend in the rates, we used the Geographically Weighted Regression [8] ap-
proach (GWR), the intercept model (using GWR package [10]) as well as a
Poisson regression model as a function of latitudes and longitudes (using SAS
[23]). The spatial autocorrelation ‘best’ model was estimated by the variogram
(using GS+ [9]) by minimising the residual sum of squares (RSS) criteria. The
second step was to explain the observed heterogeneity by the available covari-
ables through regression models. Consider the general model:

g(E[y]) = Xi β + Ui

where y is the rate variable, g is the link function and Ui , i = 1, 2, . . . , m de-

notes the area-specific random effects (spatially unstructured and structured
effects). A possible way to take into account spatial variability in a Poisson
252 T. Niyonsenga et al.

regression analysis of rates is to use the approach of Kleinschmidt et al. [11].

Their idea is the following: First, we suppose that there is no spatial auto-
correlation. We then perform an ordinary Poisson regression model (with no
correlation structure) as a function of the deprivation indices, calculate the
signed residuals, and fit a theoretical semivariogram model (using GS+) to
estimate the nugget σ12 , the sill σ 2 and the range ρ. These three parameters
are then used to define a covariance matrix V associated with the spatial
random effects:

V = Iσ12 + Fσ 2 with Fij = exp(−dij /ρ)

were dj refers to the distance between centroids of LHUs i and j, and I is an

identity matrix. We assume here an exponential model for the spatial struc-
ture, but Fij could be given by any other theoretical model adjusted to the
variogram. Then, we fit a spatially correlated Poisson-mixture model (using
the GLIMMIX SAS macro) with the covariance matrix V defined above, and
calculate a second set of signed residuals. We iterate this process until there is
convergence in the estimates. Another way of taking into account the spatial
variability is to use a GWR model [8], the idea being to allow for geographical
variation in the parameters β instead of fitting global regression models. The
Gaussian model is given by:

yi = β0 (ui , vi ) + βk (ui , vi )xik + εi
k

where (ui vi ) denotes the coordinates of the ith LHU. If we assume that LHUs
far from each other are more likely to have different coefficients, a weighted
calibration (W) is used, which is a Kernel-type function of the distance and
a varying bandwidth. The coefficients β are thus estimated by:

β̂(ui , vi ) = (X W(ui , vi )X)−1 X W(ui , vi )y

where β̂ contains k + 1 continuous functions of geographical coordinates. Be-

cause the GWR package was not yet available for binary and count data in
2003, we modeled the logarithm of the rates, and used the normal approxi-
mation for the error term.

3 Results
A total of 50,839 patients met the inclusion criteria and were listed in the
MED-ECHO register between January 1st 1996 and December 31st 1998.
Almost 10% of the individuals (n = 4, 749) died during the “index hospital-
ization”. Thirty-three patients were excluded because of their age (< 25 years
old) or because of an error in the code of residence. Furthermore, 1,516 pa-
tients had been hospitalized in the year preceding the index hospitalization
 Geomatics, Epidemiology and BioStatistics 253

Table 1. ACS incidence and early readmission rates

ACS incidence non-fatal incidence early readmission
women 17,841 15,205 1,305 (8.6%)
< 55 years 2,059 1,991 179 (9.0%)
≥ 55 years 15,782 13,214 1,126 (8.5%)
men 31,449 28,649 2,785 (9.7%)
< 45 years 2,419 2,389 247 (10.3%)
≥ 45 years 29,030 26,260 2,538 (9.7%)
total 49,290 43,854 4,090 (9.3%)

and were also excluded, for a total cohort of 49,290 patients. Among these,
43,854 (89%) were alive 30-days after discharge. At the index hospitalization,
the average age of patients was 66.2 years (± 13.3), and men represented 63.8%
(n = 31, 449) of total cohort. A total of 4,090 early readmissions (9.3%) have
been observed. The early readmission rate is higher for men but lower for
older patients (Table 1). Figure 1 shows the 1996-population within LHU of
the province of Quebec, while Figs. 2 and 3 show material and social depriva-
tion indices respectively. We can easily see a large population density in the
south part of the province, as well as in the coasts of the St-Laurent River.
In addition, as measured with deprivation indices, we can observe that the
urban regions are generally less deprived in the material sense but the reverse
is observed for the social deprivation index.
Crude ACS incidence and readmission rates are shown in Figs. 4 and 5. The
most likely clusters are highlighted as black lines. In Figs. 6 and 7 however, we
present the smooth surface estimates of ACS incidence and early readmission
rates with the GWR method (intercept model only; Monte Carlo test for
spatial variability). The parameter estimates range from 0.00086 to 0.01266

Fig. 1. Population of Quebec in 1996

254 T. Niyonsenga et al.

Fig. 2. Material Deprivation Index (MDI)

(p < 0.0001) for ACS incidence and from 0.08093 to 0.13639 (p = 0.0300)
for readmission rates. There is a North-East trend for ACS incidence rates
while the readmission rates tend to increase as we move away from the urban
centers (see smoothed surfaces) (Figs. 6 and 7).
Poisson regression models of the number of ACS and readmissions as a
function of latitude and longitude showed similar trends (predicted rates not
shown), confirming the observed trend from the GWR analysis. The method
proposed by Kleinschmidt et al. [11] was explored but the semivariogram of
the signed residuals suggested the absence of an autocorrelation structure, so

Fig. 3. Social Deprivation Index (SDI)

 Geomatics, Epidemiology and BioStatistics 255

Fig. 4. Crude ACS incidence with hot spots (black lines)

the process stopped at the first iteration. Nevertheless, to see if the deprivation
indices explain the heterogeneity in the rates, we performed a GWR on the log-
arithm of the rates as a function of these two indices. The parameter estimates
for readmission rates were significant (p < 0.0001) while for ACS incidence
rates only the SDI estimate was significant (p < 0.0001). The residuals were
the ratio of observed and expected rates. Maps of surface residuals (Figs. 8
and 9) show that the two indices explain some of the variation but the ob-
served trend in some regions remains unexplained by the deprivation indices.

