THE DETECTION OF CHANGE

IN SPATIAL PROCESSES

WITH

ENVIRONMENTAL APPLICATIONS

ELAINE B. MARTIN B .Sc. B .Sc.

A Dissertation Submitted To The

University of Glasgow

For The Degree of Doctor of Philosophy

© Elaine B. Martin, October 1992

 ProQuest Number: 13834195

All rights reserved

INFORMATION TO ALL USERS

The quality of this reproduction is d e p e n d e n t upon the quality of the copy subm itted.

In the unlikely e v e n t that the a u thor did not send a c o m p le te m anuscript

and there are missing pages, these will be noted. Also, if m aterial had to be rem oved,
a n o te will ind ica te the deletion.

uest
ProQuest 13834195

Published by ProQuest LLC(2019). C opyright of the Dissertation is held by the Author.

All rights reserved.

This work is protected against unauthorized copying under Title 17, United States C o d e
M icroform Edition © ProQuest LLC.

ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, Ml 4 8 1 0 6 - 1346
GLASGOW f
UNIVERSITY
LIBRARY
 ABSTRACT

Ever since Hailey (1686) superimposed onto a map of land forms, the direction of trade
winds and monsoons between and near the tropics and attempted to assign them a
physical cause, homo-sapiens has attempted to develop procedures which quantify the
level of change in a spatial process, or assess the relationship between associated
spatially measured variables.

Most spatial data, whether it be originally point, linear or areal in nature, can be
converted by a suitable procedure into a continuous form and plotted as an isarithmic
map i.e. points of equal height are joined. Once in that form it may be regarded as a
statistical surface in which height varies over area in much the same way as the terrain
varies on topographic maps. Particularly in environmental statistics, the underlying shape
of the surface is unknown, and hence the use of non-parametric techniques is wholly
appropriate. For most applications, the location of data points is beyond the control of
the map-maker hence the analyst must cope with irregularly spaced data points. A variety
of possible techniques for describing a surface are given in chapter two, with attention
focusing on the methodology surrounding kernel density estimation.

Once a surface has been produced to describe a set of data, a decision concerning the
number of contours and how they should be selected has to be taken. When comparing
two sets of data, it is imperative that the contours selected are chosen using the same
criteria. A data based procedure is developed in chapter three which ensures
comparability of the surfaces and hence spurious conclusions are not reached as a result
of inconsistencies between surfaces. Contained within this chapter is a discussion of
issues which relate to other aspects of how a contour should be drawn to minimise the
potential for inaccuracies in the surface fitting methodology.

Chapter four focuses on a whole wealth of techniques which are currently available for
comparing surfaces. These range from the simplest method of overlaying two maps and
visually comparing them to more involved techniques which require intensive numerical
computation. It is the formalisation of the former of these techniques which forms the
basis of the methodology developed in the following two chapters to discern whether
change/association has materialised between variables.
One means of quantifying change between two surfaces, represented as a contoured
surface, is in terms of the transformation which would be required for the two surfaces to
be matched. Mathematically, transformations are described in terms of rotation,
translation and scalar change.

Chapter five provides a geometrical interpretation of the three transformations in terms

of area, perimeter, orientation and the centre of gravity of the contour of interest and
their associated properties. Although grid resolution is fundamentally a secondary level
of smoothing, this aspect of surface fitting has generally been ignored. However to
ensure consistency across surfaces, it is necessary to decide firstly, whether data sets of
different sizes should be depicted using different mesh resolutions and secondly, how
fine a resolution provides optimal results, both in terms of execution time and inherent
surface variability. This aspect is examined with particular reference to the geometric
descriptors used to quantify change.

The question of random noise contained within a measurement process has been ignored
in the analysis to this point. However in practice, some form of noise will always be
contained within a process. Quantifying the level of noise attributable to a process can
prove difficult since the scientist may be over optimistic in his evaluation of the noise
level. In developing a suitable set of test statistics, four situations were examined, firstly
when no noise was present and then for three levels of noise, the upper bounds of which
were 5,15 and 25%.

Based on these statistics, a series of hypothesis tests were developed to look at the
question of change for individual contour levels i.e. local analysis, or alternatively for a
whole surface by combining the statistics and effectively performing a multivariate test.
A number of problems are associated with the methodology. These difficulties are
discussed and various remedial measures are proposed.

The theoretical derivation of the test statistic, both in the absence and presence of
random noise, has proved mathematically to be extremely complex, with a number of
stringent assumptions required to enable the theoretical distribution to be derived. A
major simulation study was subsequently undertaken to develop the empirical probability
distribution function for the various statistics defining change for the four levels of noise.
Also for each of the statistics, the resultant power of the test was examined.
The remaining chapter explicitly examines two case studies and how the methodology
developed in the preceding two chapters may be implemented. The first example cited
raises the question, 'Has a seasonal temperature change resulted during the fifty year
span, 1930 to 1980, within the contiguous United States of America?' The data base was
provided by the United States Historical Climatology Network (HCN) Serial
Temperature and Precipitation Data, Quinlan et al (1987).

The second problem examines whether there is an association between background

radiation levels, within three regions of the south-west England, and the location of
various forms of leukaemia or whether case location is a product of the population
distribution. Differences between this example and the previous illustration materialise in
terms of the spatial resolution of the data; the leukaemia data are defined as punctual data
points and are extremely sparse; the population distribution is defined as areal regions;
with the radiation data being of a more continuous format. The methodology developed
required modification, but aside of this a preliminary set of conclusions were reached.
 ACKNOWLEDGEMENTS

The author would like to express her sincere gratitude to her supervisor, Dr. E.M. Scott,
for her encouragement and guidance throughout the period of her research. In addition,
the author would like to thank the members of the statistics department at Glasgow
University for their support and friendship during this time.

The author is indebted to Dr, D. Sanderson for the many stimulating discussions, the
opportunity to partake in the collection of aerial survey radiation data and finally, for
allowing her to make use of previously collected radiation data.

The author would also like to thank Dr. F. Alexander and the Leukaemia Research Fund
for providing the case/control data for the leukaemia study.

This research has been supported by the Science and Engineering Research Council.

Finally, the author would like to express her heartfelt gratitude to her family for their
tolerance and understanding throughout the years.
 vi

Table of Contents

Title page i
Abstract ii
Acknowledgem ents v
Table of Contents vi
List of Figures xii
List of Tables xviii

Chapter 1 Spatial Data Analysis - Its Application

1.1 IntroductionStatistics and Spatial Data 1

1.2 Surface Mapping 3
1.3 Surface Representation 6
1.4 Surface Comparison 9

Chapter 2 Surface Mapping

2.1 Interpolation methods 13

2.2 Moving averages 15
2.3 Kriging 17
2.3.1 Semi-variogram 17
2.3.2 Various forms of kriging 19
2.4 Trend Surface Analysis 23
2.5 Density Estimation 27
2.5.1 Histogram based estimators 28
2.5.2 Kernel estimators 28
2.5.3 Fixed kernel estimators 29
2.5.3.1 Smoothing parameter selection 30
2.5.3.1.1 Likelihood 32
cross-validation
2.5.3.1.2 Least squares 32
cross-validation
2.5.3.1.3 Test graph method 34
2.5.3.1.4 Bootstrap method 35
 yn

2.5.3.1.5 Estimating 'h' from a 35

standard distribution
2.5.3.1.6 Summary 36
2.5.4Alternative kernel estimators 37
2.5.4.1 Nearest neighbour method 37
2.5.4.2 Adaptive kernel estimator 38
2,6 Summary 38

Chapter 3 Surface Representation

3.1 Contour Selection 40

3.2 Contour Notation 40
3.3 Number of Class Intervals 42
3.4 Selection of Class intervals 42
3.4.1 Idiographic group 43
3.4.1.1 Natural break methods 43
3.4.1.2 Percentiles 43
3.4.1.3 Nested means class limits 44
3.4.2 Serial class 44
3.4.2.1 Normal percentiles 45
3.4.2.2 Standard deviation 45
3.4.2.3 Equal arithmetic intervals 45
3.4.2.4 Geometric progression 46
3.4.3 Illustration of four contouring techniques 46
3.5 A Semi-automatic Contour Selection Procedure 49
3.6 Locating and Tracing The Contour 63
3.6.1 Contouring by gridding 63
3.6.2 Contouring by triangulation 65
3.7 Alternative Contouring Techniques 67
3.8 Accuracy of Statistical Maps 68
3.9 Summary 71

Chapter 4 Different Methods of Surface Comparison

4.1 Introduction 73
4.2 Subjective Analysis 73
 V lll

4.3 Global Analysis 77

4.3.1 Lorenz curve 77
4.3.2 Coefficient of areal correspondance 80
4.3.3 Correlation coefficient 82
4.3.4 Comparison analysis of trend maps 83
4.3.5 Difference maps 85
4.3.6 Pattern of differences 87
4.4 Local Analysis 87
4.4.1 Complexity index 87
4.4.2 Image registration 89
4.4.3 Image restoration 90
4.4.4 Shape change 92
4.4.4.1 Outline data 93
4.4.4.2 Landmark data 95
4.5 Summary 98

Chapter 5 The Characterisation of Spatial Change

5.1 Introduction 100

5.2 Various Measures For Describing A Contour 103
Transformation Parameters
5.2.1 Scale 103
5.2.1.1 Area 104
5.2.2.2 Perimeter 105
5.2.2 Orientation 105
5.2.3 Centre of gravity 108
5.3 Variability of The Contour Descriptor 109
5.3.1 Simulation study 110
5.3.2 Simulation results 115
5.3.2.1 Area 115
5.3.2.2 Perimeter 119
5.3.2.3 Orientation 120
5.3.2.4 Centroid displacement 122
5.3.2.5 Smoothing parameter 122
5.3.2.6 Summary 123
5.4 Simple Statistics for Describing Change 124
 5.4.1 Scalar comparators 125
5.4.1.1 Ratio 125
5.4.1.2 Differences 126
5.4.1.3 Proportion 126
5.4.1.4 Scalar summary 127
5.4.2 Rotation 128
5.4.2.1 Ratio 128
5.4.2.2 Differences 128
5.4.3 Translation descriptors 129
5.4.3.1 Centroid displacement 129
5.4.3.2 Standard displacement 131
5.4.3.3 Percentage displacement 131
5.4.3.4 Standard deviation displacement 132
5.4.3.5 Overlap 133
5.4.3.6 Standardised overlap 134
5.4.3.7 Coefficient of areal 134
correspondance
5.4.3.8 Translation Summary 134
5.4.4 Summary 135

Chapter 6 An Hypothesis Testing Approach to Surface Comparison

6.1 Introduction 136

6.2 Simulation Study 138
6.2.1 Scalar change 139
6.2.2 Orientation 139
6.2.3 Centroid displacement 140
6.2.4 Overlap 140
6.2.5 Summary 143
6.3 Simulation Results 145
6.3.1. Distribution of the test statistic under 145
the null hypothesis
6.3.1.1 Area 145
6.3.1.2 Perimeter 146
6.3.1.3 Orientation 147
6.3.1.4 Centroid displacement 148
 X

6.3.1.5 Overlap 149

6.3.2 Distribution under the alternative hypothesis 150
6.3.3 Power of the tests 152
6.3.3.1 Area 153
6.3.3.2 Perimeter 154
6.3.3.3 Orientation 155
6.3.3.4 Centroid displacement 155
6.3.3.5 Overlap 156
6.3.4 Summary of no random noise 157
6.4 Random noise 159
6.4.1 Distribution of the test statistic 159
under the null hypothesis
6.4.1.1 Area 159
6.4.1.2 Perimeter 161
6.4.1.3 Orientation 163
6.4.1.4 Centroid displacement 165
6.4.1.5 Overlap 167
6.4.2 Power curves 169
6.4.2.1 Area 169
6.4.2.2 Perimeter 170
6.4.2.3 Orientation 172
6.4.2.4 Centroid displacement 173
6.4.2.5 Overlap 174
6.4.2.6 Summary 176
6.5 Multivariate Hypothesis Testing Procedure 177
6.6 Limitations of The Technique 179
6.7 Advantages of The Technique 181

Chapter 7 The Application Of The M ethodology to Two

Environmental Case Studies

7.1 Why The Studies Were Selected 182

7.2 Case Study 1 The Investigation of 183
Climatic Change
7.3 United States Historical Climatology Network 185
7.4 Subjective Impression 186
 xi

7.4.1 Summary of subjective analysis 194

7.5 Local Hypothesis Testing Procedure 195
7.5.1 Summary of local results 200
7.6 Global analysis 201
7.6.1 Existing techniques 201
7.6.1.1 Correlation analysis 202
7.6.1.2 Paired t-interval 203
7.6.1.3 Trend surface 285
analysis
7.6.1.4 Regression analysis 204
7.6.1.5 Summary of the existing 206
techniques
7.6.2 Global hypothesis testing approach 207
7.7 Conclusions 208
7.8 Case Study 2 The Investigation of a Possible 210
Link Between Leukaemia and the Underlying Radiation Fields.
7.9 Description Of The Problem and How It differs 212
To The Preceeding Example.
7.10 Pilot study 213
7.10.1 Radiation data 214
7.10.2 Population data 215
7.10.3 Epidemiological data 219
7.10.4 Data summary 220
7.11 Analysis 221
7.11.1 Introduction 221
7.11.2 Analysis of case locations 221
and population data
7.11.3 Analysis of radiation fields 227
and case locations
7.11.4 Analysis based on leukaemia rate 234
7.12 Summary 238
7.13 General conclusions 239
References 240
Appendix 258
 X ll

List of Figures

Chapter 1 Spatial Data Analysis - Its Application

1.1 Relationship between point data and plot types

Chapter 2 Surface Mapping

2.1 Example of an hypothetical semi-variogram 18

2.2 Surfaces produced for varying levels 30
of smoothing parameter

Chapter 3 Surface Representation

3.1 Definition of various 41

contour forms
3.2 Surface contoured using the 47
equal range technique
3.3 Surface contoured using the nested-means 47
methodology
3.4 Surface contoured using the percentile 47
methodology
3.5 Surface contoured using the 48
standard deviation technique
3.6 Histogram of the % number of points 50
for four contour levels
3.7 Histogram of the % number of points 51
for five contour levels
3.8 Histogram of the % number of points 52
for six contour levels
3.9 Histogram of the % number of points 53
for seven contour levels
3.10 Histogram of the % number of 54
pointsfor eight contour levels
3.11 Histogram of the % number of points 55
for nine contour levels
 xni

3.12 Histogram of the % number of 56

points for ten contour levels
3.13 Graphical representation of the two 57
rules for selecting number of contours and
standard deviation proportion
3.14 Histograms describing the results 59
for four contours and the relevant
standard deviation proportions
3.15 Histograms describing the results 59
for five contours and the relevant
standard deviation proportions
3.16 Histograms describing the results 60
for six contours and the relevant
standard deviation proportions
3.17 Histograms describing the results 61
for seven contours and the relevant
standard deviation proportions
3.18 Histograms describing the results 61
for eight contours and the relevant
standard deviation proportions
3.19 Histograms describing the results 62
for nine contours and the relevant
standard deviation proportions
3.20 Histograms describing the results 62
for ten contours and the relevant
standard deviation proportions
3.21 Illustration of contour degeneracy 65
3.22 Illustration of contouring by triangulation 66

Chapter 4 Different Methods of Surface Comparison

4.1 Various forms of subjective analysis 74

4.2 Lorenz curve for civil parish data within 79
the south-west of England
4.3 Commonly measured shape descriptors 93
Chapter 5 The Charaterisation of Spatial Change

5.1 Illustration of the three transformations 101

5.2 Relationship between area and 103
perimeter for simple constructs
5.3 Evaluation of the area of a polygon 104
based on the trapezia rule
5.4 Definition of the angle of orientation 106
5.5 Problem concerning the theoretical and 107
empirical definition of principal axes
5.6 Definition of orientation bounds 107
5.7 Illustration of the use of the
bootstrap 114
5.8 Area results for coefficient 117
of variation
5.9 Perimeter results for coefficient 118
of variation
5.10 Variation in perimeter for a 120
given area
5.11 Coefficient of variation results for 121
orientation
5.12 Coefficient of variation results for 122
centroid displacement
5.13 Results for the coefficient of 123
variation for the smoothing parameter

Chapter 6 An Hypothesis Testing Approach to Surface Comparison

6.1 Scalar increase/decrease 139

6.2 Definition of angles and areas 141
used to evaluate overlap
6.3 Values of angular displacement, 0, 142
to achieve required level of overlap
for thecoefficient of areal correspondance
 xv

6.4 Empirical probability distribution 145

for areal change under the null hypothesis
of no change
6.5 Empirical probability distribution 146
for perimeter change under the null
hypothesis of no change
6.6 Empirical probability distribution 147
for orientation change under the null
hypothesis of no change
6.7 Empirical probability distribution 148
for centroid displacement under
the null hypothesis of no change
6.8 Empirical probability distribution 150
for overlap under the null hypothesis
of no change
6.9 Sensitivity of the scalar metrics 153
6.10 Power curve for area with 154
no random noise
6.11 Power curve for perimeter with 155
no random noise
6.12 Power curve for orientation with 155
no random noise
6.13 Power curve for centroid displacement 156
with no random noise
6.14 Power curve for overlap 156
proportion with no random noise
6.15 Empirical probability distribution 160
for areal change under the null hypothesis
of no change
6.16 Empirical probability distribution 162
for perimeter ratio under the null hypothesis
of no change
6.17 Empirical probability distribution 164
for orientation change under the null
hypothesis of no change
 6.18 Empirical probability distribution 166
for centroid displacement under the
null hypothesis of no change
6.19 Empirical probability distribution 168
for overlap under the null hypothesis
of no change
6.20 Power curves for area in presence 170
of random noise
6.21 Power curves for perimeter in 171
presence of random noise
6.22 Power curves for orientation in 172
presence of random noise
6.23 Power curves for centroid 174
displacement in presenc of random noise
6.24 Power curves for overlap in presence 175
of random noise
6.25 Problematical contour shapes 180

Chapter 7 The Application Of The M ethodology To Two Environmental

Case Studies

7.1 Box-and-whisker difference plots 187

(1980-1930)
7.2 Bivariate plots of temperature for the 188
contiguous United States of America
7.3 Spatial subjective analysis for spring 189
7.4 Spatial subjective analysis for summer 191
7.5 Spatial subjective analysis for autumn 192
7.6 Spatial subjective analysis for winter 193
7.7 Residual map for spring 205
7.8 Location of the three grids 213
7.9 Age distribution for the three grids 217
7.10 Lorenz curves for the three diagnostic 222
categories of interest for grid 1
 XVII

7.11 Comparison of civil parish population 224

and number of cases for various
disease types
7.12 Location of leukaemia cases in relation to 226
population surface
7.13 Histogram of cases and controls for 229
grid 1, for the three diagnostic
categories of interest
7.14 Cumulative distribution function for the 231
three disease categories
7.15 40k radiation surface with cases 233
superimposed
 List of Tables

Chapter 1 Spatial Data Analysis - Its Application

1.1 Spatial analysis techniques 2

1.2 Various combinations of 5
map types
1.3 Various specialised forms 7
of isolines/isarithms
1.4 Types of comparisons involving 9
the four basic data types

Chapter 2 Surface Mapping

2.1 Summary of various 14

interpolation techniques
2.2 Situations in which the 19
various types of kriging are applicable

Chapter 3 Surface Representation

3.1 Brooks and Carruthers (1953) 42

contour selection procedure
3.2 Contour levels for the various 47
methods
3.3 Summary of the standard deviation 58
proportions and number of contours
3.4 Accuracy of evaluation of a 71
unit circle, for x-data points

Chapter 4 Different Methods of Surface Comparison

4.1 Dissimilarity levels for the 79

Lorenz curve
4.2 A 2 x 2 resemblance matrix 81
 xix

4.3 Elementary measures for 94

measuring the shape of
geographical areas

Chapter 5 The Characterisation of Spatial Change

5.1 Range of smoothed 115

bootstrap parameters investigated
5.2 Results for the coefficient 116
of variation for area
5.3 Results for the coefficient 119
of variation for perimeter
5.4 Coefficient of variation in 122
the estimated smoothing parameter, h
5.5 Operators based on area and 125
perimeter for performing a
scalar comparison
5.6 Relationship between scalar 127
comparators for two surfaces
5.7 Descriptors for the various 130
forms of centroid displacement

Chapter 6 An Hypothesis Testing Approach to Surface Comparison

6.1 Summary of descriptors 137

6.2 Transformations to point P(xp,yp) 144
6.3 Percentage points for the 146
empirical probability distribution
function relating to areal change
6.4 Percentage points for the empirical 147
probability distribution function
relating to perimeter change
6.5 Percentage points for the 148
empirical probability distribution
function relating to the test statistic
describing orientation
 XX

6.6 Percentage points for the 149

empirical probability distribution
function describing centroid displacement
6.7 Percentage points for the empirical 150
probability distribution function
relating to the test statistic describing overlap
6.8 Various forms of the alternative hypothesis 151
6.9 Percentage points for areal change 161
under the null hypothesis
6.10 Percentage points for perimeter 163
change under the null hypothesis
6.11 Percentage points for orientation 165
change under the null hypothesis
6.12 Percentage points for centroid 167
displacement under the null hypothesis
6.13 Percentage points for overlap 169
under the null hypothesis
6.14 Correlations for the various 177
test statistics for all levels of noise

Chapter 7 The Application Of The M ethodology To Two Environmental

Case Studies

7.1 Global mean changes in five 184

carbon dioxide doubling studies
7.2 Summary statistics for the four seasons 187
7.3 Observed values of test statistics 195
describing the various forms
of change for the contours of interest
7.4 Results of local hypothesis testing procedure 198
7.5 Results for correlation based analysis 202
7.6 Results for paired t-interval 203
7.7 Results derived for the correlation 204
between the two trend surfaces
7.8 Results for regression slope analysis 205
7.9 Results of global hypothesis
testing procedure
7.10 Summary of breakdown of total radiation
dose received by the
population of Thurso, Dionian (1986)
7.11 L.R.F.'s diagnostic breakdown of
the various forms of leukaemia
7.12 Breakdown of leukaemia cases
into diagnostic groups
7.13 Comparison of the sizes of
population areas for three regions
7.14 Summary of the spatial resolution and
quantity of the three data sources
7.15 Results for various dissimilarity
indices for Lorenz curves
7.16 Sample sizes to detect change
of specific size
7.17 Summary statistics for grid one
7.18 Results of Kolmogorov-
Smirnov test
7.19 Example of leukaemia rates
for grid one
7.20 Results from performing an analysis
of variance on grid (1,2,3), radiation
level (1-8) and leukaemia type (1-4),
7.21 Difference in rates between the
three leukaemia types for Cornwall
 1

CHAPTER 1

SPATIAL DATA ANALYSTS - ITS APPLICATION

§1.1 Introduction Statistics and Spatial Data

The first manifestation of statistics for spatial data appears to have arisen in the form of
data maps, Hailey (1686) superimposed onto a map of land forms the direction of trade
winds and monsoons between and near the tropics, and attempted to assign them a
physical cause.

Spatial models appeared much later. Student (1907) examined the distribution of
particles throughout a liquid and instead of analysing their spatial positions, he
aggregated the data into counts of particles per unit area and found they followed a
Poisson distribution. Fisher (1935) during the 1920's and 1930‘s established the
principles of randomisation, blocking and replication. Although these control for bias
and randomisation neutralises the effect of spatial correlation, they do not neutralise the
spatial correlation at spatial scales larger or smaller than the plot dimension.

More recently nearest neighbour techniques for analysing agricultural field trials have
attempted to take spatial dependence into account by using residuals from neighbouring
plots as covariates or by differencing, Besag and Kempton (1986).

In areas such as geology, ecology and environmental science, it is not often possible (or
appropriate) to randomise, block and replicate the data. There is therefore a need for new
statistical methods and approaches that address new questions arising from old and new
techniques. Many of the resulting problems such as resource assessment, environmental
monitoring and medical imaging are spatial in nature.

The development of statistical procedures to quantify the level of change in an

environmental process or assess the relationship between associated spatially measured
variables is an area of considerable statistical and practical interest. Much of the analysis
of spatially collected data is concerned with the identification and explanation of spatial
structure, and with the analysis and explanation of possible links between two spatial
processes e.g. humans and the environment.
 2

The form of a spatial analysis undertaken is in part conditioned by the dimensionality of

the data, Unwin (1981) recognised four separate classes; point, line, area and surface.
Within each category techniques both descriptive and analytical exist, table 1.1 collates
some of these methods.

Points Lines
mean centre/standard distance random walk
standard deviational ellipse vectors
gradient analysis graph theory
nearest neighbour analysis nodality
variance/mean ratio test connectivity
quadrant analysis flow analysis
space clustering dispersion

Area Surfaces
location quotients isolines
space clustering trend surface analysis
poisson probability power series polynomials
hierarchial clustering fourier series

Table 1.1 Spatial analysis techniques

Methodological problems stem from the fact that spatial patterns are frequently intricate
and complex, and to date many investigators have grossly simplified these patterns.

An additional problem of applying statistics to spatial data concerns the problem of scale.
Often spatial patterns or variable associations show striking differences at different scales
of analysis. It can be argued that scale should be used creatively, Cleek (1979). For one
thing, replication of findings at different scales tends to confirm hypothesis.
Furthermore, White (1972) stated that one can try to identify the scale at which a certain
process is most effective: this in itself may provide clues as to how the process works. As
in all forms of analysis the reliability and validity of the data will potentially limit the
value of sophisticated statistical models and techniques. Methods which account for
random error in a spatial variable may help to extend the feasibility and power of the
technique.
 3

In different areas of application, various specialised difficulties will arise, each separate
analysis must take account of these localised difficulties e.g. in medical geography, the
spatial position of an individual at the time of diagnosis is taken to be the current place of
residence. For chronic diseases such as cancer, which have a long latent period, the place
of diagnosis may have little aetological significance.

All the problems discussed are confined to individual sets of spatial data. Effectively
performing a comparison of spatial variables increases the number of potential pitfalls.
The development of a method which invokes only automatic procedures will reduce the
possibility of attaining spurious conclusions regarding associations/differences.

The first step in developing such a framework is to convert the data, of whatever
dimensionality, to a continuous form, by one of the available interpolation techniques.
Once in this format it may then be presented as a contoured surface, the question of
contour selection and fitting is an essential stage of the analysis since incompatible
selection may result in inaccurate statements.

Once both these stages have been successfully implemented, the question of comparison
may be examined. Intuitively, the most appealing means of comparison is to superimpose
the surfaces and describe the change in terms of various geometric properties of the
contour. Based on this notion, a series of statistical tests were constructed to look at
change/association between variables.

§1.2 Surface Mapping

The definition of a map is given to be a representation of (part of) the earth's surface,
showing physical and political features. In mathematics, a mapping is described as a
translation from one vector space to another. For the practitioner, the first vector space is
the real multidimensional world and the second is the face of a sheet of paper on which
the map is drawn, From this definition, a map and surface are one and the same since a
surface is described as having length and breadth but no thickness. The paper
accordingly restricts a map to a three dimensional vector space where the length and
breadth represent the spatial co-ordinates and the z-variable, the phenomenon being
mapped, is defined in terms of lines or symbols.
 4

Representing spatial information in terms of surfaces/maps has four advantages:-

1. They provide a synoptic expression of the information represented.

2. They enable a number of important spatial properties that were not initially measured
to be isolated e.g. orientation, shape and relative location. Maps/surfaces therefore
contain spatial structure.

3. They are models of the real world.

4. Finally, they are communication devices which may be used to express ideas or
generate hypothesis.

However, maps/surfaces contain three theoretical difficulties as models of the real

world:-

1. They are static and cannot be drawn to include a time dimension.

2. They are the result of a spatial process which theoretically may be defined as:-

Zj = f (Xi,yj)

The z-values above result from a deterministic process but in practice they are the result
of a stochastic process in which chance variation, £*, is unquantifiable:-

zi =f(xi,y,)+e1

3. Maps/surfaces are constrained by humans and may be used to display information

inadequately or alternatively the data used in their construction may be of poor
quality.

The level of measurement of a spatially distributed variable is a basic control on the

choice of map type, method of analysis and ultimately on the nature of the inferences
that can be drawn from a study of a variable’s statistical structure.
 5

Stevens (1946) identified four basic levels; nominal, ordinal, interval and ratio. The
properties of which are discussed by Siegel (1956). Nominal data serve merely as
symbols and cannot be manipulated mathematically, the simplest form are dichotomous
variables e.g. yes/no. Statistical and cartographic operations are limited, but a number of
primary operations may be executed. The second form of data, ordinal, assumes
relationships are both transitive and asymmetric. Statistically the remaining two levels,
ratio and interval, may be treated in a similar manner since both allow the distance
between categories to be defined for equivalents units of measurement.

Apart from the hierarchical property of level of data, measurement has the characteristic
of dimensionality. Non-dimensional variables play an important role where comparisons
between sets of numbers are to be performed. Dimensional analysis is more restrictive in
practice but is possibly more valuable in applied work.

In theory there is almost no limit to the choice of units of measurement and dimensions
of data which may fall into any one of the four categories. However locational data is
restricted to one of four types of geographical data; point, line, area or surface. Table 1.2
summarises the various types of maps which may be produced using such information,
Unwin (1981).

Dim ension

L evel Point Line Area Surface

Nom inal Dot maps Network map Coloured area Freely coloured
map map
Ordinal Symbol map Ordered Ordered Ordered
network map coloured map chorochromatic
map
Interval and Graduated Flow map Choropleth Contour type
Ratio symbol map map map

Table 1.2 : Various combinations of map types

 6

Nowadays a surface is a starting point of an analysis. The next stage is to summarise and
describe the distribution prior to examining whether or not a recognisable pattern exists.
For each of the twelve possible map forms, various types of analysis can be undertaken.
Unwin (1981) examines some of these procedures. Dimension is the strongest single
influence on the type of statistical inference which may be undertaken.

A further feature which determines the type of analysis which is appropriate relates to
whether or not the data is isotropic or anisotropic. The former describes a data set in
which the properties are the same in any direction whilst for the later, these vary with
direction. Analysis of an anisotropic surface is more complex since the direction of
variability is not always known.

Particularly in environmental statistics, the underlying shape of the surface is unknown

hence the use of non-parametric techniques is wholly appropriate. For most applications,
the location of data points is beyond the control of the map-maker and does not
correspond to a geometric pattern. In these cases, the analyst must cope with irregularly
spaced data points. The mapped surface need not necessarily pass through the points
exactly hence the irregular location of points is not a complication. Other techniques
necessarily pass through all the points. A more detailed discussion of possible techniques
for describing a surface are given in chapter two with attention focusing on the
methodology surrounding kernel density estimation.

Density estimation provides a natural and intuitive means for describing the data, and the
technique does not differ vastly from some of the more traditional approaches for
representing spatial data such as nearest-neighbour and kriging. It is this technique which
has been used in the development of surfaces for subsequent analysis.

§1.3 Surface Representation

Viewing locational data as a statistical surface Robinson (1961) suggested that much of
morphometric analysis may be applicable to the ’topography’ of isarithmic surfaces.
Although isarithmic maps are three-dimensional in the sense that a distribution varies
continuously over an area, they are commonly referred to as single component maps
since only one component, the z-variable, is being mapped. The most common method
of showing variation in the z-dimension is through the use of isolines or perspective
plots. A wide variety of terms have been used to describe surfaces of this form, the most
general being, isoline and isarithm. Table 1.3 summarises some of the more specialised
terminology.

Isoline Quantity Isoline Quantity

represented represented

Contour Equal heights Isotherms Equal temperatures

Isovels Equal speed Isohevts Equal rainfall

Isonoets Equal intelligence Isohels Equal sunshine

Isotachs Equal wind speed Isobar Equal pressure
Isophers Equal freight rate Isochrone Equal time/distance

Table 1.3 Various specialised forms of isolines/isarithms.

Clearly such illustrative procedures are based on the assumption that the distribution is
continuous. Wamtz (1959) demonstrated that considerable theoretical and operational
gains were to be had by regarding such variations as continuous.

Two types of isolines exist, the first refers to the more conventional form where the
surface exists at all points. The second variant describes maps in which the isolines
display spatial variation in derived quantities, that are themselves related to some area
e.g. population density. These are more commonly referred to as isopleths. Problems
associated with such maps are, unlike isolines, the spaces between the isopleths do not
have any values related to them. Secondly the form of the isoline is dependent on

1. The shape of the areal units.

2. The location of the control points

In the 1950’s a number of authors advocated their use particularly Mackay (1951),
Schmid and MacCannell (1955), and Porter (1958) whilst more recently Nordbeck and
Rystedt (1970) and Tobler and Lau (1970) have also supported their application. There
are however situations where it is prudent to retain the data in their discontinuous form.
The non-overlapping cells are mapped as choropleth maps with appropriate values
assigned to each cell. Analysis should then proceed by area-based methods.
 8

Figure 1.1 summarises the relationship between measured point data, networks and
contours in surface mapping.

Measured regular
Measured randomly
grid point data
located point data

\s
Regular grid network Triangular network

Perspective plots
Contour plots

Figure 1.1 Relationship between point data and plot types.

A perspective plot takes a point, P, in three-dimensional space and transforms it to a

point in two-dimensions, Q, its two-dimensional image. The general isometric projection
is the simplest type. To evaluate Q from P, P is multiplied by;-

rcosa -cosP 0^
sin a sinp 1
, 0 0 0,

The standard isometric projection sets a - p = 30°.

Once an isarithmic surface has been produced to describe a set of data, a decision
concerning the number of contours and how they should be selected has to be taken.
When comparing two sets of data, it is imperative that the contours selected are chosen
using the same criteria. A data based procedure is developed in chapter three which
ensures comparability of the surfaces of interest, and hence avoids spurious conclusions
being reached, as a result of inconsistencies between surfaces. Contained within this
chapter are other issues which relate to how a contour should be drawn to minimise the
potential for inaccuracies in the surface fitting methodology.
 9

§1.4 Surface Comparison

A great deal of interest has focused on the detection of common or differing features
between two or more distinct variables in various fields of application e.g. image
processing, geology, medical geography etc. Invariably we are attempting to make some
objective comments concerning the existence of an ’areal' association between variables,
or alternatively, trying to discern the type of change which may have resulted due to
some, possibly unknown, underlying process.

However caution should be displayed when interpreting the findings. Where a similarity
is seen to exist between two variables, a causal link may not necessarily hold. Similar
forms can be generated by markedly different processes and only seldom will any single
analysis be able to discriminate between them. Although this observation suggests that
this mode of analysis is of little value, in many situations it is an important step to
identifying the possible processes which may be responsible for the change.

Surface comparison is further complicated by the surfaces themselves, like must be

compared with like, or the association may be a by-product of the technique and not the
processes hence the emphasis on surface construction in chapter three.

The number of types of comparisons which can be made is enormous. Restricting

ourselves to the four basic types of data, Table 1.4 shows that there are at least ten
different types of comparisons.

DISTRIBUTION A DISTRIBlUTION B
point line area surface
point 1 2 3 4
line 5 6 7
area 8 9
surface 10

Table 1.4 Types of comparison involving the four basic data types.

Comparisons between similar distributions is the most obvious e.g. area distribution of
woodland and soil type (8). Nevertheless comparisons of distributions of differing spatial
 10

dimensions may be undertaken e.g. point location of leukaemia sufferers and the spatial
dispersion of a specific radionuclide (4) but potentially different types of analysis will be
required. (This problem is dealt with more fully in Chapter 7).

Before attempting to compare surfaces, the analyst must first define their objectives and
determine what is to be compared. Also, consideration must be given to the difference
between a single-valued comparison function, that gives an overall comparison, and a
spatial comparison or function which produces a map of areal goodness-of-fit that
permits a two-dimensional similarity interpretation. Chapter four focuses on a whole
wealth of techniques, currently available. These range from the simplest method of
overlaying two maps and visually comparing them, to more involved techniques which
require intensive numerical computation. It is the formalisation of the former of these
techniques, which forms the basis of the methodology developed in chapters five and six
to discern whether change/association has materialised between variables, or over time.

One means of quantifying change between two surfaces, represented as contoured

surfaces, is in terms of the effective transformation which would be required for the two
surfaces to be matched. The most usual mathematical definition of transformation is in
terms of quantities relating to rotation, translation and scalar changes.

This approach has previously been used to collate two sets of images i.e. image
registration. In both medical and satellite imagery, these transformation measures are
generally considered to be nuisance parameters and the data is accordingly transformed
to remove these factors. The technique for isolating these change parameters is based on
the movement of easily identifiable points on a surface, landmark data. These points are
not anticipated to deviate in locality over time or between variables. Once the two
surfaces have been 'matched1, change is invariably described in a subjective rather than
quantitative manner, Coombes et al. (1991).

In environmental applications, two problems are associated with such an approach, firstly
it is unusual to have data of this format, since in field studies it is extremely difficult to
return to the same site where the initial measurement was recorded after time, t,
primarily because of practical difficulties and secondly, the type of transformation
causing the change is of considerable importance to the earth scientist since it may give
them a handle on the underlying process contributing to the change.
 11

Both the validity and the properties of the four measures area, perimeter, orientation and
centre of gravity are assessed in chapter five as to their feasibility for forming the basis
of measures used to quantify the three transformations, rotation, translation and scalar
change. The question of grid resolution is examined specifically in terms of these
descriptors. This aspect of surface fitting acts as a secondary level of smoothing, the finer
the mesh, the more smooth the resultant picture. In terms of a comparison it is essential
that compatibility across surfaces is achieved. First, we must decide whether mesh
resolution should vary with data set size and secondly what an optimal resolution is in
terms of time and resultant surface variability.

The question of random noise contained within a measurement process has been ignored
in the analysis to this point, in practice, some form of noise will always be contained
within a process. Quantifying the level of noise attributable to a process can prove
difficult since the scientist may be over optimistic in their evaluation of the noise level.
In developing a suitable set of test statistics, four situations were examined, firstly when
no noise was present and then for three levels of noise, the upper bounds of which were
5,15 and 25%.

Based on these statistics, a series of hypothesis tests were developed, both to assess
change at a local level i.e. change between individual contour levels, and on a global
scale. A number of problems are discussed and various remedial measures are proposed.

The theoretical derivation of the test statistic, both in the absence and presence of
random noise, has proved mathematically, extremely complex. A number of stringent
assumptions would be required to enable the theoretical distribution to be derived. A
major simulation study was subsequently undertaken to develop the empirical probability
distribution function for each of the various statistics defining change for the four leveis
of noise. Also for each of the statistics, the resultant power of the test was examined.

The remaining chapter explicitly examines two data sets. They illustrate how the
methodology developed in the preceding chapters may be transposed to real problems
and the problems which materialise when implementing the technique.

The first data set deals with potentially the simplest situation, the comparison of two sets
of similar measurements recorded at the same locality. The example is concerned with
the topical subject of climate change and the notion of global warming. It examines the
 12

question of whether there has been a change in the seasonal temperature during the 50
year period 1930 to 1980 within the contiguous United States of America. The data base
was provided by the United States Historical Climatology Network (HCN) Serial
Temperature and Precipitation Data, Quinlan et al (1987).

The second example deals with a more complex situation in which the spatial resolution
of the variables of interest differ, the questions of interest are therefore formulated in a
different way and the methodology was modified accordingly. The problem examines
whether there is an association between the level of background radiation, within three
separate regions of the south-west of England, and the location of various forms of
leukaemia, or whether the cases are purely a product of the population distribution.
Differences between this example and the previous illustration, materialise primarily in
terms of the composition of the data; the leukaemia data is defined in terms of punctual
data points and is extremely sparse in quantity, the population distribution is defined in
terms of areal regions i.e. civil parishes and finally, the radiation data is of a more
continuous format.

The latter example highlights the difficulties surrounding analysis of data varying in
spatial resolution and quantity, but it also emphasis the potential applicability of the
method to problems where the spatial resolution does differ between variables. Although
the sparsity of the leukaemia data serves as a restriction to any formal analysis,
alternative quantitative methods and subjective techniques founded on the methodology
were implemented to answer the question of interest.
 13

CHAPTER 2

SURFACE MAPPING

§2.1 Interpolation Methods

Most locational data, whether it be originally point, linear or areal in nature, can be
converted into a continuous form and plotted as a contour map by one of the many
available interpolation methods. Once in that form, it may be regarded as a statistical
surface in which height varies over area, in much the same way, as terrain varies on a
topographic map.

If an investigator has control over the location of data points, they would be arranged in
a regular lattice or grid with uniform spacing between points. In this case interpolation
could be accomplished by fitting a hyperbolic paraboloid to every four data points by
double linear interpolation, Switzer et al. (1964), fitting a polynomial to, up to 25
surrounding data points, by Newton’s divided difference formula, Steffensen (1927),
Berezin and Zhidkov (1965), or bicubic spline interpolation, de Boor (1962).

For most applications, the location of data points is beyond the control of the map maker
and does not correspond to a geometric pattern. In these cases, the analysist must cope
with irregularly spaced data points. Sometimes the surface to be mapped need not fit the
data points exactly, i.e, the mapped surface need not attain the value of the dependent
variable specified at the data point. Subsequently the irregular location of data points will
not present a serious complication. An example of such is a trend surface (i.e. a
polynomial of specified order) which can be fitted to the data by regression, Krumbein
(1959).

For the surface to fit the data points exactly, Bengtsson and Nordbeck (1964) grouped
data points into triangles or quadrilaterals. Shepard (1968 a,b) developed an interpolation
algorithm, SYMAP, which produced a smooth, continuously differentiable surface which
passed through all data points, the method being based on the idea of moving averages.
 14

Table 2.1, although not exhaustive, lists three groups of interpolation techniques with
examples of each type:-

Non-parametric density Trend surfaces Distance related

estimation
Naive estimator Polynomial Moving averages
Kernel density estimation Fourier series Moving median
Nearest neighbour Spatial filtering Varying Quantile Method
(balloon density)
Adaptive kernel method Linear programming Kriging
Maximum penalised method Newton's statistical prediction
technique
Orthogonal Series estimator The polygonal method
Reflection and Replication Thiessen polygon
Techniques
Transformation technique

Table 2.1 Summary of various interpolation techniques.