Fig. 5. Crude readmission rates with hot spots (black lines)

256 T. Niyonsenga et al.

Fig. 6. Smoothed ACS incidence (GWR intercept estimates)

Fig. 7. Smoothed readmission rates (GWR intercept estimates)

4 Discussion
Spatial analysis, in the sense of analysis of data in a geographical perspective,
has to be linked to geomatics and geographical information systems. Within
the context of public health issues, a map displaying geographic heterogeneity
is one of the most powerful tools for interpretation of spatial data once we
determine which surface estimate best portrays the data and which estimation
methods are appropriate for the health parameters of interest. Elliott et al. [6]
reviewed estimation methods such as Kernel-based and GAM-based methods
 Geomatics, Epidemiology and BioStatistics 257

Fig. 8. ACS incidence as a function of MDI and SDI (GWR residuals)

Fig. 9. Readmission rates as a function of MDI and SDI (GWR residuals)

as well as estimation methods involving the mixture of distributions. Lawson

et al. [13] motivated the use of model-based estimation methods, especially
methods involving random effects within a Bayesian framework. Cressie [4]
pointed out the full potential for hierarchical spatial model-based methods
using full Bayesian and Monte Carlo Markov Chain approaches. However, an
important question for future consideration is what we gain through the use
of additional complexity from the crude surface to Kernel-based to empirical
Bayes to full Bayesian methods. Future work will also explore the added value
258 T. Niyonsenga et al.

of the hierarchical setting by using other patient and LHU level covariables
as potential predictors with a hierarchy in the parameters of interest within
the Bayesian framework and in the context of no prior information (flat pri-
ors) or limited information. Another promising avenue for exploration is how
Monte Carlo simulations may be used to inform the choice of prior distri-
butions. In the context of establishing relationships between health events
and covariables (within exploratory or confirmatory spatial data analysis), it
would be quite interesting to combine both Kernel-based methods such as the
GWR approach and random effects models to deal simultaneously with large
scale spatial variability (uncorrelated heterogeneity) and small scale variabil-
ity (spatial autocorrelation). Another important consideration is the difference
between semivariogram approach to model the spatial autocorrelation struc-
ture and conditional autoregressive (CAR) models [15]. As Fotheringham et al.
[8] pointed out, the GWR approach offers an interesting framework by allow-
ing estimation of a continuous surface for each potential correlate which could
usefully be mapped. This approach offers a fertile ground for future study.

Acknowledgements
This project was subsidized by the Network of Centres of Excellence GEOIDE
and Merck Frosst Canada Ltd. At the time of the study, the first author
had the support of the Clinical Research Center of the Centre Hospitalier
Universitaire de Sherbrooke.
References

1. ARCGIS. Release 8.2. ESRI, USA

2. Ashton CM, Kuykendall DH, Johnson ML, et al (1995) The association between
the quality of inpatient care and early readmission. Ann Intern Med 122:415–21
3. Cardiologie tertiaire, situation actuelle, perspectives et propositions, 2000.
Ministère de la Santé et des Services Sociaux
4. Cressie N (2000). Spatial statistics and environmental sciences. In: proceedings
of the section of statistics and the environment. American Statistical Associa-
tion. Alexandria, VA,1–10
5. DMTI Spatial Data Delivery System (2000) CanMap Streetfiles et PostCode
files, http://www.dmtispatial.com
6. Elliott P, Wakefield JC, Best NG, Briggs DJ (2000) Spatial Epidemiology: Meth-
ods and Applications. Oxford University Press, New York
7. ESRI Date & Maps 2002, Media Kit, 2002, ESRI ArcGIS, http://www.esri.com
8. Fotheringham AS, Brunsdon C, Charlton M (2002) Geographically Weighted
Regression: The Analysis of Spatially Varying Relationships. John Wiley & Sons,
England
9. GS+ Geostatistics for the Environmental Sciences, version 5.3.2. γ Design Soft-
ware. Painwell, MI, USA
10. GWR Version2.0.5. University of Newcastle, England
11. Kleinschmidt I, Sharp BL, Clarke GPY, Curtis B, Fraser C (2001) Use of gener-
alized linear mixed models in the spatial analysis of small-area malaria incidence
rates in KwaZulu Natal, South Africa. Am J Epidemiol 153(12):1213–21
12. Kulldorff M (1997) A spatial scan statistic. Commun. Stat.: Theory
Methods,26:1481–96.
13. Lawson AB, Browne WJ, Vidal Rodeiro CL (2003) Disease Mapping with
WinBUGS and MLwiN. Wiley & sons, New York
14. Levy AR, Tamblyn RM, Fitchett D, McLeod PJ, Hanley JA (1999) Coding
accuracy of hospital discharge data for elderly survivors of myocardial infarction.
Can J Cardiol 15(11):1277–82
15. McNab Y, Dean CB (2000) Parametric bootstrap and penalyzed quasi-likelihood
inference in conditional autoregressive models. Stat. Med., 19(17–18), 2421–35
16. Mehta RH, Montoye CK, Gallogly M, et al for the GAP Steering Committee
of the American College of Cardiology (2002) Improving quality of care for
260 T. Niyonsenga et al.

acute myocardial infarction: the guidelines applied in practice (GAP) initiative.