A resume of four of the above techniques is presented focusing on the theory but also
incorporating a number of practical illustrations

1. Moving averages
2. Kriging
3. Polynomial trend surface
4. Kernel density estimation

The first three methods are more commonly associated with geological mapping,
however more recently the ideas of density estimation has been used to produce surfaces
for analysing geographical data, especially in terms of disease mapping and clustering,
Bithell (1990).
 15

§2.2 M oving Averages

In the interpolation or girding of spatial data, one of the simplest methods is that of
moving averages. A ’true value’, Zp} at point P is estimated by taking the average of all
surrounding points within a certain limiting radius. This average, Zp, is a linear estimator
with weights 1/n on the n points lying within the radius and zero on all other
observations. Clarke (1979) showed the resulting estimate of Zp^ does not necessarily
minimise the sum of squared deviation where no attempt has been made to optimise
these weights.

The reason why this estimator is poor is because the value at the centre of the search
circle is estimated from observations that lie between the two extremes, i.e. from
observations which lie close to the limiting distance and also, from those points which lie
close to the centre of the circle, all the points are equally weighted. A modification of
this approach that corrects for this failing is to allocate weights, dependent on their
distances, from the critical point P, by using a monotone decreasing function e.g. an
inverse weighting function i.e.

where z; = value of the i^1 dependent variable to be mapped

p. = distance from P to data point i

If the point P is very near data point i, then pj is small, the further away we move, the
larger p. becomes, so the weight accordingly decreases.

Therefore as the distance p, -> o the value at P is the limit of:-

 16

Thus for points near data point i, the computed value is near to value z..

It may be shown that by computing such a value z p for every point P over the map area,
a 'smooth* surface is obtained. This method has several shortcomings

1. As the number of data points becomes large this method for calculating Zp becomes
long and inefficient,

2. Direction from data points to the point P is not considered giving interpolated
values that are implausible in some situations. Under this simple method, the
computed value at P depends only on the distances to the data points 1,2, n but
not on their location relative to P. For example, assuming pj and p2 are held fixed
in length, the following configuration of data points give identical values at P:-

+ + + + + +
1 P 2 P 2 1

3. In practice we can alter the weighting. In the above example of inverse distance
weighting, use was made of the actual distance pt, more generally, we could weight
using p.b, in which the exponent b is set to some value other than unity. Values
greater than this will decrease the relative effect of distant points whilst, for values
less than one, the importance of distant points will be increased. Many commercial
computer programs use this method with b=2, giving an inverse distance squared
weighting. With inverse square weighting, the surface is level at every data point
(i.e. the derivative is zero) regardless of the location of that data point or its data
value.

A number of embellishments were made to this technique to compensate for the above
deficiencies, Shepherd (1968 a,b).

This idea is extended further in the description of kriging which takes into account both
the variation in the measurements and distances between the measurement points
 17

§2.3 Kriging

Many geological surfaces, both real and conceptual, can be regarded as regionalised
variables which have properties intermediate between a truly random variable and one
completely deterministic. Typically, regionalised variables are functions describing
natural phenomena that have geographic distributions such as the elevation of the ground
surface, or changes in grade within an ore body. Unlike random variables, regionalised
variables have continuity from point to point but the changes in the variable are so
complex that they cannot be described by any tractable deterministic function.

Even though a regionalised variable is spatially continuous, it is not usually possible to

know its value everywhere. Instead its values are known only through samples taken at
specific locations.

Geostatistics was originally developed by Georges Matheron of the Centre de

Morphologic Mathematique in Fontainebleau, France. It involves estimating the form of
a regionalised variable in one, two and three dimensions. One of the basic statistical
measures of geostatistics is the semi-variance which is used to express the ratio of change
of a regionalised variable along a specific orientation i.e. the semi-variance is a measure
of the degree of spatial dependence between samples along a specific support.

§2.3.1 Semi-variogram

For simplicity assuming the samples are point measurements and the support is regular,
the semi-variance, yh, is defined as:-

1 / \2
Yh=^ 5 (z(xi)~ z(Xi+h))
where
z(xj) = measurement of a regionalised variable taken at location i.
z(x. + h) = measurement taken h intervals away.
n(h) = number of pairs of samples at distance h apart.

If we evaluate the semi-variances for different values of h, the results may be plotted in
the form of a semi-variogram. For small h, the points tend to be similar and the semi­
 18

variance will be small, increasing h, the points being compared are less and less closely
related and their distances become larger, resulting in a large value of yh. At some
distance the points being compared are so far apart that they are unrelated and then-
squared difference becomes equal in magnitude to the variance around the average value.
The semi-variogram consequently develops a flat region, the sill. The distance at which
the sill is attained is termed the span or range. Figure 2.1 illustrates a hypothetical semi-
variogram. The nugget effect i.e. the value of the semi-variogram at h=0, corresponds to
the mean squared sampling and assaying error.

..-.Sill____

A
Nugg'et <£ . Span, .y
Effect
*
*
h

Figure 2 . 1 Example of an hypothetical semi-variogram.

In principle the experimental semi-variogram is known only at discrete points

representing distances h, whilst in practice the semi-variance may be required for any
distance. For this reason the discrete experimental semi-variogram is modelled by a
continuous function that can be evaluated for any desired distance.

Fitting a model equation to an experimental semi-variogram is a trial-and-error process,

usually done by eye. Clark (1979) describes and gives examples of the manual process
whilst Olea (1977) provides a program that computes a linear semi-variogram having the
same slope at the origin as the experimental semi-variogram, he illustrates the technique
on a set of geological data.

If the regionalised variable has been sampled at a sufficient density relative to the range,
there will be no significant difference between estimates assuming a linear semi-
variogram model and other semi-variogram models.
 19

When the form of the semi-variogram is known, it is possible to estimate the value of the
surface at any unsampled location, this estimation procedure is termed kriging. The term
kriging is derived from the name of D. G. Krige, a South African mining geologist and
statistician, who first introduced the idea to avoid systematic overestimation of reserves
in the field of mining.

Unlike convential contouring algorithms, this method provides a measure of the error or
uncertainty of the contoured surface. Kriging uses the information from the semi-
variogram to find an optimal set of weights which are then used to estimate the surface at
unsampled locations.

§2.3.2 Various forms of kriging

Six types of kriging are

1. Punctual
2. Block
3. Lognormal
4. Disjunctive
5. Universal
6. Generalised covariances

Table 2.2 describes the situation where each of the forms is most applicable.

Distribution Stationarity
of data
Stationary Simple drift Local trends Severe
anistropy
Normal Simple kriging Universal Generalised ?
(point or block) kriging covariance
Simple Lognormal ?
»
? ?
•

(known kriging
distribution)
Complex Disjunctive ? ?
«
?
kriging
Table 2.2 Situations in which the various types of kriging are applicable.
 20

Punctual kriging is the simplest form of kriging. The data consists of measurements
taken at dimensionless points and the estimates are made at other dimensionless points.

For simplicity, we will assume the variable to be mapped is statistically stationary i.e.
free from drift, for a regionalised variable z(x)

E[z(x)] = m constant
E[z(x) - m][z(x) - m] = cov(x - x )

i.e. the covariance between two points x and x' does not depend separately on x and x',
but only on the vector (x-x).

The value at an unsampled location may be estimated as a weighted average of the

known observations, with the resultant value at point P given by

A n n
XP - XwiXi where ^ w; = 1
i= i i= i

A
The estimate Xp will differ from the true, unknown value Xp by an amount termed the
estimate error, ep

ep= (xp-xAp)

There are an infinite number of possible weight combinations that can be selected, each
gives a different estimate and estimation error, however only one combination produces
a minimum estimation error, it is this unique combination of weights that kriging aims to
find,

A complete derivation of the kriging equations is provided by Olea (1975) for punctual
kriging, optimum values for the weights are found by solving a set of simultaneous
equations which include values from a semi-variogram of the variable being estimated.
The weights are evaluated such that the resulting estimates are unbiased and have
minimum estimated variance.
 21

For n data points we have (n+1) equations with only n unknowns. Since we have more
equations than unknowns, the additional degree of freedom may be used to ensure the
solution has the minimum possible estimation error. This is accomplished by including a
Lagrange multiplier, X. The complete set of equations have the following form in matrix
notation: -

y fa i) Y (h J y (0 - y (h j i' wr yK )
y (h j y(h22) y(h23) “ ■ y(h2n) i w2 y(h2p)
. 1
y (0 y(Kz) y(hn3) y (0 i Wn y(hnp)
1 1 1 ••• 1 0 _X_ i

In general we solve the equation

[A][W] = [B]

for a vector of unknown coefficients [W]. The terms in matrix [A] and vector [B] are
taken directly from the semi-variogram or the mathematical function describing it. Once
the unknown weights have been determined, the variable at location P is given by:-

x P= X w iXi
i=i

For block kriging the same form of equations as given in (1) are valid, but the y values
are replaced by y. y is the semi-variogram which expresses the variance between the
observation and the block to be estimated.

Where the distribution of the variable of interest is not even approximately normal but
strongly skewed, it is possible to fit a lognormal distribution. Krige (1978) considers this
method of kriging. The data are first transformed using an appropriate transformation

y, =l°g(xi +a)
 22

where a is an arbitrary constant to optimise the fit to a normal distribution. These values
are then used to compute the semi-variogram and generate the ordinary kriging
estimates. Link and Koch (1975) have shown various problems are associated with this
type of kriging.

Theoretically, the best possible estimator of xp is some function of the data x1,x2,...,xn
and is defined to be the conditional expectation of xp given the n observations. In order
to obtain this estimator we need to know the precise distribution of the (n+1) variables.
For the simplest case of a normal distribution and stationarity, the conditional
expectation can be defined and is identical with the best linear estimator.

In disjunctive kriging, a 'best-fit* appropriate transformation is used, consisting of a set of

Hermite polynomial functions, Journal and Huijbregts (1978). By using such
transformations, it is possible to convert the data into a form which will approximate to a
univariate normal distribution for the x[ values and bivariate normal distribution for
every pair of values xj, xj. The class of possible functions is restricted to those which are
linear combinations of univariate functions of the data. The estimator obtained is non­
linear and many of the desirable properties of simple linear kriging are discarded e.g.
unbiasedness.

A study of the sea-floor by Journal (1969) provoked Matheron's interest in the problems
of non-stationarity and the development of universal kriging. The approach parallels that
of trend surfaces where the phenomena is split into two components, a deterministic
trend (or drift) plus a random error (or fluctuation), section 2.4. The difference between
the two approaches is that the fluctuations used in geostatistics are not assumed to be
independent, as in trend surfaces. Instead of assuming the drift is a constant, it is allowed
to follow a trend which may be expressed in terms of a simple polynomial.

Although the idea is a step away from the rarely adhered to requirement of stationarity,
major problems exist with this technique, Armstrong (1984). These are primarily
associated with obtaining a valid estimation of the semi-variogram and secondly the
indeterminacy of the drift.

Finally generalised covariances rely on the theoretical fact that lowest order differences
are the strongest forms of stationarity. These differences are termed generalised
 23

increments. The concept of generalised increments leads directly to that of generalised

covariances, using this notion the kriging equations may be rewritten in terms of
generalised covariances. The result of this is to relax the stationarity requirements but the
cost is measured in terms of the complexity of the theory and the resultant computations.

Generally the ideas of geostatistics have advantages over other more conventional
techniques for analysing spatial data. They have a sound theoretical basis, they allow
some estimate of the quality of estimation procedures and finally they have some claim
to statistical properties such as unbiasedness, linearity and minimum variance. However
they require some fairly stringent assumptions to be made which in practice can rarely be
adhered to.

Regionalised variable theory is not valid under conditions where its defined form of
stationarity does not hold. However, although unproved, most users of the method
suggest that departures from stationarity are not of practical significance since local
stationarity may be assumed.

A second assumption is that of normality. The majority of data sets are not strictly
normal, transforming them to achieve normality and then applying the kriging equations
results in the estimate being sub-optimal, non-linear and biased.

Finally the major problem lies with the corner-stone of parametric geostatistics, the semi-
variogram. Whatever data set is used the semi-variogram will often depart significantly
from all the theoretical models, skilful interpretation is required to fit one or more
models to the empirical curve.

§2.4 Trend Surface Analysis

Trend surface analysis was first drawn to the attention of earth scientists in the mid
1950's by Oldham and Sutherland (1955), Miller (1956), Krumbein (1956), Grant (1957)
and Whitten (1957). These authors used the techniques for the analysis of gravity maps,
stratigraphic maps, isopach maps and maps representing specific attributes of
sedimentary and igneous rocks, respectively.

In the course of time, the number of applications has increased significantly and the
method has been refined and generalised. More recent studies in which trend surface
 24

analysis was used as a primary tool for arriving at conclusions include Anderson (1970),
Bradley (1970) and Haining (1987).

A trend is any large scale systematic change that extends smoothly and predictably from
one map-edge to the other. Examples of such systematic trends might be the dome of
atmospheric pollution over a city, the steady rise in cirque floor levels noted across many
mountain ranges and so on. The detection and separation of such trends is the object of
trend surface analysis.

Once again thinking of a surface mapping as a scalar field:-

Zi=f(Xi,yi)

A trend surface simply specifies a precise mathematical form for the function, f, and fits
it to the observed data by least squares regression.

In practice no simple function will fit the observed data. Firstly, even when the
underlying surface is simple, measurement errors will have been introduced into the
observed data. Secondly, in practice it is exceedingly unlikely that only one trend-
producing process will be in operation. It follows that as well as trend represented by a
simple mathematical function of the x and y co-ordinates, there will also be local
departure from this i.e. residuals:-

zi = f K yi)+E i

The main problem in trend surface analysis is to decide upon a particular function for the
trend part of the function. Although there is an enormous range of possible functions, the
simplest trend surface is an inclined plane i.e. a linear trend surface:-

z i = P o o + P io x i + fW ! + ei £i ' N(O,02) (2)

For the general case of a polynomial of degree m:-

z. = i l M y " + e. (3)
j=0 k =0
 25

In specific applications, p + q < m where m denotes the degree of the trend surface. A
trend surface of degree m has r = (m+l)(m + 2) coefficients pjk These are calculable
only if the number of observations satisfies the condition n > m . Where n=m a special
type of surface is fitted with no residuals.

Returning to the simplest case of a linear trend surface, (2), the constants have a simple
physical interpretation, the first /200 represents the height of the plane surface at the map
origin where (x.,ys) = (0,0). The second /310 is the surface slope in the x-direction and
/?01 gives its slope in the y-direction. For interpretation of quadratic and cubic
coefficients see Cliff et al, (1975).

The theory of least squares estimation is used to fit a surface to the data i.e. the sum of
the squares of the residuals at all data points is minimised for all possible surfaces of that
degree, determining uniquely the coefficients, Pjk, in equation (3) i.e.

z(u) = f(u)Tp + e(u)

4
1

"l
0 -T -»

c r —*

yi
X

y? •
1
where S=f (u) = x2 y2 4 yi ■
P-

_1 yn 4 yI ■
Thus P satisfies:-

P = (STS)“1STZ. (4)

In general equation (4) is valid, however (sTs) may in practice be singular hence its
inverse does not exist. Various methods have been developed to circumvent this
particular problem e.g. Whitten (1970) developed a method using orthogonal
polynomials for irregularly spaced data.
Various degrees of polynomial for a trend surface have been cited in the literature e.g.
Krumbein (1959) illustrated the use of both a linear and quadratic surface to describe the
trend in the lithologic composition (classic ratio) of Pennsylvanian rocks in parts of
Kansas and Oklahoma. Whitten (1970) employed octic surfaces in his description of the
sub-surface elevation on top of the Devonian Dundee Limestone in a 900 square-mile
area of Central Michigan. Coons et al. (1967) examined the Mid-Continent gravity high
to assess the relation between basement and Paleoxoic structures using surfaces of up to a
complex 15^ order polynomial.

The method of trend surfaces works quite well with surfaces that are of a simple form
but it is very sensitive to data point distribution and aberrant values. Gray (1972) and
Robinson (1972) illustrate the validity of such a statement, whilst Davies (1973)
demonstrates the pattern of the data is unimportant as long as the area of interest is
covered, the clustering of points within an area does not influence the final surface. A
further problem is that the number of data points should be restricted in order that rather
extensive interpolations are based on relatively few data.

A further cautionary note to the usage of this technique is that care should be taken in not
relying too heavily on the statistical significance tests which have been developed to
decide the degree of a trend surface (analysis of variance), or to check residuals for
’outliers’ (t-tests), or on confidence belts which can be calculated for the trend surfaces.
These statistical tests produce exact results only if the residuals are stochastically
independent. In reality the residuals from a trend surface are frequently subject to
significant spatial auto-correlations. In that situation, the test statistics based on the
model of uncorrelated residuals may be severely biased, Agterberg (1964), Watson
(1971).

Various other techniques have been advocated to eliminate this problem. Instead of
postulating a trend-residual model, a trend-signal-noise model may be defined. The trend
signal denotes a continuous random function with noise describing a stochastically
independent component. Haining (1987) considers alternative ways of fitting trend-
surface models which have been modified to include a random process model to
accommodate spatial autocorrelation. Haining considered two cases, firstly where the
order of the trend was assumed known and secondly when the order is unknown. For the
former the method of parameter estimation used was the Cochrane-Orcutt time series
procedure, Ripley (1981). In case of the order of trend being unknown, a first order
model is estimated and the residuals are tested for autocorrelation using the generalised
Moran coefficient (GMC), Ripley (1981) for autocorrelation. If autocorrelation is
present, an iterative scheme is performed to estimate the model parameters, otherwise
they may be estimated using ordinary least squares. This procedure is repeated for a
second order model. If none of the second order coefficients are significant or, the
increase in is not significant, the first-order model is retained; otherwise the third-
order model is estimated and the procedure repeated.

Trend surface analysis is probably most useful when trend and residuals can be linked to
separate spatial processes which were operative on a regional and local scale.

§2.5 Density Estimation

Density estimation is possibly the most important topic in applied statistics since unless

the density f (x) is known its characteristics must be inferred from a sample x1,x2,.,.,x n.
Two possible approaches to density estimation are parametric and non-parametric. The
former assumes the data are drawn from one of the known parametric family of
distributions. The density, f, underlying the data is then estimated by finding estimates of
the unknown parameters from the data and substituting these estimates into the formula
for the function, f.

The second approach, non-parametric, is less rigid in its assumptions concerning the
distribution of the observed data. The data determines an estimate of the probability
density, f. Density estimation was first proposed by Fix and Hodges (1951) as a method
for relaxing the rigid distributional assumptions inherent to discriminant analysis.

For the remainder of the section, attention will focus on the second form of density
estimators, in particular those applicable to multivariate data since the ensuing work is
spatial in nature. Additional problems arise when estimating a density function in a
multi-dimensional setting. These primarily relate to regions of the sample space which
are devoid of observations, this is referred to as the ’empty space phenomena’ by Scott
and Thompson (1983).

A variety of non-parametric probability density estimators have been proposed and

studied since the pioneering work on kernel methods by Rosenblatt (1956) and Parzen
 28

(1962). The types of estimators fall into four categories kernel, histogram-type,
orthogonal series and splines, some overlap does exist. Only the first two forms of
estimators will be discussed, Fryer (1977) reviews all four methods, not only with
reference to their theoretical properties but also to their applicability to real life
problems. Most density estimation research has dealt with the one-dimensional case and
applications to relatively small data sets. Multivariate extensions of histograms, kernel
and nearest neighbour estimators have been studied theoretically and are usually applied
to bivariate data sets.

§2,5,1 Histogram based estimators

The histogram is extremely efficient computationally compared to kernel methods, but

statistically it is quite inefficient. Relative sample sizes required for histograms to have
errors comparable to kernel estimators increases rapidly with increasing sample size and
dimension. Scott (1985a) studied the frequency polygon which is formed by the linear
interpolation of adjacent mid-bin values of a histogram, statistically it has the same order
of efficiency as the kernel estimator but still has the computational efficiency of the
histogram. Problems exist with such an approach, the achievement of the same order of
efficiency requires greater sampling density and secondly, the bins of the optimal
frequency polygon are wider than those for the optimal histogram. With large bins, bin
edge effects become more pronounced.

Scott (1985b) proposed two other variants, the average shifted histogram and a frequency
polygon of the average shifted histogram. With the latter it can be shown to be
functionally identical to a related interpolated kernel estimator of binned data.

§2.5,2 Kernel estimators

Moving away from the idea of mapping based on binned data and its related problems to
the kernel method of density estimation. The definition of a kernel estimator as a sum of
'bumps' centred at the observation is easily generalised to the multivariate case.
Rosenblatt's results for the naive estimator were extended to the bivariate case by Maniya
(1961), with Cacoullos (1964,1966) obtaining the p-dimensional multivariate equivalent
of Parzen’s work, particularly with reference to the evaluation of optimal mean square
error and other invariance properties.
 29

§2.5.3 Fixed kernel estimators

The simplest form of the multivariate kernel density estimator is the fixed kernel. Let
Xi>x2,...,x n be an independent, identically distributed random sample from a
multivariate distribution with density f. The kernel density estimate of f based on these
data is:-

f (x) = n-1h“d2 K{h_1(x- x.)}

i=l

where K { } is the kernel function defined for d-dimensional x satisfying

Jic(x)dx =l
Rd

and h is the smoothing parameter, assumed to tend to zero as n tends to infinity.

For mathematical reasons, the kernel selected is a radially symmetric, unimodal

probability function. Within this family of functions, it has been demonstrated that in
terms of efficiency, the choice of kernel is not critical. The selection should be
conditioned by the intended use to which the methodology is being applied. Various
examples of kernel functions are:-

1. Epanechnikov Kernel

K (x) = { ^ C^^d + 2^<1_- T- )

[ 0 otherwise
where cd = volume of the unit d-dimensional sphere

2. Standard multivariate normal density function

k(x) = (2n)~% exp[-^2XTx]

3. 5 V 1
0 otherwise
 30

4. Kj L ) = ]4 jt-1( l - x Tx)3 ?T? <:1

0 otherwise

The last two apply solely to the two-dimensional situation.

§2.5.3.1 Smoothing parameter selection

The smoothing parameter, or alternatively the window width, is important for its role in
determining the resultant shape of the density estimate.

Figure 2.2 displays pictorially the consequences of varying the smoothing parameter.
The estimates are based on a set of fifty data points simulated from a bivariate normal
distribution using the NAG (1984) random number generator subroutine, G05EAF.

Figure 2.2(a) displays the optimal surface as determined by one of the automatic
procedures, least squares cross-validation, described later. Selecting too small a value for
the smoothing parameter results in a more spiked representation of the surface 2.2(b), by
selecting a larger value of smoothing parameter than deemed 'optimal' by one of the
automatic procedures, results in the density estimate being oversmoothed and hence
causing any detail to be obliterated, figure 2.2(c).

•a -« • i > -J .1 » i t

Figure 2.2 (a)O p tim al procedure for Figure 2.2 (b) Under-estimation of h.
selection of h.
 31

-3
-3 -1 0 1 3

Figure 2.2 (c) Over-estimation of h.

Figure 2.2 Surfaces produced for varying levels of smoothing parameter.

A variety of techniques have been developed to overcome the subjectivity of selecting h.

Where the kernel is purely for illustrative purposes, selecting the window width in this
manner is fairly standard. Five techniques for selecting h automatically or semi-
automatically are suggested in the literature:-

1. Likelihood cross-validation
2. Least squares cross-validation
3. Test-graph method
4. Bootstrap method
5. Selecting h assuming the data arises from a standard distribution

The first two methods use the concept of cross-validation. The idea of cross-validation
was developed by Mosteller and Wallace (1963) for use in discriminant analysis. In the
context of smoothing parameter selection, the cross-validation criterion is the sub­
division of a sample size, n, into a 'construction' subsample, size (n-1), and a validation
subsample, size 1, in all n possible ways. Stone (1974) provides a detailed description of
cross-validation in various areas of application.
 32

§2.5.3.1.1 Likelihood cross-validation

This method as its name suggests uses the ideas of likelihood to assess the adequacy of
fit for different smoothing parameters.

A score function CV(h) is evaluated based on the log-likelihood averaged over each
choice of omitted x{. The value of smoothing parameter for which the function is
optimised is selected as the ’optimal' value;-

c v o o -^ ito g t^ )

Scott and Factor (1981) adjudged the behaviour of CV(h) to be unduly sensitive to
outliers and as a result tended to oversmooth the function. The problem is potentially
more worrisome with multivariate data, due to the difficulties of detecting outliers.

One of the requirements of a statistical procedure is that it should be consistent i.e. good
estimates of the quantity of interest should be obtained if very large samples are
available. If one of the tails is monotonic and dies off at an exponential rate, the use of
likelihood cross-validation leads to inconsistent estimates of the density, Schuster and
Gregory (1981).

Although simple to implement and computationally not excessive in terms of time, the
method is not ideally suited for use in real data sets

§2.5.3.1.2 Least squares cross-validation

The idea suggested by Rudemo (1982) and Bowman (1984) is based on minimising a
suitably well behaved loss function, the integrated squared error:-

I.S.E.= j f f - f f d x = J f 2 - 2 jf f + J f 2
 33

The basic principle of least-squares cross-validation is to construct an estimate of R|^f J

from the data themselves and minimise this estimator over h to give the choice of
window width. Silverman (1986) demonstrates how a score function may be evaluated
from the data to allow this value of h to be calculated. Defining

f -I (x) = (n - 1)~‘h”d£ K{h_1(x - Xj)}

Let M0(h) = J f 2 -2 n " 1^ L J(x.) (4)

i=l

M0 depends on the data. The idea of least squares cross-validation is to minimise the
score M0 over h. Equation 4 is not in a form suitable for computation. Letting K(2) be
the convolution of the kernel with itself. It may be shown that:-

jf ( x ) 2dx = n_2h“d ^ K (2){h-1(xj - xj)}

i j

and

(x, - X j } - ( n - l ) “'h -dK(o)

i i j

The resultant score function is

M,(h) = n V X X K ’I r 'f x j - x j)}+2n-1h-dK(0)

i J

where K*(t) = K2(t)-2K(t)

K2(t) is the convolution of the kernel with itself.

Stone (1984) provides a strong large sample justification of least squares cross-
validation. Stone’s theorem states that the score function M^(h) says asymptotically, just
 34

as much about the optimal smoothing parameter, from the integrated squared error point
of view, as if we actually knew the underlying density, f.

§2.5.3.1.3 Test Graph Method

The third of the methods may be described as semi-automatic. For a set of independent,
identically distributed observations from a d-dimensional density, f, the
estimator function is given by

f«(x) = j*X 'h~ dK{h-'1(x- Xj)}

j= l

The test graph method is then:-

n (x ) = i n-‘h-d-2V2K{h-‘(x- Xj)}
j=l

The window width is chosen to give the best estimate of the density such that the random
fluctuations, in the second derivative, V2, of the estimate, will be asymptotically of
maximum size irSsup|/" |, where 8 is a calculable constant and the value of sup|V2f| can
be estimated from the test graph.

The subjectivity of the method arises since test graphs are drawn of the second derivative
of the estimator for various band widths, to assess which gives the 'right size' of
fluctuations. The problem with this technique is the difficulty in assessing the graphs. In
two-dimensions, contour plots may be used, but in higher dimensions difficulties of
presentation are encountered. The second problem is the time required to produce each
graph. Finally to achieve reasonably accurate results, a fairly sizeable data set is
recommended.

Despite these problems, Silverman (1978) demonstrated the acceptability of the results in
general. In conclusion it is felt that this method is most useful as an aid to checking a
pre-selected window width.
 35

§2.5.3,1.4 Bootstrap method

The selection of bandwidths for density estimation based on the bootstrap, once again
involves the idea of minimising the integrated squared error over various values of the
smoothing parameter, h, Faraway and Jhun (1990).

The integrated mean squared error may be decomposed into a bias and variance
component Direct application of the bootstrap fails since it is incapable of estimating the
bias term in this context. To overcome this problem, an initial estimate of the density,
with the bandwidth chosen by one of the other methods, is calculated and resampling is
subsequently done from this. The bias term is then calculable.

A by-product of this technique is the construction of a confidence band for the

smoothing parameter.

Overall this method is shown to perform comparatively better than cross-validation in

terms of the integrated mean squared error. The bandwidths are generally larger than for
cross-validation resulting in smaller variation. The negative factor of this approach is the
time taken to evaluate the smoothing parameter.

§2.5.3.1.5 Estimating 'h' from a standard distribution

The final method for selecting bandwidths to be discussed is to assume the underlying
density is standard e.g. multivariate normal. Based on the ideas of bias and variance the
window width can be shown to be

h ,,= A (K )n ''>'l M (5)

where depends on the kernel and 0 is the d-

variate normal density. For specific kernels A(K) has been calculated, Silverman (1986).

The window width may then be evaluated directly from (5) if the Fukunaga estimate
(1972) is used
 36

£(x) = (detS— k { it 2(x - X; )TS_1(x - Xj)}

nh £1

where K (xTx) = K (x)

S = covariance matrix

If the kernel is radially symmetric and the data untransformed, a single scale parameter,
g , for the data is evaluated from:-

o*=d-*5>.
i

and the value oh0pt is the resultant window width.

This method gives a 'quick and dirty' estimate for the smoothing parameter simplifying
the task of finding the optimal value for the other techniques.

§2.5,3.1.6 Summary

The importance of selecting the smoothing parameter was summarised in figure 2.2.
However a whole wealth of techniques are available for selecting h. Biasing the
conclusions towards the requirements of the proceeding work, eliminates three of the
methods, firstly the test graph method is not fully automated, secondly estimating the
smoothing parameter from a standard distribution requires the assumption that the data
arose from a standard distribution, the use of a non-parametric mode of analysis was
selected to avoid such a restriction. Finally the use of likelihood cross-validation is
restricted by its sensitivity to outliers, which is especially problematical for multivariate
data.

Finally, of the remaining two methods, the bootstrap technique requires the smoothing
parameter to be selected initially by some existing technique. The overall reduction in
integrated mean squared error over that of the least squares cross-validation method does
not necessarily warrant the additional increase in computational time. In light of this,
least squares cross-validation was used for smoothing parameter selection.
 37

§2.5.4 Alternative Kernel Estimators

Although the kernel estimator can be shown to exhibit a series of desirable properties, in
terms of practicalities the main drawback of the method is its inability to deal
satisfactorily with the tails of the distribution, without oversmoothing the main part. Two
possible adaptive approaches are the nearest neighbour and adaptive kernel methods.

§2.5.4.1 Nearest neighbour method

This approach described by Loftsgaarden and Quesenberry (1965) has also been termed
the balloon density by Tukey and Tukey (1981). The estimator in d-dimensions is
defined in the following manner:-

where rk(t) - Euclidean distance from t to the k^1nearest data point

- d-dimensional volume of the d-dimensional sphere of radius

volume of the unit sphere in d-dimensions

For a sample size, n, one expects nf (t)vk(t) observations to lie within the sphere radius

rk (t), centre t .

The overall estimates obtained by the nearest neighbour approach are not satisfactory for
a number of reasons:-

1. They are prone to local noise.

2. The tails of the estimator are unduly heavy.
3. It is discontinuous and its integral over all space is infinite.
 38

As a technique it is not the most suitable method due to the above limitations. The
adaptive kernel method overcomes these difficulties while still being adaptive to the
local density.

§2.5.4,2 Adaptive kernel estimators

This method offers a combination of the desirable smoothness properties of the Parzen-
type estimators with the data-adaptive characteristics of the k-nearest neighbour
approach.

A kernel is placed at each of the observed data points but the window width is allowed to
vary according to the underlying density. Intuitively the smoothing parameter h(xj)
should vary inversely with the true density, f, i.e. in regions of low density a broader
kernel should be used and vice-versa.

Theoretical work of Abramson (1982) showed that ^f~^(Xi) is a good choice.

The density f~ ^ (x j is unknown but may be replaced with an appropriate estimate. This
estimate makes use of one of the alternative methods, the choice of which is not crucial,
since the adaptive method is insensitive to the fine detail of the pilot study.

Breiman et al. (1977) describe the technique and compare it to the Parzen estimator.
Overall, they found it was superior to the best Parzen estimate.

Silverman (1986) feels the technique has advantages over the ordinary fixed kernel in
some situations, but each case should be treated separately.

§2.6 Summary

Within this chapter four methods of surface fitting have been discussed. Each of the
techniques requires some unknown parameter to be defined:-

1. Moving average - weighting method.

2. Trend surface - order of polynomial.
3. Kriging - form of semi-variogram.
 39

4. Kernel density estimation - smoothing parameter, h.

The unknown parameters are all basically controls on the smoothness of the resultant
surface. Of the four techniques, only the fixed kernel may be defined to be fully
automatic. Currently no automatic selection procedure is available for selecting the
'optimal' order of polynomial or weighting method. In terms of kriging, the model for the
semi-variogram requires considerable skill to fit it to the empirical data.

Potential drawbacks of the fixed kernel as mentioned previously relates to their inability
to deal with long tailed distributions. However this apart, it was decided that since least
human intervention was required it was the most suitable for fitting a surface. However
any of the other methods are feasible and the ensuing work is wholly applicable using the
other techniques.
 40

CHAPTER 3

SURFACE REPRESENTATION

§3,1 Contour Selection

Before proceeding to investigate the various aspects of contour selection and fitting, it
is necessary to describe the two types of question which may arise when comparing
surfaces since they condition the form of contour selection. The most common form of
questions arising are:-

1, Are the underlying statistical distributions of the surfaces similar?

2. For a specific level of contour, does the spatial pattern differ between surfaces?

A subtle difference exists between the selection procedures for the two questions, the
latter does not require a formal selection method since the levels will be the same for
each surface. In terms of the former question, the two surfaces may not necessarily
describe the same variable, therefore a procedure which ensures contour selection is
consistent is essential to avoid reaching spurious conclusions. Motivation for a quasi­
automatic technique which allows the spacing and the number of class intervals to be
selected was prompted by the first of these questions. Later within the chapter the
question of contour accuracy is examined with its role in influencing various geometric
properties of a contour being discussed.

§3.2 Contour Notation

Before proceeding to examine specific topics relating to the description of a surface in

terms of a set of contours, various definitions and notation are introduced.

A closet^ contour is defined to be a closed plane figure bounded by three or more

straight line segments that terminate in pairs at the same number of vertices, and do not
intersect other than at their vertices. The sum of the interior angles is (n-2)*1800>where
n is the number of sides; the sum of the exterior angles is always 360°- The polygon is
 41

approximated by a set of N points, the first is arbitrary but the sequence is delineated in
an anticlockwise direction, figure 3.1 (a).

An open contour, is a contour which has two or more of its ordinates defined to be end­
points on the boundary of the region. For analysis, if the region is defined a-priori, an
open contour is considered equivalent to a closed contour, figure 3.1 (b), although the
conditions of a closed contour are not necessarily satisfied.

A disjoint contour is one which comprises more than one element, these may be either
open or closed, figure 3.1 (c).

Figure 3.1 (a) Closed Figure 3.1 (b) open Figure 3.1 (c) Disjoint
contour. contour contour

Figure 3 . 1 Definition of various contour forms.

A summary of the contour notation used subsequently is given below

Cjj contour j, in surface i, i = 1, Ns , j = 1 , Nc

Ns total number of surfaces

Nc
<c total number of contours

N total number of pointtidescribing contour j ,

A

where Cjj = (xfllJyfll),(x^,yfl2),...,(xaN,yflN)

xjjk k^1x-ordinate of contour j , in surface i

yijk kth y-ordinate of contour j, in surface i
 42

§3.3 Number of Class Intervals

There is no definitive answer to the number of contours which might be used for
constructing a map. A finely contoured map based on a few control points gives an
impression of accuracy unwarranted by the primary information, conversely it is
wasteful to use very few contours when in practice much is known about the inflexion
points in the surface under investigation.

Dobson (1973) stated that in absolute terms the human eye can confidently identify
only a small number of shadings or symbols i.e. between four and ten. This view was
supported by Monmonier (1973), Robinson and Sale (1969) and Jenks and Caspell
(1971).

One guide as to the choice of the number of contours was given by Brooks and
Carruthers (1953) who advocated the number of classes in a histogram should satisfy:-

number of classes < 5*logio (member of observations)

In terms of a contour map based on n observations, the surface should not consist of
more than (51ogio(n)-l) contours, table 3.1.

Number of 50 100 150 200 500

observations
Maximum
number of 7 9 9 10 12
contours

Table 3 . 1 Brooks and Carruthers (1953) contour selection procedure.

§3.4 Selection of Class Intervals

A second question pertaining to the display of spatial data in terms of isolines concerns
the selection of class intervals. Numerous class interval selection methods have been
proposed, Evans (1976) categorised these on the basis of four groups:-
 43

1. Exogenous
2, Arbitrary
3, Idiographic
4. Serial

The first of these groups, exogenous, selects the intervals on some meaningful baseline
level which relates to, but is not derived from the data to be mapped e.g. a critical
population density threshold. As its name suggests, those methods categorised under the
arbitrary heading are founded on a series of numbers of no particular significance.
Usage of this method is indefensible and should never be implemented.

§3.4.1 Idiographic group

More specific to the details of the data set to be mapped, the idiographic group may be
divided into three sub-groups:-

§3.4.1.1 Natural break methods

Two groups fall under this heading multi-modal and multi-step. Intervals for the former
are based on 'natural breaks' in regions where the frequency is low. Jenks and Coulson
(1963) found that clear natural break classes do not occur and subjective judgements
vary greatly between people. It is usually possible to find apparent breaks but these are
often the result of small sample size and their significance should not be exaggerated.

In a similar way the multi-step methodology divides the cumulative distribution into a
series of 'treads' and 'rises’. The same set of problems exist as for multi-modal plus
double the number of contours are required.

Neither of these methods should be utilised in quantitative mapping unless significant

multimodality has been demonstrated statistically,

§.3.4,1.2 Percentiles

The percentile system of subdividing the frequency distribution is very useful in

ensuring equal representation of classes. It is invariant to scalar, rotational and
translational changes. The main disadvantage is that class intervals vary irregularly in
 44

different parts of the measurement scale and between maps of the same variable, for
different areas or times. Furthermore a percentile based map provides no information
on the underlying frequency distribution. It is therefore incumbent that a histogram is
positioned alongside the map with the class boundaries accordingly marked. This
facilitates interpretation in relation to the measurement scale but Schultz (1961) still
advocates very careful reading of the scale. In terms of map comparisons it is not
advisable unless the underlying statistical distribution is uniform.

§3.4.1.3 Nested means class limits

Two classes are formed about the mean with each class then being subdivided at its
own mean, Scriptor (1970). This method balances the desirable properties of equal
numbers per class against that of equal class widths. Class intervals are narrow in the
modal parts of the frequency distribution and broad in the tails. Extreme values
influence the positions of the means of various orders so that the less closely spaced the
values are in a given magnitude range, the broader the class.

For a rectangular frequency distribution, nested means approximate the equal interval
or percentile solutions. For a normal distribution, it will approximate a standard
deviation based method and finally for a 'j* shaped distribution, a geometric
progression, see section 3,4.2.2 and section 3.4.2.4, respectively.

Nested means provide a generally applicable, replicable yet highly inflexible interval
system. They do not form a numerical series independent of the data and do not allow
numbers of classes other than 2m, where m is an integer.

§3.4.2 Serial class

The serial class encompasses all methods with limits which are mathematically derived
and fixed in relation to some statistical descriptor e.g. mean, median, range or standard
deviation. Once again there are a number of techniques specific to this category, many
of which are variants of a theme.
 45

§3.4.2.1 Normal percentiles

This method of standardised class intervals based on the normal probability distribution
was developed by Armstrong (1969). For normal data this method is preferable to true
percentiles in being uninfluenced by minor details of the frequency distribution. Evans
et al. (1975) mention the desirability of transforming the data to achieve normality and
then applying this method. The same set of problems arise with the implementation of
this method as for true percentiles.

§3.4.2.2 Standard deviation

The class width is defined as a proportion, Ks, of the standard deviation. Class intervals
are centred on the mean, which is a class mid-point if the number of classes is odd and
a class boundary if the number is even; the highest and lowest classes are necessarily
open-ended.

This method is particularly suited to data sets which display approximate normality or
are fairly symmetric with a pronounced mode near the mean. The internal intervals are
equal and for skewed unimodal distributions a transformation may be applied to
achieve approximate normality

Use of a standard deviation basis does not imply a class interval of one standard
deviation which is too coarse if more than four contours are employed. Selecting a
proportion too large results in the tail classes being under-utilised, alternatively too
small a value reverts the method to that of a percentile based solution.

§3.4,2,3 Equal arithmetic intervals

Contour maps constructed by dividing the range into the desired number of levels fall
under the serial group heading. The contours are selected as for topographic maps, but
this is rarely feasible for statistical maps due to the structure of the data set and the
possible existence of outliers.

An extreme example is cited in Jenks and Coulson (1963). They studied an area in
Central Kansas where the population ranged from 1.6 to 103.4 persons per square mile.
 46

On a map with seven equal intervals only four classes were represented with over 90%
of the map lying within a single band.

§3,4.2,4 Geometric progression

This method has been developed to deal with 'j' shaped distributions. The geometric
progression is set to pivot about some measure of central tendency e.g. median. The
median is then made the geometric mid-point of the class, if an odd number of classes
is desired, or the boundary between two central classes if an even number are required.

The sequence of class widths are delineated as:-

a ax ax^ ax3 ........... ax(N_1)

where a = size of first term

X = base of the geometric progression
N = number of classes

The major drawback ofthis technique is its severe limitations of use. The base, x,
requires trial and error to determine its ideal value for each data set, hence data sets
contoured using this method are not easily automated.

§3.4.3 Dlustration of four contouring techniques

Figures 3.2 to 3.5 provide illustrative examples of four of the contouring methods
described above. The surface, based on 3432 points, describes the distribution of
potassium 40 in the south-west of England around the towns of Yeovil, Taunton and
Illminster.

Table 3.2 summarises the contour levels for the four methods of contouring. The
original data set, although not reproduced, is characterised by fairly long tails, taking
logs of the data draws the tails in slightly.
 47

Contouring Contour Level

technique 1 2 3 4 5 6
7
Equal range 32.85 65.70 98.55 131.4 164.2 197.1 229.9
Nested 33.05 59.02 82.02 94.88 106.7 129.4 177.8
means
Percentile 39.96 63.33 85.67 95.40 103.1 111.0 146.2

Standard 7.11 36.37 65.24 94.88 124.1 153.4 182.6

deviation

Table 3.2 Contour levels for the various methods.

1 2900

12000

11500-

30000 50500 31000 315 0 0 33000 335 0 0 33000 33500 34000 34500 35000 35500 560 0 0

Figure 3.2 Surface contoured using the equal range technique.

1 2500

12000 -

1000
50000 30500 31000 31500 33 0 0 0 33500 33 0 0 0 33500 34000 34500 35000 35500 36 0 0 0

Figure 3.3 Surface contoured using the nested-means methodology.

50000 50500 3 )0 0 0 31 5 0 0 33000 52 5 0 0 33 0 0 0 33500 34000 34500 35000 35500 36 0 0 0

Figure 3.4 Surface contoured using the percentile methodology.

 48

$0000 $0500 llooo $1500 $2000 33S00 33000 33500 34000 34500 35000 35500 33000

Figure 3.5 Surface contoured using the standard deviation technique.

The contour levels for percentile and nested means methods are similar, this was
reflected in the resultant surfaces. The problem as mentioned previously with the nested
means method is the restriction on the number of classes selected whilst for the
percentile technique comparability between surfaces is not necessarily assured.

For this example, the equal range method and standard deviation techniques produce
similar results. The major drawback of the former of these methods concerns the
distortion of results if an outlier is present. The levels selected may then be severely
biased towards one of the extremes. Furthermore the equal range method does not
guarantee comparability of surfaces. The standard deviation method is a more robust
method which works equally well for data which is originally near normal in
distribution, or alternatively, which requires to be transformed to achieve normality.
One drawback of the standard deviation and equal range methods is that for
distributions with long tails, one or both of the extreme levels may not be represented in
the final surface due to a lack of information in these regions. This problem does not
affect the comparison of surfaces since comparability is achieved between the
remaining levels.

This problem apart, it was decided that in view of the versatility of the standard
deviation method and the knowledge that comparability is assured between surfaces of
approximately normal data, contour selection should be based on this method.

A word of caution at this point is that many commercially available contouring routines
use the equal range or percentile technique for automatic contouring. Comparisons
using this form of contour production should be treated with caution.
 49

§3.5 A Semi-Automatic Contour Selection Procedure

A simulation study was carried out to investigate whether it was possible to create an
empirical rule which allowed the number of levels, and the standard deviation
proportion to be selected on a consistent, and automatic basis, rather than from
examining the resultant distribution for various permutations of the two variables, until
a satisfactory solution was derived. The study was based on how well the spatial
distribution of various data sets were reproduced for different numbers of contours and
standard deviation proportions.

Initially a sequence of 50, 100 and 200 points were simulated from a standard normal
distribution using the random number generator from within Minitab, release 7.1. The
response of the standard normal distribution was examined for a range of standard
deviation proportions, Ks, 0.3 - 1.4, and number of classes, Nc, 4 to 10. The response
of each data set was gauged in terms of a series of histograms, figures 3.6 - 3.12,
constructed from the percentage number of points contained within each band for the
various proportions, and how well the known underlying distribution was reproduced.
If the resultant histogram satisfied the following criteria the combination of the two
factors was deemed satisfactory to represent the spatial distribution of the data:-

1. No excessive pooling in the open-ended classes.

2. No under-utilisation of the open-ended classes.

3. The resultant histogram mirrored the underlying distribution of the data i.e. normal.

From these diagrams, an ad-hoc selection rule, similar to that developed by Brooks and
Carruthers (1953) was formulated to initially select the desired number of contours for
the size of the data set. A second rule was constructed to select an appropriate value for
the standard deviation procedure based on the number of contour levels selected.
 50

Figure 3.6 (a) Kg - 0.3 Figmre 3.6 (b) Kg = 0.4 Figye 3.6 (c) Kg = 0.5

Figure 3.6 (d)-------

—tig---------- K«--------
= 0.6 Figure 3.6 (e) Kg = 0.7 Figye 3.6 (f) Kq = 0.8

■—i

Figye 3.6 (g) Kg = 0.9 Figye 3.6 ( h ) K g = 1.0

Figure 3.6 (j) Kg = 1.2 Figure 3.6 ( k ) K g = 1.3 Figure 3.6 (1) Kg = 1.4
Key ------------- 50 points 100points 150points
Figure 3.6 Histogram of the % number of points for four contour levels.
 51

50 50

40

90 9

» 20

10 10
Q Q
0 I I I 4 S 0 2 I 4 ! 0 ? I 4 I

Figure 3.7 (a) Kg = 0.3 Figure 3,7 (b) KQ= 0.4 Figure 3.7 (c) Kg = 0.5

9 1 1 1 V 5

Figure 3.7 ( d ) K . = 0.6 Figure 3.7 ( e ) K , = 0.7 Figure 3.7 ( f ) K g = 0.8

30

90 9

30 30

10
0 0
0 2 t 4 s t t I 5

Figure 3.7 (g) Kg = 0.9 Figure 3.7 ( h ) K g = 1.0 Figure 3.7 (j) Kg = 1.1

Key ------------- 50 points 100 points 150 points

Figure 3.7 Histogram of the % number of points for five contour levels.
 52

Figure 3.8 (a) Kg = 0.3 Figure 3.8 (b) Kg = 0.4 Figure 3.8 (c) Kg = 0.5

Figure 3.8 (d) Kg = 0.6 Figure 3.8 (e) Kg = 0.7 Figye 3.8 (f) Kg = 0.8

Figure 3.8 (g) Kg = 0.9 Figure 3,8 (h) Kg = 1.0 Figure 3.8 (i) Kg = 1.1

Key ------------- 50 points 100 points 150 points

Figure 3.8 Histogram of the % number of points for six contour levels.
 53

t t > 1 « s « r 9 1 2 1 4 5 4 7

Figure 3.9 (a) Kg = 0.3 Figure 3,9 ( b ) K g = 0.4 Figure 3.9 (c) Kg = 0.5

30

0
9 1 2 1 4 S 4 7 9 2 9 4

Figure 3,9 (d) Kq = 0.6 Figure 3.9 (e) Kg = 0,7 Figure 3.9 (f) Kg = 0.8

r*“ 'i ! i
! i

t i r- i * j « t

Figure 3.9 (d) Kg = 0.6 Figure 3.9 ( e ) K g = 0.7 Figure 3.9 (f) Kg = Q.8

Key ------------- 50 points 100 points 150 points

Figure 3.9 Histogram of the % number of points for seven contour levels.
 54

0 t i 3 t ] 4 r i

Figure 3.10 (a) Kg = 0.3 Figure 3,10 (b) Kg = 0.4 Figure 3.10 (c) Kg = 0.5
90 SO

90 90
30 30

10 10

0 0 3ESL
0 1 2 1 4 S 4 0 2 I 4 0 1 2 I * 5 4 r I

Figure 3,10 (d) Kg = 0.6 Figure 3.10 ( e ) K g = 0,7 Figure 3.10 (f) Kg = 0.8

90

30

10

JZl a 0 r“
0 f 2 1 4 » 4 7 0 1 2 - 1 4 4 4 3 1

Figure 3.10 (g) Kg = 0,9 Figure 3.10 (h) Kg = 1.0 Figure 3,10 (i) Kg = 1.1

Key ------------- 50 points 100points 150points

Figure 3.10 Histogram of the % number of points for eight contour levels.
 55

.= h .
Q I 9 3 t S * ? l »

Figure 3.11 (a) Kg = 0.3 Figure 3.11 (b) Kg = 0.4 Figure 3.11 (c) Kg = 0.5
90
40

90

—
20
p -
—
I'TI

□ sl o . hb —
0 2 1 I s i r ■ f

Figure 3.11 (d) Kg = 0.6 Figure 3.11 (e) Kg = 0.7 Figure 3.11 (f) Kg = 0.8

i i

=j=L Q a h CL
0 I 1 I « ] t T I t

Figure 3.11 ( g ) K g = 0.9 Figure 3.11 (h) Kg = 1.0 Figure 3.11 (i) Kg = l .l

Key ------------- 50 points 100points 150points

Figure 3 . 1 1 Histogram of the % number of points for nine contour levels.

 56

01 i I t 1 < r I 1 10 • i i t t s o r i t i o o i i 1 t 1 4 r i o id

Figure 3.12 (a) Kg = 0.3 Figure 3.12 (b) Kg = 0,4 Figure 3.12 (c) Kg = 0.5

SSL zU Jz
o i o i t s t r i t t o o i i s t s t r i t i o 0 1 1 J 4 J « 7 » » 1 0

Figure 3.12 (d) Kg = 0.6 Figure 3.12 (e) Kg = 0,7 Figure 3,12 (f) Kg = 0.8

0 I
n I 1 I I
EL
o i i i t i o f i f i o

Figure 3.12 (g) Kg = 0.9 Figure 3.12 (h) Kg = 1.0 Figure 3.12 (i) Kg = 1.1

Key .. ------------- 50 points ................... 100points 150 points

Figure 3.12 Histogram of the % number of points for ten contour levels.
 57

The result of the study was the formulation of the following two very simple rules

1- {loge(N )<C 0 S21og„(N)}

2- i W 5K.S 1 ^
log,(Cc) s loge(C „/ 2 )J

where Cc = number of contour classes

Ks = standard deviation proportion
N = number of data points

These are displayed graphically in figure 3.13.

2.2

2.0

=r 1.2

1.0
iL—i
l- 6

0.8

0.*
0.2

0.0
0 50 100 150 200 250 300 350 *00 450 500
0 2 * 6 a to <2 U
Nu m ber o f d a ta p o i n t s
NUMBER OF CONTOUR CLASSES

Figure 3.13 (a) Selection of number of Figure 3.13 (b) Selection of standard
contour classes. deviation proportion._________________

Figure 3.13 Graphical representation of the two rules for selecting number of
contours and standard deviation proportion.

The above rules were validated on three different data sets, a N(4,10), a Un(7,4) and a
Ga(l,0,4), solely for those values of standard deviation proportion and number of
contours, determined from the rules. Table 3.3 summarises the relevant combinations of
 58

interest e,g, for six contours, standard deviation proportions of 0.6-0.9 were examined
in steps of 0.1 for 50,100 and 200 points.

Number of 4 5 6 7 8 9 10
contours
Standard
deviation 0.8-1.4 0.7-1.0 0.6-0.9 0.6-0.7 0.5-0.7 0.5-0.6 0.5-0.6
proportion (Ks)
Number 50 yes yes yes yes no no no
of data 100 no yes yes yes yes yes no
points 200 no no .... yes yes yes yes yes

Table 3.3 Summary of the standard deviation proportions and number of contours.

From the histograms of the results, figures 3.14 - 3.20, a well defined rule of thumb
materialised as to how the standard deviation proportion should be selected from within
an interval. The greater the tendency for the original data to emulate an uniform
distribution the closer to the lower bound Ks should be selected; the more peaked the
data set becomes, the nearer to the upper bound the ideal value lies. Applying these
directives ensures the three criteria mentioned earlier are satisfied.

The rules are not binding and values lying outwith the recommended ranges should still
produce satisfactory results. The main advantage of using the above empirical rules is
that for data sets differing in size, a number of classes appropriate to all the data sets
may be selected. Secondly, the rules are easily implemented with values being defined
a-priori, this dispenses with the previous time absorbing trial-and-error method .

No matter what means of contouring interval selection is used, the same technique
should be implemented for all surfaces and any parameters required as input defined on
the same criteria where a comparison is to be undertaken.
 59

i r » i i « i i t i i

Figure 3.14 (a) Figure 3.14 (b) Figure 3.14 (c) Figure 3.14 (d) >
Kc = 0.8 N = 50 Ke = 0.9 N = 50 Kg = 1.0 N = 50 Kg =1.1 N = 50

• t I I 4 t I t I I I I I

Figure 3.14 (e) Figure 3.14 (f) Figure 3.14 (g)

Kg = 1.2 N = 50 Ks = 1.3 N = 50 Kg = 1.4 N = 50

Figure 3.14 Histograms describing the results for four contours and the relevant
standard deviation proportions.

• i r i

Figure 3.15 (a) Figure 3.15 (b) Figure 3.15 (c) Figure 3.15 (d)
K^= 0.7 N = 50 K^= 0.8 N = 50 K<^= 0.9 N = 50 K<^= 1.0 N = 50

« * * % i i r i « i

Figure 3.15 (e) Figure 3.15 (f) Figure 3.15 (g) Figure 3.15 (h)
Kg = 0.8 N = 100 Kg = 0.9 N = 100 Kg = 1.0 N = 100 Kg = 1.1 N =100

Key N(4,10) Un(7,4) G a(l,0.4)

Figure 3.15 Histograms describing the results for five contours and the relevant
standard deviation proportions.
 60

i i i i I I t I * I 4 « I I I 4 I I

Figure 3.16 (a) Figure 3.16 (b) Figure 3.16 (e) Figure 3.16 (d)
Ks = 0.6 N = 50 Ks = 0.7 N = 50 Ks = 0.8 N = 50 Ks = 0.9 N= 50

I I I I l I 4 I I I I 4 % I

Figure 3.16 (e) Figure 3.16 (f) Figure 3.16 (g) Figure 3.16 (h)
Ks = 0.6 N = 100 Ks = 0.7 N = 100 Ks = 0.8 N = 100 Ks = 0.9 N = 100

* 1 * i • s «
a I I I I I 1 4

Figure 3.16 (i) Figure 3.16 (j) Figure 3.16 (k) Figure 3.16 (1)
Ks = 0.6 N= 200 Ks = 0.7 N = 200 Ks = 0.8 N = 200 Ks = 0.9 N= 200

Key:- ------------- N(4,10) Un(7,4) Ga(l,0.4)

Figure 3.16 Histograms describing the results for six contours and the relevant
standard deviation proportions.
 61

I I I 1 • I 4 * \ I S • I I t

Figure 3.17 (a) Figure 3.17 (b) Figure 3.17 (c) Figure 3.17 (d)
K^= 0.6 N = 50 K^= 0.7 N = 50 Ks = 0.6 N = 100 Ks = 0.7 N = 100

CL

Figure 3.17 (e) Figure 3.17 (f)

Ks = 0.6 N = 200 Ks = 0.7 N = 200
Figure 3.17 Histograms describing the results for seven contours and the
relevant standard deviation proportions.__________

—I

I » 2

Figure 3,18 (a) Figure 3.18 (b) Figure 3.18 (c)

Ks = 0.5 N = 100 Ks = 0.6 N = 100 Ks = 0.7 N = 100

a I I 2 I 4 * *
Q.
t •

Figure 3.18 (d) Figure 3.18 (e) Figure 3.18 (f)

Ks = 0.5 N = 200 Ks = 0.6 N = 200 Ks = 0.7 N = 200
Key:- N(4,10) Un(7,4) --------- G a(l,0.4)

Figure 3.18 Histograms describing the results for eight contours and the relevant
standard deviation proportions.
 62

« i > 3 i s i r i i o i a s i $ t r i f

Figure 3.19 (a) Ks = 0.5 Figure 3.19 ( b ) K s = 0.6

Number of points = 100 Number of points = 100

Figure 3.19 (c) Ks = 0.5 Figure 3.19 (d) Ks = 0.6

Number of points = 200 Number of points = 200
Key N(4,10) Un(7,4) -------------- Ga(l,0.4)
Figure 3.19 Histograms describing the results for nine contours and the
relevant standard deviation proportions.

0 1 I I * S « T t * IS
13 ttl

Figure 3.20 (a) Ks = 0.5 Figure 3.20 ( b ) K s = 0.5

Number of points = 200 Number of points = 200
K ey:- ------------- N(4,10) Un(7,4) Ga(l,0.4)
Figure 3.20 Histograms describing the results for ten contours and the relevant
standard deviation proportions.
 63

§3.6 Locating and Tracing The Contour

The next step in producing a pictorial representation of the underlying spatial distribution
is to physically draw the contours using the a-priori criteria. The form of the data
determines the most plausible method for contouring. Routines which contour data on a
rectangular grid are the simplest and many algorithms exist. Alternative methods use the
ideas of triangulation.

§3.6.1 Contouring by gridding

Gridding is the process of determining values of the surface at a set of locations that are
arranged in a regular pattern which completely covers the mapped area. In general,
values of the surface are not known at these uniformly spaced points and so they must be
estimated from the irregularly located control points where the values are known. The
locations where estimates are made are referred to as ‘grid points' or 'grid nodes'. The
methods discussed in sections 2,3 to 2.5 enable a regular grid to be evaluated e.g. kernel
density estimation, trend surface analysis, kriging.

A contour of a given height may be produced in two ways. First each rectangle of the
grid is examined in turn and the sections of the contour within that rectangle drawn.
Alternatively, once part of a contour has been located within a grid rectangle, the rest of
the contour may be traced through the whole grid. Further contours are then sought and
when found, they too are traced through the grid.

Broadly speaking, a contour of height h, crosses one of the grid lines (lines between two
adjacent grid points, ht(A),ht(B) if:-

ht(A) < h < ht(B)

The exact point of intersection of the contour with the gridline must be calculated. If the
grid points are sufficiently close that a contour does not cross the line joining them more
than once, then a good approximation to the point of intersection may be made using
inverse linear interpolation. Having found a starting point, the next point must be found
and in a similar manner further points, so the contour may be traced through the grid.
 64

In the 1960's Dayhoff (1963) and Cottafava and le Moli (1969) developed various
contouring algorithms. The embryonic ideas of Cottafava and le Moli were brought to
maturity in conjunction with those of Dayhoff by Heap and Pink (1969).

Heap and Pink realised that all contours must either cross the boundary or a horizontal
grid line hence it was only necessary to record the intersections of the contour with
horizontal lines i.e.

ht(A) < h < ht(B) (1)

where A is immediately to the left of B is sufficient

All such interval lines are marked during a preliminary scan. The grid points on the
boundary are then examined to see whether they satisfy (1). Each time one is found, the
contour is traced through the grid. This accounts for all open contours. When all open
contours are drawn, the closed contours are sought. A search is first made for a marked
grid line, when found it serves as the starting point, the contour is traced removing the
appropriate marks, until the starting point is reached. This process is repeated until no
marks remain.

Two forms of degeneracy are present in contouring, the first which is not applicable in
the case of Heap's and Pink's routine, refers to the case where the grid point has the same
height as the contour of interest. Cottafava and le Moli (1969) encountered this problem
but by 'virtually' altering the value of the height at the grid point by a small amount this
problem is eradicated. Rothwell (1971) and Crane (1972) adopted a similar approach.

Secondly if we consider a rectangle where the height bears the relationship to the contour
height as shown, figure 3.21(a). Given entry on a particular side, all three sides appear to
be the exit side. This suggests that another contour passes through the rectangle hence
three solutions are plausible, figure 3.21 (b) - 3.21 (d)
 65

Figure 3.21 (a) Illustration of a situation Figure 3.21 (b) Solution A.

where a degeneracy may arise.

Figure 3.21 (c) Solution B. Figure 3.21 (d) Solution C.

Figure 3 , 2 1 Illustration of contour degeneracy.

Heap and Pink used the idea of Dayhoff to circumvent this problem. They approximated
the height at the centre of the rectangle by averaging the height at the four grid points.
The rectangle was then divided into four triangles which cannot be degenerate.

Cottafava and le Moli (1969) argue the situation only occurs rarely and select either
solutions B or C depending on how the rectangle is encountered. Rothwell (1971) adopts
a more complex solution involving directions.

§3.6.2 Contouring Bv Triangulation

This method avoids the necessity for initially interpolating the grid points onto a regular
grid. Control points are assumed to be located without any particular regularity, these are
initially connected by straight lines. This forms a mesh of triangles that covers the
 66

surface. By interpolating down the sides of the triangles, locations can be found where
the elevation is a constant, specified value.

It is apparent that if the control points are connected in a different manner, a different set
of triangular plates will be defined and a different sequence could result in conspicuously
different-appearing contour lines. A set of unique, 'optimal' triangles are therefore
required. Gold et al. (1977) suggested possible solutions, the individual triangles should
be as near equilateral as possible, they should have the minimum possible height or
alternatively the longest leg of each triangle should be the shortest possible.

This problem impeded the development of triangulation based contouring algorithms

since an iterative process was required to obtain the optimal configuration. Using a
network referred to as Delaunay triangulation Gold et al. (1977) and McCullagh and
Ross (1980), developed an algorithm which produced almost optimal networks on the
first pass. Delaunay triangles are defined uniquely for a given set of data points. The
triangles formed being as nearly equiangular as possible with the longest sides of the
triangles being as short as possible.

Figure 3.22 (a) Illustration of Thiessen Figure 3.22 (b) Delauney triangulation
polygon.___________________________ network.

Figure 3.22 Illustration of contouring by triangulation.

The idea behind a Delaunay triangular network is as follows; in a field of scattered

points, each point is surrounded by an irregular polygon such that every location within a
polygon is closer to the enclosed part than it is to any other, figure 3.22 (a). Conversely,
every location outside a specific polygon is closer to some other point than it is to the
point within the polygon. This is the most compact division of space. These are
commonly referred to as Thiessen, Dirichlet or Voronoi polygons. Immediately
 67

surrounding the Thiessen polygon enclosing a specific point A, are other Thiessen
polygons, each of which encloses a single point. If these points are connected by a
straight line, the result is a Delaunay triangular network, figure 3.22 (b).

The assumption that the triangles represent tilted flat plates is a very crude
approximation of a surface. A better approximation can be achieved using curved or bent
triangular plates, particularly if these can be made to join smoothly across the edges of
the triangles. Tipper (1979) described how the three-dimensional equivalent of the spline
function may be used to ensure that abrupt changes in direction where contour lines cross
from one plate to another are avoided. McCullagh (1983) uses a tricubic polynomial to
avoid this problem, the details are given by Birkhoff and Mansfield (1974).

The finished map, contoured by triangulation is very similar to a map produced using a
gridding algorithm. The main difference arises where the surface slope changes abruptly,
the triangulation method makes the surface look sharper than it is in reality. Along the
margins of maps this problem may be corrected by the judicious insertion of pseudo­
points, but these cannot be inserted into the body of the map.

§3.7 Alternative Contouring Techniques

Other forms of contouring are available. Powell (1973) divided each rectangle into eight
triangles and approximated the surface by piecewise quadratic functions. A contour line
may then be approximated as a sequence of pieces of conic sections which can be drawn
easily, since conic sections have a convenient parametric form.

McLain (1974) describes a method for arbitrary data points, but advocates a two-stage
process, first interpolating heights to a regular mesh and then contouring the rectangular
mesh. The method suggested was that of bicubic splines. Once the intersection of a
contour with a grid line has been located, the contour is traced through the rectangle by a
series of short steps in one of eight directions, N, NE, E, SE, etc. The direction of the
next step is selected from one of three directions which depend on the previous step i.e.
one in the same direction and two at 45 degrees to either side of the previous step. The
step selected is that which is closest in value to the contour height.

Many other techniques are available, but the methodology of Heap and Pink (1969) is
simple to understand and has stood the test of time. Coding has been made available,
 68

Heap and Pink (1969), and after minor modifications and conversion to Fortran 77, this
method was used to produce all contours.

A further advantage of implementing this sub-routine and not using one of the
commercially available packages for producing contours e.g. Ghost 80, UNIRAS, S-Plus
etc. was the extra feature that contour co-ordinates were accessible to the user. This is
not a feature of the majority of graphical packages.

§3.8 Accuracy of Statistical Maps

Within the previous sections, methodology was developed for selecting the number and
bounds of class intervals and a number of techniques were described as to how a contour
may be drawn. However a question fundamental to the construction of a statistical map is
how accurate is the resultant surface.

The question of accuracy arises when an isarithmic map has been constructed from data
which do not cover the entire domain of the mapped area, these data points are subject to
observational and/or location errors and finally to errors generated by the method of
interpolation between punctual sample points. In such instances the map is referred to as
an estimated map as opposed to the underlying error free map.

Accuracy is then a measure of the displacement between the estimated and true map, the
latter which is rarely, if ever, available. Various attempts have been made within the
literature to quantify the accuracy of maps.

Blumenstock (1953) discussed the problem of assessing reliability of isarithmic maps for
meteorological maps but his analysis is of wider significance for all maps of
geographical phenomena. The method he initiated was to first take the basic data and
estimate the magnitude of all the sources of unreliability. The data was then corrected for
bias and thirdly, the standard error was computed in terms of the observational and
sampling error. The final step required the determination of the chance that any one
particular plotted value will be in error by x units, solely on the basis of the unreliability
of the data. This final step enabled an estimate to be obtained as to how many of the
plotted values lie outside their correct isarithmic intervals.
 69

The method is constrained in its usefulness since it is reliant on the person implementing
the technique to define levels of reliability of their data and secondly it assumes the
errors are normally distributed.

Switzer (1975) examined the question of error induced as a result of interpolating

between punctual sample data. He considered a map to be a partition of a domain, R, into
k sub-domains, k being the number of colours used in a map. The sub-domains are
A A A

denoted R^, R2 ,...Rk in the true map and are denoted Ri, R 2, . . Rk. in the true map. The
n data points underlying the estimated map are at locations sj, S2 , S3, sn, which may
or may not be at the centres of basic sampling cells, S^, S2 , S3 , ...., Sn. The estimated
A.
map is constructed by assigning a cell to the F 1 sub-domain Ri if the data point inside
that cell is observed to fall in Rj, the n basic sampling cells being a relatively fine
partition of the domain, R.

As an index of precision, Switzer examined the discrepancy between the true and
estimated maps as measured by the mismatch areas :-

i.e. Lu - \x Ri Rj

\x = area

Where Lqj, the precision index, can be expressed in terms of Lebesgue integrals:-

Ly = X S j(sh) j 5 i(s)d|a(s)
h-l St,

where Ly = area of that region which belongs to true subdomain i but is

represented as subdomain i on the estimated map.
8 . = indicator function for the true subdomain i
Sh = sampling cells
sh = data points

Problems with such a technique relate to the difficulty of modelling the form of the
location error, differences in the index will materialise depending on the model selected,
e.g. spherical, square, hexagonal, etc.
 70

Other techniques although developed for choropleth maps may be extended to contour
maps, Jenks and Caspall (1971). Problems with attempting to evaluate an accuracy index
are that the indices themselves require the estimation of various quantities, or
alternatively, certain assumptions require to be satisfied for the index to be applicable, so
far no method has been developed to circumvent these problems.

Mackay (1953) showed problems in accuracy may not necessarily originate from the
original data but from the locational pattern of the recording points or 'control' points.
Where choropleth maps of discrete counties are concerned, the framework of control
points is fixed; where isarithmic maps are being developed from point observations, the
framework problem may be more complex. The implementation of Heap and Pink's
triangulation method for contouring takes account of this possible error source.

Hsu and Robinson (1970) and Morrison (1971) illustrate that both sample size and the
method of areal sub-division affect the quality of data and hence of isopleth maps.

A further source of error is attributable to how the points on a contour are joined. They
may be joined as straight lines or by a curve fitting algorithm. Although the latter
produces aesthetically pleasing contours, they are not necessarily accurate. Too often the
contours produced by this method reflect the curve fitting algorithm used rather than the
data being contoured, especially if local curve fitting algorithms are used. As a
consequence, when a curve fitting algorithm is used there is no guarantee that contours
of different heights will not cross, the situation is consequently totally unacceptable.

In practice to avoid this problem straight lines should be used to join the points on a
contour. Not only does it have the advantage of simplicity of implementation but it also
gives the viewer of the resultant map a good appreciation of how coarse or fine the map
really is.

The utilisation of straight lines to join points of equal heights is of less consequence
when the grid over which we interpolate is extremely fine. Neither the shape of the
contour nor various geometric attributes associated with it, such as area, perimeter, will
be unduly affected. For a coarse grid differences will arise in the measures and possibly
shape. A very simple example illustrates this point.
 71

For a simple construct, the circle, the question of what kind of accuracy is achievable
with x-data points was examined. A circle of unit radius, centre the origin was depicted
using 4, 8 , 16, 32 and 64 data points. Table 3.4 evaluates the accuracy of the result by
taking the ratio of the grid value to the true value for area and perimeter.

Number of Angle subtended Ratio of grid Ratio of grid

points between area to true area perimeter to true
successive data perimeter
points

4 90 0.6366 0.90031
8 45 0.9003 0.97449
16 22.5 0.9745 0.99358
32 11.25 0.9936 0.99839
64 5.625 0.9984 0.99960

Table 3.4 Accuracy of evaluation of a unit circle for x-data points.

The results of the table show that changing the number of points to specify the area of
interest has less effect on perimeter than area. For both measures 99% accuracy was
achieved using 32 points, specifying additional points served only to increase the
computational time for minimal gain in accuracy. By increasing the number of points a
more aesthetically satisfactory picture results.

The question of grid resolution is of considerable interest since it may inadvertently have
an effect on the outcome of the analysis. This is examined in detail in chapter 4.

In reality, it is extremely difficult to place any sort of numerical bounds on the accuracy
of the resultant statistical map since the true surface is never known.

§3.9 Summary

The construction of a contour plot from a set of spatially referenced data follows a
sequence of clearly defined steps:-
 72

1. If the data are irregularly spaced, an interpolation technique which enables a regular
grid to be estimated is implemented. The methodology used was that of kernel density
estimation with smoothing parameter selection undertaken using least-squares cross
validation.

2. Select number and bounds of class intervals using the following empirical rules and
by studying the histogram of the data to decide whether the variables should be selected
closer to the upper/lower bound:-

!• {loge(N )<C 0 S21oge(N)}

2.

where Cc - number of contour classes

Ks = standard deviation proportion
N = number of data points

3. Using Heap's and Pink's contouring methodology construct the contour plot using
straight lines to join the contour nodes.

By implementing the three steps described on a consistent basis, the error attributable to
the procedure should be of the same order for all surfaces.
 73

CHAPTER 4

DIFFERENT METHODS OF SURFACE COMPARISON

§4.1 Introduction

Many methods are available for effecting a comparison between two or more spatially
referenced data sets, these range from the very simple to the more complex and
computationally intensive approaches. The form of the analysis depends primarily on
four factors:-

1. The manner in which the data is presented e.g. point, line, areal or surface.
2. Whether variables to be compared relate to the same measurement.
3. Whether the comparison is inter-regional.
4. Whether the observations are recorded at the same locality.

A simple way of categorising the various methods available for performing a comparison
is to group them according to their mode of interpretation:-

1. Subjective comparison.
2. Comparison technique for a global analysis.
3. Comparison technique relevant to a local analysis,

§4.2 Subjective Analysis

The simplest and oldest method of comparison is to overlay the maps of interest and
describe how their distributions differ.

Where information has been recorded on variables at the same locality i.e. a repeated
measures format, an isopach/residual map may be constructed to describe the spatial
arrangement of the differences between the surfaces.

Other subjective approaches include plotting dependent versus independent variables and
making a visual assessment of the strength of the relationship. McGlashen (1972)
conducted a survey in Central Africa of fifty five diseases and twenty environmental
 74

factors that might be associated with the diseases. Data came from patient records at 84
hospitals. From visual examination of these disease and factor plots, McGlashen was
subsequently able to perform a contingency table analyses. For example, he compared
the number of annual cases of diabetes mellitus with whether or not cassava was the
staple food eaten by hospital patients. In a similar manner Prentice et al. (1991)
subjectively assessed the goodness of fit between observed and simulated isopall maps
i.e. maps of equal pollen count, by plotting the observed levels of pollen count against
those values simulated from a response surface model.

These three techniques of subjective comparison, superposition of maps, isopach maps

and dependent versus independent variables are illustrated in figure 4.1. They examine
the relationship between the dispersion of two measures of ambient radioactivity, beta
and gamma radionuclides, within the south-west of England. The majority of external
radiation dose to the population originates from radionuclides in two of the three natural
radioactive decay series i.e. 238jj an(j 2 3 2 ^ The decay of these sources of gamma
radiation causes the emission of both alpha and beta particles into the atmosphere. It is
therefore anticipated that a strong association will exist between the two radiation
variables.

4 .0

5.5

5 .0

£ 2 .5

i-2.0
m
LU

1X5

0.0
0.0 at as as a7
GAMMA (n g y /a )

Figure 4.1(a) Plot of Gamma (mgy/a) versus Beta (mgy/a) for the south-west of
England.
 75

12500

12000

11500 - -

11000
36000 35000 34000 33000 32000 51000 30000

KEY
Beta Gamma
(mgy/a) (mgy/a)
1 5% percentile 2 25% percentile
3 50% percentile 4 75% percentile
5 95% percentile

Figure 4.1(b) Spatial distribution of beta and gamma radiation variables within the
south-west of England.

12500 56 7
\w\
12000 -

11500 - -

11000
36000 35000 34000 33000 32000 31000 30000
(Beta - Gamma)

KEY
1 -1.5 ------- 5 0.5
2 - 1.0 ------- 6 1.0
3 -0.5 --------- 7
1.5
4 0.0
Figure 4.1(c) Residual/isopach map of the two radiation variables, beta and gamma,
in the south-west of England.
Figure 4.1. Various forms of subjective analysis.
 76

Subjectively, all three methods suggest the existence of a strong areal dependence
between the two radiation fields, the information imparted by each diagram is different.
The first diagram ignores the spatial dimension of the data. A simple graphical
presentation of the relationship between the beta value at station i and the gamma value
at station i is illustrated in figure 4,1 (a). This plot is only applicable when the
measurements are recorded at the same locality, thus restricting its usefulness in
presenting comparative information.

The remaining two methods incorporate the spatial dimension of the data. Figure 4.1 (b)
illustrates the superposition of the two surfaces representing the spatial distribution of the
variables, A strong spatial association is apparent between the two variables, as indicated
from the preliminary discussion. Where differences do materialise between the variables,
this mode of presentation enables those areas of greatest change/similarity to be both
located and also, a description given as to the structure of the change. One method of
describing change is in terms of various mathematical descriptors e.g. change in
orientation, location and/or size between a set of comparable contours. This idea forms
the basis of the methodology developed in the following two chapters to describe and
quantify change. For this example, differences between the two data sets are minimal, it
primarily arises in terms of the size of the contours; contours relating to the gamma
radiation field extend over a smaller area than those depicting beta radiation, particularly
in the west, the converse is true in the east.

The specialised nature of the data i.e. observations recorded at the same locality, enables
a residual map to be constructed. Before the residual map could be produced, it was first
necessary to normalise the two data sets. Interpretation of a residual map of this form is
not as intuitive to non-statisticians, however it localises those areas of greatest change in
both a positive and negative direction. In the west, the beta values tend to be higher than
those of the gamma radiation field. In the east, with the exception of the south-east
comer, the reverse is true.

The overlaying of the surfaces is the most versatile of the methods since it firstly,
incorporates the spatial dimension of the data and secondly, variables recorded at
different locations may be examined using this method, this is not plausible for the other
two methods.
 77

A fundamental drawback to all the above methods lies in their subjectivity. Confronted
with the same two maps, not everyone will agree that an areal association exists, or will
assess the degree of association as the same.

§4.3 Global Analysis

In terms of an overall analysis, a number of numerical measures have been proposed

which attempt to eliminate the uncertainty attributable to subjective techniques.

§4.3.1 Lorenz curves

The Lorenz (1905) curve is a diagrammatic tool which allows for the visual and
quantitative comparison of the cumulative relationship between two variables. The
Lorenz curve may be defined mathematically as the curve whose ordinate and abscissa
are O and F, respectively, such that:-
x

F(x) = jf(x)dx

where v is the population mean of x. Convexity to the F-axis is a necessary condition for
all Lorenz curves.

The commonest equality measure is the Gini (1913-1914) coefficient which is a direct
function of the Lorenz curve. It can be shown, Kendall and Stuart (1958), that the Gini
coefficient is equal to twice the area between the line F = O and hence the Lorenz curve
is defined to be

G=odi = ^ j J|x-y|dF(x)dF(y)

A Lorenz curve is an area-by-area plot of the ratios of the two variables made in order to
indicate similarities of the distribution. The calculations involved for two variables
are:-
 78

1 . evaluate - z, where Zy - variable i for region j

= ratio for region j

i = 1,2 j = 1„..,N

2 . rank z^, smallest proportion given rank one.

3. evaluate

where sz.. = standardised variable i for region j

4. Maintaining z^, the szy are accumulated, %zy, i.e. cumulative percentages
for variable i, region j

5. plot % Z jj against % z 2j

Where the distributions are proportionally identical in each area, the resultant plot is a
straight line through the origin with slope one, with completely separate distributions
resulting in a line which would follow the z3 axis. In a real world situation the curve will
lie between the two extremes.

The difference between the plotted curve and the theoretical optimal is a measure of the
dissimilarity of the distributions. Various indices are available for quantifying the level
of dissimilarity

1.. Dm= m a x |% Z j . -% z 2j

area between the curves

area below diagonal
 79

The latter index appears in the literature in a variety of guises and has been given
different names according to the nature of the two variables. In population studies where
Zj is the population and z2 the land area, it is termed the index of population
concentration. For the same z2, but zp the land area under a particular use, it is described
as the coefficient of areal localisation.

An example of a Lorenz curve is given in figure 4.2. The curve examines the relationship
between the area of civil parish districts within the south-west of England, around the
towns of Illminster and Yeovil, and the population contained within each of these
parishes.
IUU

C£

UJ

UJ

o
0 20 60 50 10D
CUMULATIVE PERCENTAGE OF PO PULATIO N

Figure 4.2 Lorenz curve for civil parish data within the south-west of England.

Dm Di Df
Dissimilarity results 51.10 51.10 71.04

Table 4.1 :- Dissimilarity levels for the Lorenz curve

Table 4,1 reports the results for the various measures of dissimilarity. Df is numerically
greater than the other two but it makes use of all the available information; in all three
cases the level of dissimilarity indicates that the distributions do not correspond very
closely. Generally small civil parishes are associated with high populations and vice-
versa. The small parishes tend to be located in urban areas with the geographically much
larger parishes within rural areas.
 80

The Lorenz curve is a useful graphical and numerical tool for making comparisons, but it
suffers from a number of disadvantages:-

1. Its use is constrained to choropleth maps,

2. It should not be used with negative numbers or density ratios.
3. It will not discriminate between different arrangements of the areal units.

§4.3.2 Coefficient of areal correspondence

Minnick (1964) constructed a cartostatistical method for determining areal

correspondence that was simple, meaningful and clearly interpretable. In terms of basic
algebra:-

_ A nB
A uB

~ area °ver which phenomena are located together

total area covered by the two phenomena

Completely separate distributions give a value of zero whilst for exactly coincident
distributions, a value of one is reported for the coefficient of areal correspondence, Ca.

Returning to the example cited on the relationship between the two radiation fields,
section 4.2, a value for the coefficient of areal correspondence was calculated to be
0.974, when the two fields were compared at their respective median values.
Interpretation of this value confirms analytically what had been expressed in terms of our
subjective beliefs i.e. the two radiation variables display a strong association.

An equivalent measure, the proportion of total area of the observed map categorised
correctly by a simulated map, PCC, was used by Webb et al. (1987) to quantify the level
of association between observed and simulated fossil-pollen maps. They acknowledged
that the significance of measures of association is difficult to assess analytically since the
map patterns are spatially highly-correlated, Cliff and Ord (1981). In an attempt to
compensate for this drawback, an empirical reference distribution for each of the
measures was computed for all possible comparisons among the observed maps for all
pollen types and times, but those comparisons within each type at different times were
excluded. This reference distribution gave an indication of how large PCC could be when
 81

no association between maps was expected. For a specific comparison of an observed

and simulated map, a large value for one of the measures relative to the values in the
reference distribution, indicates a stronger association between the simulated and
observed map than that expected for the comparison of two observed maps for different
pollen types.

A further modification and development of the method was introduced by Court (1970).
For the simplest case of one isopleth, representing the median value of the distribution,
areally weighted, the two surfaces were superimposed and the resultant diagram
described four regions, table 4.2

REGION A
above median below median
REGION B above median (+ » + ) (-, + )
below median ( + ,- )

Table 4.2 A 2 x 2 resemblance matrix.

For such a 2 X 2 resemblance matrix, the coefficient of medial correlation qm is defined

to be:-

(like area) - ^ (unlike area)

qm“ v
2,area

This coefficient has three advantages over Ca, the coefficient of areal correspondence,
first its limits are -1 to +1, with a perfect negative areal correspondence, giving a value
of -1. Secondly, the method may be extended to deal with an even number of classes.
According to Court (1970) the sampling distribution of this coefficient is roughly normal
hence statistical significance may be tested, although care should be taken due to the
presence of spatial auto-correlation.

Hugg (1979) used this method to compare the geographic distribution of work disability
and poverty status for persons aged 18-64 using the fifty states of the United States as
units of observations.
 82

§4,3.3 Correlation coefficient

An approach suggested by Robinson and Bryson (1957) and Mirchink and Bukhartsev
(1959) was the use of Pearson's product moment correlation coefficient: -

XtZu-s^Mza-Zfe)
r = —
nSziSZ2

\2
where s . = sqrt I v v
n

This coefficient is equally applicable for interval or ratio scattered data. Pyle (1973)
compared census tract maps of measles incidence in Akron, Ohio for 1970-1971 against
various demographic and socio-economic variables. Ecological correlation was also used
by Gesler et al, (1980) to compare maps of community characteristics to disease
reporting and hospital use by census tract in Central Harlem Health District, New York
City.