JAMA 287(10):1269–76
17. Ministère de la Santé et des Services Sociaux, http://www.msss.gouv.qc.ca
18. Pampalon R, Raymond G (2000) Un indice de défavorisation pour la planifi-
cation de la santé et du bien-être au Québec. Maladies Chroniques au Canada
21(3):113–22
19. Petersen LA, Wright SM, Normand SLT, Daley J (1999) Positive predictive
value of the diagnosis of acute myocardial infarction in an administrative
database. J Gen Intern Med 14:555–8
20. Pilote L, Califf RM, Sapp S, et al (1995) Regional variation across the United
States in the management of acute myocardial infarction. GUSTO-1 Investi-
gators. Global utilization of streptokinase and tissue plasminogen activator for
occluded coronary arteries. N Engl J Med 333(9):565–72
21. Pilote L, Granger C, Armstrong PW, et al (1995) Differences in the treatment
of myocardial infarction between the United States and Canada. A survey of
physicians in the GUSTO trial. Med Care 33(6):598–610
22. Ryan TJ, Antman EM, Brooks NH, et al (1999) 1999 update: ACC/AHA guide-
lines for the management of patients with acute myocardial infarction. A re-
port of the ACC /AHA Task Force on practice guidelines. J Am Coll Cardiol
34(3):890–911
23. The SAS systems for windows. Release 8.02. SAS Institute Inc. Cary, NC, USA
24. Thomas JW, Holloway JJ (1991) Investigating early readmission as an indicator
for quality of care studies. Med. Care 29:377–94
25. Van der Werf F, Topol E, Lee K, et al (1995) Variations in patient treatment
and outcomes for acute myocardial infarction in the United States and other
countries. JAMA 273:1586–91
26. Weissman JS, Ayanian JZ, Chasan-Taber S, et al (1999) Hospital readmissions
and quality of care. Med. Care 37:490–501
27. Wenneberg JE (1984) Dealing with medical practice variations: a proposal for
action. Health Aff (Millwood) 3:6–32
28. Williams JI, Young W (1996) Appendix – A Summary of Studies on the Qual-
ity of Health Care Administrative Databases in Canada. In: Goel V, Williams
JI, Anderson GM, Blackstein-Hirsch P, Fooks C, Naylor CD (eds) Patterns of
Health Care in Ontario : The ICES Practice Atlas. 2nd ed. Ottawa, Ontario :
Canadian Medical Association, pp 339–45
Interactive Cumulative Curves for Exploratory
Classification Maps

Gennady Andrienko and Natalia Andrienko

Fraunhofer Institute AIS Schloss Birlinghoven, Sankt Augustin, Germany

gennady.andrienko@ais.fraunhofer.de

1 Classification as an Instrument
for Exploratory Analysis

Representation of values of a numeric attribute referring to geographical ob-

jects, in particular, areas of territory division, is often done in cartography
using the technique of classification. According to this technique, the value
range of the attribute is divided into intervals. Then different colours are
chosen to represent values from each of the intervals on a map.
Historically, classification was indispensable due to technical limitations
involved in production of paper maps as well as in display of maps on early
graphical computer screens [6]. With appearance of modern computer hard-
ware these limitations were eliminated, and it became possible to produce
unclassed maps. In such maps numeric values are encoded by proportional
degrees of darkness.
The merits and flaws of classed and unclassed maps have long been a
topic of hot debates in cartography. It is not our intention to recite here this
discussion or to take either side in it. Although classed maps are in the focus of
this paper, this does not mean that we regard them as superior to unclassed
maps. Our interest is the use of maps in exploratory analysis of spatially
referenced data. In such analysis each type of maps plays its own role, and,
hence, there is no question whether to select a classed or an unclassed map.
An analyst should use both because these are complementary instruments of
analysis.
Our opinion can be substantiated as follows. The goal of exploratory
analysis is to gain understanding of given data, that is, to derive a short
(compressed) description of their essential characteristics. Thus, according to
Bertin, understanding is “discovering combinational elements which are less
numerous than the initial elements yet capable of describing all the infor-
mation in a simpler form” ([3], p. 166). With regard to spatial distribution
of values of a numeric attribute, an analyst initially has one value per each
262 G. Andrienko and N. Andrienko

spatial object. The description of this data set could be substantially simplified
if there were a directional trend in value distribution, for example, increase of
values from the north to the south or from the centre of the territory to its
periphery. Alternatively, a shorter description could be derived if the territory
could be divided into possibly smaller number of coherent regions with low
variation of attribute values within the regions. This technique is known as
regionalisation. Unclassed maps are better suited for detecting trends because
they do not hide differences. Classification discards differences between values
within a class interval and gives the corresponding objects similar appearance
on the map. When these objects are geographical neighbours, they tend to
be visually associated into clusters. This property makes classed maps well
suitable for regionalisation. Which of the two ways to simplification occurs to
be possible or more effective in each specific case, depends on the data and
not on the preferences of the analyst. Therefore it is necessary to have both
an unclassed choropleth map and a classification tool in order to investigate
properly data with previously unknown characteristics.
It is clear, however, that a single static classed map cannot appropriately
support regionalisation. It is well known in cartography that different selection
of the number of classes and class breaks may radically change the spatial
pattern perceived from the map [5, 7]. There is no universal recipe of how to
get an “ideal” classification with understandable class breaks, on the one hand,
and interpretable coherent regions, on the other hand. Therefore when we say
that classification may be used as an instrument of data analysis, we mean
not a classed map by itself but an interactive tool that allows the analyst to
change the classes and to observe immediately the effect on the map.
The exploratory value of classification was recognised in cartography only
relatively recently. Initially classification was regarded as a tool for conveying
specific messages from the map author to map consumers. Thus, the paper
[8] considers various possible intentions of the map designer and demonstrates
how they can be fulfilled through application of different classification methods
and selection of the number of classes.
In early nineties Egbert and Slocum developed a software system called
ExploreMap intended to support exploration of data with the use of classed
choropleth maps [4]. The most important feature of the system was a pos-
sibility to interactively change the classes. Another implementation of this
function based on direct manipulation techniques is the “dynamic classifica-
tion” tool incorporated in the system CommonGIS [1, 2]. Exploration on the
basis of classification is additionally supported in CommonGIS by the function
of computing statistics for the classes: the range of variation and the average,
median and quartile values of any selected attribute for each class.
In this paper we describe a recently developed extension of the dynamic
classification tool that exploits the properties of the cumulative frequency
curve and generalised cumulative curves. In the next section we define the
relevant notions and explain the use of cumulative curves in classification and
data exploration.
 Interactive Cumulative Curves for Exploratory Classification Maps 263