A non-parametric form of the correlation coefficient, Goodman and Kruskal's gamma,

Goodman and Kruskal (1954), was also used by Webb et al. (1987) to assess the
association between the pollen-fossil maps. The level of agreement or disagreement
between the observed and simulated fossil-pollen maps was assessed by mapping the
results on a large spatial scale. The authors agreed that although some spatial smoothing
was introduced and the pollen counts were recorded for only four levels, the resultant
effect was desirable since it reduced palynological noise, i.e. small scale spatial and
temporal variability and secondly it reduced the scale discrepancy between the coarse
spatial scale of the model and the generally finer scale of the data. An ordered
contingency table based on these levels was constructed using the interpolated values
from the contouring step. The coefficient values were calculated:-

Y = -2 ^ 0 < y ^ 1
p + q
where p = number of concordant pairs of observations
q = number of discordant pairs of observations
 83

This coefficient has the advantage of having a direct probabilistic interpretation, it is the
difference in probability of like rather than unlike orders for the two variables when two
individuals are selected at random, y takes the value one when the data are concentrated
in the upper-left to lower-right diagonal and the value zero in the case of independence,
however zero need not imply independence. This coefficient has similar
drawbacks/limitations in its use and interpretation as the PCC described earlier. A
procedure for formal testing also requires a reference distribution.

Tests of statistical significance may be performed to test whether the correlation

coefficient is significant, these are fundamentally incorrect since various assumptions
inherent to such a test cannot be satisfied in such a situation, i.e. data points are rarely
independently distributed, and distributions to be compared will invariably demonstrate
significant spatial autocorrelation. The correlation coefficient as a measure of similarity
is appropriate as it stands but further inferential analysis should be avoided.

In many situations, subtle changes in the distribution of a variable will occur over a time
span, a correlation analysis will potentially report a high level of correspondence, but in
reality it imparts no information on the complexities of these changes.

§4.3,4 Comparison analysis of trend maps

Merriam and Sneath (1966) developed a very simple procedure which allowed trend
surfaces to be compared. Data points which represent a surface are not independent
hence they believed it may be more prudent to estimate the similarity between surfaces in
terms of the coefficients of the trend.

A model for surface i, assuming a linear trend is given by

z ij = P o o + P i » x ij + PoiyIJ+eiJ F.i r N (°,a2)

where zjj - value of dependent variable at point j, for surface i.

(xij >yij) = spatial location of dependent variable j , for surface i.
 84

The values of the coefficients of the fitted surface may then be used as mathematical
descriptors of the observed surface. In practice the base is excluded. The similarity
between the surfaces may be expressed in one of two ways :~

1. A correlation coefficient

•v/varpi varpt

where
covp ik = covariance between the coefficients for surface i and k.
varpi = variance of the coefficients for surface i.

Instead of comparing observations xl andx2, the coefficients (/31s) of the equations for
trend surface i and trend surface k are compared.

2. The taxonomic distance

d ik “ fyk )

where j = 1, s where s = order of the trend surface

i, k s surface of interest

This measure is the square root of the mean of the squared differences between
equivalent coefficients for the same order surfaces to be compared. The taxonomic
distance is no more efficient than the correlation coefficient, it may be easier to interpret
since it is always positive and is not constrained to values less than one. Identical trend
surfaces give taxonomic distances of zero, with increasingly dissimilar trend surfaces
having increasingly greater taxonomic distances between them.

For trend surfaces of order s, the higher order polynomial terms will generally take small
values, to use the values as they stand amounts to estimating the variance of the
differences corrected for height. In order to ensure all the terms make an equal
contribution to the overall similarity measure, the coefficients are normalised prior to
 85

calculating either of the measures. The implementation of this method requires that the
surfaces to be compared are of the same order.

Classification or comparison of regional trend surfaces by grouping on the basis of the

calculated coefficients of the polynomial terms is restricted by two factors. First, the
trend surface parameters are not invariant under certain changes of scale and orientation
of the co-ordinate system and secondly, the estimators are still correlated. Miesch and
Conner (1967) showed that shifting the origin of the co-ordinate system changed both the
values of the estimated coefficients and the percentage explanation of individual terms,
the overall percentage explanation by terms for a given order remains unchanged. The
absolute values of the regression coefficients, will occasionally be subject to extreme
fluctuations in the higher powers as a result of machine rounding and truncation.
Variation will also arise within the data due to measurement error.

These factors all serve to restrict the applicability of this method of comparison to those
situations where a comparison is to be effected between quadrat systems of similar size
and shape. An example of its implementation is given in chapter 7 where the surface
temperature for two years are compared for the contiguous United States of America.

§4.3.5 Difference maps

The idea of an isopach/residual map has been extended and formalised to situations
where the observations have not necessarily been recorded at the same locality. In such
cases an intermediate grid is required to be evaluated before the difference between the
two matrices of grid values is taken, the resultant matrix is then contoured.

The most complicated scenario is where interest is in two different variables each of
which has been recorded at differing locations. Davies (1973) believes that to directly
compare two maps under these conditions, one must be expressed in terms of units of the
other or alternatively, both converted to standardised unitless forms.

Expressing one variable in terms of the other allows the user to perceive where the
mapped variable is 'greater/smaller’ than predicted on the basis of the other variable. The
ideas of least squares regression, Seber (1977), enables the implementation of this
procedure. For the vector of grid values for variables X and Y, we compute;
 86

Yi-Po+PA

Typically linear regression is used, but low order polynomial regression is equally
A A

applicable. The result of the above is a vector of predicted values of Y. The Y's are
based solely on values of the second variable X, hence a residual surface (Y. - Y.) may
be regarded as a map of the differences between X and Y. No statistical assessment of
the regression of Y on X is possible where the two variables were not measured at the
same control points, since the regression is based on the estimates of X and Y at the grid
points. The difference/residual surface may be displayed in terms of positive and
negative residuals.

Problems with the methodology exist:-

A
1. The estimates Y account for only a part of the variation in Y. Unless the
correlation of X and Y is ’high' serious errors will be introduced by the substitution of the
A
estimate Y for X.

2, It may not be possible to decide which variable should be used as the

estimator. If the correlation between X and Y is 'high' the two regression lines will nearly
coincide. If the correlation is not pronounced, the two lines may produce radically
different results, Mills (1955). In practice, it may be argued that in the absence of a high
correlation, it is pointless to compare the two by a predictive model.

The problems inherent to difference maps based on estimated or predicted variables can
be avoided if the original two maps are converted to a standardised format. After the data
on each map have been standardised, they may be contoured in the conventional manner.
However the contour values will be in units of standard deviation above or below the
mean. The resultant contour map is liable to contain ambiguous areas, i.e. an expected
positive difference can result from subtracting a low positive area from a high positive
area, however it is also achievable by subtracting a large negative area from a low
positive area. Similar ambiguous cases result for negative differences.
 87

§4.3.6 Pattern of differences

Cliff (1970) pioneered a method based on the analysis of the pattern of differences
between two maps. The criteria for executing such a technique was that the variables of
interest were measured in the same units and related to the same areas. Cliff then
constructed a three-colour map, coloured according to the relation between the variables.
The fundamental proposition in Cliff's method was that if the two maps did not differ
significantly, in a spatial sense, then the distribution of the three colours will not be
significantly different from random. On the other hand, any spatial pattern in the colours
is indicative of some unknown spatial process. The test statistic is a simple joins-count
approach: -

_ observed number of joins - expected number given by a random process

standard deviation of the expected values

Expressions for the expected number and standard deviation of joins of both the same
color and different colors can be found in Cliff's (1970) work.

§4.4 Local Analysis

The final category of surface comparison techniques relate to those methods which fall
under the heading of local analysis. The term local analysis refers to those techniques
where specific aspects of the spatial distribution are examined, in terms of a contoured
map, this may be the 75th percentile, for example.

One feature of many of these methods is that invariably it is convenient to regard the
underlying qualitative variation as a multi-colour pattern or, where appropriate as a two-
colour black-white pattern. In general the techniques in this group are computationally
more intensive and specialised.

§4.4.1 Complexity index

Complex spatial geologic patterns may be regarded as realisations of random processes,

Dacey (1964), Watson (1971), Matem (1960). The estimated parameters of such
processes serve as convenient summary characteristics of the observed spatial patterns
 88

e.g. patchiness and prevalence and provide a basis for their classification and
comparison, Switzer (1973).

These ideas formed the basis of an approach which attempts to describe the level of
pattern complexity of a map. By complexity, Switzer meant the spatial scale of variation.
A pattern that has a self-contained area is less complex than the same proportion of area
distributed in many scattered smaller areas. This notion of complexity as a scale
measurement may also be viewed as a measure of patchiness of the pattern. One intuitive
index cited was

xv _ total length of boundaries

_ ------------------------------------------------------------------------_ ---------------

(area of region)72

The larger the value of X, the more complex the pattern. X is also invariant to the choice
of measurement unit. We shall need some convention on how to measure boundary
length, which is assumed to be finite, section 3.8.

It can be shown, Matem (1960), that if the pattern is regarded as a realisation of a

random process, then

mean of X = —7t(area of region)^ Q' (0)

2

Q’ (0) = derivative of Q(d) at d=0

Q(d) = probability that two colours distance d apart are of different
colours.

Because we are concerned here about the estimation of 'pattern properties', from
discretely spaced data, we may wish to know how the complexity parameter, X, might be
estimated from a square grid, say. Basically we require to estimate Q'(0). Switzer
proposes one method for evaluating Q'(0) for a square grid, By altering the sample space
and shape of the grid and the sampling density, the complexity index, X, will be
changed.
 89

§4,4.2 Image registration

In image-analysis where the analysis of two or more images of the same scene is to be
undertaken, registration is required. Image registration is the process of determining the
position of corresponding points in two images of the same scene. If the difference
between the images is any combination of translation, rotation and scaling then by
determining the positions of a minimum of two corresponding points, control points, in
the image, the images may be registered.

Extraction of control points is an almost impossible task. If it is possible to find straight

lines within the image, the intersection of the lines produce control points, Stockman et
al. (1982), alternatively the image can be segmented and closed-boundary regions
defined within the image, the centre of gravity of these regions will then produce control
points. Goshtasby et al (1986) used the idea of centres of gravity of closed boundary
regions as control points. Various segmentation techniques are available, the main
objective of these is to produce a desirable number of closed boundary regions within
the image.

A point pattern matching technique is required to establish correspondence between

control points in an image. One method is to match point patterns by a clustering
approach, Stockman et al (1982). Using the clustering technique, matching is carried out
between all possible pairs of points in the two sets, When matching point pairs, the
translational, rotational and scaling differences between them are determined and a point
entered into a parameter space showing the parameter values. Correct matches tend to
make a cluster whilst mismatches randomly fill the parameter space. The parameter
values corresponding to the most dense cluster are used to map one set to another and
determine the correspondence between the two sets of points.

By knowing corresponding control points in the images, corresponding regions may be

identified. For comparison of two regions in image analysis, it is desirable to refine the
regions to become as similar as possible. If two corresponding regions are more similar,
it is anticipated their centres of gravity correspond to each other more closely. Region
similarity is then obtained by comparing the shapes of the segmented regions. Since the
images have translational, rotational and scaling differences, the shape measures must be
invariant to these transformations.
 90

Some of the techniques suggested by Goshtasby et al. for defining shape similarity are
Fourier descriptors, shape signatures, centroidal profiles, invariant moments and shape
matrices. In terms of the latter, the shape is transformed into a binary matrix by polar
quantization of the shape. The zeros and ones in the matrix show points that belong to
the outside and inside of the shape respectively. The dimensions of the matrix that
determine the quantization steps are determined by the user. The fewer the number of
ones in the obtained matrix, the more similar the shapes. By choosing differing
dimensions the results will not necessarily exhibit robustness.

In reality we are not always interested in removing the scaling difference, rotational and
translational changes since it is these factors which give us a handle on the possible
sources of the underlying processes which are responsible for the change.

§4.4.3. Image restoration

A further set of spatial comparison techniques contained within the image processing
literature have arisen through the assessment of image processing algorithms for the
restoration of images. A number of numerical measures are available for quantifying the
discrepancy or 'distance' between two images: -

1. The distance between two grey scale images x and y, measured by the root mean
squared difference between corresponding pixel values is given by :-
i
i S ( x ( t ) - y(t))2] 2
_ teT

where x(t) denotes the brightness value of image x at pixel t. This is an example of an L2
metric. It has many mathematical advantages and is the basis of the optimal linear
(Wiener) filtering theory, Hamming (1983). A further example using distance between
corresponding pixel values is that of:-

l X |x ( t ) - y(t)|
IN te T
 91

2. In classification problems, where the pixel values are class labels, image distances
can be measured by the pixel disagreement rate i.e. the proportion of pixels given
conflicting class labels:-

—number{tsT: x(t) ^ y(t)}

These two measures involve a pixel by pixel comparison, although widely used, these
measures are generally recognised as unsatisfactory, Besag (1986), since they ignore the
spatial context and are inadequate in expressing human perceptions of similarity.

Baddeley (1987) has suggested a modification which attempts to combine the ideas of
the L2 metric with those of the Hausdorff metric. The Hausdorff distance between two
sets of pixels X , Y c T is:-

V IV v\
H(X,Y) I
= max^ SaP d(t>Y) ’ SUP d(t-Xh ^
[t s X te Y J

where d(t,X) is the shortest distance from a pixel t e T to a subset X c T

inf d (s,t)
d(t,X) =
seX

i.e. H(X,Y) is the largest distance from a point in one set to the nearest point in the other
set.

A major drawback of the Hausdorff metric is its susceptibility to outliers and its lack of
robustness. The metric is equal to the distance from one white pixel, say, to the
remaining white pixels hence the alteration to the value of a single pixel can markedly
affect the value of the metric.

For X>0, the X metric between images x,y, defined by Baddeley, is denoted as

A^x.y) = sup max{8w (t);8Xiyj[(t)}

 92

where for each t e T

5U y(t) = inf{a>0:3 se T ,d T(s,t) < aX,dv(x(t),y(s)) <a}

In other words, two images are closer than a units in the metric if, for every pixel in the
x-image, there is a pixel in the y-image less than aX units away, with a brightness value
differing by less than a and vice versa. Intuitively X is the 'rate of exchange' between
errors in pixel brightness and errors in pixel distance.

The LP metrics compare the image brightness values of x and y at the same pixel
position, t, i.e. a 'vertical' comparison whilst the Generalised Hausdorff metric basically
performs a lateral comparison since H(X,Y) equals the maximum distance from a pixel
in one image to the nearest pixel with the same value in the other image. Finally the X-
metric equals the maximum height of a rectangle needed to touch the graph of y from
any point on the graph of x and vice-versa and is effectively a trade-off between vertical
and lateral comparisons.

§4.4.3 Shape change

Scientists and philosophers in many disciplines have long recognised the potential for
gaining insight into spatial processes by studying spatial form. Study of the distinctive
shape of a distribution occupied distinguished biologists like Thomson (1917), who saw
in changing forms clues to biological growth and evolution. Scholars in all disciplines
who have attempted to formalise process-form arguments have found it necessary to
devise adequate methods for describing shape.

If shape is stable through time the effect of the conflicting forces has been resolved and
an equilibrium state has been achieved. However continued growth or atrophy is
indicative that such a state has not yet been obtained; the forces operative on the object
still remain unresolved. If one considers processes to be a collective term for all the
unresolved forces continuing to shape the object in question, studies of form and process
are legitimate.

Generally shape is defined to be the set of properties possessed by any closed figure of at
least two-dimensions which has a planar representation, and which possesses precise
boundaries i.e. form with size removed.
 93

The statistical analysis of shape data has a vast range of applications in biology,
archaeology, geography and chemistry for example. Two main classesof shape analysis
exist, those relating to outline data and secondly, those methods based on landmark data.

§4.4.4.1 Outline data

Shape factors are very simple to derive computationally and require only basic
information about the shape. In trying to measure shape, two questions must be asked:-

1. What characteristics do we measure?

2. How do we combine them into an effective index?

In addressing the first, we assume that the shape of each spatially discrete area is being
studied separately. The attributes most commonly measured are area, perimeter, major
axis, radii of internal and external enclosing circles, figure 4.3. In order of increasing
dimensional scale these are

1. Points within the closed figure e.g. centre of gravity.

2. Lines within the closed figure e.g. perimeter.
3. Area of the closed figure.

KEY
A Area
P Perimeter
L Longest axis
Ri Radii of internal enclosing
circle.
Re Radii of external enclosing
circle.
Figure 4.3:- Commonly measured shape descriptors.

For all three situations, many different 'shapes' will share similar numerical values. The
number of indices based on these primary measures is very large, Boots and Lamoureaux
(1972). Table 4.3 describes some of the more commonly cited measures for describing
shape, particularly in geographic applications.
 94

Measure M athem atical R eference

form ulation

Elongation ratio L/L1 L'- length minor axis Werrity (1969)

Form ratio A/L2 Horton (1932)
Circularity ratio (4A)P2 Miller (1953)

Compactness ratio (2-7JtA) / P Richardson (1961)

A/A' A' - area of smallest Cole (1964)
enclosing circle Gibbs (1961)
1.273A/L2

Ellipticity index L /2 { A /[ ti(L/2)]} Stoddart (1965)

XUOOd^Xd.MlOO/n)
Radial shape index i=l i=l Boyce and Clarke (1964)
di = radial distance from a
point to the circumference
of the circle.
Table 4.3. Elementary measures for measuring the shape of geographic areas.

An interesting shape measure was described by Young et al. (1974), to circumvent some
of the potential problems associated with the above indices i.e. lack of robustness and to
a lesser extent, the possibility that some of the properties may be altered in the transition
from analysis of shapes of a continuous form to analysis on a discrete grid.

The measure described by Young et al, was based on the notion of bending energy. They
suggested a two-dimensional outline made out of a homogeneous material, if allowed to
adopt its 'free form', would assume the shape of a circle since a circle minimises stored
energy. More convoluted outlines require additional work in the form of bending energy.
The curvature calculation is simple. The shape is divided into n small regions and the
curvature Kn js defined to be
 95

The total ’bending energy' is given by ^ K j 2 over the whole region. The measure is
i*i
invariant to position and rotation but is affected by size as well as shape differences.

Other techniques based on the definition of the curvature at sample points on an outline,
or within a surface, have been used to compare biological forms. One method is to
describe the continuous curvature at sample points on the outline. The outline can then
be considered to be a continuous curve. Implicit in this function, describing the curve,
are the actual Euclidean locations of any sample points that are required, (each sample
point can be described by its tangent angle and arc-length from an arbitrary start point).

The local curvature at any point on the outline of a shape can be calculated from the
'chain-code' directions from pixel to pixel, of an outline. By repeated averaging of
adjacent curvatures, a smooth graph can be drawn. Graphs of different forms may then
be compared.

Bookstein has suggested it is possible to compare forms by sampling the tangent angle
function, arc-length and tangent angle at landmarks and analysing these values by a
multivariate statistical technique.

A mathematical technique used to describe the entire facial surface shape and the
changes occurring in the face has been developed based on a classification system
inaugurated by Besel and Jain (1988). By decomposing the surface into fundamental
shape patches, an objective, quantitative and qualitative description of the face can be
produced. Each surface point on the face is classified as belonging to one of eight surface
types by computing values of the Gaussian and mean curvature, Coombes et al (1991).
Gaussian curvature is a measure of the curvature at a point on the space surface, given as
a ratio of the discriminant of the two fundamental forms of the surface, the first describes
the metric and the arc length and the second, defines the direction cosine of the normal to
the surface. Points on the surface may then be classified as flat, elliptic, parabolic or
hyperbolic. The mean curvature is the sum of the principal curvatures. These two
curvatures are independent and both are needed to describe a surface unambiguously.
 96

The advantage of this method of describing the surface is that it is independent of

orientation, rotation and displacement. Thus the description of the face will be the same
from any viewpoint.

In order to produce a classification, the signs of the Gaussian and mean curvatures are
used. These are computed by passing a local neighbourhood operator over a depth map.
The points on the face may then be colour-coded according to the surface type to which
they belong and a 'surface type1 image may be produced. As all data has some random
variation, it is necessary to set thresholds on the curvatures, below which the surface is
classified as flat.

For the clinician, the advantage of such a method lies in the fact that the facial surface is
described quantitatively but at the same time, retains a reasonable amount of
descriptiveness, allowing rapid appreciation of the differences between surfaces.

§4.4.4.2 Landmark data

Morphometries, the study of geometrical form of organisms, combines themes from

biology, geometry and statistics. Data for morphometric studies usually include
geometric locations of landmarks i.e. points that correspond biologically from form to
form.

There are two general methods for characterising landmarks, firstly those which locate
landmarks by the juxtaposition of different identifiable structures e.g. in fish, the
'anterior fin base' and 'posterior fin base' delimit the body outline. Secondly, those
located using geometric properties e.g. the tip of a tooth may be taken at the point where
the curvature of the edge is greatest. Landmark locations may be augmented by
information about the curving of external or internal boundaries between landmarks.

Two approaches to the investigation of group differences/associations in size and/or

shape or between size change and/or shape change are

1. Multivariate morphometries.
2, Deformation analysis.
 97

In multivariate morphometries, configurations of landmarks are measured one at a time

in collections of 'morphometric variables'. Some of these measures are size variables e.g.
distances between landmarks, whilst other variables have a value independent of
geometric scale, such as ratios of distances or other shape variables, as well as functional
transforms of these ratios.

Generally, Bookstein (1986) concluded that although ’size1 and ’shape’ are verbally
orthogonal, computationally and conceptually they are inextricably entangled.

The second morphometric tradition concentrated on the theme of deformation. This was
introduced into descriptive biology by Thompson (1961) under the heading of 'Cartesian
Transformation'. A deformation is a mapping which takes neighbouring points to
neighbouring points and which alters lengths of short segments by factors which never
get too large or too small. The notion is an informal version of what mathematicians call
a diffeomorphism: e.g. the reals and the interval (0 ,<») are diffeomorphically equivalent,
since the diffeomorphism

f: > (O.oo) : f(x) = ex

g: (0 ,°o) -> 9? : g(x) = logx

i.e. a one-to-one transformation which, along with its inverse has a derivative at every
point of a region and its image.

Most techniques for the geometric study of mappings, choose to model the configuration
of landmarks by a map from some algebraic simple family, and then interpret either the
coefficients of the fitted map or else, its distributed 'error of fit'.

Sneath (1967) expresses the Cartesian co-ordinates of the landmarks of one form by a
cubic bivariate polynomial in the x and y ordinates of the same landmarks in another
form. The resultant coefficients are impossible to interpret directly. Sneath’s purpose was
instead to summarise them in a single net measure of dissimilarity between forms.

In general these analysis suffer from the dependence on landmarks being readily and
unambiguously identifiable. On gently curving surfaces such as the human back or face
this is not the case, hence it serves as a constraint on the accuracy of the technique.
 98

Many papers have since suggested methods for the analyses of shape in two-dimensions
and almost all have been based on the movement of homologous landmarks, Bookstein
(1978,1984 a,b,1986), Siegel and Benson (1982).

A whole wealth of other techniques for describing shape and hence allowing
comparisons to be effected are summarised in O'Higgens and Johnson (1985) and include
medial transforms Blum (1973), and fourier analysis, Erlich et al. (1983).

§4.5 Summary

The advancement of technology has enabled spatial comparisons to move in leaps and
bounds over the past two decades. Prior to this time, techniques available were simplistic
and the numerical measures defined were invariably difficult to interpret and information
was ignored.

Major strides have since taken place especially within the field of image analysis with
research initially focusing on the development of good restoration algorithms. A by­
product has been an attempt to develop statistics which express the deviance from the
’true' picture and hence allow the performance of algorithms to be assessed and
compared. Ideas based on distance metrics have been the main contributors in this area.

Within the field of medical imagery, particularly reconstructive surgery, interest has
focused on change through surgery or growth. The use of landmarks has been paramount
in this field but the question of selection of such points has raised a number of questions
as to the suitability and robustness of this approach. Other methods such as the idea of
curvature have been suggested to describe a surface, however "for comparison a
subjective-based ordering mechanism is used to discriminate between images.

Much work has been done in the field of shape analysis to describe surfaces as diverse as
drainage basins, central business districts and human faces. These measures have been
used to assess similarities/differences between the variables of interest. The major feature
common to all these methods is that they are invariant to translational, rotational and
scaling differences. In the field of environmental sciences this aspect of change
potentially enables the scientist to diagnose the process which is responsible for the
change. The next two chapters will concentrate on the development of test statistics
 99

based on these three modes of transformation to examine the question of change or

similarity.
 100

CHAPTER 5

T H E CHAR ACTERTSATTON OF SPAT!AT, CHANGE

§5.1 Introduction

Within this chapter, the framework of the methodology used to describe change between
two spatial processes represented as contoured surfaces, is developed. The contoured
surface being produced as described in chapter 3.9. Although the procedure is explicitly
formulated in terms of expressing change, it is equally applicable for investigating
associations between spatial processes.

Subjectively change may be described by evaluating those features of the contoured

surface which the eye perceives as having changed when the two surfaces are
superimposed. These changes will invariably be summarised in terms of one or all of the
three transformations: -

1. Scalar
2. Translation
3. Rotation

Figure 5.1 illustrates five simulated examples of two contours where various
transformations have been imposed. The first shows two identical contours i.e. the shape,
size and spatial location are coincident. Figure 5.1(b) is an example of two contours
relating in spatial locality but differing in size. A change in the orientation of one of the
contours has occurred due to some external process in figure 5.1(c). The penultimate
diagram describes the situation where contour A has been displaced with respect to
contour B. Finally figure 5.1(e), illustrates the most complicated scenario where
translational, rotational, scalar and shape change have all resulted. The value of
undertaking a statistical procedure to assess change of this magnitude is questionable due
the primary differences which exist between the two surfaces.
 101

Figure 51(a) No change. Figure 5.1(b) Scalar change.

^ \
/ \ / \
/ \
J— i

Q _D ''s> J

Figure 5.1(c) Angular change. Figure 5.1(d)L ocation shift.

KEY

------------ SURFACE A
------------- SURFACE B

Figure 5.1 (e) Complex change.

Figure 5.1 Illustration of the three transformations.

 102

Three stages were required for the formalisation of a set of test statistics which would
allow the various transformations to be quantified in terms of potential change statistics

1. The definition of various geometric properties of a contour i.e. area, perimeter, centre
of gravity and orientation, A simulation study was performed to examine how the
variability of these measures was affected by changing the fineness of the underlying
mesh.

2. Based on these quantities, potential statistics were examined to assess which enabled
the three transformations, scalar change, rotation and translation to be quantified
satisfactorily. A whole series of plausible test statistics can be listed, ranging from those
involving ratios or differences to those which are specific to a particular area of
application. The method of assessment for each of the proposed test statistics was based
on the following criteria

1. Comparability of results.
2. Bounds of -°° and +00 are unattainable in practice,
3. Robustness, i.e. ordering of the surfaces is irrelevant.

The first and third conditions are of particular relevance in constructing a statistical
distribution since the resultant distribution should be globally applicable.

3, Those measures deemed to describe most satisfactorily the three transformations

from part 2 were then formalised. These provided the basis of the test statistics for
describing change and their distributions are derived in chapter 6 .

§5.2. Various Contour Descriptors For Characterising Transformation

Parameters

§5.2.1 Scale

The most common form of transformation liable to be found in the environment is scalar
change i.e. expansion/shrinkage of a process. Examples include the decay of a
radionuclide and the deforestation of an area.
 103

The simplest criteria for describing scale is in terms of the relative size of the contour.
Two measures which describe the specific size of an object are area and perimeter.
These two measures quantify different aspects of size; area refers to the part of a two-
dimensional surface enclosed within a specified boundary or geometric measure, whilst
perimeter relates to the length of a 'curve' enclosing a region of a space. For simple
surfaces, such as a circle or rectangle, the two measures are related, figure 5.2.

*0

ss

90
CL x
N.
DC
fc?*o
UJ 20

£ tr xn
ilJ »
a '5
LU
Cl

10
5

0
0 50 100 150 200 !50 300 950 <00 0 50 100 150 200 250 300 350 *00 <50 500
AREA / PI AREA

Figure 5.2 (a) Circle Figure 5.2 (b) Rectangle

Figure 5.2 Relationship between area and perimeter for simple constructs.

In practice surfaces will be more complex hence a simple relationship as illustrated

above is unlikely to arise, although the assumption of independence may potentially still
be violated. This question is investigated further in chapter 6 in terms of a global
approach for detecting change.

§5.2.1.1 Area

The evaluation of an areal quantity is simply achieved.

Let Ay = area of contour j, in surface i

 104

where Ajj = JJdxdy Cjj - region of interest

cij

The analytical evaluation of the region of interest will invariably prove to be impossible
to achieve apart from for simple regions. A numerical procedure was therefore required
for quantifying area.

The area of any polygon represented as a vertex list may be calculated by summing the
areas of the trapezia under each side, down to the axis. The direction of the sides must be
taken into account, so that sides on the bottom of the polygon are subtracted from the
total, figure 5.3. Care must be taken to ensure the polygon is stored in an anti-clockwise
direction, if stored in a clockwise direction, the absolute value should be taken.

Figure 5.3 Evaluation of the area of a polygon based on the trapezia rule.

A- = |[(xijiyij2 + W ip + - + x<i(N-»y«N+ w J -

(xpy,! + xpyp + ••• + V i m + xsiy«N>3

j |T n- i r
= ,1 - (x»+i>yijk)} + [(wiii - xiiiyijN)]
 105

The major problem with this approach is that if the polygon is sited a long way from the
x-axis, the area of the trapezia will be much larger than the area of the polygon and
accuracy will be lost. Temporarily making one vertex the origin will avoid this problem.

§5.2.1.2 Perimeter

A similar formulation for perimeter is possible.

Let Py = perimeter of contour j, in surface i,

where Pjj = X { (x ijk-x wit+1))2 + (yijk- y u(k+]))2} + {(x^ - Xijl)2 + (y,N- y J 2p

k-1

§5.2.2 Orientation

One possible definition to be used in assessing rotation is the orientation of a contour i.e
the angle subtended by the major axis and the x’-axis. The origin being defined by the
ordinates of the contour's centre of gravity, C(XCGjj,YCGjj). The principal axis, PP', is
delineated by the line which describes the maximum distance between two points on the
contour's boundary, which lie on a straight line, pass through the centre of gravity and is
wholly contained within the bounds of the contour, figure 5.4

The theoretical derivation of the angle of orientation is simply a function of the distances
PC and AC. Empirically it is not so straightforward. The crucial step is to identify the
principal axis. To simplify the calculations the contour of interest should first be
translated so that the centroid becomes (0,0). The orientation of the figure will
accordingly be preserved.
 106

Figure 5.4 Definition of the angle of orientation.

One method for locating the principal axis is

1. Ascertain whether the mesh points, A, B where A^B, comprising part of the contour
boundary are in diagonally opposite quadrants. Simply checking the following truth
statement enables verification of the above statement:-

J(sn(xljk).NE.sn(x2j.k)). AND.(sn(yljk).NE,sn(y2jk))j => TRUTH where sn = sign

2. If step 1 is true, then the angles subtended by points A and B with the x-axis are
calculated, 0 ijA, 0 ijB, respectively, where:-

modulus*; tan' S = A, B.

3. if e1JA= ep, then the two points lie at the extremes of a straight line which passes
through the origin and the distance djjg may be calculated.

4, Steps 1 to 3 are repeated for S=1,N where N = Number of interpolated

contour points.
 107

5. The angle of orientation is the angle which subtends the maximal distance between
two end points:-
ORj = {8 ijs/max(dijs)}

A number of problems arise with this technique. Theoretically the test 0ijA = 0ijB is
valid, but in practice is unworkable. Two reasons for this are:-

1. Inaccuracies due to rounding errors. When working in real space it is seldom possible
to evaluate two numbers which coincide to more than two or three decimal points.

2. Theoretically a closed contour is defined to be continuous, in practice the ordinates of

a contour's boundary are only known at discrete points, the mesh intersections. The finer
the mesh, the more valid the assumption of continuity. This feature complicates the
location of the principal axis, figure 5.5. The 'theoretical true' principal axis may not
necessarily be described by the defined mesh points, hence physically it will be
impossible to locate the true principal axis.

CA- 0
J«

---------- True principal axis

Empirical principal axis

Figure 5.5 Problem concerning the Figure 5.6. Definition of orientation

theoretical and empirical definition of bounds,
principal axis.
 108

The points which lie closest to the 'true' empirical principal axis may not always be
strictly diametrically opposite. The simplest method to account for these discrepancies,
due to the discretisation of the contour's boundary, is to define bounds for the angle of
orientation, figure 5.6

0
eijA—(e*A-«2)
iT — q

ijA
. A ^ ijA

T~

where - mod ulusJ tan"1

V X ij(A+l) J

r \
= modulus tan -i y ij(A-l)
V X ij(A-l) J

By defining the angle of interest in terms of a closed interval, we are theoretically

eliminating the problem of discretisation by defining the contour boundary in terms of
'continuous' segments.

Inherent within the definition of the principal axis was the assumption that the axis was
contained wholly within the bounds of the contour. The simplest way of verifying this
assumption is not violated is to perform the ray test i.e. only one value of djjg is
calculated for each mesh point; where more than one value results, the ray test will be
negative and the combination of points do not define the major axis. Computationally,
time is increased since it is not possible to stop the procedure once a diametrically point
is located, (N-S) calculations have to be performed, for S=l, N.

A final problem is where the contour of interest is almost cyclic in definition. Definition
of the angle of orientation will then present problems, unless the cyclic property is
recognised at the subjective stage.

§5.2.3. Centre of gravity

One possible measure which may form the basis of translational movement is that of the
displacement of the centre of gravity.
 109

Analytically the centre of gravity may be found by:-

XCGy = ~:JJxf(x,y)dxdy
Mc

YCG5 = ^:jjyf(x,y)cixciy
M c

where M = JJf(x,y)dxdy
Cij
Cy - the region of interest.
XCGy - x-ordinate of the centre of gravity
YCGy - y-ordinate of the centre of gravity

As for area, it is seldom feasible to define the region Cjj mathematically, however a
simple numerical procedure will enable the critical point to be located.

XCGii = f

YCG« = i f f * * " ]
where

x v = X [ (V -XijkKy-ok—Vijn)+ (xijk - x ijN)(yiik- y liN)]{(xiik + xiJ(k. 1)-)}

k-2
N
YV = I [ ( x ijN- x ijk)(yijk- y ijN) + (xijfc- x ijN)(yijk - y ijN)]{(y1Jk+ yl)(t.„)}
k-2
Ay = area for contour j in surface i (§5.2.1.1)

§5.3 Variability Of The Contour Descriptors

Section 3.8 mentioned the importance of mesh resolution in influencing the results for
area and perimeter for a very simple example. In this section, the level of variation
 110

attributable to different levels of grid resolution, in each of the four primary statistics is
examined. The reason for such an investigation is that for comparative work it is
essential the level of noise resulting from the surface fitting procedure is minimised, and
where data sets differ in size, the noise contributions should be of a similar order. A
simulation study was undertaken to examine the level of grid resolution which both
minimised the error due to the surface fitting technique and secondly did not increase the
computational time for minimal gain in accuracy.

§5.3.1 Simulation study

A study was performed on three sizes of data set 50, 100 and 150 points, these were
believed to be representative of sizes likely to be encountered in environmental
applications. Each set of observations were generated by simulating a bivariate normal
distribution, using the NAG (1984) subroutine G05EAF:-

f (x) = (27t)"1|l | 2 e x p |- i ( x - f ;i)TS~1( x -p ;)J

where \i = (0, 0)T

The fixed kernel method of density estimation was used for surface evaluation, the kernel
function was defined to be that of the standard multivariate normal density function and
the smoothing parameter was evaluated using least squares cross-validation.

The measure used to define the levels of variability introduced into the contour
descriptors as a result of the mesh resolution was the coefficient of variation

standard deviation
mean

The coefficient of variation is dimensionless hence it enables the relative variability

between the data sets to be compared.

The effect of mesh resolution on the contour descriptors was investigated by examining
the effect of changing the number of grid points over which the raw data was
 Ill

interpolated to produce the surface. Based on a square grid, nine sizes of grids were
investigated; (222), (322), (392), (452), (502), (552), (632), (672) and (712). The
theoretical distribution of each of the primary measures was unknown, hence the
standard bootstrap, Efron (1979), was used to evaluate the coefficient of variation. The
bootstrap is an easily implemented although computationally intensive device which
allows the sampling distribution of some statistic to be estimated; this parametric statistic
frequently being a parameter estimate.

Suppose for a sample of N independent observations x 1,x 2,x 3,---,xN, we are required to
A

estimate a parameter 0. The underlying distribution, F, of the observations is unknown

but it is assumed to be both unimodal and well behaved. Let Fn be the empirical
probability distribution of F, having mass 1/N at each observed x[ (i = 1,.... ,N). A
random sample, herewith known as a bootstrap sample is drawn with replacement from

( 1)

and the characteristic of interest 0 * is calculated;

where (2)

Steps (1) and (2) are repeated S times to give:-

The mean and standard deviation of this sample can then be evaluated;

where

The more re-samples performed the closer the parameter of interest tends to its true
value. In reality the process has to be terminated at some stage, hence a compromise
 112

between time and accuracy has to be reached. The number of replicates performed for
this study was 250 since initial results indicated reasonable stability at this level.

The simulation study took the following form:-

1. Simulate data set x 1Jx 2,x 3,-*-,xN from a standard bivariate normal distribution.

2. Generate bootstrap sample ..,xN*.

3. Calculate the smoothing parameter using least squares cross validation.

4. Construct surface from bootstrap sample.

5. Evaluate area, perimeter, centre of gravity and orientation for the contour depicting
the upper quartile.

6 . Repeat steps (2) to (5) 250 times.

The usage of the standard bootstrap in step 2 resulted in the contour breaking down into
subsidiary components for some combinations of data. The problem was traced to one of
the bootstraps peculiar properties. Every value within a bootstrap sample is drawn from
the original data with replacement, some values will be repeated several times causing
the sample to become over-discretised. In addition the surface constructed using the
standard bootstrap appeared to deviate from the original data, especially for 50 data
points.

The simplest remedial action was to replace the standard bootstrap by its smoothed
counterpart, samples generated using this technique do not possess this property. Instead
of resampling with replacement from the original observations, x 1,x 2,x 3,---,xN, a non-
A
parametric estimate f h of the underlying density is evaluated and resampling is
A

performed from fh, the smoothed version. Effectively each point in the new sample is a
perturbation of the original data point randomly selected as before i.e. the bootstrap
A
sample is obtained by independently sampling the distribution f h with density:-
 113

1111 W l hbs J
The function K is assumed to be a bounded density function symmetric about zero and
with unit variance. The density is a kernel estimator of density, f, of the population F,
with smoothing parameter h^g.

The following procedure may then be used for evaluation of a smoothed bootstrap:-

1. Choose I uniformly with replacement from i = l .,N.

2. Generate e to have probability density function K.

3. Let Y = XI+he

If realisations Y are required to have the same first and second moment properties as
those of the observed sample then, by using the following transformation which scales
the kernel to have the same variance matrix as the data, this requirement will be satisfied

_ (x -X + h ^ e )
Y = X+ ~ y

Hall et al (1989) demonstrated that the benefits to be derived from smoothing will
diminish with increasing sample size, as N hte —> 0. This theoretical statement
confirms the behaviour witnessed in terms of the over discretisation problem which was
most noticeable for 50 data points. For 150 points, this problem was less serious.

The effect of using the smoothed version of the bootstrap as opposed to the standard
form is expressed diagramatically in figure 5.7. Once again a set of 50 data points were
simulated from a bivariate normal distribution, the resultant density being expressed in
figure 5.7(a). Figure 5.7(b) is an example of a resample using the standard bootstrap,
whilst the third diagram illustrates the smoothed bootstrap.
 114

Figure 5.7(a) Original data._________ Figure 5.7(b) Standard bootstrap.

Figure 5.7 (c) Smoothed bootstrap.

Figure 5.7 Illustration of the use of the bootstrap.

The question of what value of bootstrap smoothing parameter should be used is a

research topic in its own right. Within the simulation study a series of values were
examined. The sequence for the study was as before but with step two now reading:-

2. Generate smoothed bootstrap data set with smoothing parameter, h^s

An additional step was added at the end to examine the question of the smoothing
parameter for the smoothed bootstrap.
 115

7. Repeat steps (2) to (6 ) for the range of h^g under consideration.

The range of bootstrap smoothing parameters considered was determined by simulating

200 bivariate normal distributions and evaluating the smoothing parameters for each
separate distribution. Based on the evidence of Hall et al (1989), the range for the
smoothed bootstrap parameter was taken to be slightly wider than that of the smoothing
parameter for the original data. Increments of 0.1 were examined for the smoothed
bootstrap parameter. Table 5.1 collates these results.

Smoothing parameter for Bootstrap smoothing

the original data parameter
No. of points h (lower) h (upper) hbs (lower) hbs (upper)
50 0.20 0.65 0.15 0.80
100 0.21 0.57 0.15 0.8
150 0.22 0.55 0.15 0.8

Table 5 . 1 Range of smoothed bootstrap parameters investigated.

§5.3.2 Simulation results

§5.3.2.1. Area

Increasing the size of the data set has negligible effect on the coefficient of variability at
each level of grid resolution. However across levels a trend is evident; increasing the grid
fineness, the level of variability decreases, figure 5.8. Table 5.2 summarises the results
from figure 5.8 for the mean level of the original smoothing parameter, calculated from
the original 200 simulations used to produce the ranges for the smoothing parameter in
table 5.1.
 116

No. of Grid Size

points
222 322 392 452 502 552 592 632 672 7 i2

Coefficient of varialtion
50 1.340 0.742 0.561 0.461 0.398 0.340 0.300 0.250 0.278 0.235

100 1,340 0.801 0.616 0.465 0.398 0.324 0.313 0.275 0.250 0.241

150 1.279 0.789 0.545 0.465 0.385 0.335 0.328 0.283 0.275 0.220

Table 5.2:- Results for coefficient of variation for area.

The table indicates the presence of a strong exponential trend with a tailing off of
variability occurring for grid levels of 63 and upwards. In terms of related data sets size,
this corresponds to 80,40 and 26 times the original sizes for 50,100 and 150 data points,
respectively. In real terms variation of the order of approximately 0.3% is introduced
when interpolating over a grid of 4000 mesh points.

For 50 data points, a bootstrap smoothing parameter of less than 0.2 produced fairly
unstable results, this behaviour was explained by the findings of Hall et al. (1989) who
advocated a suitable value for the smoothed bootstrap should be calculated using the
original data set, the lower bound for 50 data points, from our simulations, table 5.1,
being 0.2 hence the instability. For values greater than 0.2, a remarkable degree of
consistency was established across all values of h^s.

If interest is solely in area, then a mesh of 4000 points and a value for h^s selected equal
to that calculated for the original data will produce a high degree of accord across
surfaces, Variability of the 0.3% will be introduced into any ensuing analysis. This will
invariably be masked by the other forms of error.
 117

5>* 1
.
4
3
b « b b ’*
5 '• *
2 t..
u
p u.

4
.
4
§

BOOTSTRAP SMOOTHIMG PARAICTER bootstrap smoothing paraicter

BOOTSTRAP SMOOTHING PARAMETER

Figure 5.8 (a) Coefficient of variation for data set comprising 50 points.
».»

5 i. i s
<
5
9
> 9>•
5 b
U
s-ClI K
S'
I..
bootstrap stoth ing paratcter
bootstrap smoothing paraicter
BOOTSTRAP SMOOTHING PARAICTER

Figure 5.8 (b) Coefficient of variation for data set comprising 100 points

5
s
5
b
S A* ■- A»
3
sCl ci
u.

Ik

Sab
ci

At Al At At A* At At At At

BOOTSTRAP SKXJTHING PARAICTER

bootstrap SMOOTHI * paraicter

Figure 5.8 (c)C oefficient of variation for data set comprising 150 points.

KEY KEY KEY

— 2 2 x 22 ---------- 45x45 ----------- 59x59
- - 32x32 ------------50x50 ------------63x63
39x39 55x55 ---------- 67 x 67
..............71 x 71

Figure 5.8 Area results for coefficient of variation.

 118

BOOTSTRAP smoothing parameter BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAItTER

Figure 5.9 (a) Coefficient of variation for data sets comprising 50 points.
IB IB !■
1 5
\
5“
IB £ .\ «- 9
IS
r* 1 ;;
ufb &u s &US
LB
5 um 5 »•»
us u us -------- ------ u us
LB
jh '-** jh u "
in 3 in 3 in
IB . fc* ua
IS IS us
IN

BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARA/tTER BOOTSTRAP SWOTHING PARAfCTER

Figure
•w 5.9 (b) Coefficient of variation
TV
for data sets comprising 100 points
gw 5 **
^ IB
<
3>• L«
5 “h m
>
& US 6 u*

su i. a
B u*
G i. a

^ in
IB
ta

BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAICTER BOOTSTRAP SMOOTHING PARAICTER

Figure 5.9 (c)C oefficient of variation for data sets comprising 150 points.

KEY KEY KEY

— 2 2 x 22 45 x 45 ----------- 59x59
- - 32x32 - - 50x50 ------------63x63
39x39 — 55 x 55 ..............67 x 67
---------- 71 x 71
Figure 5.9 Perimeter results for coefficient of variation
 119

§5.3.2.2 Perimeter

The pattern of the levels of variability due to the surface fitting regime for perimeter
differ to those of area, figure 5.9. In terms of the original data set size, a considerable
difference is expressed between the results for 50, and 100 and 150 points. Table 5.3
reports the results for the mean value of the original smoothing parameter:-

No.of Grid Size

points
222 322 392 452 502 552 592 632 672 712

Coefficient of variation
50 1.87 2.54 1.84 2.02 2.05 1.90 1.56 1.78 2.42 2.53

100 1.53 1.76 1.49 1.46 1.33 1.36 1.49 1.56 1.73 1.61

150 1.60 1.64 1.53 1,44 1.47 1.32 1.47 1.64 1.51 1.64

Table 5.3 Results for the coefficient of variation for perimeter.

Firstly, across grid resolutions differences of 20% and upwards are reported between the
results for 50 data points and the other two data sets. Secondly, a trend emerges in terms
of the smoothed bootstrap parameter, the results do appear to stabilise from the apparent
mean value of the smoothing parameter upwards. This levelling off of variability is
masked by fluctuations in the results, particularity for the coarser grids. Increasing the
number of simulations may smooth out of some of these fluctuatiuons.

A final feature of these results is the consistency of the results for all the values of mesh
resolution considered, the range being slightly tighter for 150 points than for 100 points.
In terms of selection of grid size, the variability in the perimeter results indicates that the
effect on perimeter need not be considered, as long as the value of bootstrap smoothing
parameter selected, lies close to the value of smoothing parameter determined from the
original data.
 120

In terms of the overall variability, approximately 1.5% variation is seen in the results,
five times that for area. The reason for this five-fold increase may be explained by
considering fractals.

Figure 5 . 1 0 Variation in perimeter for a given area.

Figure 5.10 displays two shapes which have the same resultant area but whose perimeters
differ considerably. These two shapes may easily be the result of two simulation runs. It
is this feature of fractals which contributes to the differences in levels of variability for
area and perimeter,

§5.3.2.3 Orientation

For the remaining geometric measurements, orientation and centroid displacement, the
variability introduced into the results due to the surface fitting procedure was examined
for each of the grid resolutions, but the various levels of bootstrap smoothing parameter
were not examined. Computationally this procedure was extremely intensive. Utilising
the results for the earlier measures, area and perimeter, the evidence strongly supported
the case for choosing the smoothing parameter, h^s, to take the same value as evaluated
for the original data.

Figure 5.11 illustrates the results for orientation. Orientation appears to react in a similar
manner to that of perimeter. Changes in grid resolution have a minimal effect on the
level of variability introduced into the system for 100 and 150 points and for grids above
32x32 for 50 points. Increasing the sample size sees a marked reduction in the coefficient
of variation between data sets of 100 and 150 points and 50 points.
 121

0.8

0.7
2T
O
f— KEY
$
ou_ - 50 points

" 100 points

- 150 points
0.0
0 <0 30 30 40 50 00 70

GRID RESOLUTION

Figure 5.11 Coefficient of variation results for orientation.

One possible explanation is that for larger data sets, the contour shape is more clearly
defined hence less variation will be introduced. Approximately half of the variability is
seen in 100 points compared to that for 50 points, with a further reduction of a sixth for
150 points.

Although not reported, increasing the correlation from 0.5 to 0.9 resulted in a marked
decrease in the coefficient of variation, particularly for 50 points. This raises the question
of whether orientation should be used as a descriptor particularity for small data sets
where orientation may be ill-defined. Generally there will be little change in the angular
direction but where it does occur, the two contours to be compared will possibly be
analogous in shape hence changes will be easily identifiable. There will be situations
where the definition of orientation is questionable, as mentioned previously e.g. cyclic
contours.

For orientation to be a satisfactory descriptor, the analysis of large data sets is required to
ensure variation introduced as a result of the surface fitting technique is minimised, so
that it can be separated from change due to some unknown underlying process.
 122

§5.3.2.4 Centroid Displacement

The effect of mesh resolution on centroid displacement is shown in figure 5,12,

effectively it is similar to that of area. Changes in grid resolution affect the level of
variability, whilst changes in the size of the data set has no real effect.

KEY
DC

- 50 points
LU

Ll_
- 100 points
U_

0 .3 - 150 points
0. 6
0 10 20 30 50 60 70
G R ID R E S O L U T IO N

Figure 5.12 Coefficient of variation results for centroid displacement.

As for orientation, considerable variation is seen in the results. Potentially this may be a
result of the variability introduced from evaluating the centre of gravity, since for the
evaluation of both orientation and centroid displacement it is required to be evaluated.

§5.3,2.5 Smoothing Parameter

Although not strictly a geometric measurement of interest, it was decided for

completeness to examine the level of variability introduced into the selection of the
smoothing parameter when the bootstrap smoothing parameter and data set size were
varied, the results are presented in table 5.4 and figure 5.13.

50 points 100 points 150 points

Coefficient of Variation 0.100 0.087 0.072

Table 5.4 Coefficient of variation in the estimated smoothing parameter, h.

 122

§5.3.2.4 Centroid Displacement

The effect of mesh resolution on centroid displacement is shown in figure 5.12,

effectively it is similar to that of area. Changes in grid resolution affect the level of
variability, whilst changes in the size of the data set has no real effect.

2U 2.2
CD
£ 2.0
KEY

50 points
IU

ti­
- 100 points
ll-
aO
- 150 points
0.6
0 10 20 30 40 SO 60 70
G R ID R ES O LU T IO N

Figure 5.12 Coefficient of variation results for centroid displacement.

As for orientation, considerable variation is seen in the results. Potentially this may be a
result of the variability introduced from evaluating the centre of gravity, since for the
evaluation of both orientation and centroid displacement it is required to be evaluated.

§5.3.2.5 Smoothing Parameter

Although not strictly a geometric measurement of interest, it was decided for

completeness to examine the level of variability introduced into the selection of the
smoothing parameter when the bootstrap smoothing parameter and data set size were
varied, the results are presented in table 5.4 and figure 5.13.

50 points 100 points 150 points

Coefficient of Variation 0.100 0.087 0.072

Table 5.4 Coefficient of variation in the estimated smoothing parameter, h.

 123

s S
< <

%
g •><*
b h
Bl I t
i
U

Li 0
.
1
BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAMETER

Figure 5,13 (a) 50 points Figure 5.13 (b) 100 Figure 5.13 (c) 150
points points

K E Y M e s h Resolution ---------- 222 452 632

Figure 5.13 Results for coefficient of variation for the smoothing parameter.

When evaluating a surface, it is the original localities of the points which are of
importance for selecting the smoothing parameter, when using least squares cross-
validation, hence the superposition of the results for different levels of mesh resolution.
In terms of different values for the smoothed bootstrap, a similar pattern emerges to that
of before. For the range of values which were defined for the original 200 simulations,
the results were extremely consistent, particularily for the lower quartile upwards. This
explains the sharp increase in the coefficient of variation for 50 points for valuesof the
smoothed bootstrap less than 0.47, the lower quartile,and 0.39 for 100 points and 0.35
for 150 points.

§5.3.2.6 Summary

Returning to the original very simple example, section 3.8, which illustrated the
importance of grid resolution, the results obtained in terms of the strong dependence of
the areal results on grid size and the minor role it plays in influencing perimeter
variability was confirmed.

The size of the data set was of crucial importance in terms of variability introduced into
the perimeter results. In practice comparability of data set of 100 points and upwards of
varying or equal size poses no problems. However for smaller data sets, comparisons
between these and large data sets will display a lack of concordiality in variability due to
the surface fitting mechanism. It is physically impossible to reduce the level of
 123

5 o.ts

0
.
19 x 0,10

0.0 9,1 Os* 0.9 0.* 0.5 0 .9 f tf fcl

BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAMETER BOOTSTRAP SMOOTHING PARAHETER

Figure 5,13 (a) 50 points Figure 5.13 (b) 100 Figure 5.13 (c) 150
points points

K E Y M e s h Resolution ---------- 222 ------------- 452 632

Figure 5.13 Results for coefficient of variation for the smoothing parameter.

When evaluating a surface, it is the original localities of the points which are of
importance for selecting the smoothing parameter, when using least squares cross-
validation, hence the superposition of the results for different levels of mesh resolution.
In terms of different values for the smoothed bootstrap, a similar pattern emerges to that
of before. For the range of values which were defined for the original 200 simulations,
the results were extremely consistent, particularily for the lower quartile upwards. This
explains the sharp increase in the coefficient of variation for 50 points for values of the
smoothed bootstrap less than 0.47, the lower quartile, and 0.39 for 100 points and 0.35
for 150 points.

§5.3.2.6 Summary

Returning to the original very simple example, section 3.8, which illustrated the
importance of grid resolution, the results obtained in terms of the strong dependence of
the areal results on grid size and the minor role it plays in influencing perimeter
variability was confirmed.

The size of the data set was of crucial importance in terms of variability introduced into
the perimeter results. In practice comparability of data set of 100 points and upwards of
varying or equal size poses no problems. However for smaller data sets, comparisons
between these and large data sets will display a lack of concordiality in variability due to
the surface fitting mechanism. It is physically impossible to reduce the level of
 124

variability for 50 points to that for 100 points. Realistically we are talking of variation of
the level 1% to 2% in the results, this will generally be of lesser importance than
variability due to physical measurements.

Returning to the areal measurement, it is this one measure which controls the level of
grid over which to interpolate. A grid comprising 4000 points appears to minimise the
levels of variability and hence ensure comparability.

For orientation and centroid displacement, the bootstrap smoothing parameter was
selected to be equal to that for the original data. Orientation responded in a similar
manner to that of perimeter i.e. data set size was the most influential factor in
determining the level of variation due to the surface fitting technique whilst for centroid
displacement, grid resolution was the controlling factor.

The other aspect which emerged during the examination of grid resolution concerned the
choice of smoothing parameter when executing the smoothed bootstrap. An interval for
the range of smoothing parameters calculated using least square cross-validation was
constructed by a simulation study. By selecting any value which was contained within
the interval which has as its lower bound the lower quartile, consistency was achieved.

In practice, a simulation study may be performed on the data sets of interest and the
mean value taken for the smoothing parameter for the smoothed bootstrap.
Computationally this is not excessive, unless the data set of interest is very large.
Selecting the same value as for the original data set will be satisfactory, unless a high
degree of accuracy is necessitated.

§5.4 Simple Statistics For Describing Change

The previous section examined the effect of grid resolution on various geometric
descriptors and secondly, the level of bootstrap smoothing parameter required to
minimise variability from the surface fitting procedure to ensure comparability was
achieved across surfaces, generated from data sets of possibly differing size. We now
move on to consider the comparison of two surfaces, constructed using the methods
described in chapter 3.8.
 124

variability for 50 points to that for 100 points. Realistically we are talking of variation of
the level 1% to 2% in the results, this will generally be of lesser importance than
variability due to physical measurements.

Returning to the areal measurement, it is this one measure which controls the level of
grid over which to interpolate. A grid comprising 4000 points appears to minimise the
levels of variability and hence ensure comparability.

For orientation and centroid displacement, the bootstrap smoothing parameter was
selected to be equal to that for the original data. Orientation responded in a similar
manner to that of perimeter i.e. data set size was the most influential factor in
determining the level of variation due to the surface fitting technique whilst for centroid
displacement, grid resolution was the controlling factor.

The other aspect which emerged during the examination of grid resolution concerned the
choice of smoothing parameter when executing the smoothed bootstrap. An interval for
the range of smoothing parameters calculated using least square cross-validation was
constructed by a simulation study. By selecting any value which was contained within
the interval which has as its lower bound the lower quartile, consistency was achieved.

In practice, a simulation study may be performed on the data sets of interest and the
mean value taken for the smoothing parameter for the smoothed bootstrap.
Computationally this is not excessive, unless the data set of interest is very large.
Selecting the same value as for the original data set will be satisfactory, unless a high
degree of accuracy is necessitated.

§5.4 Simple Statistics For Describing Change

The previous section examined the effect of grid resolution on various geometric
descriptors and secondly, the level of bootstrap smoothing parameter required to
minimise variability from the surface fitting procedure to ensure comparability was
achieved across surfaces, generated from data sets of possibly differing size. We now
move on to consider the comparison of two surfaces, constructed using the methods
described in chapter 3.8.
 125

§5.4.1 Scalar comparators

Three of the more intuitive comparators based on area, perimeter and area/perimeter
combined are described in table 5.5.

Area Perimeter Area/perimeter

Ratio
A y /Py
iu . p ii

A 2j A 2 j / P 2j

Difference A y ”A 2j
V P2, A yP sj “A 2]P y

A 2j-

Proportion
^ li A y (P y + P 2 j)
A y 4- A 2j
Py (A y +A 2j )

Figure 5.5 Operators based on area and perimeter for performing a scalar comparison.

Similar comments are applicable to area, perimeter and area/perimeter, only those
properties relating to area will be discussed in detail.

§5.4.1.1 Ratio

The first of the statistics was that of the ratio of the two areas of interest. In terms of a
good comparator, ratio concurs with the three definitions

1. For differing areal sizes, the results are scalar independent. :

2. V C{j}such that Ay > 0, where Cy is the region of interest, then

if A. = A~
*} A
AP = —> + » if A^ » A2j where AR = —
0 ifAlj« A 2j Asj

From the definition of Ay 0 and 00 are unattainable in practice.

 125

§5.4.1 Scalar comparators

Three of the more intuitive comparators based on area, perimeter and area/perimeter
combined are described in table 5.5.

Area Perimeter Area/perimeter

Ratio
A j.
A i j / P ij

A 2j P 2, A 2 j / P 2j

Difference A lj ” A 2j P U ~ P 2i A P —A P
1j 2j 2j lj

A 2j

Proportion
A li PU A lj ( P lj + f*2j )
A l j + A 2j P l i + P 2j
P lj ( A lj A 2j )

Figure 5.5 Operators based on area and perimeter for performing a scalar comparison.

Similar comments are applicable to area, perimeter and area/perimeter, only those
properties relating to area will be discussed in detail.

§5.4.1.1 Ratio

The first of the statistics was that of the ratio of the two areas of interest. In terms of a
good comparator, ratio concurs with the three definitions

1. For differing areal sizes, the results are scalar independent.

2. V Cy^ such that Ay > 0, where Cy is the region of interest, then

if A.,
"j = A,£
A,,
A r — - » +oo if A„AJ » A „*=0i where AR = — -
A.,
if A„ « A2j

From the definition of Ay 0 and 00 are unattainable in practice.

 126

Ay 1
3. — - = -------— => ordering is irrelevant to the outcome, since the reciprocal
A Jj A 2 j / A lj

enables the same result to be achieved, when the two surfaces are transposed.

§5.4.1.2 Differences

For the difference operator all three conditions are not satisfied.

1. A very simple example illustrates the failure of the first property

Aj = 200 units A2 = 100 units

A} = 2 units A2 = 1 unit

In terms of areal differences,there is a lack of consistency between the sizes of the

differences hence comparability is impaired i.e. it is scale dependent.

2, V Cy such that Ay > 0

0 if Ay = A2j
a D = —> +oc i f A y » A 2 j where AD= A lj” A2j
—> -00 i f A y « A 2j

In reality no surface is infinitely large or small hence the bounds are unattainable.

3. Ay - A2J = jA2j - Ajjj => ordering of the surface is irrelevant, if the absolute value is
taken.

§5.4.1.3 Proportion

The idea of using proportion to assess the level of change has its foundation in
probability. The conditions are satisfied as follows
 126

Ai} l
3. — - ~ --------------- orderingisirrelevanttothe outcome, since the reciprocal
A-2j
enables the same result to be achieved, when the two surfaces are transposed.

§5.4.1.2 Differences

For the difference operator all three conditions are not satisfied.

1. A very simple example illustrates the failure of the first property

Aj = 200 units A2 = 100 units

Aj = 2 units A2 = 1 unit

In terms of areal differences,there is a lack of consistency between the sizes of the

differences hence comparability is impaired i.e. it is scale dependent.

2. V Cy such that Ay > 0

0 ifA ij= A 2j
Ad = —> + 0 0 i f A i j » A 2j where AD= A lj - A 2J
— > - 0 0 ifA ]j«A 2j

In reality no surface is infinitely large or small hence the bounds are unattainable.

3. |A jj - A2j| = |a 2J - Ay| => ordering of the surface is irrelevant, if the absolute value is
taken.

§5.4.1.3 Proportion

The idea of using proportion to assess the level of change has its foundation in
probability. The conditions are satisfied as follows
 127

An
1. -— satisfies the assumption of comparability.
+ A2.

2. V Cy such that Ay > 0

1
if Ay
2 L2j A
Ap = < i if Alj A2j where Ap=------ -
A u+ A 3j
—^ 0 ifA l; L2j

Although unattainable, the limits are more acceptable to field scientists.

A. A
3. -— = 1---------— => since the measure is based on a probabilistic definition,
Aij + A2J Ay + Ajj
the problem of ordering is avoided by subtracting one from one of the measures.

§5.4.1.4 Scalar summary

Both ratio and proportion measures are applicable for describing scalar changes, since
both satisfy the three conditions described in section 5.1. Returning to the two simple
areal regions described earlier, the circle and the rectangle, it can be shown that there is a
strong association between the two operators. By defining the original dimensions to be a
and b and r for the rectangle and the circle respectively and ka, lb and kr for the
dimension of the new surface, where k and 1 are real in definition, the relationship
between the measures are given in table 5.6.

Ratio Proportion
Rectangle 1 1
kl (1 +kl)
Circle _1_ 1
k2 (l+ k 2)
Figure 5.6 Relationship between scalar comparators for two surfaces.
 127

1. — — — satisfies the assumption of comparability.

A l j + A 2j

2. V Cy such that Ajj > 0

1
if A Xj - A 2-
2
A p —- —> 1 if A ^ » A 2j where AP=— ^ —
—» 0 Alj+ A 2j
>
A
A

to

Although unattainable, the limits are more acceptable to field scientists.

A. A
3. — = 1 -------- — => since the measure is based on a probabilistic definition,
A l j + A 2j A l j + A 2j

the problem of ordering is avoided by subtracting one from one of the measures.

§5.4.1.4 Scalar summary

Both ratio and proportion measures are applicable for describing scalar changes, since
both satisfy the three conditions described in section 5.1. Returning to the two simple
areal regions described earlier, the circle and the rectangle, it can be shown that there is a
strong association between the two operators. By defining the original dimensions to be a
and b and r for the rectangle and the circle respectively and ka, lb and kr for the
dimension of the new surface, where k and 1 are real in definition, the relationship
between the measures are given in table 5.6.

Ratio Proportion
Rectangle 1 1
kl (1+kl)
Circle 1 1
k2 (l+ k 2)
Figure 5.6 Relationship between scalar comparators for two surfaces.
 128

Realistically the numerical results for the ratio and proportion between two surfaces
differ only by a constant. Although, the surfaces of interest will tend to be more complex
in nature and associations between the operators may not be as straight forward as those
displayed in table 5.6, it was decided only to investigate the test statistics relating to
ratio. Similar comments apply to the test statistics when applied to the measure of
perimeter,

§5.4.2 Rotation

The simplest means to describe the change in the orientation of a contour after a time
period, t, or between two variables is in terms of an angular metric.

15.4.2.1 Ratio

The metric describing angular change as a ratio is

0 ,. fO < 0 .. < 2n
ORr = —- where <
02. \0 < O R r < ~

Although the form of the operator is dimensionless, the inclusion of infinity as one of its
bounds is undesirable. Unlike previous situations this bound is attainable in practice i.e.
when the contour relating to surface two is parallel to the x-axis. Although ordering of
the surfaces does not affect the results since the reciprocal provides an equivalent result,
when the two surfaces are transposed, in practice the concept of ratio is unworkable.

§5.4,2,2 Differences

The second form is that of differences

-271 < ORd < 271 non - symmetric contour

ORd = 02j - 02. where <- n < ORD< 7t symmetric contour
0 < 0 jj < 2k
 129

The three properties relating to ordering, dimensionality and the bound conditions are all
satisfied for angular differences.
Although unrealistic in practice, the bounds of the operator ORq have been split
according to whether the contour is symmetric.

For all work on angular change, measurements will be made in terms of radians as
opposed to degrees. The major difficulty in orientation lies in the problems incurred in
defining the principal axis, once these have been overcome the choice of operator is
unquestionable i.e. differences.

§5.4.3 Translation Descriptors

The final transformation is that of translation. Two possible statistics will be examined

1. The distance between two uniquely identifiable points in the area of interest i.e.
centroid displacement.

2. The area of overlap.

A shift in location is liable to arise in many different situations e.g. in the marine
environment, the transfer of oil slicks, saline or debris by the prevailing wind and tidal
conditions.

§5.4.3.1 Centroid displacement

Three possible means of evaluating displacement are

1. Standard displacement

2. Percentage displacement

3. Standard deviation displacement

Before evaluating displacement, rotating the contour of interest such that its major and
minor axis lie parallel to the x and y axis respectively, simplifies the ensuing analysis.
Displacement may be defined more simply for an isotropic surface i.e. one in which the
 130

variation is approximately constant over the surface compared to an anisotropic surface,

where variability is dependent on direction and therefore interest is in analysing the
directions separately. The two directional components are defined to be in the x and y
directions. Prior knowledge of a process may enable the true directions of variability to
be determined and displacement analysed accordingly. Table 5.7 identifies the form for
each of the descriptors based on the centre of gravity for each surface.

Standard Percentage Standard

Displacement Displacement Deviation
Displacement
Isotropic Surface Dro<T Undefined SDcog
Anisotropic
XCGtj -X C G 2j XCG13-X C G 2j
Surface XCG-n - XCG2i
major axis SD .
X- Direction

YCGjj - YCG2j YCGjj -Y C G 2j

Y- Direction Y C G u-Y C G a minor axis SDyj

where DC0G = ^(xCGy -X C G 2j)2 +(YCGlj - YCG2j):

t/(xCGh -X C G 2i)2 +(YCGlj - Y C o J

SD C O G “

SDxyj = pooled standard deviation in x and y direction, for contour level j.

SDxj = pooled standard deviation in x-direction, for contour level j.
SDyj = pooled standard deviation in y direction, for contour level j.

Table 5.7 Descriptors for the various forms of centroid displacement.

§5.4.3.2 Standard displacement

The first of the comparators suggested was standard displacement. Advantages associated
with this measure are

1. Ease of definition in x and y directions and if necessary x-y direction.

 131

2. Intuitively, it is easily interpreted.

3. The ordering of the surfaces is not a criterion which has to be considered since by
taking the absolute value the same result is obtained. The sign of the difference is an
indicator as to the direction of the change, this is not an issue at this juncture.

4. The achievable bounds are contained within the limits of and +°°.

The major problem with such a metric is that the measure is dimension dependent. For
some surfaces interest may be in small measures e.g. millimetres, whilst in other
examples, particularly geographical applications, we may be working in terms of
kilometres. This disadvantage unfortunately negates the usefulness of the metric as a
comparator unless some form of standardisation is performed.

§5.4.3.3 Percentage Displacement

A number of difficulties are encountered when defining percentage displacement:-

1. For isotropic surfaces, the definition of a combined measure is not intuitive. This
feature is not necessarily a major drawback, there are many situations where the analyst
is more concerned with examining the separate components of location shift.

2. The condition of robustness is violated, the major and minor axis are defined to be
expressions of one of the surfaces. Changing the baseline surface is liable to affect the
percentage displacement, unless the two surfaces are coincident in shape. Where interest
is in change across time periods then a natural ordering is imposed, however for
associations between variables, an ordering is much harder to implement.

3. A final problem associated with percentage displacement is concerned with the more
fundamental interpretation of the resultant value. A 50% shift in the x-direction may
relate to two units, whilst a displacement of 50% in the y-direction may be of the order
of ten units, depending on the size of the contour under question. In reality the two unit
shift may not be capable of being detected over and above the inherent natural variability
attributable to random noise in the spatial measurements, see chapter 6 .
 132

Linked to this problem is that of the size of the contour i.e. a displacement of 50%, say,
in one of the lower order contour levels is much less likely to be achieved due to the
bounds imposed by the region of interest than for a higher order level when the area
covered will be correspondingly less.

Although subjectively, we may state contour A has been displaced by 50%,

mathematically this presents a number of subtle complications. A way round this
problem is to measure location shift as a proportion of the standard deviation of the
contour.

§5.4.3.4 Standard deviation displacement

The last of the descriptors for translational change is standard deviation displacement.
The idea behind this potential test statistic is to express the level of change as a
proportion of the standard deviation of the data points defining the contour of interest.
Invariably, we will be comparing two contours which may be matched in terms of shape
or location, hence it is feasible to assume the standard deviation of the two contours are
similar so that pooling the variances is a valid assumption. It is under this premise that
the definition of standard deviation displacement is made. For an isotropic surface, the
variances in the x and y directions are pooled satisfying the property of robustness.

Using this formulation, it is hoped that consistency can be achieved in terms of

describing displacement between surfaces with differing dimensions. Further to the
definition of a good test statistic, the metric described is both dimensionless and the
achievement of infinite bounds is impossible, where the two surfaces exist.

§5.4.3.5 Overlap

The second translation statistic is the overlap function, kj, between two contours for
surfaces, i and i \ t say,at level, j, kj, can be expressed as

kJ = Cli 0 ufr‘i Where j = 1,NC

i,il = 1, Ns i^ i
where Nq = number of contours
Ng = number of surfaces
 133

Computationally, the method for evaluating kj is as follows

1. Define those points of Qj which are contained entirely within or make contact with

2 . Repeat step 1, transcribing Cy with Qj.

3. Combine the two data sets.

4. Order the co-ordinates sequentially in an anti-clockwise direction.

5. Evaluate the overlap area.

Two alternative forms exist for describing overlap, the coefficient of areal association,
Ca, and standardised overlap, Cs.

q - area over wk*ch the two phenomena are located together

a total area covered by the two phenomena

A
0 < C <1
A lj+ A 2 j ~ A c

_ area over which the two phenomena are located together

Co =
s baseline area
A
= 0<C <1
A li
Both descriptors range from zero to one with completely separate distributions giving a
value of zero whilst exactly coincident surfaces are described by a value of unity.

§5.4.3.6 Standardised overlap

The applicability of standardised overlap is restricted to situations where a natural

ordering may be imposed on the surfaces so that the definition of a baseline surface is not
 134

arbitrary. The other two conditions of, comparability of units and finite bounds are both
satisfied.

§5,4,3.7 Coefficient of areal correspondence

As for standardised overlap, the two properties concerning bounds and dimensions are
satisfied.

The main restriction of this descriptor relates to the special case of surface A being
contained entirely within surface B. The measure then reduces to the ratio of the areas of
the two surfaces and as a consequence imparts no information on the level of shift which
has been incurred.

This situation will not always arise and in instances where a shift is not wholly contained
within one of the contours, the metric Ca, reveals as much information about the level of
displacement as measures founded on centroid displacement and no prior assumptions
concerning the variance of the contour need be stated.

§5.4.3.8 Translation summary

Translation may be described in terms of the displacement of the centre of gravity or

alternatively by the amount of overlap. Of the three metrics described to analyse the
question of translation, only one turns out to be practically viable, standard deviation
displacement. Intuitively the concept of standard displacement is the most satisfactory
but the property of dimension dependence nullifies its global applicability.

Percentage displacement relies heavily on the natural ordering of surfaces to enable the
metric to be evaluated, this feature is not satisfactory especially where comparisons are
between different variables e.g. beta and gamma radiation levels.

The usage of standardised overlap as a possible test statistic is restricted by its reliance
on ordering, hence the coefficient of areal correspondence will be used for quantifying
overlap.
 135

§5.4.4 Summary

Development of any statistical procedures to assess the level of change between two
spatial processes will invariably necessitate some form of subjective pre-analysis. By
examining a simple overlay of the two surfaces, a judicious choice of comparators may
be selected.

Although this may be believed to bias the results since it is effectively performing a
comparison based on prior information, it may be stated that in all situations a test of
translation, scaling and rotation will be performed, it is only the format of the test which
is under consideration, all tests should produce equivalent results. The main question
which arises concerns the duplication of the test for scale by using the coefficient of areal
correspondence for testing for translation, where surface A is enclosed by surface B. In
this instance standard deviation displacement is a more appropriate measure. In all other
instances, the overlap metric is possibly more valid as prior assumptions concerning the
variability of the contour need not be made.

In terms of rotation, the choice of comparator is unambiguous i.e. differences in

orientation of the two contours of interest, whilst for scalar change, interest will focus on
the ratio of both area and perimeter.
 136

CHAPTER 6

A N H Y P O T H E S IS T E S T IN G A P P R O A C H T O S U R F A C E
C O M P A R IS O N

§6.1 Introduction

Within the previous chapter attention focused on a possible approach to describing

change. The motivation for this approach arose from the mathematical description of
those features which man perceives when assessing subjectively whether change has
occurred between surfaces represented as contours, which have been overlain. The
factors considered were area, perimeter, centre of gravity and orientation. Various
combinations of the above factors allowed different aspects of change to be quantified in
terms of rotational, translational and scalar change.

In terms of a scalar change, ratios, proportions and differences of area, perimeter and
area/ perimeter combined were discussed. Theoretically, the three measures, area,
perimeter and area/perimeter, are related hence the selection of potential test-statistic is
not clear-cut. From the ensuing work on mesh resolution, the behaviour of the
area/perimeter measure mirrored that of perimeter, hence only area and perimeter
formulations were used as the basis of the subsequent test statistics. The test statistics
selected for quantifying scalar change were based on the ratio of area and perimeter.
Differences between the areas or perimeters of comparable contours are inappropriate
since they lack global applicability, whilst measures formulated from proportions were
shown to differ from ratio, only by a constant.

For rotation, only one plausible descriptor was suggested, namely angular change i.e. the
difference between the angles of orientation for the contours of interest. Finally for
translation, overlap and centroid displacement were shown to be applicable in different
situations. Overlap generally being the better all round descriptor since no prior
assumptions are made concerning its usage, unlike centroid displacement where various
assumptions concerning the equality of the variances of the contours are made. Where
contour A is contained entirely within contour B overlap reduces to the ratio of the two
areas, hence to avoid the problem of correlation between this measure and that of areal
 137

ratio, centroid displacement should be used. Table 6.1 summarises the descriptors which
will be used as the basis of the test statistics.

Transformation (1) (2) (3) (4)

X C G j j - X C G 2j Y C G lj- Y C G 2 j
Translation S^COO overlap
SDXj SD yj

kj
Scalar
Ay. Jdi
A 2j p2j

Rotation 0 lj “ %

Table 6 .1 :- Summary of descriptors.

A series of hypothesis tests based on these descriptors are developed within this chapter
to examine the question of change between spatial processes. Interest will mainly focus
on a local form of analysis, by local we refer to specific contour levels.

It is well known that one of the main problems associated with environmental research
concerns the large amount of variation encountered in the measurements, whether it be
attributable to field measurements or alternatively, measurements taken in the laboratory.
In some cases, it is recognised that the level of variation in a measurement may result in
the available analytical capacity being exceeded. It also follows that if data are to be used
in a modelling exercise, realistic estimates of overall error must be given if a situation is
not to arise where the errors so compound one another that totally unrealistic results are
obtained.

Initially in this chapter we examine the simplest situation possible, no random noise is
present in the system. For this situation, the question 'Does contour A of surface 1, differ
from that of contour A of surface 2, according to characteristic x?' is examined.

Investigation of this question takes the form of a series of hypothesis tests using the
descriptors summarised in table 6.1 as the basis for the test statistics. The distribution of
the test statistics under the null hypothesis will be developed analytically, by way of a
 138

simulation study, since the theoretical derivation of the distribution of the characteristics
of interest is extremely complex with a large number of assumptions being required.

Following on from this, the distribution of the test statistic under various alternative
hypothesis is examined and from the results, a set of power curves are constructed.

The second part of the chapter will follow the same procedure but this time varying
levels of random noise will be introduced into the data, the levels of noise having upper
bounds of 5%, 15% and 25%. These three values being selected as representative of
noise levels which may arise in practice. Attention will focus on the ability of various
test statistics to distinguish between change due to some underlying process and that due
to inherent random noise.

Finally the results will be drawn together and a global approach to the analysis of change
will be proposed. Problems concerned with the implementation of both the local and
global approaches will be discussed and the limitations of the technique described.

§6.2 Simulation Study

A major simulation study was undertaken to derive the empirical probability distribution
function for the test statistics of interest, table 6.1. The null hypothesis, HO, takes the
form of a simple hypothesis and effectively describes the situation of no change,
Translating this to the parameters the null hypothesis are:-

1. Scalar HO : A^j = A2j

2. Rotation HO : Oy = 02j
3. Centroid displacement HO : =0
4. Overlap HO : area of overlap =100%

Before the power of the statistics under various alternative hypothesis could be assessed,
it was necessary to introduce various levels of known change into the analysis. The form
of the alternative hypothesis, HI, is composite and depending on the definition is either
one or two-sided, this is discussed in detail in section 6.3.2.
 139

§6.2.1 Scalar Change

For both area and perimeter, a similar procedure was implemented to introduce varying
levels of change into the statistic of interest. The scale of a contour may be changed by
multiplying each of the points by a constant, k, figure 6 .1.

Key

k = constant > 1
x’p= kxp
y'P= kyP

Figure 6 , 1 Scalar increase/decrease

When changing the scale of a pattern of points by a factor k, the points move k times as
far away from the origin, unless the origin is (0,0). Although theoretically this is not
undesirable, as it is the area and perimeter which are of interest, a number of practical
considerations arise. Firstly, the contour of interest may be shifted outwith the region of
investigation, resulting in a different contour being depicted as the i^1 quartile, which
would not necessarily be k times larger than the original contour as intended. The
simplest means of avoiding this situation is to first translate the centre of the original data
set to the origin (0,0) and then apply the transformation. The area/perimeter of the region
should then be transformed by the corresponding factor. Running, a test program to
validate this procedure confirmed its appropriateness.

A further simplification implemented was to describe contours circular in shape since

changes in areas and perimeters are simple to implement. This should not affect the
resultant distribution under the null hypothesis which is shape independent.

§6.2.2 Orientation

In terms of the metric involving orientation change was introduced and controlled by the
following process.
 140

Suppose P is a point with co-ordinates (xp,yp) and that it is rotated about the origin of
the co-ordinates through an angle 0. The resultant position of P, P* is given by:-

fcos 0 - s in 0 Y x n^
p * -

sin 0 cos 0

When rotating a pattern of points through an angle 0, the rotation swings about the origin
so that the pattern is moved elsewhere in space, as well as being rotated. This
displacement may be avoided by initially translating the pattern so its centroid is at the
origin (0,0) and then applying the rotation. As for the scalar metric, the assessment of the
test statistic was carried out via a simulated example whose initial orientation was
known.

§6.2.3 Centroid Displacement

A very simple transformation is required to achieve a known locational change:-

X* = X + cG x (X,Y) original data point

Y* = Y + cGy (X*,Y*) new data point

where c = constant of proportion

g . = standard deviation of the original data set in direction i, i=x,y

The only minor modification required was to shift the region of interest by the same
amount as the simulated data to ensure comparability of the contours of interest. By
neglecting to make this translation, the shifted contour potentially failed to correspond in
definition to the original contour unless coincident in size and shape.

§6.2.4 Overlap

Utilising the properties of a circle, the following method was devised to allow varying
levels of change in overlap to be evaluated. Changes in overlap when dealing with two
similar circles can be equated in terms of the difference between the sector area and the
segment area, $A, figure‘6.2.
 141

KEY

ZAOA' = ZAO'A' = 0 = overlap angle

AOA’B' = AO’A'B = SA= sector area
AB'A’C = ABA'C = $A = segment area
ABA'B' = Aov = overlap area

Figure 6.2:- Definition of angles and areas used to evaluate overlap.

The above definitions may be described mathematically as follows

SA= ^
A 2

$A = Sa -AAOA’

0r 2 If . 0 0
= 2 rsin—rcos—
2 2l 2 2

= i r 2(0-sin 0)
2

Aqy —2$A

Aov = r 2(0-sin0)

From figure 6.2 the critical parameter for varying the overlap is the overlap angle, 0. For
a specific level of overlap, the angle required to achieve this change for the coefficient of
areal correspondance is as follows:-

C„ =
ACi + ACj Aov

r 2(0 - s in 0 )
7tr2 + 7ir2 - r 2(0 - sin 0 )
 142

C, _ 9 -sin 9 0<C „<1

1+C, “ 2 jt 0< 9< 2tc

Figure 6.3 expresses the above relationship diagramatically.

100

110
120
X
to 130

MO

150

170

ISO
100 90 00 70 00 50 40 30 20 10 0
OVERLAP P R O P O R TIO N

Figure 6.3 Values of angular displacement, 0, to achieve required level of overlap for
the coefficient of areal correspondence.

A very simple angular transformation applied to the original co-ordinates allows the new
set of co-ordinates of the displaced circle to be evaluated.

^new ~ ^old + -^shift

^new =

where Xshift = 2 |r c o s ||

An additional check was included in the program to ensure the radii of the two contours
did not vary considerably and hence destroy the property of symmetry enjoyed when the
two radii are equal and on which the whole argument for controlling overlap was
founded.
 143

§6.2.5 Summary

The investigation of the various test statistics postulated in the preceding sections was
undertaken via a major simulation study. The form of the study was similar for the
various measures with the two situations of no random noise and random noise being
analysed separately.

The following sequence of steps were implemented to examine the effect of change on
the various test statistics.

1. Simulate data points xltx2,..., xN, where N = 50,100,150, using the NAG subroutine
G05EAF with

"0 . 0 " "1.0 k '

1 =
_0 . 0 _ _ k 1. 0 _

0.0 area
0.0 perimeter
where k = 0.9 orientation
0.5 centroid displacement
0.0 overlap

2. Generate a smoothed bootstrap sample Xj* x2* ...,x N*, where the smoothing
parameter is taken to be the value evaluated by least squares cross validation for the
original data set xlyx2,...,xN.

3. Apply the appropriate transformations to the bootstrap sample, table 6.2. Consider a
point P(xp,y p):-
 144

x-transformation y-transformation Conditions

H0 Hi
Area/ kxp kyP k=l k>l
Perimeter
Orientation xncos
P 0 - yJnsin
P 0 xpsin 0 + ypcos0 0=0 O<0<2tc

Centroid Xp+ca, Tp c=o c>0

Displacement

Overlap x„+ 2 jrco s —1 0=71 O<0<Jt

P 1 2j yP

Table 6.2 Transformations to point P(x ,y ).

4. Calculate the smoothing parameter for the bootstrap sample using least squares
cross-validation.

5. Construct the surface from the bootstrap sample using kernel density estimation and a
grid resolution of 63 X 63.

6. Evaluate area ratio, perimeter ratio, location shift, orientation and overlap for the
contour depicting the upper quartile.