2 Cumulative Frequency Curve and Its Use

in Classification
In classification of spatially referenced data an analyst needs to consider the
data from two perspectives, statistical and spatial, and take into account the
peculiarities of both the statistical and the spatial distributions of the data.
This means that the analyst needs to pursue at least two concurrent goals.
The first is to minimise variation of data within each class and to maximise
differences between classes. The second goal is to divide the territory into the
smallest possible number of coherent regions with low data variation within
the regions. Additional goals may emerge in particular application domains.
Thus, in demographic applications it may be necessary to minimise differences
between the classes in total population or total area. The analyst needs such
tools that would allow her/him to balance between these goals in search of an
acceptable compromise solution.
A visual representation of the statistical value distribution can help the
analyst to meet the statistical criteria. Yamahira et al. [8] suggested that
classification could be supported by a frequency histogram. However, a his-
togram represents results of prior classification and therefore can hardly serve
as a tool for it. The dynamic classification tool of CommonGIS includes a dot
plot, or point graph. This kind of graph does not require prior classification
and is well suited for including in the interactive classification device due to its
modest requirements to screen space. On the other hand, an obvious problem
is overlapping of point symbols that obscures the understanding of the distri-
bution. Slocum [7] uses so called dispersion graphs (for illustration purposes)
in his discussion of the existing classification methods. This representation is
based on classification into a large number of classes. Values fitting in a class
are shown by dots stacked at the class position. Since dispersion graphs al-
ready involve classification, they are not so good as tools for producing other
classifications.
One more method for graphical representation of statistical distribution is
the cumulative frequency curve, or ogive. In such a graph the horizontal axis
represents the value range of an attribute. The vertical position of each point
of the curve corresponds to the number of objects with values of the attribute
being less than or equal to the value represented by the horizontal position of
this point. The method of construction of the ogive is demonstrated in Fig. 1.
The curve represents the distribution of values of the attribute “Number of
cars per capita” over districts of Leicestershire (this and further examples
refer to the Leicestershire sample of the 1991 census data available at the
URL http://www.mimas.ac.uk/descartes/).
Peculiarities of value distribution can be perceived from the shape of the
ogive. Steep segments correspond to clusters of close values. The height of
such a segment shows the number of the close values. Horizontal segments
correspond to “natural breaks” in the sequence of values.
264 G. Andrienko and N. Andrienko

100%

Number of objects(districts)
with values no more than 0.5
(72% of all objects)
values between 0.3 and 0.5
Number of objects with

Number of objects(districts)
with values no more than 0.3
(14% of all objects)

CARSCAP (N)
Minimum value of Value 0.3 Value 0.5 Maximum value of
the attribute (0.13) the attribute (0.63)

Fig. 1. A cumulative frequency curve

It is important that the cumulative frequency curve does not require prior
classification. However, it can represent results of classification by means of
additional graphical elements, and we used this opportunity in the latest ex-
tension of CommonGIS. Thus, the horizontal axis of the graph may be suited
to show class breaks. In the interface adopted in CommonGIS (Fig. 2) we use
for this purpose segmented bars with segments representing the classification
intervals. The segments are painted in the colours of the classes. The positions
of the breaks are projected onto the curve, and the corresponding points of
the curve are, in their turn, projected onto the vertical axis. The division of
the vertical axis is also shown with the use of coloured segmented bars. The
lengths of the segments are proportional to the numbers of objects in the
corresponding classes. With such a construction it becomes easy to compare
the sizes of the classes. For example, the class breaks shown in Fig. 2 divide
the whole set of objects into 3 groups of approximately equal size that is
demonstrated by the equal lengths of the bar segments on the vertical axis.
The overall interface for classification is shown in Fig. 3. Besides the cumu-
lative curve display exposing statistical characteristics of the current classifi-
cation, it includes a map showing geographical distribution of the classes. For
the use of the cumulative curve as a tool for classification it is important that
its display immediately reacts to any changes of the classes, as well as the map
does. The user can change class breaks by moving the sliders (double-ended
vertical arrows) along the slider bar (on the upper right of the window). In
 Interactive Cumulative Curves for Exploratory Classification Maps 265

Fig. 2. Representation of classes on a cumulative frequency curve display. Values of

the attribute “Number of cars per capita” are divided by 2 breaks (0.404 and 0.491)
into 3 classes with approximately equal sizes (33, 33, and 34% of the whole set).
The positions of the breaks on the vertical axis are indicated below the graph: 33
and 66% (of the total number of the classified objects)

the process of moving the slider the map and the cumulative curve graph are
dynamically redrawn. In particular, changing are the relative lengths of the
coloured bars on the axes and the positions of the projection lines. Clicking on
the slider bar introduces a new class break, bringing a slider close to another
slider results in the corresponding break being removed. The map and the
cumulative curve display immediately reflect all these changes.
It is important that the cumulative frequency curve does not require prior
classification. However, it can represent results of classification by means of
additional graphical elements, and we used this opportunity in the latest ex-
tension of CommonGIS. Thus, the horizontal axis of the graph may be suited
to show class breaks. In the interface adopted in CommonGIS (Fig. 2) we use
for this purpose segmented bars with segments representing the classification
intervals. The segments are painted in the colours of the classes. The positions
of the breaks are projected onto the curve, and the corresponding points of
the curve are, in their turn, projected onto the vertical axis. The division of
the vertical axis is also shown with the use of coloured segmented bars. The
lengths of the segments are proportional to the numbers of objects in the
corresponding classes. With such a construction it becomes easy to compare
the sizes of the classes. For example, the class breaks shown in Fig. 2 divide
266 G. Andrienko and N. Andrienko

Fig. 3. The interface for classification provided in CommonGIS allows the user to
account both for statistical and for spatial distribution of values

the whole set of objects into 3 groups of approximately equal size that is
demonstrated by the equal lengths of the bar segments on the vertical axis.
The overall interface for classification is shown in Fig. 3. Besides the cumu-
lative curve display exposing statistical characteristics of the current classifi-
cation, it includes a map showing geographical distribution of the classes. For
the use of the cumulative curve as a tool for classification it is important that
its display immediately reacts to any changes of the classes, as well as the map
does. The user can change class breaks by moving the sliders (double-ended
vertical arrows) along the slider bar (on the upper right of the window). In
the process of moving the slider the map and the cumulative curve graph are
dynamically redrawn. In particular, changing are the relative lengths of the
coloured bars on the axes and the positions of the projection lines. Clicking on
the slider bar introduces a new class break, bringing a slider close to another
slider results in the corresponding break being removed. The map and the
cumulative curve display immediately reflect all these changes.

3 Generalised Cumulative Curves

In particular application domains additional classification criteria may come
into play in combination with the statistical and geographical ones. For ex-
ample, in demographic analyses it may be important to produce classes of
districts that do not differ too much in total number of population that lives
in them.
 Interactive Cumulative Curves for Exploratory Classification Maps 267

It is possible to generalise the idea of the cumulative frequency curve and to

build similar graphs summarising values of arbitrary quantitative attributes.
Examples of such attributes are area, population number, gross domestic prod-
uct, number of households, etc. A generalised cumulative curve is built in the
following way. Let the horizontal axis correspond to attribute A and the ver-
tical to attribute B. The curve matches each value x of A with the sum of
values of B computed for objects with the values of A being less than or equal
to x. So, the maximum value of A will correspond to the total sum of val-
ues of B for all the objects of the sample. Let, for example, the districts of
Leicestershire be classified according to the number of cars per capita, and a
generalised cumulative curve represent the attribute “Total population”. Then
the vertical position corresponding to x cars per capita would reflect the total
number of population living in districts with no more than x cars per capita.
The classification tool of CommonGIS allows the user to add a generalised
curve for any quantitative attribute to the cumulative frequency curve display.
The curves are overlaid, i.e. drawn in the same panel (see Fig. 4). To be
better discriminated, the curves differ in colour. The horizontal axis is common
for all of them. The vertical axes are shown beside each other on the left
of the graph. Each of the vertical axes is divided into the same number of
segments, but positions of the breaks are, in general, different. This is clearly
demonstrated in Fig. 4. It shows that 34% of all districts fit in the lower class
of the classification. They occupy only 10% of the total area but contain 50%
of total population.

Fig. 4. Generalised cumulative curves are built for the attributes “Area” and “Total
population”. The classification is done on the basis of the attribute “Number of cars
per capita”
268 G. Andrienko and N. Andrienko

Having such a tool, it is easy to account in classification for such criteria

as even distribution of population among the classes, or approximately equal
total areas occupied by the classes, or other specific criteria that may emerge
in this or that application domain. Thus, to make classes approximately equal
in total population, the analyst should focus in the process of slider movement
on the axis corresponding to total population and try to position the sliders
so that the axis is divided into segments of equal length.
Besides this opportunity, generalised cumulative curves may be used for
exploring relationships between various characteristics of the classified objects.
Let us demonstrate this on the example of exploration of unemployment in
Leicestershire. We used the attributes “Number of unemployed” and “Total
population” to calculate percentage of unemployed in total population in each
district. Then we took this new attribute as the basis for classification. The
classification tool showed us that proportion of unemployed in population
varies from 0.9 to 13.62%. We considered values above 4% to be very high
and wondered in how many districts this threshold is exceeded and where
these districts are located. We entered 4 as a class break and in this way
divided all districts into two classes: with up to 4% of unemployed persons in
population and with more than 4%. The cumulative curve display showed us
that only 18% of all districts fit in the upper class (Fig. 5). The map shows
a vivid spatial cluster of such districts in the centre of the area. It is seen
that these districts occupy a rather small part of the whole area. However,
when we selected the attribute “Total population” for representation on the
cumulative curve display, we found that the districts with high unemployment
contain 33% of the total population of Leicestershire.
We became interested whether there is a link between unemployment and
distribution of national minorities. We added to the cumulative curve display
the attribute BLAFR representing total numbers of people originating from
Africa by districts. In Fig. 5 it is vividly seen that the curve for this population
group radically differs from that for the whole population. It is also seen

Fig. 5. The use of generalised cumulative curves for exploration of unemployment

in Leicestershire
 Interactive Cumulative Curves for Exploratory Classification Maps 269

that the axis corresponding to this attribute is divided in quite a different

proportion than those for the frequency and for the whole population. Only
28% of people with African origin live in districts with lower unemployment
and, hence, 72% live in districts with more than 4% unemployed in total
population. The difference is even more dramatic for people originating from
India (represented by the attribute EGINDIAN). One can see that 81% of
these people live in the areas with high unemployment.
We continued our investigation of unemployment by moving the class break
so that the population was divided into two equal parts. Figure 6 shows the re-
sult of this operation. The new value of the class break is 2.89. This means that
50% of total population lives in districts with more than 2.89% of unemployed.
These districts, as it is seen from the map, constitute a rather small part of the
whole territory of the county. Hence, the population density in them is higher
than in the rest of the districts. Apparently, these are mainly urban districts.
The map shows also that the districts with higher unemployment are spatially
clustered. The national minorities considered above are now distributed be-
tween the classes of districts in the following way: only 14% of Indians and
18% of Africans live in the districts with lower unemployment, and, hence, 86
and 82%, respectively, live in the areas with more than 2.89% unemployed in
population.
Classification with multiple criteria involved but also to reveal significant
relationships in data. However, it should be borne in mind that this technique
is suitable only for attributes the values of which can be summed up over the
set of objects they refer to. For example, it would be wrong to apply it to
percentages, averaged values, rates, values per capita etc.
The implementation of the interactive classification tool based on the use
of cumulative curves is done in Java. This allows the use of it on different
platforms and in the WWW. The new version of the system CommonGIS
that includes the tool described can be run in the WWW at our homepage.