7. Repeat steps (2)-(6) 250 times for the derivation of the empirical probability
distribution function under the null hypothesis of no change. For the alternative
hypothesis, 100 simulations were performed and the observed values for the appropriate
test statistics calculated.

The reduction in the number of simulations under the alternative hypothesis was forced
by the CPU time required to obtain the desired information. The resultant loss in
accuracy was not critical since it was the general shape that was of interest, the power
curves were possibly less smooth than might have been expected as a result.
 145

§6.3 Simulation Results

§6.3.1 Distribution of the test statistic under the null hypothesis

The form of the null hypothesis for all the transformations is simple and in accordance
with many applications is the hypothesis of no change.

§6.3.1.1 Area

The unsmoothed empirical probability distribution function for the test statistic relating
to the ratio of two areas is given in figure 6.4.

0.20
KEY
0.10
---------------- 50 points
0.12
LU
aU=5 ------------- 100 points
oun.i 0 .0 6

-------------- 150 points

0.00
0 .9 6 5 0 .9 9 0 0 .995 1.000 (.0 0 5 t.0 1 0 1.015

AREA R A T I O

Figure 6.4 Empirical probability distribution for areal change under the null hypothesis
of no change.

The most apparent feature is the consistency across the results for differing sizes of data
set. This was anticipated from the earlier work concerning the investigation of the level
of variability introduced as a result of the surface fitting procedure. In this case, data set
size did not appear to unduly influence the variability of the results.

A numerical summary of the empirical probability distribution function in the form of

the percentage points, q = 60, 80, 90, 97.5 and 99, is given in table 6.3, i.e.
 146

P(Aj>Aj(n;q))=q/100. The points tabulated correspond to area A2j > area A2j where the

A,,
test statistic is defined to be A. = — L.
1 A,.
q 50 points 100 points 150 points

60 1.00316 1.00353 1.00385

80 1.00584 1.00485 1.00654
90 1.00691 1.00646 1.00839
95 1.00889 1.00803 1.00904
97.5 1.009815 1.00979 1.01000
Table 6.3 Percentage points for the empirical probability distribution function relating
to areal change.

§6.3.1.2 Perimeter

Figure 6.5 illustrates the empirical probability distribution of the following test statistic
for perimeter: -

Pj= 3 0 < Pj < «»

J P2j

KEY

---------- =— 50 points
U 0.12

-----------------100 points
UJ 0.06

----------------- 150 points

i—

0.97 0.90 a 99 1.00 i.QI 1.03 1.03 I. 04 t. 05

P E R IM E T E R R A T I O

Figure 6.5 Empirical probability distribution for perimeter change under the null
hypothesis of no change.
 147

q 50 points 100 points 150 points

60 1.00551 1.00478 1.00398

80 1.01002 1.00704 1.00624
90 1.01623 1.00963 1.01138
95 1.02099 1.01607 1.01337
97.5 1.02929 1.01889 1.01412
Table 6.4 Percentage points for the empirical probability distribution function relating
to perimeter change.

Table 6.4 summarises the numerical values obtained for the percentage points described
for perimeter change based on the same criteria as for area i.e.
perimeter P^ > perimeter P2j. Increasing the sample size sees a corresponding reduction
in the range of the distribution function, with the tails of the distribution being shorter for
the larger data sets. The shape of the distribution remains constant for all three data sets.

§6 .3.1.3 Orientation

Under the null hypothesis of no change, the distribution of the test statistic is as
described in figure 6 .6 ..

0 .2 4

0.20
KEY
0.16

50 points

aLU o.06
cc — 100 points

0 .0 4

— 150 points
- 0 .1 0 .0 0 .1 0 .2

CHANGE IN O R IE N T A T IO N

Figure 6.6 Empirical probability distribution for orientation under the null hypothesis
of no change.
 148

q 50 points 100 points 150 points

60 0.04786 0.03877 0.02690

80 0.06770 0.07449 0.04333
90 0.08780 0.09621 0.05465
95 0.10333 0.09966 0.06046
97.5 0.12561 0.10580 0.06777
Table 6.5 Percentage points for the empirical probability distribution function relating
to the test statistic describing orientation.
The distributions for orientation are similar for various sizes of data set, however for 50
data points, an occasional more diverse result is recorded. From the plot, there appears to
be minimal bias in the overall distribution with the mode centred approximately at zero.
As for the earlier test statistics, the percentage points for the empirical probability
distribution function are defined in table 6.5.

§6.3.1.4 Centroid Displacement

The empirical probability distribution function for centroid displacement is illustrated for
the positive set of values, figure 6.7. We are only interested in the distance between two
points, not the displacements, the direction vector being analysed in terms of the
orientation of the contour. Although the value zero is plausible within the definition of
centroid displacement, within the numerical accuracy of the procedure it is never
achieved.

KEY

50 points

-- 100 points

0.00
0.1 0.2
150 points
CENTR0I0 DISPLACEMENT

Figure 6.7 Empirical probability distribution for centroid displacement under the null
hypothesis of no change.
 149

The distribution illustrated is equally applicable to the x, y and x-y distances since the
assumption of equality of variances has to be satisfied before the x and y distances can be
combined.

A difference exists between the distributions for 50, and 100 and 150 points. For the
former the results recorded are more diverse, with a shift of approximately 0.6 standard
deviations being recorded under the null hypothesis, at a percentage level of 97.5 and the
mode of the distribution occurring at 0.2 standard deviations, the maximum displacement
for 100 and 150 points is approximately one third less with the distribution peaking at
approximately 0,1 standard deviation. A numerical summary of the empirical probability
distribution, in the form of a set of percentage points is given in table 6 .6 . The table
being symmetric about zero.

q 50 points 100 points 150 points

60 0.35560 0.23950 0.19566

80 0.42299 0.29472 0.22753
90 0.45234 0.34955 0.27755
95 0.53601 0.35761 0.30000
97.5 0.60862 0.38940 0.34023
Table 6.7 Percentage points for the empirical probability distribution function
describing centroid displacement.

§6.3.1.5 Overlap

The empirical probability distribution function of the test statistic under the null
hypothesis of no change is described in figure 6 .8 , with the salient percentage points
tabulated in table 6.7.

The results reported display greater diversity for 50 points, than for 100 and 150 points,
suggesting that the measures of interest are more concisely defined for larger data sets. A
lateral shift in the distributions away from the optimal value of one results with
decreasing sample size.
 150

q 50 points 100 points 150 points

60 0.87185 0.90219 0.91391

80 0.91106 0.92295 0.93814
90 0.94053 0.93507 0.95754
95 0.95742 0.95045 0.96398
97.5 0.96436 0.96339 0.97271

Table 6.7 Percentage points for the empirical probability distribution function relating
to the test statistic describing overlap.

0.20 KEY
0.16

50 points
C_> 0.12
.ZT
UJ
aLU
q ;
0.06 100 points

150 points
0.00
0 .5 0.* as 0 .6 0 .7 0.6 0.9 1.0

O V E RL A P P R O P O R T IO N

Figure 6.8 Empirical probability distribution for overlap under the null hypothesis of
no change.

§6.3.2 Distribution under the alternative hypothesis

The alternative hypothesis, HI, can either be of a simple or composite form depending
on the nature of the analysis. Typically it will be of a composite form and will refer to
there either being change of any form i.e, a two-sided test or alternatively, the analyst
may believe that a specific type of change is most probable hence a one-sided test is
applicable.
 151

The standard situations of both one and two-sided tests are relevant for test statistics
pertaining to scalar and angular change. However for the test statistics describing
translation i.e. overlap and centroid displacement, only one form of the alternative
hypothesis is applicable i.e. one sided tests, since the statistics have one finite bound.

Once the alternative hypothesis have been stated, the relevant rejection regions, table 6.8
may be described.

H0 Hi Rejection region

A, = A 2j Aj. < A2.

jArA j < a(N,max(rn):q )|

Alj = A 2j A, > A2j. {AjiA. > A(N,max(rn):q)}

II
>

Alj * A2j
to

|Aj:nniax > A(N,max(iJ:q)j

< } Aj,
A

OR, - OR2j OR, < OR2j {oRjiORj < ~OR(N,max(i;);q)}

to&
Jp

OR, > OR2.

II
o

{OR^OR. > OR(N,max(i^);q)}

&
CO

O R ,*O R 2j
II
o

{OR/.IORJ > OR(N,max(rn);q)}

Plj “ Plj < P2j

| P3’P] < p(N,max(i;);q)}

Plj= Plj > P2j {pj:Pj < p (N>maxW ;q)}

Plj = P2j Plj * P2j
jp/.max f p . i l >P(N,max(r0);q)J

OVj = l o v , < o v 2j {oVpOVj < OV(N,max(i;);q)}

8

OD, * CD2j
II

{cD.ilCDj > CD(N,max(rn);q)}

where N = number of points
max(rn) = maximum level of random noise incorporated into the system
Table 6.8 Various forms of the alternative hypothesis.
 152

By tabulating the numerical values of the empirical probability distribution function for
the percentage points 60, 80, 90, 95 and 97.5, the critical values of the distribution for
various significance levels are easily interpreted from tables 6.3 - 6.7.

§6.3.3 Power of the tests

The ability of the test statistic to correctly reject the null hypothesis for various values is
described in terms of the power function of the test procedure.

Suitable choices for the alternative hypothesis were defined by initially testing a coarse
grid of values and on the basis of these results, the critical regions were examined more
closely. This was especially critical for assessing the test statistics based on area and
perimeter.

Small changes in the area or perimeter ratios are seen to have a major effect on the
resultant distribution. This extreme sensitivity of the test statistic may be shown to be a
result of the underlying distribution of the test statistic.

Asymptotically, it may be assumed that both area and perimeter have an approximate
normal distribution. On this premise and the assumption of independence between the
two measures, Geary (1930) showed the ratio of two independent normal variables,

v = —, has frequency function:-

y

1 |i A ! +H ,o,2v 1 (m -H yv)2
f(v) = exp<
V2K
(<Jx2+crV)2 2 c l 2 +o,V

For constant variance and small changes in the mean of the two distributions i.e. small
scalar changes, major shifts in the resultant distribution function occur for the ratio of the
two normal distributions. Figure 6.9 illustrates the empirical distributions for the ratio of
two normal distributions for 50 data points, whose variances are constant but means
differ by 2 and 4% respectively. This illustration confirms the apparent sensitivity of the
statistics to change. Using this information, an appropriate set of scalar changes were
investigated.
 153

i i i i h ---- (a) No change

+-------- +-------- +---------+---------+--------- (b) 2% mean change

+---------+-------- +-------- +---------+------------ (c)4%mean change

0.960 0.970 0.980 0.990 1.000 1.010

Figure 6.9 : Sensitivity of the scalar metrics.

For the remaining three statistics, the levels of change were constrained by their natural,
numerical bounds, i.e. for orientation (0,rc], centroid displacement (0,2a] and overlap,
(0,100%), where 0% is total separation and for 100%, contour A encompasses contour B.
For each of these ranges an initial coarse set of values was investigated; for overlap,
90%, 70%, 50%, 30%. Once the critical zones were located, additional points were
selected in that region.

§6.3.3.1 Area

Figure 6.10 describes the power curve for area. The areal constant, k, relates to areal
change in the following mannen-

( l+ (k x 0.002))2 where k> 0

In real terms, the range of change examined was up to 2.4%.

An areal expansion/shrinkage of greater than 2%, for a 95% level of confidence, will be
discerned to be attributable to some physical process and not simply a by-product of the
surface fitting technique. The results for 50,100 and 150 points are of a similar order.
 154

t.i
s.o
0.9

KEY

0.3 - 50 POINTS
£I* 0.4
aa .
0.5

0.2
- 100 POINTS
0 .1

0.0
0.0 0.5 1.0 1.5 3.0 3.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
- 150 POINTS
AREA CONSTANT flO

Figure 6.10 Power curve for area with no random noise.

§6.3.3.2 Perimeter

A similar procedure to that for area was performed. In this instance, k effectively
introduces changes of the order 0 to 3% i.e.

(1 + (k x 0.002)) where k>0

0.8
KEY

0.3 - 50 POINTS

- 100 POINTS
0 .1

0.0
0.0 2.5 5.0 10.0
- 150 POINTS
PERIMETER CONSTANT (K)

Figure 6 . 1 1 Power curve for perimeter with no random noise.

From the diagram of the power curves, figure 6.11, the results for 100 and 150 points are
reasonably similar, however the power function, for 50 points, shows that greater
 155

changes are required for change to be defined as statistically significant and not to be
solely a facet of the surface fitting procedure.

§6.3.3.3 Orientation

Throughout the chapter, examination of the various test statistics has been based on the
bivariate normal distribution. Depicting the distribution as a contour plot, the contours
are theoretically symmetric about the major and minor axis when no random noise is
present in the system. On this assumption the change in orientation should be assessed
7t
between (0,—) radians.
2

The response of the statistic for orientation is similar to that for area, angular change of
between 20° and 25° is attributable to the surface fitting procedure, for a power of 0.05.
Changes outside this range may be considered to be a result of external factors, figure
6.12

KEY
§
°* Ol S - 50 POINTS
0.3

0 .3
- 100 POINTS

0.0
0 S 10 IS » a JO 35
- 150 POINTS
CHANGE IN ORIENTATION

Figure 6.12 Power curve for the change in orientation with no random noise.

§6.3.3.4 Centroid Displacement

The resultant power curves are shown in figure 6.13. The measurement of centroid
displacement along the x-axis is defined in terms of standard deviations i.e. k=l,
describes a shift of one standard deviation.
 156

In terms of the metric describing centroid displacement, it is apparent that the larger the
data set, the more precise the results i.e. less variability is introduced due to the surface
fitting methodology. A marked drop in the power for 50 data points from that recorded
for 100 and 150 points is seen from figure 6.13.
1.0

0.0

KEY
0.6
0£
- 50 POINTS
0.4

0.3 - 100 POINTS

0.0 £
<L2 0.4 0.4 at
- 150 POINTS
CENTROID DISPLACEMENT

Figure 6.13:- Power curve for centroid displacement with no random noise.

§6.3.3.5 Overlap

For overlap, the power curves illustrated in figure 6.14, display differing behavioural
patterns with 50 points having the weakest power, as expected.

1.0

0.9

0 .7
KEY
a
at
as - 50 POINTS
U
> J at
a
a.
as

a? - 100 POINTS

ao
- 150 POINTS
OVERLAP PROPORTION

Figure 6.14 Power curve for overlap proportion with no random noise.
 157

For the null hypothesis to be rejected, at the 95% confidence level, an overlap of the
order of at most 50% is required for 50 data points and 60% and 75% for 100 and 150
points respectively.

It appears that problems may arise with the introduction of random noise, particularly for
50 points, resulting in situations where noise attributable to the surface fitting procedure
will dominate the result, hence noise introduced from the surface fitting procedure will
be inseparable from changes due to physical processes.

§6.3.4 Summary of no random noise

The initial part of the chapter has examined the simplest situation where no noise is
present in the measurements. Even for this simple scenario, there are indications that the
statistics describing orientation and overlap, may only be applicable where noise levels
are firstly minimised and secondly, the original data set is as large as practically viable,
both in terms of the economics and further, the experimental effort required.

Even where noise is absent and no ambiguity arises in defining the angle of orientation
or the area of overlap, changes of the order of 20° and an overlap of 50% are required
before we can be certain that change is a result of some underlying process, for 50 data
points.

For the two scalar measures area and perimeter and a mesh network comprising
approximately 4000 points, the level of change which signifies external processes are
acting is surprisingly small e.g. for a contour of area 100 units, an increase to 102 units
leads to the rejection of the null hypothesis for a significance level of 0.05. The
introduction of noise into the system is liable to result in a much greater change being
required before the null hypothesis is rejected.

Finally centroid displacement responds in a fairly stable manner with values of over 0.4
standard deviations being significant in the case of 100 and 150 points and 0.8 standard
deviations for 50 points.
 158

§6.4 Random Noise

Section 6.3 examined the simplest scenario where the measurements were not subject to
random noise. In practice this is an unrealistic situation as measurements will always
contain some form of inaccuracy.

This may be due to measurement error, bias or human error. Bias or persistent error is
outwith the control of the statistician. The operator should recognise the symptoms when
recording the measurements. Failure to do this will lead to an over/under estimation of
results. The overall shape of the contour surface should be unaffected, other than be
raised or lowered by a constant factor, metaphorically speaking. The worst scenario is
the case where the operator recognises the occurrence of bias, recalibrates the instrument,
continues making measurements but fails to correct the earlier results for bias. The
surface is then a mis-match of results and worthless.

The only form of error in which we are interested is that due to measurement error,
which is controlled by the accuracy of the instrument and human limitations, typically
this form of error will be quantifiable, although over or under estimation may
occasionally result. Typically in field measurements, the level of inaccuracy will be of a
higher order than that pertaining to measurements recorded in the laboratory.

The noise levels which were believed to be most typical are those of 5, 15 and 25%. In
this section, the analysis described in section 6.3 will be repeated but for the three levels
of noise.

Two possible forms of noise may be introduced into a spatial system, either in terms of:

1. z-ordinate
2. x, y-ordinate.

Interest is specifically in the former of these error forms. A similar procedure to that
described in section 6.2.5 was undertaken, step 1 was replaced by
 159

1. Simulate data points Xir,X2r>*-->^nr f°r N = 50, 100 and 150 points using the NAG
subroutine G05EAF
2iR - n (o, x ) + n ( o, e ,)

1.0 k ' s 0
where Si =
k 1.0 0 s

0.0 area
0.0 perimeter 0.05
where k = 0.9 orientation s = 0.15
0.5 centroid displacement 0.25
0.0 overlap

§6.4.1 Distribution of the test statistic under the null hypothesis

As for the case of the null hypothesis where no random noise is present, the statement for
the null hypothesis is the same i.e. no change.

§6.4.1.1 Area

The underlying empirical probability distribution function for the areal ratio is given in
figure 6.15 with the relevant percentage points being tabulated in 6.9.
 160

0.24

0.20 0.20

a 12
& 0.06

V
0,00
0.985 0.990 0.995 1.000 1.005 1,010 1.015 1.020 0.900 0.905 0.990 0.995 1.000 1.005 1.010 1.015 1.020

AREA R A TIO AREA R A T IO

Figure 6.15 ( a ) 5% random noise Figure 6.15 (b) 15% random noise

KEY

O 0.32 - 50 POINTS
LU
=3 0.24
C5 - 100 POINTS
CC
LU

- 150 POINTS

.0 0.2 0.4 0.6 D.B 1.0 1.2 1.4 1.6 1.B 2,0
AREA RATIO

Figure 6.15 (c) 25% random noise

Figure 6.15 Empirical probability distribution for areal change under the null
hypothesis of no change.

q rn 50 points 100 points 150 points

60 5% 1.00432 1.00361 1.00365

80 5% 1.00718 1.00554 1.00556
90 5% 1.00896 1.00781 1.00708
95 5% 1.01126 1.00945 1.00781
97.5 5% 1.01191 1.01177 1.00947

©
 60 15% 1.00451 1.00451 1.00422
80 15% 1.00704 1.00660 1.00598
90 15% 1.00933 1.00835 1.00814
95 15% 1.01198 1.00923 1.00923
97.5 15% 1.01301 1.01006 1.01162

60 25% 1.00446 1.00452 1.00419

80 25% 1.00840 1.00742 1.00667
90 25% 1.01290 1.00972 1.00829
95 25% 1.58624 1.01202 1.01163
97.5 25% 2.88234 1.15785 1.06788
Table 6.9 : Percentage points for areal change under the null hypothesis.

A strong similarity is displayed between the results for 50, 100 and 150 points.
Increasing the level of random noise, results in an elongation of the tails of the
distribution. This is particularly prevalent for noise levels of the order 25%. For the case
of 50 points, aberrant values are noted, the 97.5 percentage point for 50 points is almost
double that for 95%. Increasing the number of simulations will possibly draw the value
in slightly, but it will still be some distance away from the 95% point, since the
distribution is naturally long tailed.

§6.4.1.2 Perimeter

The distribution under the null hypothesis in the presence of noise responds in a similar
manner to that under the hypothesis of no change in the absence of no noise, figure 6.16.

The difference again materialises in the length of the tails of the distributions between
data sets of differing sizes. Much closer accord is displayed between 100 and 150 points
than for 50 points for 15% and 25% random noise. As for the case of areal change, we
see that particularly for 50 data points, the distribution is elongated, this is accordingly
reflected in the values of the percentage points, table 6.10. The results reported for 5%
noise levels correspond closely to those evaluated for the case of no random noise in
section 6.3.1.1. For 25% noise, the percentage points display a greater level of contrast
between 100 and 150 points than for either 5% or 15%, in the tails of the distribution.
 162

The introduction of noise into the system does not appear to bias the results, the mode of
the distribution, for all permutations of size and noise, is still in the vicinity of one.

The shapes of the distributions remain constant for all permutations of noise level and
data set size.

0.20 0.20

CJ 0.12
LU 0.06

0.00
0.00 ,.v y , I
0.96 0.98 1.00 1.02 i.oa
0.08 0.9 2 0 .9 6 1.00 1.04 1.00 1.12 1.10 1.20
PER IM ETER RATIO
P E R IM E T E R R A T IO

Figure 6.16 (a) 5% random noise Figure 6.16 (b) 15% random noise

0.20
KEY
0.10
0.12
- 50 FOINTS
aw 0.06 - 100 POINTS

■ 150 POINTS
0.00
0.7 0 .0 0 .9 1.0 1.1 1.2 1 .5 1.4 (.5

PE R IM ET E R R ATIO

Figure 6.16 25% random noise

Figure 6.16 Empirical probability distribution for the perimeter ratio under the null
hypothesis of no change.
 163

q % 50 points 100 points 150 points

60 5% 1.00669 1.00789 1.00404

80 5% 1.01354 1.01191 1.00722
90 5% 1.01793 1.01563 1.01025
95 5% 1.02031 1.02303 1.01316
97.5 5% 1.02677 1.02663 1.01774

60 15% 1.01346 1.00972 1,00912

80 15% 1.03051 1.02123 1.01897
90 15% 1.04924 1.03320 1.02340
95 15% 1.07506 1.05203 1.02880
97.5 15% 1.15606 1.06293 1.05491

60 25% 1.06922 1.02781 1.02380

80 25% 1.17126 1.05261 1.05058
90 25% 1.24585 1.07865 1.06834
95 25% 1.58095 1.24007 1.08585
97.5 25% 1.65233 1.34295 1.17154
Table 6.10 : Percentage points for perimeter change under the null hypothesis.

§6.4.1.3 Orientation

The empirical probability distribution functions for orientation are presented in figure
6.17 with the analytical values given in table 6.11 for various percentage points.

Once again the main contrast between different sizes of data set is in the tails of the
distribution. For increasing sizes of data set, the tails are terminated more sharply. This
feature extends over all levels of noise. The mode of the distribution occurs at
approximately zero for all the distributions, indicating that bias has not been introduced
into the system.

The distribution of the test statistic for 25% noise is wider than for either of the other two
levels and the tails are generally fatter.
 164

0.24

0.20 0.20

0. I d 0.14

O
>-
zLU 0.12
ra
a
LU

C
LU
K 0.08 S 0.06
Ll- U.

0.00 0.00
- C .6 - 0 . 5 - 0 . 4 - 0 . 3 - 0 .2 - 0 .1 0 .0 0.1 0 .2 0 .3 0 .4 0 .5 0.6 -0. 6 - a 5 - 0. 4 -0. 3 - 0. 2 - 0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6
O R IE N T A T IO N CHANGE O R IE N T A T IO N CHANGE

Figure 6.17 (a) 5% random noise Figure 6.17 (b) 15% random noise

0.20 KEY

- 50 POINTS
>-
(_> 0.12
u - 100 POINTS
3LU 0.08
cr
L_
150 POINTS

0.00
- 0 .6 - Q . 5 - 0 . 4 - 0 . 3 - 0 . 2 -O .1 0 .0 0,1 0 .2 0 .3 0 .4 0 .5 0.6

O R IE N TA T IO N CHANGE

Figure 6.17(c) 25% random noise

/
Figure 6.17 Empirical probability distribution function for orientation change under
the null hypothesis of no change.
 165

q % 50 points 100 points 150 points

60 5% 0.063614 0.055197 0.040060

80 5% 0.082298 0,078591 0,056560
90 5% 0.102002 0.097638 0.066350
95 5% 0.108241 0.117996 0.071623
97.5 5% 0.235798 0.124846 0.096091

60 15% 0.113325 0.051748 0.050417

80 15% 0.210964 0.833154 0.076049
90 15% 0.281352 0.105859 0.091681
95 15% 0.348310 0,119322 0.109681
97.5 15% 0.503490 0.155222 0.112509

60 . 25% 0.225310 0.059556 0.080961

80 25% 0.303834 0.093346 0.103648
90 25% 0.450519 0.145384 0.133834
95 25% 0.550824 0.169104 0.143449
97.5 25% 0.689743 0.205192 0.161665

Table 6.11 Percentage points for orientation change under the null hypothesis.

§6.4,1,4 Centroid Displacement

Once again for 5% and 15% random noise, a strong similarity in behaviour to that of
the case for no random noise emerges. The results for 50 points being more diverse,
figure 6.18.
 166

0.20 a 20

y-
a 16
z(_]LU
=3
a
CX.
LU 0 .1 2

0 .0 6 a oe

‘.v
0.0 at as as ao a2 at t .o t.2

C EN T R O ID DISPLACEM EN T CEN TROID D ISPLA C EM EN T

Figure 6.18 (a) 5% random noise Figure 6.18 (b) 15% random noise

a20 KEY

- 50 POINTS
=3
□
QLa
lu 12
- 100 POINTS

• 150 POINTS

a oo
a2 t.o 1.2 1.4
C EN T R O ID D ISPLACEM ENT

Figure 6.18 (c) 25% random noise

Figure 6.18 Empirical probability distribution for centroid displacement under the
null hypothesis of no change.

By the time, the noise level attributable to the data is of the order of 25%, differences
emerge between the three sizes of data set. The percentage points, table 6.12, confirm
this behaviour. The tails are accordingly long for 50 data points. Although the
distributions are long-tailed, the percentage points are contained within the limits of
two standard deviations for all levels of random noise.
 167

q % 50 points 100 points 150 points

60 5% 0.38080 0.20772 0.20296

80 5% 0.48054 0.25640 0.25621
90 5% 0.51861 0.27862 0.30236
95 5% 0.53936 0.32125 0.34702
97.5 5% 0.61929 0.36598 0.39409

60 15% 0.35560 0.23949 0.19566

80 15% 0.42299 0.29472 0.22753
90 15% 0.45234 0.34955 0.27755
95 15% 0.53601 0.35761 0.30000
97.5 15% 0.60862 0.38940 0.34024

60 25% 0.78525 0.67000 0.52101

80 25% 0.94089 0.75300 0.62418
90 25% 1.02929 0.90167 0.73855
95 25% 1.20269 0.94069 0.79367
97.5 25% 1.33353 1.07630 0.85094
Table 6.12 Percentage points for centroid displacement under the null hypothesis.

§6.4.1.5 Overlap

In terms of the overlap function, the empirical probability distribution function is

marginally different for each size of data set, figure 6.16. The saliant values being
recorded in table 6.19.

The tails of the distribution are longer and fatter with increasing levels of noise. The
mode of the distribution does not lie at 1.0 , the expected value, if no change has
resulted. The underlying reason for this potential form of bias is because one is the
upper bound of the distribution and with the introduction of noise, it would be
anticipated that the two contours could be entirely contained within each other. For 5%
noise and all three data sets, the mode lies at 75%, similarily for 100 and 150 points for
15% noise. For both 15% and 25% noise, for 50 data points, the mode is more difficult
to select as the distribution is much flatter in shape, but it lies approximately at 40%.
 168

0.20 0.20

0. Id 0.16

a 06

0lO4
0.00 ) --- 1--- »—T=aF=fr'crtf|*^__,--- 1---1--- 1---h 0.00
0 .0 0.1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8 0 ,9 t.O 0 .0 0.1 0 .2 0 .5 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1.0

OVERLAP OVERLAP

Figure 6.19 (a) 5% random noise Figure 6.19 (b) 15% random noise

KEY

0 .(6 - 50 POINTS

- 100 POINTS
CK
- 150 POINTS

0.00
0 .0 0.1 0 .2 0 .5 0 .4 0 .5 0 .6 0 .7 0 .8 0 .9 1.0

OVERLAP

Figure 6.19 ( a ) 25% random noise

Figure 6.19 Empirical probability distribution for overlap under the null hypothesis
of no change.
 169

q % 50 points 100 points 150 points

60 5% 0.833994 0.883402 0.898660

80 5% 0.887435 0.914999 0.922489
90 5% 0.921110 0.932646 0.940530
95 5% 0.938217 0.948872 0.956583
97.5 5% 0.950966 0.960851 0.962937

60 15% 0.714501 0.789122 0.826559

80 15% 0.821925 0.854482 0.869114
90 15% 0.869977 0.887272 0.886470
95 15% 0.905789 0.899752 0.912062
97.5 15% 0.924986 0.920836 0.925357

60 25% 0.672920 0.723764 0.732612

80 25% 0.742420 0.800562 0.787022
90 25% 0,798821 0.854832 0.818849
95 25% 0.834341 0.892620 0.843812
97.5 25% 0.862852 0.909092 0.863835

Table 6.13 Percentage points for overlap under the null hypothesis.

§6.4.2 Power Curves

§6.4,2.1 Area

For areal change, the power curves, figure 6.20, illustrate a strong similarity between
the results for 100 and 150 points for 5% and 15% random noise, whilst for 25%,
differences materialise between all three data sets. The results for 50 points in particular
are indicative of weaker power. Increasing the noise level results in a falling off of the
overall power i.e. a lateral shift in the curves has resulted, hence greater areal changes
are required for differences, between contour areas, to be defined as statistically
significant and not simply a result of the surface fitting procedure.
 170

1.1

1 .0

0.9

0.8

0.7 0.7

0.5 0,5

at a*
o 0.4
a.
a, J

0 .2 0.2

0.1
aa 0.0 -i- -
o I 3 5 i 7 « 9 10 0 2 4 4 8 10 12 I* 10 10 20

AREA AREA

Figure 6.20 (a) 5% random noise Figure 6.20 (c) 25% random noise

i.a

0 .8 KEY

- 50 POINTS

5 - 100 POINTS

■ 150 POINTS

0.0
2 3 4 5 0 T 8 t 10
AREA RATIO

Figure 6.20 ( b ) 15% random noise

Figure 6.20 Power curves for area in presence of random noise.

§6.4.2.2 Perimeter

A strong pattern emerges in the power curves for perimeter, figure 6.21. The curves for
50 points consistently display a weaker level of power than either 100 or 150 points,
whilst only marginal differences emerge between the results for 100 and 150 points.
 171

Increasing the noise level, results in a sharp decrease in the power. For noise levels of
25% compared to those of 15% an approximate doubling of the perimeter ratio is
recorded for a power of 95%.

1.1
t.O

0 .9 0.9

0.0 ao

0 .7 a7

CL
0£
0.4

0 .3

0.2

0 .t 0.1

0 .0 ao
a 10 20 30 40 50 40 70 00 90 100 110 120
0 4 0 12 10 20 24 20 32 30 40 44 40

PERIM ETER PERIM ETER

Figure 6.21 (a) 5% random noise Figure 6.21 ( b ) 15% random noise
I.!
t.O
0 ,9
KEY
0.0
0 ,7
- 50 POINTS
O ^*4
^ 0 .5 - 100 POINTS
o
- 150 POINTS
0 .2

0.1
20 40 40 80 100 120 140 140 100 200 220

PE R IM ET E R

Figure 6.21 (c) 25% random noise

Figure 6 . 2 1 Power curves for perimeter in presence of random noise.

 172

§6.4,2,3 Orientation

The power curves for orientation change are given in figure 6.22 for all three levels of
noise. The behaviour for 5% differs to that for both 15% and 25%. A fairly strong
accord exists between the results for 0% and 5% random noise. Generally, a slightly
greater shift is required for change to be recognised as the result of external influences
for 5%. The trends across the data sets are similar.

For 15% and 25% noise a different pattern emerges, the behaviour of the results for 100
and 150 points being similar, whilst for 50 data points, a reduction in the power occurs.
Rotations of the order of 90^ and 45°' for 50 and 100 and 150 points respectively, are
required, for a power of 0.05, before It is safe to say that change is not solely a facett of
the surface fitting procedure.

1.1

t.O
0 .9 0.9

0 .7 0.7

0 ,5 0 .5

O
a.
0.1 0 .J

0 .2

0.1
0
O R IE N T A T IO N CHANGE O R IE N T A T IO N CHANGE

Figure 6.22 (a) 5% random noise Figure 6.22 (b) 15% random noise
 173

i.i
1 .0

KEY

50 POINTS

100 POINTS

150 POINTS
0.0 v .
10 2 0 3 0 * 0 5 0 6 0 7 0 8 0 9 0

O R IE N T A T IO N CHANGE

Figure 6.22 ( c ) 25% random noise

Figure 6.22:- Power curves for orientation change in presence of random noise.

§6.4.2.4 Centroid displacement

The ability of the test statistic to reject the null hypothesis was examined for various
alternative hypotheses, the resultant power curves are given in figure 6.23.

Noise levels of 5% resulted in power curves which display little difference to those for
the situation of no noise; the results for 100 and 150 points display strong accord whilst
for 50 points, a much greater shift is required before the null hypothesis is rejected.

Increasing the noise level to 15% and 25% confirms the earlier suspicions that
complications may arise in separating change attributable to the surface fitting
procedure from that due to 'real' change, especially for the latter case where
displacement of the order of two standard deviations is required before we can
categorically state that an external process has been acting in effecting a change.
 174

1 .0

0. 8 0. 8

o
CL
0. 4

O :*
UJ
o
a
0.4

0. 2

0. 0 0. 0 4-
1.4 0. 0 0. 2 0. 4 0. 4 0. 8 1. 0 1.2 1.4

C EN TR O ID DISPLACEMENT CENTRO ID DISPLACEM ENT

Figure 6.23 (a) 5% random noise Figure 6.23 (b) 15% random noise
1.0

0 .8
KEY

ctr
0. 4 - 50 POINTS
UJ

0. 4 - 100 POINTS

■ 150 POINTS

0. 0
0. 0 &2 0. 4 0. 4 0. 8 1.0 1.2 1.4 1.4 1.8 2. 0

CENTRO 10 DISPLACEMENT

Figure 6.23 (c) 25% random noise

Figure 6.23 Power curves for centroid displacement in presence of random noise.

§6.4.2.5 .Overlap

The behaviour of the power curves given in figure 6.24 is much less well defined than
for any of the previous examples. For 5% noise, the results for 100 and 150 points are
well defined, with an overlap percentage of the order of 60%, for a confidence level of
95%, required to ensure Hq is rejected, whilst for 50 points, a value of the test statistic
in the region of 40% would result in the rejection of the null hypothesis.
 175

Noise levels of 15% and 25%, give much less smooth pictures, the results for 100 and
150 points still respond in a similar manner, but the power falls away considerably. For
50 points, the two contours have to be virtually disjoint before we can be certain that
change is not solely indicative of variability attributable to the surface fitting procedure.

§ at

ae

OVERLAP PERCENTAGE

Figure 6.24 (a) 5% random noise Figure 6.24 (b) 15% random noise

KEY

• 50 POINTS

100 POINTS

150 POINTS

OVERLAP PERCENTAGE

Figure 6.24 ( c ) 25% random noise

Figure 6.24 Power curves for overlap in the presence of random noise.
 176

§6.4.2.6 Summary

Overall the introduction of random noise resulted in the underlying probability

distribution function of the test statistic behaving as anticipated, increasing the level of
random noise resulted in longer-tailed distributions and hence greater changes were
required before change due to the external process could be separated from those due to
noise in the variable of interest.

In terms of the power curves, the behaviour of the data sets for 5% random noise,
responded in a similar fashion to those of no random noise i.e. where the data set size
was influential in controlling the probability distribution function for no noise, a similar
pattern was reported in the presence of low levels of noise. Also the percentage points
recorded were similar in value.

This generalisation does not carry through for 15% and 25% noise. With increasing
noise, the power curves for 50 points generally divest themselves from those for 100
and 150 points. The main exception to this was for the test statistic describing centroid
displacement where a collation of the curves appeared to result.

One observation to emerge from the preceding analysis was the question of the validity
of defining a translation statistic, particularily for 50 points where the level of noise was
of the order of 25%. The variability due to the level of measurement noise was liable to
dominate any measure of change recorded.

§6.5 Multivariate Hypothesis Testing Procedure

Univariate techniques may be utilised to analyse changes between single contours but
we may wish to consider several test statistics simultaneously or to evaluate the same
test statistic on a number of different contours. A univariate analysis effectively
assumes the parameters are mutually independent between contours or that inter­
relationships are unimportant. In the case of overlap and areal ratio, there are situations
where the two are identical, hence correlation exists between the measures, unless the
analyst is aware of the situation at the subjective stage and substitutes the statistic based
on the centre of gravity.
 177

Generally the measures described will display only weak correlation because of the
complex nature of the contours. Table 6.14 describes the correlations between the
various statistics, for the three levels of noise. On this premise, the assumption of
independence is not unrealistic except possibly between the scalar measures area and
perimeter and location shift and orientation..

Area Perimeter Orientation Location

shift
Perimeter 0.949
Orientation -0.027 -0.049
Location shift -0.066 -0.077 0.585
Overlap -0.010 -0,014 -0.114 -0.061
Table 6.14 (a) Correlations for the various test statistics for 5% noise.

Area Perimeter Orientation Location

shift
Perimeter 0.960
Orientation -0.239 -0.224
Location shift -0.078 -0.041 0.500
Overlap 0.346 0.349 -0.274 -0.121
Table 6.14 (b):- Correlations for the various test statistics for 15% noise.

Area Perimeter Orientation - Location

shift
Perimeter 0.951
Orientation -0.060 -0.025
Location shift -0.144 -0.174 0.078
Overlap 0.161 0.210 0.074 0.033
Table 6.14 (c) Correlations for the various test statistics for 25% noise.

Table 6.14 :- Correlation between the various test statistics for all levels of noise.
 178

The strength of the relationship between location shift and orientation may be
attributable in part to the evaluation of the centre of gravity for each of the measures,
hence a mathematical relationship exists between the two. Area and perimeter are
known to be strongly related for simple structures and in the simulation study, the
contours are always defined to be pseudo- elliptical or circular hence the resultant,
possibly misleading, strength of the relationship.

A multivariate test may effectively be performed by evaluating for each level of

interest, the appropriate measure of overlap, centroid displacement, scalar change and
angular change for each set of comparable contours. Each individual contour in surface
A being assessed against the 'corresponding' contour from surface B. An order of
subjectivity is introduced at this stage, interest only being in those contours which are
recognisably comparable in some form size, shape or position.

For a global analysis, the parameter values are not necessarily assumed to be
independent. An analysis of this form will be based on a multivariate technique which
explicitly models the covariance structure of the data. Theoretically the results are more
powerful, but a number of restrictions limit the potential power of the test, section 6.6.
A parametric form of analysis for a global test of change is Hotelling's one sample T^-
test, where we are testing

HO : p. = m
HI : |i,X unconstrained

The vector of fixed constants will depend on the form of investigation. Generally we
will be concerned with comparing all aspects of the structure of the two surfaces,
hence:-

rno change in area >

1 no change in perimeter
0 ie. no change in orientation ,say
0 100%overlap
,0 , vno change in centroid displacement^
 179

Where interest is only in specific, pre-defined aspects of a surface, m can be

accordingly modified. An illustration of the implementation of a global test is given in
chapter 7.

§6.6 The Technique

1. Interpolate grid points on a regular grid comprising 4000 nodes.

2. Select number and bounds of class intervals using empirical rules described earlier.
3. Construct contour plot.
4. Calculate the basic geometric quantities area, perimeter, centroid ordinates and
orientation.
5. Evaluate the required test statistics.
6. (a) For a local analysis carry out the corresponding hypothesis test.
(b) For a global approach collate the required information and perform a
Hotelling's one - sample T^.
7. Interpret the results.

§6.6 Limitations of the Technique

A number of problems are currently associated with both the univariate and
multivariate forms of the analysis described in this chapter,which may bias the results.

Restriction 1 How should the test statistics be defined for a contour which was whole
for surface 1 and becomes disjoint for surface 2 and vice-versa.

Solution 1 For area and overlap the total area for each level of interest may be
summed for the subsequent analysis, for both the univariate and multivariate
approaches.

One plausible method for dealing with the statistic for centroid displacement is to
evaluate the statistic for each separate contour and then take a global average. This is
particularity suited to the situation given in figure 6.25 (a).

In terms of orientation, the angular change may be equated for each of the disjoint sets
of contours and hence compared. Realistically, for perimeter, there is no valid
alternative which enables a valid perimeter ratio to be evaluated; combining the
 180

perimeters will result in an over-estimation of the ratio whilst evaluating the ratio
separately for each of the areas, tends to cause an under-estimation in both cases. The
complexity of the shapes of the contours hinders any proportional combination being
undertaken. However as described in the introduction to the section, area and perimeter
may be strongly related, hence the inclusion of perimeter as a measure may not be
essential.

Figure 6.25 (a) Disjoint contour Figure 6.25 (b) Split contours

Figure 6.25 Problematical contour shapes.

Restriction 2 :- The second restriction arises where an additional disjoint contour

appears in one of the surfaces, figure 6.25 (b).

Solution 2 For the univariate approach it may be included in the total area whilst for
the multivariate case, unless interest is solely in the entire contour level and not
individually comparable contours, it should be omitted from the analysis, hence
potentially biasing the result.

Restriction 3 :- A further limitation with the methodology is linked to the general

criticism of hypothesis testing, it only answers the question, 'Has there been a change?'
or 'Is there an association between variables A and B?'. No account is taken of the size
of the difference or asssociation.

Solution 3 For practical problems, alternative approaches may enable these questions
to be answered, once we have assessed whether an association/change either exists or
 181

has occurred. Secondly, having established a hypothesis testing procedure, the close
links between it and interval estimation will enable a confidence interval to be placed
on the level of change.