Fig. 6. The cumulative curve display was used to divide the districts into two classes
with equal total population
270 G. Andrienko and N. Andrienko

4 Conclusion

Interactive, dynamic classification can be a valuable instrument in exploratory

analysis of spatially referenced data. Representation of classes on a map by
colouring facilitates perception of patterns of spatial distribution. Therefore
we included a dynamic classification tool into the array of exploratory facilities
offered by our system CommonGIS.
In classification of geographical objects according to values of a numeric
attribute it is necessary to take into account peculiarities of both spatial and
statistical distributions of values. In search for a suitable representation of the
statistical distribution we studied the properties of the cumulative frequency
curve and found it to be a good solution. This representation allows a two-
way use. On the one hand, one can visually evaluate statistical characteristics
of a given classification. On the other hand, one can produce classifications
with desired statistical characteristics. In our implementation a cumulative
curve display is included in the interface for classification together with direct
manipulation controls and a map showing the results of classification in the
geographical space. In this interface the user can gradually shift class breaks
and immediately observe the effect on the map and on the cumulative curve
display. This dynamic link between the components of the interface allows
the user to evaluate “on the fly” lots of variants of classification from the
perspective of satisfaction of geographical and statistical criteria and to arrive
eventually at a good compromise solution.
In the process of exploration of the properties of the cumulative frequency
curve we came upon an idea that this representation could be extended to
arbitrary attributes that allow summing over a set of objects. Just as the fre-
quency curve accumulates the number of objects, the generalised curve would
accumulate the values of an attribute. The use of such curves in classification
offers additional opportunities. One of them is accounting in classification for
such criteria as even distribution of population or area among classes. An-
other opportunity is investigation of relationships between the attribute used
for classification and various quantitative characteristics of the objects being
classified.
We believe that inclusion of cumulative curves in the tool for classification
available in CommonGIS significantly increases its exploratory value. Thus,
this tool has got a high appraisal of an expert in statistics and statistical
graphics and a professional in analysis of geographic information, also with a
solid statistical background. However, we have a concern that the users with
less expertise in statistics may find cumulative curves difficult to understand.
We plan to perform experiments in order to evaluate how people can han-
dle the classification tool and how much time they need to comprehend the
cumulative curve display and learn to use it.
References

1. Andrienko, G. and Andrienko, N. (1998) Dynamic Categorization for Visual

Study of Spatial Information. Programming and Computer Software, 24 (3),
108–115.
2. Andrienko, G. and Andrienko, N. (1999) Interactive maps for visual data explo-
ration. International Journal Geographical Information Science, 13 (4), 355–374.
3. Bertin, J. (1965/1983) Semiology of graphics. Diagrams, networks, maps. The
University of Wisconsin Press, Madison WI.
4. Egbert, S.L. and Slocum, T.A. (1992) EXPLOREMAP: an exploration sys-
tem for choropleth maps. Annals of the Association of American Geographers,
82, 275–288.
5. MacEachren, A.M. (1994) Some truth with maps: a Primer on symbolization
and design. Association of American Geographers, Washington.
6. Robinson, A.H., Morrison, J.L., Muehrcke, P.C., Jon Kimerling, A., and Guptil,
S.C. (1995) Elements of cartography. Wiley, New York.
7. Slocum, T.A. (1999) Thematic cartography and visualization. Prentice-Hall,
New Jersey.
8. Yamahira, T., Kasahara, Y., and Tsurutani, T. (1985) How map designers can
represent their ideas in thematic maps. The Visual Computer, 1, 174–184.
Combining REmbeddedPostgres and PostGIS

Albrecht Gebhardt

Department of Statistics, University of Klagenfurt, Klagenfurt, Austria

agebhard@uni-klu.ac.at

1 Introduction
This work presents some results of combining several pieces of Open Source
software, including R1 (see [6]), PostgreSQL2 (see [2]), and some appropriate
extension packages, see [3]. A central building block is REmbeddedPostgres,
an extension of the PostgreSQL RDBMS, which basicly delivers the possibility
to use R syntax within SQL queries. Using the work of Duncan Temple Lang
[4] as starting point, the idea arose, to implement more complex statistical
database queries. Finally the focus has been put on queries involving spatial
data. At this point PostGIS comes into action. It is another PostgreSQL exten-
sion and implements OpenGIS functions. This extension enables PostgreSQL
to process spatially referenced data.
An implementation of a “linear model” query will be shown. It involves
several modifications of REmbeddedPostgres and needs some extra SQL func-
tions written in R and Perl. This can easily be combined with OpenGIS SQL
functions. As a result a “spatial statistical SQL query” becomes possible.

2 GIS, RDBMS and Statistical Software

GIS, database systems and statistical software fulfill different tasks in the
analysis of spatial data. GIS are used for collecting and editing data, genera-
tion of maps, transformation of and operations with maps. Database systems
hold data, are used for indexing and selection by queries and can combine dif-
ferent portions of data via joins between tables. Finally the tasks of statistical
software are exploratory analysis, modelling and estimation.
Several combinations of GIS, database systems (DBMS) and statistics
software are possible:
1
http://www.r-project.org/
2
http://www.postgresql.org/
274 A. Gebhardt

• DBMS as database engine for a GIS

• Database interfaces for statistical software or GIS
• GIS interfaces for statistical software
• statistics extensions for GIS
Figure 1 tries to summarize those combinations for R, PostgreSQL and the
Open Source GIS GRASS.3 R can interact with both GRASS and PostgreSQL
by means of appropriate R libraries. GRASS GIS can access data stored in
a PostgreSQL database. PostgreSQL delivers GIS related functions via its
PostGIS extension and it can provide statistical functions via an embedded R
interpreter. During the following sections we will only focus on the relations
between R and PostgreSQL.

OpenGIS PostGIS

PostgreSQL
v. 5.1
GRASS

library(RPgSQL)
library(GRASS)
REmbeddedPostgres

Fig. 1. GRASS, R and PostgreSQL

3 PostgreSQL

PostgreSQL is an open source object relational database server. It is freely

available under a BSD style licence. Its roots reach back to 1977 where the
developement of INGRES started as a research project at Berkeley Univer-
sity, California. In 1986 Michael Stonebreaker started a new project based on
INGRES (which had been converted into a commercial product) and called
it Postgres. Meanwhile it has been redesigned at least two times and changed
its name via Postgres95 to PostgreSQL in 1996. At this time the query lan-
guage of Postgres was changed to SQL. Now PostgreSQL almost completely
implements the ANSI SQL/92 standard. Ordinary queries have the form

3
http://grass.itc.it
 Combining REmbeddedPostgres and PostGIS 275

SELECT x,y,z FROM table WHERE condition;

PostgreSQL is available for most UNIX/Linux systems. Also a native port
for Win32 systems exists. One of the advantages is the extensibility of Post-
greSQL. A first possibility to extend PostgreSQL is the use of its PL/PGSQL
procedural language. This makes it easy to develop new database functions
using the SQL language. But it is also possible to write C code and to make
these C functions available in SQL as a new procedural language. This is the
way how PostGIS and REmbeddedSQL have been implemented.

4 PostGIS

PostGIS is an OpenGIS implementation for PostgreSQL and follows the “Sim-

ple Features Specification for SQL”,4 see [5] . The OpenGIS5 standard defines a
minimum set of geometric data types, operations and functions for a database
to become useable as backend for a GIS. Most commercial databases contain
such OpenGIS extensions, examples are Oracle Spatial, IBM DB2 Spatial
Extender and Informix Spatial DataBlade.
PostGIS implements additional data types like point, line, polygon etc.
It stores geometry information in a so called geometry table. New functions
like “Distance”, “Area2d”, “Box3D”, to mention just a few of them, are now
available. A full list of functions and detailed introduction can be found in
the online documentation6 of PostGIS.
PostGIS functions can be used to extend ordinary SQL queries to “geo-
metric” queries:
SELECT x,y,z FROM table
WHERE Distance(GeomFromText(’POINT(x0 y0))’,
SRID(geopoint)), geopoint)<r;
In this simple example we would select all values x, y, z within a circle (x, y) ∈
B((x0 , y0 ), r) ⊂ R2 .

5 REmbeddedPostgres
REmbeddedPostgres7 is part of the omegahat8 project. The development of
REmbeddedPostgres was driven mainly by the two ideas. Running an R inter-
preter within the database saves much data transmission time because com-
putation takes place at the server and not at a database client. Additionally
4
http://www.opengis.org/techno/specs/99-049.pdf
5
http://www.opengis.org/
6
http://postgis.refractions.net/docs/
7
http://www.omegahat.org/RSPostgres/
8
http://www.omegahat.org
276 A. Gebhardt

database users could easily adopt to a statistical language if it is following the

SQL syntax. Reference [4] gives a technical overview and discusses examples
like prediction using an imported linear model.
REmbeddedPostgres implements both record and aggregate functions
(also referred to as PL/R functions). Record functions can be used to apply
simple R scripts to several arguments. Aggregate functions are more complex.
They determine the result iterative by presenting the data sequentially to an
“update function” which saves its results in a state variable.

6 Combination

6.1 Simple Queries

After installing both components into a PostgreSQL server it is easy to build

simple “spatial statistical” SQL queries. E.g. if the mean of some variable of
interest within some circular region is to be retrieved one could use
SELECT mean(z) FROM table
WHERE Distance(GeomFromText(’POINT (x0 y0))’,
SRID(geopoint)), geopoint)<r;
Other examples can be created easily by combining different R functions for
univariate statistics with different OpenGIS geometry functions.

6.2 More Advanced Functions

It would also be desirable to have the possibility to execute more complex

queries like
SELECT lm(’z~x+y’) FROM table
WHERE Distance(GeomFromText(’POINT(x0 y0))’,
SRID(geopoint)), geopoint)<r;
Because SQL aggregate functions can only be applied to a single argument we
have to introduce new data types which combine the arguments x, y, z into a
vector (x, y, z) . If we restrict the linear model to multiple linear regression,
we can implement this using user defined float8 based vector types. In this
case it is necessary that REmbeddedPostgres can handle these new data types
and has access to appropriate type conversion routines.

7 Technical Details
The idea is to extend REmbeddedPostgres in the following way to be able to
handle user defined data types:
 Combining REmbeddedPostgres and PostGIS 277

• Create user defined types (needs input and output functions written in C
for PostgreSQL internal use).
• Write type conversion routines similar to the conversion routines for float8
already contained in REmbeddedPostgres.
• Introduce an additional table repg_utypes which contains the type OID,
type name, the names of to/from R converter functions and pathname of
a shared object file containing the shared library code which implements
the converter functions.
It is also necessary to add a “user type registration function” to the R initializa-
tion in REmbeddedPostgres which reads the table repg_utypes, dynamically
loads the shared library, accesses the converter routines (via the dlopen call)
and adds registration info to the converter routines structure.
A first prototype of this approach is working. It reads the table repg_utypes
within the C code by connecting back to the data base engine using the libpq
C interface.
Current development focuses on improving the above mentioned REmbed-
dedPostgres extension. This makes use of PL/PERL as another PostgreSQL
extension module, which implements an internal interface to the Perl9 script-
ing language. Combining PL/R and PL/PERL can help in simplifying the
notation of more complex PL/R queries.
The following problem arises. User data types have to be of a fixed length.
That means different types for several vector dimensions together with several
converting functions have to be written. This leads to a more complex syntax
of the PL/R queries, because calls to the type conversion routines have to
be added. Using PL/PERL can help here to hide these details from the user.
Perl scripts analyze the arguments and construct the PL/R queries. The ap-
propriate vectorized query can than be executed by using the DBD::PgSPI10
Perl module for internal database access from PL/PERL (see [1]).
Currently we can apply e.g. R’s linear model function lm() to a spatial
subset selected by means of OpenGIS functions in two steps. First creating
a SQL view and then applying the PL/PERL function lm to this view. The
PL/PERL function performs parsing of the linear model formula given in
S notation and then calls appropriate type conversion routines and finally
executes the PL/R aggregate function containing the call to the R function
lm().
The result of the following query would be the estimated parameter vector
θ of a linear model z = θ0 + θ1 x + θ2 y + ε.
CREATE VIEW spatial_view AS
SELECT X(geopoint),Y(geopoint),z FROM table
WHERE Distance(GeomFromText(’POINT (x0 y0))’,
SRID(geopoint)), geopoint)<r;
SELECT lm(’z~x+y’,’spatial_view’);
9
http://www.perl.org/
10
http://jamesthornton.com/postgres/7.3/postgres/plperl-database.html
278 A. Gebhardt