Restriction 4 A further limitation of the technique is in terms of the size of the data
set and the level of random noise present within the system, these factors apparently
control how small a change may be detected over and above that of ’natural' variability.

Solution 4 The only viable solution is for the practitioner to collect more data, or
alternatively ensure the data collected is of a high quality and errors are minimised.

Restriction 5 Related to the previous limitation of noise level is that of the incorrect
assessment of the level of noise within a series of measurement. Over or under­
estimation of the noise factor may inadvertantly alter the result.

Solution 5 One possible method to eliminate this potential error source is instead of
defining a p-value or other appropriate value, the maximum level of noise at which the
statistic is significant may be quoted, for a particular significance level. Although
unorthodox, it eliminates the potential for incorrectly defining the noise level.

§6.7 Advantages of the Technique

The main advantages of the methodology are

1. Its versatility and intuitiveness. The technique may be applied to a whole wealth of
situations ranging from detecting changes in pollen levels, to assessing the relationship
between the geology of an area and its background radiation levels.
2. The non-reliance on dimensionality ensures that two or more non-similar measures
may be examined for changes.
3. It is simple to implement and the test statistics are easily calculable, once the
surfaces have been constructed.
4. The lack of restriction placed on the methodology required to develop the surfaces,
bar that of consistency, widens its appeal.
5. In terms of a local approach it enables the internal spatial complexities to be
investigated more specifically.
 182

CHAPTER 7

T H E A P P L IC A T IO N O F T H E M E T H O D O L O G Y T O T W O

E N V IR O N M E N T A L C A SE S T U D IE S

§7.1 W hv The Studies Were Selected

Within this chapter, two examples have been selected to illustrate the potential diversity
and complexity of the applications to which the methodology developed in chapters five
and six may be applied.

The first example deals with one of the simplest situations; the investigation of climatic
change given the results from a monitoring network. The observations were recorded at
fixed localities and different time points, t. The example addresses the question of
whether a seasonal temperature change has resulted within the contiguous United States
of America during the fifty year span, 1930 to 1980. Interest has been expressed in such
a question due to the high public profile of global warming and its possible
consequences. Attention will focus on the implementation of the hypothesis method, on
both a local and global scale, with the results of the global test being compared to those
of other global techniques, described in sections 4.3.

The second example was primarily selected to illustrate that the methodology was not
only applicable to simple problems, but also to others where the level of spatial
variability differed between variables. The motivating question was whether background
radiation is a causative factor in the induction of leukaemia. A pilot study, commissioned
by the Leukaemia Research Fund, examined this question for a region of south-west
England. In addressing this problem, a secondary issue was examined; ‘Was the
underlying population the sole factor in controlling the distribution of cases?' The data
comprised leukaemia cases, population figures and radiation values which all differed in
spatial resolution and quantity. The sparsity of the leukaemia data served to curtail the
implementation of the methodology directly but it provided the motivation for the
implementation of alternative forms of analysis.
 183

§7.2 Case Study 1 The Investigation of Climatic Change

In recent years there have been numerous press releases and technical articles concerned
with the ideas of climatic change and potential global warming, including Woodward
and Gray (1992), Karl et al. (1991), Tsonis and Eisner (1989). Of central focus has been
the phenomenon called the 'greenhouse effect’.

The planet is made habitable by the presence of certain gases. These gases trap long­
wave radiation emitted from the Earth’s surface and result in a global mean temperature
of 15°C, as opposed to -18°C, in the absence of an atmosphere. By far the most
important greenhouse gas is water vapour. However there is a substantial contribution
from carbon dioxide and smaller contributions from ozone, methane and nitrous oxide.

The concentrations of carbon dioxide, methane and nitrous oxide are believed to be
increasing and in recent years other greenhouse gases, principally chlorofluorocarbons
(CFC's), have been added in significant quantities to the atmosphere. There are many
uncertainties in deducing the consequential climatic effects. Typically it is estimated that
increased concentrations of these gases since 1860, may have raised global mean surface
temperatures by O.S^C or so, and the projected concentrations could produce a warming
of about 1.5°C over the next 40 years.

Numerical climate models indicate that other changes in climate would accompany the
increase in globally averaged temperatures with potentially serious effects on many
social and economic activities.

Much of the evidence for a global warming effect has been based on large-scale Global
Circulation Models (GCM's). Table 7.1 summarises the results from five of the most
commonly cited methods.

Within these studies the global and annual average warming varies from 2.8K to 5.2K.
The warming is accompanied by an increase in evaporation and precipitation. The
enhanced radiative heating of the surface due to increases in carbon dioxide and water
vapour is balanced by increased cooling due to evaporation, producing a more intensive
globally averaged hydrological cycle. The change in temperature is not uniformly
distributed in time or space. A concise review of how these changes may be effected is
given by Mitchell (1989).
 184

Study Source Surface Precipitation

Temperature Change
Change (K) (%)
GISS Hansen et al. 4.2 11.0
(1984)
NCAR Washington and 4,0 7.1
Meehl (1984)
GFDL Wetherald and 4.0 8.7
Manabe (1986)
MO Wilson and Mitchel 5.2 15.0
(1987)
OSU Schlesinger and 2.8 7.8
Zhao (1987)

Table 7 . 1 Global mean changes in five carbon dioxide doubling studies.

The GCM's are all based on multi-level mathematical representations of the atmosphere.
Given the complexity of the environment and the relative simplicity of the models there
is much controvorsy concerning their validity.

Other numerical studies have focused on the idea of climate change and long-term
patterns in temperature and precipitation, (Karl, Heim and Quayle (1991), Jones et al.
(1986), Diaz and Quayle (1980)).The critical question of climate variability has also
been addressed, (Karl et al. (1984), Agee (1982)).

A more recent study by Woodward and Gray (1992) raises the question of whether trend
based analysis is valid for temperature data since for data of this form, it is common to
observe trends that increase over one time span and then decrease over the next, more
than likely because of the correlation imposed on the data by the physical phenomena
that drive them. Although all the preceeding analysis invoking the use of a trend based
procedure produce statistically significant results, the authors ask, 'If conditions remain
the same, should we predict the temperature to increase in the future for an extended
period of time?'
None of these studies have incorporated the spatial aspects of the temperature or
precipitation field. Early papers which focused on the spatial-temporal change of
 185

temperature and precipitation over the U.S.A. were either regional in nature, Sellers
(1968), or restricted to relatively short time periods of the order of a decade or two,
Skaggs (1975).

More recently Handcock and Wallis (1990) developed a comprehensive model for the
spatial dimension in conjunction with the temporal component. Since the model was for
the meteorological field as a whole, this facilitated its direct comparison with GCM's and
allowed prediction of derived quantities throughout the region and over time.

For a gradual increase of 5°C over 50 years, it has been suggested that it will take
between 20 and 30 years before the change is discemable above the natural variation in
temperature.

The quality of all these studies depends on the climatic records being both reliable and
accurate. Many of the data bases used in the past were not believed to be either free of
bias and error or suitable for long term climate studies.

Some of the more recent studies have made use of what is believed to be an accurate,
unbiased, modem historical climate record set up by the Carbon Dioxide Research
Program of the United States Department of Energy and the National Climatic Data
Centre (NCDC) of the National Oceanic and Atmospheric Administration (NCAA),
Quinlan et al. (1987).

Utilising the above data set, investigation of a very simple question was undertaken to
illustrate the applicability of the methodology developed. The question of interest
examined the seasonal changes in temperature during the 50 year span 1930 to 1980 and
was 'Has a change occurred in temperatures between 1930 and 1980?'. The results
obtained were compared to those evaluated on the basis of a number of other existing
techniques. The seasons were defined as spring:- March, April and May; summer.- June,
July and August; autumn:- September and November; and finally winter- December and
January and February of the next/ upcoming year.

§7.3 United States Historical Climatology Network

A network of 1219 stations (HCN network) within the contiguous United States of
America was set up for the specific purpose of compiling an accurate, serially complete,
 186

modem historical climate data set suitable for detecting and monitoring climate change
over the past two centuries. The data base comprises station histories, monthly
temperature (maximum, minimum and mean) data and total monthly precipitation.
Potentially it is the most reliable and continuous sequence to have been collected in
recent times, a whole wealth of sources gave rise to the final data set including
climatological publications, universities, federal agencies, individuals and data archives.

All stations were quality controlled by NCDC with the use of outlier and areal edits, each
station being corrected for time of observation differences, instrument changes and
moves, station relocation and urbanisation effects, Karl et al (1986), Karl and Williams
(1987), Karl et al (1988). A number of features associated with the data are

1. Confidence factors for each adjusted estimate.

2. Only a small portion of the data is missing with some missing values being estimated
from neighbouring stations.

3. The data is constantly being updated and enhanced and hence provides a unique
source for the evaluation of greenhouse effects.

§7.4 Subjective Impression

Before proceeding to the subjective analysis of comparison, a few basic summary

statistics are cited in table 7.2 for the four seasons and the two years in question.

The first tool for assessing subjectively the question of change was based on an
univariate form of analysis, the box-and-whisker plot. Specifically the plot examines the
differences between the 1980 and 1930 temperatures, figure 7.1. Spring and winter
display a similar trend across the results, the mean temperature generally being cooler in
1980 than 1930 whilst the reverse is true in summer and autumn. In terms of range
greatest variability amongst stations is demonstrated in the winter months where
considerable differences are displayed for the two years in question.
 187

Year Season Number Number Mean Stdev. Min. Max.

(missing)

1930 Spring 1174 25 51.400 8,380 30.29 75.74

1980 Spring 1184 15 50.862 7.956 21.24 77.81

1930 Summer 1170 29 71.994 6.612 50.05 97.72

1980 Summer 1172 27 72.563 7,830 48.81 96.87

1930 Autumn 1169 30 53.294 8.187 33.76 78.78

1980 Autumn 1161 38 53.895 7.970 35.83 80.22

1930 Winter 1170 29 34.141 9.959 10.74 67.30

1980 Winter 1191 8 34.028 10.572 5.790 65.98

Table 7.2 Summary statistics for the four seasons.

O * * I + 1------** Figure 7,1 (a) Spring

I +1------** Figure 7.1 (b) Summer

*** i *** Figure 7.1 (c) Autumn

------------ 1 + i------------****** * * o O Figure 7.1 (d) Winter

—I----------h---------- h-------- +--------- +-------- +— Temperature change

- 12.0 -6.0 0.0 6.0 12.0 18.0 (°C)

Figure 7.1 Box-and whisker difference plots (1980-1930).

The major problem is as for all univariate approaches to spatial statistics, the spatial
context of the data is ignored and no feel is attained for how these differences are
distributed over the United States of America.
 188

Bivariate plots are an alternative means for describing the data. These enable a better
assessment of the relationship between the two data sets to be made. For spring and
summer a fairly narrow ellipse encloses all the results, whilst for autumn the points are
less well confined especially at the lower range of temperatures. Finally for winter, a
separation of the enclosing ellipse occurs at the cooler end of the scale, indicating a
greater diversity of results between the two years in question, this confirms the maximal
change of 18°C reported in the box-and-whisker plot for the differences.

However, although the bivariate plot is slightly more satisfactory than the univariate
approach, the spatial dimension is still ignored.
100

ca*
so
$
ae
i 45

SPRING 1980
SUMMER 1980

80
75

70

85
40
ae
55

50

*5

40
15 ,
35

30

AUTUMN 1980 WINTER 1980

Figure 7.2 Bivariate plots of temperature for the contiguous U.S.A. for 1980 and 1930.
 189

We will now consider two of the spatial techniques which formed the cornerstone of the
methodology developed in the preceding chapters, to test whether change has resulted
between the two variables of interest. Figures 7.3 to 7.6 illustrate the superposition of
the two surfaces of interest and the differences between the two data sets i.e. isopach
maps, for a selected set of contours for each of the four seasons. The selection of the
contours was based on a set of pre-defined temperature levels since interest was in
whether change, in the form of an increase or decrease, had resulted during the two
years in question, not whether changes had occurred in the spatial distribution of the
two surfaces, i.e. we are comparing the same temperature contours and assessing by eye
any apparent differences over time. For clarity only five contours have been depicted
for each surface.

Based on a recommended grid of 4000 points, the surfaces were plotted using kernel
density estimation, with selection of the smoothing parameter undertaken using least-
squares cross validation.

45

40

35

30

120 110 100 90 80

LONGITUDE

KEY
1980 1930

■1------ 40°F 2 45°F 3------ 50°F

4------ 55°F 5------ 60°F

Figure 7.3(a) Superposition of temperature surfaces for spring.

 190

45

55 - M ; f p ^ :

50 ■
6^
120 110 100 90 80
LATITUDE
KEY
-------30F --------2------- - 2 ° F -------- 3-----------1°F ip------- 0°F
5------ 10F b— 2 ° F ----------1------ 3°F

Figure 7.3(b) Isopach map for Spring (1980 -1930)

Figure 7.3 Spatial subjective analysis for spring.

For temperatures up to 50°F almost total accord is expressed between the isotherms,
above this level indications are that 1980 tended to be slightly cooler, this is indicated
by the contours depicting higher levels of temperature generally being smaller in areal
dimension for 1980 than 1930.

Moving to the residual map, a difference in the geographical dispersion of temperature

change is suggested. A region in the south-west of the United States appeared to
experience an overall increase in temperature levels during the 50 year span whilst the
rest of the continent was undergoing a reversal of this apparent trend. The situation is
relatively more complex to analyse than suggested from the superposition of the two
surfaces.

The two techniques describe change in a different way, the first specifically examines a
temperature level x°F, for the whole of an area, whilst the isopach map ignores the
underlying level of temperature and considers differences at each specific station.
 191

45

40

55

50
120 110 100 90 80
LONGITUDE
KEY
1980 ______ 1930
■1------ 65°F--- 2------ 70°F 3------ 75°F
4------ 80°F--- 5------ 85°F

Figure 7.4 (a) Superposition of temperature surfaces for summer.

8- T v

120 110 100 90 80

LONGITUDE
KEY
1 -----_30F ------ 2------- - 2 ° F ------- 3- -1°F 1 ------ 0°F
5--- 1 ° F ---------fc------ 2 ° F --------- =/---------- 3°F

Figure 7.4 (b) Isopach map for summer (1980 - 1930)

Figure 7.4 Spatial subjective analysis for summer.

 192

Once again greater concordance is displayed between the temperature contours at the
cooler end of the range whilst at the upper end the area covered by the contours is
correspondingly greater for 1980 than 1930, indicative of the climate becoming warmer
over time.

Moving to the isopach map for summer, an approximate north/south divide is indicated.
Generally for the more southerly latitudes the temperatures recorded are higher in 1980,
whilst the reverse is true for the northerly latitudes. Anomalies do appear to dispel the
overall validity of such generalisations, however, this general trend is in accord with
other papers in this field.

35

30 4^4- a
120 110 100 90 80

KEY

-------------- 1980 1930

1---- 45°F 2- 50°F 3---- 55°F

4---- 60°F 5- 65°F

Figure 7.5 (a) :-Superposition of temperature surfaces for autumn.

 193

45

40

55

50

120 110 100 90 80

LONGITUDE
Key -1-----------3<>F 2 -------- -2°F 4 -------- -10f
fp 0°F 5 ----- 1°F b----- 2®F
Figure 7.5 (b) Isopach map for autumn (1980-1930)

Figure 7.5 Spatial subjective analysis for autumn

Examining figure 7.5 (a) closely, it can be seen that no clear pattern emerges, with
trends being less apparent than for either spring or summer. For some contours with
temperature level x^C, the areal range is greater for 1980 than 1930, whilst for others,
the reverse is true. This extremely complex set of changes is borne out by the isopach
map, figure 7.5 (b).

40

55

50

120 110 100 90 80

LONGITUDE
KEY:- 1980 1930
1----
25°F 2---- 30°F 3---- 35°F
4 45°F 5---- 45°F
Figure 7.6 (a) Superposition of temperature surfaces for winter
 194

35

30
120 110 100 90 80
LONGITUDE
KEY
----------- _30p ------------ _2°F -1°F 0°F
------------ 1 ° F ------------ 2 ° F -------------- 3°F
Figure 7.6 (b) Isopach map for winter.

Figure 7.6 Spatial subjective analysis for winter

Of all the seasons the greatest temperature changes are apparent during winter for the
two years in question. There appears to be a strong indication that temperatures have
become more extreme during this time. This is particularily apparent from the isopach
map. The north-west appearing cooler during 1980, whilst the south-east has witnessed
a slight increase in the temperatures during the 50 year span.

§7.3.1 Summary of subjective analysis

From the various forms of subjective analysis, temperature change may be summarised
as follows:-

1. Temperatures appear to have risen during the 50 year span for summer and autumn
but the cause of the change is unknown i.e. whether it is due to climatic changes or
variability in the data.
2. For spring and winter, in general, a drop in the overall temperature has been
recorded with the latter season recording the greatest fall of all the seasons.
3. For autumn, an extremely complex set of changes have occurred over time.
 195

In terms of a subjective analysis, a contoured map illustrates the underlying reality of the
complexities of change whilst the univariate or bivariate approach mask many of the
interesting features and fails to provide the analyst with a feel for possible causative
factors,

§7.5 Local Hypothesis Testing Procedure

Many of the techniques described in chapter 4.4 for performing a local analysis are
highly convoluted and unsuitable for this type of analysis, the main thrust of this
example is to illustrate the apparent simplicity of the hypothesis technique and how it
circumvents the problem of pre-defining the inherent level of random noise within the
data.

The basis of the local hypothesis testing procedure are the diagrams illustrating the
superposition of the temperature surfaces for the two years in question, figures 7.4(a),
7.4(b), 7.4(c) and 7.4(d), for spring, summer, autumn and winter, respectively. For each
of the diagrams, the various geometric properties for quantifying the test statistics were
evaluated i.e. area, perimeter, orientation and centroid displacement. The last of which
was broken down into latitude and longitude displacement. Appendix 1 details the
individual results for each contour. The test statistics evaluated using these values are
given in table 7.3.

The disparity in the number of entries for each of the temperature levels is solely a
feature of the overall contour structure which for some levels is more disjoint than for
others. In terms of the test statistics, only those contours within each level which were
recognisably comparable were included. For this particular problem, only a couple of
small contours were omitted from the analysis.

Season/ Area Perimeter Orientation Standard deviation

temp. displacement
X y
Spring 1.0019 1.00218 -5.8xl0-4 -3.635x10-3 0.04086
40°F
0.93136 0.97296 0.0355 0.13955 -5.2071
45°C 0.9780 1.11360 -1.8X10-3 0.30704 0.25152
1.0284 1.02239 -0.0621 -0.70175 0.28630
 196

50°F 0.94278 0.96553 0.3292 0.06645 0.26064

0.81413 0.96790 8.33x10-3 0.37520 0.416832
55°F 1.0757 0.91162 -0.0405 -0.64094 -0.63024
0.82558 0.58925 -0.0261 0.51950 3.0044
1.93404 1.2025 1.30525 -0.1060 -0.22030
60°F 0.9524 0.95500 0.14496 0.21021 7.6538
0.6826 0.81996 -0.28203 2.5099 2.0117

Summer 0.6316 0.79948 0.02231 4.88674 27.3777

65 °F
0.9710 0.99610 -0.626x10-3 0.01747 0.29005
70°F 0.9054 0.77191 0.41104 -2.97786 1.34336
75°F 1.1460 1.04100 -5.70x10-3 -0.28812 -0.43094
0.4004 0.96455 -0.0205 0.25537 0.99326
1.0830 0.61902 -0.3170 3.92562 -4.40771
80°F 1.9531 1.54463 -0.0670 -3.9866 0.393133
85 °F 2.205 1.49793 -0.0645 -13.213 -9.72238

Autumn 2.0384 1.4272 2.1746 -19.988 -30.0176

45 °F
1.10813 1.0375 -0.03423 -0.7295 -2,95231
0.97157 1.00264 -6.82X10-3 -2.20x10"4 0.04103
50°F 1.01993 1.00522 0.34103 -0.08018 -0.07099
1.1015 1.05752 0.07033 -0.35134 -1.6495
1.2139 1.07564 0.02168 -0.36916 -8.3451
55°F 0.9679 1.07260 -0.29228 -6.02269' -20.1170
0.96135 0.89430 -9.28x10-3 0.51844 0.33500
1.53413 1.2233 0.02482 1.17628 -3.86028
60°F 1.02073 1.0220 -0.05399 -0.94467 -1.78106
0.71339 0,79512 0.39070 -0.13559 3.3182
65°F 0.97220 0.98311 -0.02078 -11.4984 -36.5372
1.08070 1.04255 0.02806 -1.13102 -1.47291
1.2358 1.04862 -0.07368 -7.8949 -1.81121
1.9770 1.53175 0.09076 -27.65119 -39.220
 197

0.74826 0.826653 0.06799 2.6625 7.2238

0.79049 0.89743 -3.257x10-3 2.73579 3.09780

Winter 0.97678 0.93888 -3.02x10-3 -0.04095 -0.20802

25 °F
0.66336 0.74910 -2.32220 3.36146 10.6727
0.03390 0.19516 -0.22166 137.29 471.32
30°F 1.25450 1.8898 -0.02368 0.50791 -1.176
35 °F 0.93717 1.00449 0.0486 0.36145 -2.29x10-3
1.21001 1.02836 -0.02758 -0.65218 -7.0473
57.7506 6.69628 -0.02585 -86.3656 -15.8296
40°F 0.92218 0.9789 -0.01228 0.21380 -0.03105
3.0485 1.85556 -0.95204 -8.52874 -137.92
1.1272 1.2917 0.05763 -2.51320 -17.1397
1.00273 0.97947 0.11944 0.54803 -1.58112
45 °F 0.34335 0.44868 -0.51460 15.8644 19.2410
1.10012 0.94118 0.30433 10.93002 -10.9958
1.1510 0.9855 -0.0714 -0.4002 -5.2148
1.6535 1.51163 0.2633 -20.1513 -7.8087
11.000 2.6401 -0.5273 -55.131 -956.13

Table 7.3 Observed test statistics describing the various forms of change for the
contours of interest.

For the local analysis, the null hypothesis for each test statistic describing the situation of
no change in the geometric quantity, whilst the alternative was that change of some form
had occurred.

The results of these hypotheses tests, table 7.4. were reported in terms of the upper
bound of the random noise at which the observed value of the test statistic was
significant for a 5% significance test i.e. N.S. (non-significant), 0%, 5%, 15% and 25%
i.e. for a reported value of x%, say, the null hypothesis may be rejected in favour of the
alternative if the upper bound of noise in the original data is less than or equal to x%.
Alternatively if the data has an inherent noise level greater than x%, then change maybe
due to random noise and not some underlying physical process.
 198

Season Tempera Area Perimeter Orienta­ Centroid

ture tion Displacement

X y
Spring 40°F NS NS NS NS NS
15% 15% NS NS 25%
45 °F 15% 5% NS NS NS
15% 5% NS 15% NS
50 °F 15% 5% 25% NS NS
25% 5% NS 5% 15%
55°F 25% 15% NS 15% 15%
25% 25% NS 15% 25%
25% 25% 25% NS NS
60°F 15% 15% 15% NS 25%
25% 25% 25% 25% 25%

Summer 65 °F 25% 25% NS 25% 25%

15% 5% NS NS NS
70°F 25% 25% 25% 25% 25%
75°F 25% 5% NS NS 15%
25% 15% NS NS 25%
25% 25% 25% 25% 25%
80°F 25% 25% NS 25% 15%
85°F 25% 25% NS 25% 15%
25% 25% NS 25%. 25%

Autumn 45 °F 25% 25% 25% 25% 25%

25% 5% NS 15% 25%
15% NS NS NS NS
50°F 15% NS 25% NS NS
25% 15% NS NS 25%
25% 15% NS 25% 25%
55°F 15% 15% 25% 15% NS
15% 25% NS 25% 25%
 199

25% 25% NS 25% 25%

60°F 15% 5% NS NS 25%
25% 25% 25% 25% 25%
65°F 15% 5% NS 25% 25%
25% 5% NS 25% 25%
25% 5% NS 25% 25%
25% 25% NS 25% 25%
25% 25% NS 25% 25%
25% 15% NS 25% 25%

Winter 25°F 15% 15% NS 25% 25%

25% 25% 25% 25% 25%
25% 25% 25% 25% 25%
30°F 15% 25% NS 25% 25%
35 °F 15% NS NS 5% NS
25% 5% NS 15% 25%
25% 25% NS 25% 25%
40°F 25% 5% NS NS NS
25% 25% 25% 25% 25%
25% 25% NS 25% 25%
NS 5% 15% 15% 25%
45 °F 25% 25% 25% 25% 25%
25% 15% 25 15% 25%
25% 5% NS 15% 25%.
25% 25% 25% 25%- 25%
25% 25% 25% 25% 25%

Table 7.4 Results for local hypothesis testing procedure.

 200

§7.5.1 Summary of local results

The following general comments may be drawn from the results of the local analysis. For
the separate test statistics, greatest change is witnessed in those statistics expressing
scalar change i.e. area and perimeter.

The orientation of the contours appear fairly static for all seasons and temperature levels,
although a number of anomalous results are recorded. Referring to table 7.3, these
changes appear to be linked to those situations where either, the areal dimension of the
contour is small, the contour is near cyclic in shape or finally, more seriously, where the
contour describing a specific temperature level has become disjoint over time between
1980 and 1930 or vice-versa. This observation is potentially worrisome and an
alternative method for collating contours may be required.

Finally in terms of centroid displacement, although not geographically universal, the

results indicate a general cooling which is most pronounced in the east and a general
warming in a southerly direction, this bears out the findings of Diaz and Quayle (1980).

More specifically, examining each of the seasons in turn

Spring The changes in the scalar statistic may be succinctly summarised by noting the
more disjoint appearance of the contour structure at higher temperature levels for 1980.
In accordance with the previous findings this suggests that a drop in temperature has
resulted. In general a higher level of agreement is observed at cooler temperatures where
the areal dimension of the contours is greater than for physically smaller contours at the
same level.

Overall the changes reported in the other statistics are not of great interest with no trends
evident. The changes in spring are not very strong and this confirms the subjective
impressions.

Summer The scalar trends for summer are similar to those for spring with warmer
temperatures once again being the source for greatest change. For both area and
perimeter the changes are significant for noise levels of 25% and possibly greater.
 201

The centroid displacement results appear to suggest the possibility of a north/south shift
in temperature level. Referring specifically to the test statistics, the indications are that an
increase in temperature has occurred for the north between 1930 and 1980. In a similar
vein, an increase in temperature is more liable to have been reported in the west during
this time span, this is less apparent than for the latitude results.

Autumn Greater changes are evident during the autumnal period than for the previous
seasons. In terms of scalar change no pattern emerges, an areal increase is indicated for
some contours during the 50 year span whilst, for the same temperature level, the reverse
is true.

A semblance of a trend in terms of directional change is indicated. Once again the south
and west both appear to have witnessed a warming of the temperatures during the fifty
years of the study.

Winter Overall the changes are the most diverse for all of the seasons. In terms of the
scalar changes a very complex structure emerges and generalisations are not plausible.
The north/south, east/west divide once again appears to materialise as described for the
preceeding seasons.

§7.6 G lobal Analysis

§7.6.1 Existing Techniques

Some of the more common methods cited in the literature for comparing two sets of
spatial data on a global basis were used to examine the question of change between the
two data sets

1. Correlation analysis
2. Paired t-interval
3. Trend-surface analysis
4. Regression analysis

Before proceeding to execute any of the above techniques, the assumption of normality
was verified for all permutations of season and year. The second assumption of equality
 202

of variances across temperatures for each season was checked using the standard F-test,
the results indicated pooling of variances was valid for the data.

Each of the above methods may be implemented using one of the standard statistical
packages e.g. Minitab, hence their wide appeal to users in other areas.

§7.6.1.1 Correlation analysis

The first method used the ideas of correlation to examine the strength of the relationship
between the variables. The question of independence between the two data sets was
examined by the standard hypothesis procedure. ^7 &ch. ',

HO: p = 0
H I: p * 0

and secondly, a confidence interval was derived for the correlation coefficient using the
following result:-

zW - z (p) n(o 1 )
IK
n -3

The implementation of the above two procedures enables the analyst to formulate an
opinion as to the strength of the relationship. Table 7.5 collates the results :-

Season Correlation Hypothesis Confidence Interval for ( p)

(P) test
(p-value)
Spring 0.975 0.00 0.9721 0.9776
Summer 0.969 0.00 0.9654 0.9721
Autumn 0.963 0.00 0.9587 0.9667
Winter 0.926 0.00 0.9170 0.9381

Table 7.5 Results for correlation based analysis.

 203

The results display a strong association between the two variables. The presence of a
strong physical connection reduces the potential for a weak association between the two
variables.
§7.6.1.2 Paired t-interval

With the underlying assumptions of normality and the pooling of the variances having
been checked, paired t-intervals were calculated for the four seasons. The hypothesis of
interest being :-

HO : M-d = 0
HI : p-d * 0

i.e. under the null hypothesis, no change occurs in the mean value of the differences for
the two data sets whilst under the alternative, some form of change has taken place over
time. Table 7.6 summarises the results.

Season Paired t- interval

Spring (-0.792, -0.574)
Summer (0.264, 0.512)
Autumn (0.310, 0.571)
Winter (-0.575, -0.115)

Table 7.6 Results for the paired t-interval.

From the results it is apparent that overall the temperatures in 1980 are cooler than for
1930 for spring and winter, the greater changes being in the mean value for spring. A
reversal of the pattern is seen for the other two seasons, with little disparity between the
mean changes.

§7.6.1.3 Trend surface analysis

The penultimate technique compared the trend surface for the two sets of data using the
correlation coefficient basis for comparison, section 4.3.4. The most appropriate form of
trend surface was that of a quadratic i.e. of order two
 204

Y j = P 0 + P ix i + P zY i + P s ^ Y i + P<x i + PsY i

i.e. temperature is equal to the sum of a constant term related to the means of the
geographic co-ordinates, plus a polynomial expansion of degree two of the geographic
co-ordinates, plus a randomly distributed measurement error.

The same criteria as described in the section on correlation analysis allows the analyst to
judge the strength of the relationship between the two surfaces. A large value of p is
indicative of a low level of change and vice-versa, the results are given in table 7,7.

Season Correlation value

Spring 0.977
Summer 0.891
Autumn 0.999
Winter 0.981

Table 7.7 :-Results derived for the correlation between the two trend surfaces.

A higher level of accord is suggested by the results of the trend surface analysis than for
the standard correlation analysis. This is a manifestation of the type of analysis, the two
surfaces will inevitably have similar trends due to the underlying physical phenomena
which intrinsically control the temperature levels e.g. altitude, latitude, longitude etc. and
these are constant for the two data sets.

§7.6.1.4 Regression analysis

In terms of a regression, two approaches are possible. Both methods are based on the
model for simple linear regression :-

temPi9so = a + Ptempi930 +Ei £j - N (0, a 2) (1)

Method 1 On the assumption that no temperature change has resulted if the slope (3, of
the fitted regression line equates to one, when a, the intercept, is forced to equal zero, a
standard interval for P was constructed, table 7.8.
 205

i.e. temp1980 = Ptemp1930 + Ei ej _ N(0,a2)

The point estimates for the slope confirm the similarity of the trends for spring and
winter, and summer and autumn, i.e. for the former, the temperatures have fallen during
the fifty year span, the reverse is true for the latter. However in terms of the confidence
interval for the slope, the results for both summer and winter are indicative of no change
over the time period of 50 years.

Season Slope (3) Interval estimate for 3

Spring 0.9850 0.9829 0.9871
Summer 1.0065 0.9997 1.0133
Autumn 1.0064 1.0023 1.0106
Winter 0.9890 0.9712 1.0067
Figure 7.8 : Results for regression slope analysis.

Method 2 The second technique fits regression equation (1) and from the fitted values,
A

y , a surface is produced which superimposes these values against the actual values of y.
Alternatively, a residual map may be used to describe the results, figure 7.7 describes the
surface for spring.

120 - -

Vasaaaamssamma n t - . v
110

100 k.

80 | r

50 40
KEY
1 .................... 6 2 ................ 3.0 3------ 0.0
4 ............... 1.5 5-......... 6.0
Figure 7.7 Residual map for spring
 206

This approach combines a theoretical and subjective based analysis for assessing change.
The advantage of this technique is that the residual surface attempts to incorporate the
spatial dimension which is ignored in the preceding methods. The plot underlines the
simplifications imposed by utilising such techniques and that change cannot be truly
expressed in terms of one parameter.

§7.6.2 Summary of the existing techniques

Correlation based methods provide a measure of the association/change between data sets
but are generally uninformative. For data of this form a strong natural association exists
due to the physical conditions which intrinsically control temperature. In terms of the
trend surface analysis, the resultant measure suffers from the same drawbacks as for a
standard correlation analysis.

The paired t-interval provides greater insight into the mean change particularly in terms
of the direction of change. A potential problem related to ignoring the spatial dimension
is that of auto-correlation and the possibility that the results are more extreme than in
reality. The overall findings confirm the observations postulated in the subjective
analysis: 1980 temperatures have fallen in spring and winter whilst the reverse has
occurred during the other two seasons.

The final method based on the notion of regression saw two possible modes of analysis;
the first gauged change as a function of the slope of the regression line. A slope of one,
indicative of no change when the intercept was forced to equal zero. Once again the same
reservations expressed for a correlation approach hold here. The second of the regression
methods, although not providing a finite solution for change does consider the spatial
dimension, although only subjectively.

Attempting to produce a 'true1 global value for change is liable to be difficult using the
standard statistical methods. Utilising the methodology developed in chapters four and
five, to produce a global test goes part of the way to accounting for the spatial
complexities of change, illustrated so clearly in figure 7.7.
 207

§7,6.3 Global Hypothesis Testing: Approach

The global hypothesis testing procedure used the test statistics describing angular,
translational and scalar change namely angular change, latitude and longitude
displacement and areal change respectively. The minimal number of contours of interest
was five because of the condition imposed when implementing a Hotelling's One-Sample
T^ test, n > p, where n is the number of contour levels and p, the number of variables of
interest. The number of temperature levels examined for each season was five. Within
each temperature level, the surface was represented by more than a single contour.

In terms of the structure of the hypothesis testing procedure, the null hypothesis related
to the situation of no change in all the measures described. The alternative was expressed
in general terms of change.

As mentioned in the previous chapter (§6.7), a number of complications arise when

implementing the multivariate testing procedure. These relate mainly to the definition of
the various descriptors, where contour A has become disjoint over time or alternatively
where two contours have amalgamated. A consistent approach to this problem was
adopted:-

(a) For area, for the specific contour level of interest, the sum was taken.

(b) A global mean was taken for the centroid displacement.

(c) In terms of orientation, a mean difference was effectively used. The whole contour
being split into its two component parts and the orientation for each segment evaluated,
the differences with the original disjoint contour were then taken and'finally, the mean
calculated.

A collation of the necessary mathematical quantities are given in appendix 1, for all the
contours depicting the various temperature levels of interest for each season with table
7.9 describing the results of the global analysis.
 208

Season Hotelling's One P-value

Sample Test
Statistic
Spring 0.93476 0.465
Summer 0.31140 0.820
Autumn 1.04009 0.400
Winter 1.16014 0.360

Table 7,9 Results of global hypothesis testing procedure.

For all seasons, the null hypothesis cannot be rejected in favour of the alternative.
Differences in the results therefore appear to be solely attributable to natural variability.
This statement should be tempered due to a number of drawbacks of the methodology
cited i.e. only those contours which are visually comparable are incorporated and
secondly, restraints are imposed by the selection of the contour level. Selecting contours
in the zone of greatest change i.e. the upper range for each of the seasons may cause a
reversal of the results. Selection based on a pre-defined criteria eliminates this
methodological drawback and area of potential bias.

§7.7 Conclusions

The local approach provides a more detailed insight into the changes which have resulted
over time than the global analysis. However in general it may be stated that some form of
change has occurred for all four seasons. This change is most apparent for winter and
then on a sliding scale, autumn, summer and finally spring. This generalisation glosses
over many of the intricacies of change and how regional differences are evident, the
question of micro-climate possibly influencing the results.

The changes are primarily of a scalar nature and to a lesser extent, translational. In terms
of orientation, it may be reported that almost no change has occurred.

This example has illustrated the simplicity of implementing both approaches and how the
results generally conform to what has been postulated in the subjective impression. The
main point of interest is in terms of the centroid displacement, where some of the test
statistics are highly inflated. This primarily occurs when dealing with areally small
 209

contours where the standard deviation is less than one, as a result, the statistic describing
centroid displacement is considerably inflated. The resultant conclusions are unaffected
since if the two standard deviations are comparable then this problem is eliminated in the
ensuing test statistic, whilst if the two contours differ considerably, the results confirm
the anticipated behaviour.

The drawback as indicated earlier is that we have no estimate of the size of the
difference. This is a problem with all hypothesis based tests. Overall on a local scale,
change of some form is indicated, even if the level of noise, inherent to the data set is
fairly large. On this premise, it is essential that further investigations should be carried
out to see whether this change may be linked to a physical process. Although Handcock
and Wallis (1989) concluded 'It will be 20 to 30 years before change is discernible from
the natural variability in temperatures', the use of methodology of this form shows that
change can potentially be detected before this duration in a good quality data set.

The other factor to emerge from this example relates to the ability of the global test to
take account of the spatial dimension, which to date the majority of techniques have
ignored. Figure 7.7 illustrates the consequences of ignoring this factor. By adopting a
consistent approach to the problematical contours, the global analysis should be a good
indicator of change.
 210

§7.8 Case Study 2 The Investigation Of A Possible Link Between

Leukaemia And The Underlying Radiation Fields

For some years there has been considerable public concern about radiation in general
and man-made radioactivity in particular. Public awareness of the impact of
radionuclides in the environment has been heightened even more by the Chernobyl
reactor accident.

Radiation of natural origin is widespread in the environment. The earth itself is

radioactive and naturally occurring radionuclides are present in the air we breathe, in
the food we eat and in our own bodies. Everyone is exposed to natural radiation and for
most people it is the highest contributor to total dose. Table 7.10 provides an estimate
of the breakdown of the total radiation dose received by the people of Thurso provided
by the NRPB, Dionian (1986),

Source Percentage of Total Dose

Natural Radiation 79.0
Fallout 12.0
Medical 7.5
Dounreay discharges 1.2
Sellafield discharges 0.3

Table 7.10 Summary of breakdown of total radiation dose received by the people of
Thurso, Dionian (1986).

Man-made radionuclides have been distributed throughout the world as a result of

nuclear weapons testing in the atmosphere. These radionuclides are inhaled, deposited
on the ground giving rise to external exposure, and they can also be transferred through
food-chains to our diet. Even though the period of intensive weapon testing occurred
more than twenty years ago, residual activity from these tests and from occasional more
recent explosions, still give rise to some small exposure of the population.

Radioactive materials are discharged from nuclear installations, some industrial

premises and from medical and research institutes. Accidental releases of radionuclides
 211

may also occur and, as Chernobyl has shown, a severe accident at a nuclear power
station can lead to widespread contamination of the environment.

Radionuclides are subject to all the physical, chemical and biological processes of
environmental transfer. No matter how complex the pathway by which the activity may
reach man, the actual routes of human exposure are limited to: external irradiation;
inhalation of airborne material; ingestion of activity in food or water. Measurements are
related to these mechanisms of exposure. Thus, in the environment we measure external
dose rates and activity concentrations in air, food and water.

In this section we are specifically interested in whether background radiation is a

causative factor in the induction of various forms of lukaemia. To date, the spatial
analysis of disease patterns has been used by investigators as one tool with which to
address problems of disease causation. Four main approaches to this type of analysis
have been taken

1. Ecological analysis
2. Mapping and estimation patterns of disease
3. Clustering
4. Regression/correlation analysis

An ecological study is one in which the unit of analysis is a group of individuals, often
defined geographically, and their relationship between the incidence of a disease in
spatial units and other covariates is examined. Making aetiological inferences about
individuals from data on groups is potentially hazardous. However where data on
individuals is unavailable it is useful to express group relationships between areas. The
role of ecological studies in epidemiological research is discussed by Morgenstem
(1982).

The second group, estimating and mapping disease rates, have generally been
concerned with the production of rates relating to the incidence of cancer with the end-
product being a cancer atlas. The rates are usually computed using mortality or
morbidity data as the numerator, with census data providing the denominator. Walter
and Bimie (1991) have reviewed the techniques of analysis, presentation and
interpretation of atlases from various countries.
 212

A more limited approach to the problem of spatial analysis of disease patterns is that of
disease clustering. Hills and Alexander (1989) examine some of the approaches taken
and the associated problems. Generally testing for clustering is aimed at tackling two
issues. First, is there a tendency for clustering to occur and, if so, where? Second, do
clusters occur in specific areas e.g. near suspected environmental hazards.

Finally the use of correlation and regression have featured strongly in testing whether
natural radiation, radon in particular, is a causitive factor in the induction of particular
forms of leukaemia. This has generated two schools of thought, firstly, those who
advocate that ecologically low levels of ionising radiation is harmful to human beings.
Henshaw et al (1990) suggested that in the United Kingdom, 6-12% of myeloid
leukaemias may be attributed to radon. In Cornwall where radon levels are higher, this
increases to 23-43%. This view supported the findings of Kneale and Stewart (1987)
who demonstrated the existence of a correlation between childhood cancer and
background indoor gamma radiation. The second school advocate the hormesis effect,
i.e. low levels of radiation are beneficial. This has been substantiated by both animal
and biological experimentation and to a lesser extent in a number of studies on humans
in Japan, Ujeno (1983), the U.S.A., Hickey et al (1983) and more recently in a larger
study in India, Nambi and Soman (1987,1990).

§7.9 Description Of The Problem And How It Differs To the Previous

Example

The effects of pollution, industrialisation, social deprivation and so forth are

increasingly perceived as the causes of geographical anomalies in health. Within this
section we examine the effect of background radiation on man i.e. 'Is there a link
between the disease pattern of certain forms of leukaemia and the underlying radiation
fields?’