This application is mainly based on the open programming interface of Post-

greSQL, which made it possible that several PostgreSQL extensions exist:
an OpenGIS implementation, an embedded R interpreter within the database
and a Perl interface. The combination of these building blocks made it possible
to implement some “spatial statistical” functions directly within the database
server by using synergy effects. The fact that all parts consist of freely avail-
able software is also very important.
References

1. A. Descartes and T. Bunce (2000) Programming the Perl DBI. O’Reilly &
Associates, Inc.
2. J. Hartwig (2001) PostgreSQL Professionell und praxisnah. Addison-Wesley.
3. K. Hörhan (2002) Integration von Statistiksoftware in Datenbanksystemen und
deren Anwendung in der räumlichen Statistik. Thesis, University of Klagenfurt
4. D. T. Lang (2001) Scenarios for using R within a relational database manage-
ment system server. Technical report, Bell Labs.
5. Open GIS Consortium Inc. The OpenGIS abstract specification (1999). http:
//www.opengis.org/techno/abstract.htm
6. R Development Core Team (2004) R: A language and environment for sta-
tistical computing. R Foundation for Statistical Computing, Vienna, Austria.
3-900051-07-0.
Index

Acute Coronary Syndrome, 249–258 Environmental data, 57–63, 194, 233,

Anisotropy 237–238
geometric, 31, 35 Environmental mapping, 58
ANNEX models, 57–63 Epidemiology, 194, 235–236, 241, 242,
ArcGIS, 12, 164, 219, 250 249–258
External drift, 3, 5, 12, 57–63
Bayesian hierarchical model, 196, 197,
198, 240, 241, 245 Fractional Brownian motion, 137, 147
Bayesian Kriging, 30 Fuzzy model, 153–159
Bayesian prediction, 33
Generalised additive model (GAM),
240, 256
Change of support, 5, 12
Generalised linear model (GLM),
CommonGIS, 262, 263, 264, 265, 266,
240–241
267, 269, 270
Geodetic network, 173, 174, 184,
Conditional simulation, 17, 26, 65, 124,
185, 186
126, 209
Geomatics, 249–258
Contrasts, 10–12, 13, 17, 18, 26, 107
Geostatistical Analyst, 12, 164, 205, 219
Covariance function, 17, 19, 20, 21, 22,
Geostatistical model
23, 24, 29, 30, 31, 32, 33, 34, 39,
linear, 3, 4
51, 54, 55, 68
Geostatistics, 3, 4, 5, 12, 17, 22, 26, 29,
Cross correlation, 8–10, 214
45, 58, 59, 105, 106, 107, 145, 162,
Cross validation, 36, 214
204, 206, 212, 242
Gibbs distribution, 17
Data bases GIS, 4, 13, 93, 153, 155, 157, 159, 161,
relational, 229–230 201–280
Disease mapping, 235, 236, 240, 243, GRASS, 132, 219, 274
244, 245 GSLIB, 12
Drift estimation, 110 Gstat package, 4, 12
GWR model, 252
Empirical Bayesian approach, 33
Energy functional, 17, 18–20, 22, 23, Indicator simulation, 65, 69–70, 211
24, 26 Indicator transform, 66, 67, 208, 209
Entrepreneurship data, 189–198 Indicator variography, 67–68, 75
282 Index

Information systems Public health, 65, 215, 235,

epidemiological, 235–245 249, 256

Kriging Random field (r.f.)

Bayesian, 30 Gaussian r.f., 21, 30, 31, 34
external drift, 3, 5 non-Gaussian r.f, 17–26
indicator, 65, 68, 70, 208–211 spartan r.f., 17–26
ordinary, 4, 32, 35, 106, 107, 108, 110, RDBMS, 201–280
124, 145, 209 Regional classification, 65–76
universal, 3, 124, 209 RembeddedPostgres, 273–278
Risk assessment, 161, 162, 236, 243, 244
Land use, 91–101, 223, 224, 225,
226, 233 Singular value decomposition, 173, 178,
Least squares, 4, 11, 29, 34, 103, 110, 182–183, 184, 185
173–186 Soil polygon, 153–159
Linear regression, 3, 8, 11, 12, 276 Spatial association, 212–214, 280
Link function, 195, 196, 251 Spatial correlation, 3, 5, 8–10, 12,
58, 60, 114, 137–151, 163, 164,
Monte Carlo Markov Chain (MCMC), 165, 169
198, 208, 241, 242, 245, 257 Spatial modeling, 94, 189–198
Moran’s I, 212, 213, 214 Spatial prediction, 5, 10, 33, 45–55, 57,
Multivariate prediction, 12, 58 58, 195–196, 212
Spatial sampling, 93–95
Neural networks, 35, 57–63 Spatio-temporal prediction, 10
Non-Gaussian models, 17–26 Spectral density, 18, 19, 21, 22
Non-stationary data, 137–140 Splancs package, 217, 219
Splines, 35, 45–55, 123
OpenGIS, 273, 274, 275, 276, Standardised mortality ratio, 240
277, 278 Support vector machines, 35, 65–76

PL/PERL, 277 Temporal cross correlation, 8–10

PL/R, 276, 277 Tracer data, 137–151
Poisson regression, 197, 251, 252, 254 Traffic data, 105, 107, 109, 110, 113–117
PostGIS, 273–278 Trans-Gaussian Kriging, 29–41
Postgres, 273–278
PostgreSQL, 273, 274–275, 276, Variational method, 23–24, 26
277, 278 Variogram
Precipitation data, 122, 125, 128, anisotropic, 12, 109, 110
131, 132 relative, 139, 140, 145, 147, 149, 150
Predictive distribution
posterior, 35, 37, 38 Wavelets, 79–87