In 1989 the Leukaemia Research Fund commissioned a pilot study to investigate the
feasibility of aerial radiometric survey for generating information on background
radiation levels. The information garnered was to be used in juxtaposition with the
epidemiological data to examine the pre-stated hypothesis. This problem differs from
the previous climatic illustration in the following ways:-
 213

1. It seeks to establish whether an association exists between two sets of differing

spatial variables, cases and population and secondly, radiation levels and cases.

2. The spatial variables of interest differ in spatial resolution. This point is expanded
upon in section 7.10.

Furthermore it differs in form to the more theoretical types of problem which

underwrote the methodological development i.e. the comparison of two sets of punctual
data points. This apart, the analysis addresses two separate issues, firstly the
relationship of the case locations to the underlying population and secondly, the
relationship between the case locations and the radiation fields. Before proceeding to
the analysis stage a general overview of the pilot study is given with a short description
of the three data sources

1. epidemiological data
2. population data
3. radiation data.

§7.10 Pilot study

The pilot study commissioned by the Leukaemia Research Fund in July 1989 required
an aerial survey to be flown over three disjoint regions in South-West England covering
approximately 2 2 5 0 k m 2 in total. In September of that same year the survey was
undertaken. The location of the three grids are shown in figure 7.8.

Figure 7.8 Location of the three grids.

 214

Grid one encompasses approximately 900km^. The main population centres being
Yeovil, Wellington, Illminster and the southern edge of Taunton. The remainder of the
area consists of small hamlets with the Black Down Hills forming the main
geographical feature.

The second grid lies to the north of grid one and is approximately 20kmxl5km. Part of
the north section lies in Bridgewater Bay. The area is predominately rural with two
main towns, Bridgewater and Bumham-on-Sea. Crossing this region is part of the
Sedgemoor Drain system.

Finally grid three which is located to the south-west of grids one and two and west of
Plymouth has Liskeard, Saltash, Torpoint and Launceston as its principal towns. The
southern fringes of Bodmin Moor encroach into this grid as does the estuary of the
River Tamar.

§7.10.1 Radiation Data

The first primary source of information was that of the radiation data. Aerial radiation
survey methods are based on the ability of gamma radiation to propagate up to a few
hundred metres in air from the originating source of radioactivity. It is possible to
monitor the flux of gamma radiation above ground or sea level using highly sensitive
gamma ray spectrometry equipment mounted in aircraft flown close to the ground.
Environmental radioactivity measurements using portable field based spectrometers
may take from 15 minutes to 30 minutes per sampling site, or environmental soil cores
extracted from single sites may take several days each to analyse. High volume aerial
survey equipment can make sensitive readings every few seconds while moving on
preset paths above the land surface.

The survey was conducted with 1km line spacing and 500m resolution along each flight
line using an aerospatiale squirrel helicopter flown at 120km/hr. Raising a detector
above ground opens the detection geometry so that the area on the ground being
sampled increases very rapidly. Typical areas of investigation are such as to give 90%
of the detected signal from a circle diameter 4-5 times the height above the ground.
Thus each observation at a survey height of 100m is averaged over a circle of diameter
of 500m. Details of how the analysis of the data set proceeds to produce values for the
 215

various, radiation fields is given in Sanderson et al (1990). A summary of the set of

steps undertaken are as follows: -

1. Generation of summary files which involves collation of individual readings and

their corresponding positions and altitudes along each flight line. Checks are
incorporated at this stage for any anomalies.

2. Detector background rates are subtracted from the readings.

3. The counts are stripped i.e. the spectral interferences between adjacent channels is
removed.

4. Altitude corrections are made to the stripped count rates.

5. The stripped counts are converted to calibrated data - so that for each location,
equivalent uranium, eU, and thorium, eTh, and potassium, in kBq/kg, and alpha,
beta and gamma dose rates in mGy/a are calculated - using linear equations derived by
regression analysis of ground level concentrations against aerial observations.

§7.10.2 Epidemiological Data

The epidemiological data provided by the Leukaemia Research Fund, (L.R.F.)

comprised a set of all the recorded incidences of leukaemia during the five year period
1984 to 1988 for the three grids surveyed. The case locations were identified by the
Ordinance Survey co-ordinates of the postal code area in which the person resided at
the time of diagnosis. Additional information available on these people included age,
diagnostic code and date of diagnosis.

The diagnostic codes of which there are ninety, serve as identifiers for the type of
leukaemia contracted. Early in 1990, a Leukaemia and Lymphoma Atlas was launched
by the L.R.F. for selected regions of England and Wales. The Atlas was based on ten
medically defined categories of leukaemia of which the diagnostic codes formed the
basis. These ten groups can be split into two broader categories, all lymphoproliferative
diseases and all myeloproliferative disorders. The first group, all lymphoproliferative
diseases contains six of the ten categories whilst the remaining four fall into the
category of all myeloproliferative diseases, table 7.11.
 216

A ll Lymphoproliferative Diseases All Myeloproliferative Diseases

Acute lymphoblastic leukaemia Acute myeloid leukaemia

Chronic lymphocytic leukaemia Chronic myeloid leukaemia
Hodgkin's disease Myeloid dysplasia
Low-grade non-Hodgkin's lymphoma Other myeloproliferative disorders
High-grade non-Hodgkin's lymphoma
Multiple Myeloma

Table 7.11 L.R,F,'s diagnostic breakdown of the various forms of leukaemia.

A total of 377 cases were reported for the three grids. The breakdown of the cases for
the ten diagnostic groupings is given in table 7.12, Neither cases of multiple myeloma
nor myeloid dysplasia were extracted by the L.R.F. from their original records for
analysis.

Diagnostic Group Gridl Grid 2 Grid 3

All lymphoproliferative diseases 155 33 74
Acute lymphoblastic leukaemia 3 ‘3 5
Chronic lymphocytic leukaemia 53 15 22
Hodgkin's disease 22 5 13
Low-grade non-Hodgkin's 46 5 19
lymphoma
High-grade non-Hodgkin's 29 5 10
lymphoma
Multiple Myeloma 0 0 0
Acute myeloid leukaemia 31 10 19
Chronic myeloid leukaemia 8 2 5
Myeloid dysplasia 0 0 0
Other myeloproliferative disorders 18 6 13
All myeloproliferative disorders 57 18 37

Table 7,12 Breakdown of leukaemia cases into diagnostic groups.

 217

By means of a chi-squared test of homogeneity there was no apparent significant

difference between the grids as to the frequency of occurrence for each disease type.

20
60

50

40

30

30

0 10 30 30 40 50 60 70 B0 90 100 0 10 20 30 40 50 60 70 80 90 100

AGE (YEARS) AGE (YEARS)

Figure 7.9(a):- Age distribution for grid 1 Figure 7.9(b):- Age distribution for grid 2
30

20

0 10 20 30 40 50 60 70 60 90 100

AGE (YEARS)

Figure 7.9(c):- Age distribution for grid 3

Figure 7.9 :- Age distribution for the three grids.

 218

For the three grids, the distribution of age was similar with over 70% of the cases aged
fifty plus. In all three instances the distribution was strongly skewed, figure 7.9. This is
a facet of leukaemia since in general the risk for the majority of leukaemia types
increases with age and for some strains, only those persons in their late middle and old
age are afflicted e.g. chronic lymphoblastic leukaemia.

On the basis of previous experience within the L.R.F. particular interest was centred on
three of these disease groupings namely,

1. all lymphoproliferative diseases

2. all myeloproliferative disorders
3. acute myeloid leukaemia.

The main reason for opting for these three categories was firstly that, high dose studies
have tended to identify myeloid leukaemia as one consequence of exposure. Other
reasons for the separation into lymphoproliferative and myeloproliferative groups is
because, aeteologically, they are easily separable. The extraction of acute myeloid
leukaemia is because it is a well recognised group which is highly malignant. Further
breakdown was avoided since some of the categories were already very small.

For the latter two disease categories, a set of controls were supplied by the Leukaemia
Research Fund. The all lymphoproliferative disease category was not supplied with a
distinct control set but it was suggested, (F. Alexander, per comms.), the all
myeloproliferative and acute myeloid leukaemia controls should be combined and serve
as the controls. For categories two and three, controls were produced using a matching
ratio of 3:1 for each grid and a multinomial allocation was used to assign the
appropriate number to the available post-codes with probabilities proportional to

X v(i)n(i)
i

where i - age stratum 0-14,15-64, 65-79

v(i) - overall L.R.F. age-specific incidence
n(i) - estimated stratum population for the post codes
 219

§7.10.3 Population Data

Although the occurrence of the leukaemia cases of interest spanned the five years 1984-
1989, the most readily available population data was from the 1981 small area statistics
(SAS).

A wide range of different areas are available for describing the population base of
interest including enumeration districts, electoral wards, local government districts,
counties and civil parishes. Optimally the unit selected should be small enough to reveal
patterns of interest and yet large enough to present reasonably stable data. Table 7.13
illustrates the difference in size between three areal divisions

Districts Wards Enumeration

Districts
Cornwall 7 130 1109
Devon 10 256 2334
Somerset 5 153 1046

Table 7.13 Comparison of the sizes of population areas for three regions.

Civil parishes as defined at the local government reorganisation in 1974 are either
single enumeration districts or the amalgamation of two or more. An enumeration
district, the smallest areal descriptor covers approximately 150 households. Finally a
ward describes a local government district electoral ward as existed at census day (5^
April 1981).

The nature of the ensuing analysis prompted the utilisation of both civil parishes and
enumeration district data. The enumeration district data gave access to the centroids of
each district, hence it was more suitable for the generation of a population surface. In
rural regions, enumeration districts are much larger than in an urban environment,
however this apart, they enabled a better defined surface to be produced due to the
larger number of districts within the area of interest as opposed to either wards or civil
parishes.
 220

§7.10.4 Data Summary

The proceeding three sections have described various aspects of the data. The most
apparent feature being the difference in the spatial resolution for each of the data sets
and the contrast in the sizes of the data sets, table 7.14.

Data Source Spatial Data Quantity

Resolution
Grid 1 Grid 2 Grid 3
Radiation 1km x 1km 3240 1116 2867
data
Leukaemia point location 212 51 101
data
Population civil parishes 60 32 47
data

Table 7.14 Summary of the spatial resolution and quantity of the three data sources.

The epidemiological data is extremely sparse by nature and concentrated in urban areas,
depicting it as a continuous surface is liable to introduce large amounts of inherent
error. The use of enumeration districts rather than the more sparse civil parish data
enables a plausible surface to be constructed. Finally the radiation surface is generated
using transect data and is near continuous in description.

These differences cause problems in the ensuing analysis since the methodology
developed earlier requires the data to be not too sparse and preferably depicted as a
series of point values rather than areal regions as in the case of the population data.
However certain aspects of the methodology developed earlier provide possible
directions of investigation.
 221

§7.11 Analysis

§7.11.1 Introduction

As mentioned previously two basic comparisons were of interest, firstly the relationship
of the case locations and the overall population and secondly, the relationship between
the radiation fields and case locations.

The format of the data as summarised in the preceeding section required that in addition
to spatial techniques, i.e. the mapping of case locations on the radiation fields and the
population surface to assess their comparability, other approaches were required,
namely,

1. Histograms of the number of cases and controls in the different radiation field
levels.

2. Kolmogorov-Smimov tests to compare the cumulative distribution functions for

both cases and controls.

3. Description of various summary statistics for the different radiation variables for
both cases and controls.

4. Basic statistical analysis of 'risk' against radiation level.

§7.11.2 Analysis of case locations and population data

Before proceeding to examine the main hypothesis of interest, 'Is there an association
between the spatial dispersion of background radiation fields and the location of
leukaemia cases?', it was essential to examine whether any anomalies existed between
the location of leukaemia cases and the underlying population for the three regions of
interest. The interpretation of areas of 'high' radiation levels where a large number of
cases are located depends on the underlying population density of the area. The
following graphical procedures examined this preliminary question:-
1. Lorenz curves.
2. A map of the population density with cases superimposed.
3. A weighted difference histogram.
 222

Graphical results are presented for grid one for each of the three leukaemia groups. For
the remaining two grids the results will be briefly summarised. Full details of the pilot
study and the statistical analysis are given in Sanderson et al. (1992).

The first technique was that of the Lorenz curve, figure 7.10, described in section
4.3.1. There appears to be some disparity between the case population and the
underlying general population. This increases from disease categories one to three.
However caution should be expressed in the interpretation of these results. The major
difficulty when implementing this and subsequent techniques stem from working with a
grid whose bounds are clearly defined by straight lines. How should areal regions
which are only partially contained within the grid be dealt with? Three plausible
approaches are:-

1. Only include those parishes which are wholly contained within the grid.

2. Estimate the proportion of the population which falls within the grid.

3. Incorporate all civil parishes but display caution in the interpretation of the results.

The plots have all been presented with all the parishes contained within the bounds,
either partially or entirely.

>00 , 100
in in
LJ UJ
in in
o eo ■ u 80
Lu tu
a o
UJ LU
u 13
-< 60 ■ -C 60
t—
JZ 3T
LU LU
CJ
or £
at 40 ■ o’ CO ■
LU UJ
>* >*

j 20 ■ j 20 ■
s: 33:
U CJ
0 0 ■
0 20 CO 60 00 100 0 20 CD 60 80 100
c u m u l a t iv e p e r c e n t a g e o f p o p u l a t io n EMULATIVE PERCENTAGE OF POPULATION

Figure 7.10 (a) All lymphoproliferative Figure 7.10 (b) All myeloproliferative
cases. cases.
 223

cn
ft
(5 80

ft
ft
-< 60
•3T

ft 40

>■
UJ

*—

ft 20
CJ
o

0 20 40 60 00 100

CUMULATIVE PERCENTAGE OF POPULATION

Figure 7.10 (c) Acute myeloid

leukaemia cases.
Figure 7.10 Lorenz curves for the three diagnostic categories of interest for grid 1.

The three measures cited in chapter 4.3.1 for quantifying the apparent level of
dissimilarity within a Lorenz curve are summarised in table 7.15. As anticipated from
the preceeding discussion, the levels of dissimilarity are fairly high with acute myeloid
leukaemia displaying the weakest association between cases and population.

Dissimilarity All All Acute myeloid

measure. lymphoproliferative myeloproliferative leukaemia
diseases. diseases
^m 32.2668 46.2941 57.0049
Dl 32.2671 46.2942 57.0051
Df 44.3800 61.6900 73.8960
Table 7.15 Results for various dissimilarity indices for Lorenz curves.

Figure 7.11 illustrates a histogram based approach to describe the potential

association/differences between the relative frequency of cases and population for the
civil parishes contained within the region. A second histogram explicitly depicts the
differences in the two frequencies which are shown superimposed in the first diagram.
The problem of incorporating parishes only partially enclosed in terms of the number of
cases still materialises, but unlike for Lorenz curves, it is possible to identify those
parishes where the greatest disparity occurs and check their location. Any anomalies are
 224

easily identifiable using this mode of presentation. A similar set of conclusions are
reached as for Lorenz curves.

0.2 -- KEY

UJ
ZD POPULATION
CD
UJ
Eo. ------------ CASES
UJ
>►

QCu>0
UJ f) Lr\ ,4 J n n , .0 . tRt1j J1 J L xEL-EDOtflm L f l f t , . , jlo u jtl J Frtf n J L iTijJlrh-rfffkn

0 10 20 30 40 50 60 70 80 90 100 110

C IV IL PARISHES
cn
0. 1 ■■ DIFFERENCE IN
LlJ
CJ
RELATIVE FREQUENCY
LjJ „
cc 0 .
LU
0 lP n ■ ij iJ l j rKji JuJli ILflrw-JljTiJjj.,
BETWEEN CASES
0 10 20 30 40 50 60 70 80 90 100 110
3- 0 . 1 4-
r .i AND POPULATION

Figure 7.11(a) Comparison of civil parish population and number of all

lymphoproliferative cases.

KEY
t_3
LU
POPULATION
C3
11i
a:
CASES
UJ

0 10 20 30 40 50 60 70 80 90 100 110

C IV IL PARISHES
cn0' 1 DIFFERENCE IN
LU
CJ
RELATIVE FREQUENCY
a 0.0
LU
tT- A -J
Ll. BETWEEN CASES
0 10 20 30 40 50 60 70 80 90 100 110
S-0 . 1 4- AND POPULATION

Figure 7,11 (b) Comparison of civil parish population and number of all
myeloproliferative cases.
 225

^0.2 + KEY
LJ
zLU
ZD ----------- POPULATION
C3
LU
£l0. 1 + -------------- CASES
LU
>■
rt
^0.0 riL--.}P4l-f JJJ1
0 10 20 30 40 50 60 70 80 90 100 110

C IV IL PARISHES
co
LU
0 . 1 ■- DIFFERENCE IN
LJ
RELATIVE FREQUENCY
□2*0 . 0
LU
jitit1Jk j u J
BETWEEN CASES
t 0 10 20 30 40 50 60 70 80 90 100 110
° - 0 . 1 4- AND POPULATION

Figure 7.11 (c) Comparison of civil parish population and number of acute myeloid
leukaemia cases.

Figure 7.11 Comparison of civil parish population and number of cases for various
disease types.

The preceeding two approaches were applicable due to the choropleth style of the
population data. The availability of centroid data for enumeration districts enabled a
spatial based approach to the problem to be adopted. The population surface was
constructed using the a-priori criteria of section 3.9. The sparsity and local
concentration of cases in a small number of regions restricts the feasibility of fitting a
surface through these points. Representing the cases as points conveys a good
impression of the spatial dispersion of the cases in relation to the underlying population,
figure 7.12, but curtails the implementation of the local hypothesis testing procedure.

Once again there do not appear to be any anomalous results, the majority of the cases
being located in towns, suburbs or for the single cases in some of the smaller rural
villages which typify this area.
 226

KEY

W - WELLINGTON

T - TAUNTON

Y - YEOVIL

I - ILMINSTER
50000 50500 5(000 5(500 52000 52500 55000 53500 54000 54500 35000

Ordinance Survey Co-ordinates

Figure 7.12 (a) Location of all lymphoproliferative leukaemia cases and population
distribution

KEY

W - WELLINGTON

T - TAUNTON

Y - YEOVIL

I - ILMINSTER
50000 50500 5(000 51500 52000 33500 55000 53500 54000 54500 35000

Ordinance Survey Co-ordinates

Figure 7.12 (b) :- Location of all myeloproliferative leukaemia cases and population
surface

KEY

W - WELLINGTON

13000 T - TAUNTON
1
11500 Y - YEOVIL

I - ILMINSTER
30500 3(000 31500 52000 52500 53000 53500 54000 54500 35000

Ordinance Survey Co-ordinates

Figure 7.12 (c) :- Location of acute myeloid leukaemia cases and population surface

Contour Level
1 2 3 4 5 6

Population level 1200 1700 2100 2600 3100 3600

Figure 7.12 :- Location of leukaemia cases in relation to population surface.

 227

Unsurprisingly for all three methods, the preliminary analysis of case and population
densities in the main is controlled by the population density. A similar set of general
trends are apparent for grids two and three hence it may be concluded, that the
behaviour for all the grids is similar in terms of case locations and population with no
obvious differences apparent.

§7.11.3 Analysis of radiation fields and case locations

The previous section highlighted the difficulty of separating the case and population
distributions. The next step was to examine whether background radiation is
instrumental in determining the location of leukaemia cases i.e. are the cases of
leukaemia primarily located in areas corresponding to high/low radiation levels.

One of the major drawbacks of an analysis of this type is the paucity of the case data set
and secondly the matching of the cases and controls to a radiation grid cell. Due to the
shortcomings of the epidemiological data, the grid cell is the one in which the case was
located at the time of diagnosis, it may not represent the value of the radiation field at
the time of conception when the cells are most sensitive nor will it describe the total
radiation life history of the case.

In terms of the restricted size of the data set and the breakdown into three separate
grids, major differences would be required between the case and control radiation levels
for any change to be detected.

40 r Gamma
Difference cases controls Difference cases controls
(kBqm‘2) (ni) (n?) (mGy/a) (ni) (n?)
5 12383 37149 0.01 530 1590
10 3096 9287 0.015 236 708
15 1376 4128 0.02 133 399
20 744 2322 0.05 22 66
25 496 1488 0.07 11 33
30 344 1032 0.10 5 15

Table 7.16 Sample sizes to detect change of specific size.

 228

Table 7.16 summarises the approximate sample sizes which would be required to detect
a change in the radiation level between a group of cases and controls for a 5%
significance level, assuming both normality and a value for the standard deviation.

In terms of typical values for natural variability over Britain, a change of O.OlmGy/a in
the background level of gamma radiation corresponds to an increase/decrease of 3% in
the natural variability, whilst the detection of a change of O.lmGy/a equates to a 30%
change in the natural levels over Britain.

Within this section the underlying population is represented by a set of control locations
selected as described in section 7.10.1. Each case and control location, identified by its
ordinance survey co-ordinate, was matched using a nearest neighbour technique to the
appropriate radiation field.

Figure 7.13 displays the raw histograms of the radiation features, for the nuclide ^Or ,
for the three disease categories for grid one. The bins were selected in accordance with
a scheme implemented by Sanderson et al (1990). For the case and, to a lesser extent,
the control distribution, two features were apparent, bimodality and skewness.

For the remaining five radiation fields similar trends were reported. Also for grids two
and three the range of radiation values was comparable between the cases and controls.
Bimodality is weaker in grid two than for either of the other two grids, however the
lack of symmetry property is present in both grids.
 229

0 300 *00 600 600 1000 1300 1400 1600 0 200 400 600 600 1000 1300 1*00 1600

40K (kBqm-2> 4°K (kBqm-2>

Figure 7.13 (a) All lymphoproliferative Figure 7.13 (b) All myeloproliferative
diseases. diseases.

30 KEY

Cases

Controls

0 300 400 600 BOO 1000 1300 1400 1600

4°K (kBqm-2)
Figure 7.13 (c) Acute myeloid
leukaemia.

Figure 7.13 Histogram of cases and controls for grid one for the three diagnostic
categories of interest.

For both sets of information there is evidence of bimodality and lack-of-symmetry,

hence performing the standard parametric statistical tests for comparing univariate data
 230

sets would be inappropriate, due to the violation of the basic assumptions. On this
premise, table 7,17 collates a set of summary statistics for grid one for the three
conditions of interest and the six radiation fields.

All Ivmphonroliferative diseases

N Summary 40k eTh eU Alpha Beta Gamma
statistic
Cases 154 median 529.1 39.98 49.78 18.225 2.260 0.400
minimum 174.6 18.88 25.64 8.96 1.10 0.200
maximum 1243.3 60.70 68.5 24,7 4.30 0.660
Controls 300 median 514.1 38.87 49.46 17.99 2.230 0.395
minimum 195.0 18.44 23.11 8.74 0/930 0.190
maximum 1422.2 66.61 72.23 25.04 4.68 0.670

N Summary 40K eTh eU Alpha Beta Gamma

statistic
Cases 57 median 517.1 39.17 48.5 17.60 2.20 0.400
minimum 184.2 23.74 28.61 12.03 1.31 0.270
maximum 1223.4 52.47 62.47 22,68 4.21 0.640
Controls 216 median 516.2 38.81 49.13 17.86 2.23 0.395
minimum 195.0 19.63 23.11 8.92 0.95 0.200
maximun 1422.4 66.61 72.23 24.70 4.68 0.670

Acute Mveloid Leukaemia

N Summary 40k eTh eU Alpha Beta Gamma
statistic
Cases 31 median 517.1 38.7 48.0 17.60 2.160 0.380
minimum 331.0 23.74 35.7 12.03 1.470 0.290
maximum 1190.3 52.28 61.3 21.96 4.210 0.640
Controls median 503.6 39.06 50.66 18.42 2.225 0.395
minimum 198.6 18.44 25.3 8.74 0.930 0.190
maximum 1192.0 56.62 69.81 25.04 4.230 0.640

Table 7.17 Summary statistics for grid one.

 231

For both all lymphoproliferative diseases and acute myeloid leukaemia the median
value for 4^K is noticeably lower for the controls. For the other radiation fields,
differences are less noticeable although, in general, for all lymphoproliferative
diseases, the median values for the controls are lower. One potentially interesting
feature arises for acute myeloid leukaemia where, with the exception of 4^K, the cases
all have a lower radiation value than the controls.

These results indicate that the collation of the data for all three grids would possibly
mask some interesting features, however as emphasised earlier to detect significant
differences, the changes are required to be fairly large.

1 .0 ** 1.0 t *
0 .9 0 .9

0.8 0.8

|o , S 0 .7

£0.6
till* 1°-6
* 0 .5 * o .s
I—

~ 0 .5

i° .
0.1

0 .0
200 400 600 000 1000 1200 H0Q 1600 200 400 000 NO 1000 1200 1400 1600

4°K (kBqnr2) 4QK (kBqm-2)

Figure 7.14 (a) All lymphoproliferative Figure 7,14 (b) All myeloproliferative
diseases. diseases.
1.0 t • • •

0.9

0.8 KEY
|o ,

g0.6
-+* Cases
0.5

io.4
£ 0 .3 Controls
i 0,2
0.1

0.0
200 400 600 NO 1000 1200 1400 1600

40r (kBqm-2)
Figure 7.14(c) Acute myeloid
leukaemia.
Figure 7.14 Cumulative distribution function for the three disease categories.
The third technique implemented was to compare the case and control figures using the
Kolmogorov-Smirnov test, the cumulative distribution functions are given in figure
7.14. For all three grids and disease characteristics none of the results are significant,
table 7.18.

Once again the bimodality feature described earlier is apparent from the diagrams.

Disease Group Value of Test Statistic Critical Value for a 5 %

significance level
All lymphoproliferative 0.60526 1.36
diseases
All myeloproliferative 0.67155 1.36
diseases
Acute myeloid leukaemia 0.47585 1.36

Table 7.18 Results of Kolmogorov-Smimov test.

The preceeding techniques in this section have all ignored the spatial context of the
radiation data and as seen from the previous example on climatic change, this may
result in some of the underlying complexities being smoothed over. Within the
remainder of the section, a spatial based approach will be adopted to assess the cause of
the bimodality.

Initially, the generation of a radiation surface was hindered due to a problem in the
selection of a suitable smoothing parameter, h, since minimisation of the score function
was not possible. The source of the problem was traced to the near repeatability of
points within the data set. Silverman (1978) stated that real data are nearly always
rounded or discretised to a greater or lesser degree. For a discretised data set
x1,x2,...,x n, let m be the number of pairs i < j for which = xr If a data set of size n
is discretized to a grid of k points, no matter how the data points fall, it can be shown
by Jensen's inequality, Feller (1966);-
 233

If the data are all nonuniform, m will generally be much larger than the minimum value

given by (1). If — is larger than some theoretical value p, depending on the kernel k,
n
then the least squares cross validation score function tends to minus infinity as h —» 0
and hence least squares cross validation as it stands will choose the degenerate value
h=0 for the window width.

Not only is it dangerous to use least squares cross validation on discretised or repeated
data but Silverman (1978) emphasises that the behaviour of the score function for small
h is highly sensitive to a very fine small scale effect in the data.

One approach to avoid this problem is to perturb the data points by a small amount. The
data points were gradually perturbed by an increasing amount until a realistic estimate
of the smoothing parameter was achieved. This amount is individual to each data set. In
practice the aim is to perturb the data set sufficiently that a value of h is selected which
is non-infinite. For the case of the radiation surface, it was decided to ensure that by
perturbing the data, the point did not encroach into a neighbouring grid cell, i.e. the
limits of perturbation were 1km X 0.5km, the spatial resolution of the data.

In an attempt to locate the cause of the bimodality, the information garnered from the
histograms andcumulativedistribution functions was used.The case data was split into
two groups at the pointwhere the two distributions appear to split. For 40K this was
taken as 500kBqm_2- Figure 7.15 (a) describes the radiation surface for grid one, with
the location of the acute myeloid leukaemia cases superimposed, whose levels were less
than SOOkBqnr^- Figure 7.15 (b) describes the scenario for cases with levels greater
than SOOkBqnr^.

KEY

W - WELLINGTON
12500

T - TAUNTON
•O
Y - YEOVIL
11500-

I - ILMINSTER
i iooo !
30000 30500 31000 31500 32000 32500 33000 33500 3*000 3*500 35000 35500 33000

Figure 7.15 (a) radiation surface with cases whose levels < SOOkBqm"^
superimposed.
 234

KEY

V - WELLINGTON

T - TAUNTON

Y - YEOVIL

I - ILMINSTER

Figure 7.15(b) radiation surface with cases whose levels > 500kBqm'2
superimposed.

Figure 7.15 radiation surface with cases superimposed.

A fairly-conclusive explanation of the bimodality emerges. The first set of cases (^ K <
500kBqm"2) were centred around Yeovil, where the radiation levels are generally
lower than those in the Taunton region and the second group were focused in the
Wellington region, where levels were higher. A number of anomalies are apparent but it
appears the geology of the area is a major factor in causing the bimodality. Although
not so pronounced the same pattern was repeated for the controls.

A similar diagnosis may be placed on the bimodality for grid three, but it is less
striking. The cases with lower levels of occur in the Torpoint, Saltash area, whilst
the cases with higher levels are scattered in the north of the region.

§7.11.5 Analysis Based On A Leukaemia Rate

Preliminary investigation of whether there is a link between the location of the cases
and the radiation field has been graphical in nature to this point. For a more formal
approach to the investigation of the hypothesis of interest, a leukaemia rate was
defined

y . . Number of cases in region x

Leukaemia rate = --------------------------- ------
Population in region x

As mentioned previously, three sources of information are available:-

 235

1. Case data for the three grids surveyed, identified by the ordnance survey co­
ordinates of the postal district for that case.

2. Radiation data, in integrated form, typically corresponding to a spatial resolution of

500m x 1000m. Each radiation value was associated uniquely with the ordnance survey
co-ordinates of the centroid of the grid cell.

3. Population data from the 1981 census, for each of the enumeration districts
contained within the grids.

For this problem, the region of interest was sub-divided into grid cells of size 500m x
500m. For each cell a value was calculated for :-

1. number of cases
2. radiation level in region x
3. population in region x

The number of cases was easily derived for each cell by simply checking whether a
point lay within the cells bounds. The radiation levels were also simple to evaluate
since from the aerial survey values had been calculated for a resolution of 1km x
0.5km, by using a nearest neighbour technique, an appropriate value was calculated for
the centroid of the cell.

For the population in region, x, a more complicated approach was required since an
enumeration district spanned both complete and segments of grid cells. As an initial
solution to the problem, a surface produced using kernel density estimation was fitted to
the data and the population for a given small area was evaluated by numerically
integrating under the estimated surface.

For the analysis based on a leukaemia rate, the radiation data was categorised into eight
groups defined a-priori. For these groups, the population and case figures for all grid
cells of a particular radiation level were summed. The leukaemia rate was then
estimated. The resultant data was of of the form, (Xj,^): where xi is the leukaemia
rate, t i the radiation value (mid-value) in category i. For some of the analysis, it was
necessary to collapse the eight cells down to four due to the sparsity of the leukaemia
 236

data, table 7.19 describes the results for grid one for 40k, for all lymphoproliferative
diseases, (ALL).

Radiation Population No. of cases Rate Combined

value total (ALL) (xi) rate
(kBqm_2)
100 13.77 1 72.60 x lO'3
300 3518 18 5.117x10-3 5.380 x lO'3
500 18813 84 4.465 x lO'3 4.465 x lO'3
700 995 5 5.025 x lO’3 5.025 x lO'3
900 5041 16 3.174 x lO'3 3.174x10-3
1100 4187 24 5.732 x 10-3 5.732 x lO'3
1300 712.7 3 4.209 x lO'3 3.854 x lO'3
1500 65.8 0 0

Table 7.19 Example of leukaemia rates for grid one.

Performing any statistical test on a maximum of eight data points does not provide
conclusive evidence of an association, but it may indicate that an analysis of this type is
plausible for data sets of this form, or alternatively, there may be preliminary
indications of an association, which may advocate the implementation of a larger study.

A regression analysis leukaemia rate against radiation level displayed a general lack of
consistency between both grids and leukaemia types. For some situations a quadratic fit
was the best, whereas a linear trend sufficed for others. The major problem lies in
attempting to summarise a complex and variable situation in a simplistic manner. The
error rate on the variables are known to be large especially for the case data, where
error rates follow a Poisson distribution i.e. Vn. The use of Poisson regression may be
more applicable in this instance and it has been used in the study of cancer incidence,
Gail (1978).

The second analysis of rate performed was an analysis of variance. For each of the six
different radiation variables (eU, eTh, K^O, a, P and y), the factors examined were grid,
leukaemia type and level of radiation value. The response variable was leukaemia rate.
 237

Model:- Xljk = n + a i +Pj + y k+(aP)lj (1)

X “ i=0 X P i= 0 2 v k =0 K ap^O Vi X (« P )9 = ° VJ
i j k i j

where Xijk = leukaemia rate for grid i, leukaemia type j, and

radiation level k.

By adopting model (1), we are attempting to assess which factors are important in
influencing the rate. Table 7.20 reports the results.

Radiation Leukaemia Grid Radiation Leuk.

Field type Level type/grid
r 40 Significant Non- Significant Non-
Significant Significant
eU Significant Significant Significant Non-
Significant
eTh Significant Non- Non- Non-
Significant Significant Significant
a Significant Significant Non- Significant
Significant
B Significant Significant Non- Significant
Significant
V Significant Significant Significant Significant

Table 7.20 Results of performing an analysis of variance on grid (1,2,3), radiation

level (1-8) and leukaemia type (1,2,3,4).

From table 7.20, differences exist between the leukaemia types, but also between the
grids for some of the radiation fields. The first of these results is unsurprising since it is
well known that the rates for the three types of leukaemia differ, table 7.21, but the
second result suggests that the rates differ between grids, this may indicate contrasting
environmental reasons. The radiation variable was found to be significant for total
gamma, and eU the last two being of likely geological origin, this again indicates
differences between the grids. The confounding nature of the analysis clearly makes it
 238

difficult to disentangle the individual effects, and the apparent differences amongst
grids argue against increasing the sensitivity of the analysis by pooling the data from
the three grids.

Disease Type Relative risk for Cornwall

All lymphoproliferative diseases 102.5 -110.0
All myeloproliferative diseases 97.5 -102.5
Acute myeloid leukaemia 90.0 - 97.5

Table 7 . 2 1 Differences in rates between the three leukaemia types for Cornwall.

§7.12 Summary

The study and its analysis has demonstrated the potential for linking epidemiological
data and radiation fields. The interpretation of the results is not however clear-cut, the
three geographical regions demonstrate interesting differences, with grid two showing
anomalous features, not least an indication that cases appear to be associated with lower
radiation levels than controls.

Difficulties encountered include the paucity of the leukaemia data which serves as a
restriction to the implementation of the hypothesis methodology and other forms of
statistical analysis. The usage of three grids as oppoosed to one, dimished the potential
power of the study, hence only tentative conclusions could be reached. The spatial
framework of the analysis which supported many of the ideas in the preceeding sections
would be plausible for diseases which are more common. By utilising a spatial analysis,
a clearer picture of what is going on is obtainable. Addressing the problem, by ignoring
the spatial dimension leaves many questions unanswered, as demonstrated by the
location of the apparent source of bimodality using a spatial approach.

Appropriate developments for further work would be to

1. select two regions with very different radiation features

and/or
2. increase the number of cases.
 239

§7.13 General Conclusions

The two case studies have demonstrated the pros-and-cons of the methodology
developed. The climatic example has shown the simplicity of implementing the
hypothesis transformation technique at both a local and global level. It also emphasises
the need for a spatial framework on which to base an analysis. Many of the analyses
carried out ignored the spatial dimension, illustrated a strong relationship between the
two sets of data, but on closer examination, it was only to be anticipated from the
nature of the data. The local hypothesis testing procedure enabled a much more detailed
resume of the changes to be portrayed with some interesting features concerning
differences between regions being highlighted.

Although for the second example no formal analysis was undertaken to establish
whether a link exists between radiation levels and case locations in the south-west of
England, this was primarily a result of the design of the pilot study, however various
graphical and semi-quantitative approaches were undertaken. Firstly the analysis of
three disjoint regions resulted in the already limited number of cases being reduced to
sizes less than desirable for any formal statistical analysis. The selection of one
complete grid would have increased the power of the study and the problem of
bimodality would possibly not have manifested itself. This apart, the methodology of
hypothesis testing prompted the form of the subjective spatial analysis undertaken.

In general, from the examples examined, it appears that the local hypothesis testing
procedure is a step in the right direction when comparing two sets of data which may or -
may not be compatible in terms of spatial resolution. It addresses the complexity of the
data introduced by the spatial dimension and may also give the analyst a pointer as to
the cause of the change, if they are unaware of the form of the change i.e. scalar,
rotational or translational, A number of problems, especially in terms of disjoint
contours, still exist with the method but hopefully these can be resolved .
 240

REFERENCES

Abramson, I.S. (1982). On bandwidth variation in kernel estimates - a square root law.
Ann. Statist, 10,1217-1223.

Agee, E.M. (1982). A diagnosis of twentieth century temperatures at West Lafayette,

Indianna. Climate Changes 4, 399-418.

Agterberg, F.P. (1964). Methods of trend surface analysis. Quarterly Colorado of Mines,
59,111-130.

Anderson, P. (1970), The uses and limitations of trend surface analysis in studies of
urban air pollution. Atmospheric Environment, 4,129-147.

Armstrong, M. (1984), Problems with universal kriging. Mathematical Geology, 16(1),

101-108.

Armstrong, R.W. (1969). Standardised class intervals and rate computation in statistical
maps of mortality. Annals Association American Geographer, 59, 382-390.

Baddeley, AJ, (1987). A class of image metrics. Proc. ANZ AAS Congress, Townsville,
Queensland, Australia.

Bengtsson, B.E. and Nordbeck, S. (1964). Construction of isarithms and isarithmic maps
by contour. Report BIT4, University of Lund, Lund, 87-105

Berezin, I.S. and Zhidkov, I.P. (1965). Computing Methods, Vol. 1, Chapter 2. Addison-
Wesley, Reading, Mass.

Besag, J. (1986). On the statistical analysis of dirty pictures (with discussion). Journal
Royal Statistical Society B, 48, 259-302.

Besag, J. and Kempton, R.A. (1986). Statistical analysis of field experiments using
neighbouring plots. Biometrics, 5, 351-360.
 241

Besel, P.J. and Jain, R.C, (1988). Segmentation through variable order surface fitting.
IEEE transactions on pattern analysis and Machine Intelligence, 10,167-192.

Birkhoff, G. and Mansfield, L. (1974). Compatible triangular finite elements. Journal

Mathematical Analysis and Applications, 47(3), 531-553.

Bithell, J.F. (1990). An application of density estimation to geographical epidemiology.

Statistics in Medicine, 10,691-701.

Blair, D J. and Bliss, T.H, (1967). The measurement of shape in geography. University of
Nottingham, Department of Geography, Bulletin of Quantitative Data for
Geographers, 11.

Blum, H. (1973). Biological shape and visual science. Journal Theoretical Biology, 38,
205-287.

Blumenstock, D.I. (1953). The reliability factor in the drawing of isarithms. Annals of
the Association of American Geographers, 290-304.

Bookstein, F. L. (1978). The measurement of biological shape and shape changes. In:
Levin S. (ed) Lecture notes in biomathematics 24. Springer-Verlog New York.

Bookstein, F. L. (1984a). A statistical method for biological shape comparisons. Journal

of Theoretical Biology, 107,475-520.

Bookstein, F. L. (1984b). Tensor biometrics for changes in cranial shape. Annals of

Human Biology, 11,413-437.

Bookstein, F.L. (1986), Size and shape spaces for landmark data in two-dimensions.
Statistical Science, 1,181-242.

Boots, B.N. and Lamoureaux (Jr.), M.S. (1972). Working notes and bibliography on the
study of shape in human geography and planning. Council of Planning
Librarians, Exchange Bibliography, 346,1-22.
 242

Bowman, A.W. (1984). An alternative method of cross-validation for the smoothing of

density estimators. Biometrics, 71, 353-360.

Boyce, R. R. and Clark, W.A.V. (1964). The concept of shape in geography.

Geographical Review, 54, 561-572.

Bradley, R. (1970), The excavation of a beaker settlement at Belle Tout, East Sussex.
Proceedings of the Prehistoric Society, 36, 312-379.

Breiman, L,, Meisel, W. and Purcell, E. (1977). Variable kernel estimates of multivariate
densities. Technometrics, 19(2), 135-144.

Brooks, C.E.P. and Carruthers, N. (1953). Handbook of Statistical Methods in

Meteorology, Her Majesty's Stationary Office, London.

Cacoullos, T. (1964). Technical Report No, 40, Dept of Statistics, University of

Minnesota,

Cacoullos, T. (1966). Estimation of a multivariate density. Annals InstStatist, Math,, 18,

179-189.

Clarke, I. (1979). Practical Geostatistics. Applied Science Publishers, London.

Cleek, R.K. (1979). Cancers and the environment: The effect of scale. Society Science
Medicine, 13D, 241-247.

Cliff, A.D. (1970). Computing the spatial correspondence between geographical patterns.
Transactions, Inst, British Geographer, 50,143-154.

Cliff, A.D., Haggett, P., Ord, J.K., Bassett, K. and Davies, R.B. (1975). Elements of
Spatial Structure: A Quantitative Approach, Cambridge University Press, London.

Cliff, A.D. and Ord, J.K. (1981). Spatial processes: Models and applications. London:
Pion.
 243

Cole, J.O. (1964). Study of major and minor civil divisions in political geography. Paper
presented at 20th International Geographical Congress, Sheffield, England.

Coombes, A., Linney, A.D., Richards, R. and Moss, J. (1991). A method for the analysis
of the 3-D shape of the face and changes in the shape brought about by facial
surgery. Biostereometrics Technology and Applications, ed. Herron R., Procc.
SPIE 1380,180-189

Coons, R.L., Woolard, G.P. and Hersherg, G. (1967). Structural significance and
analysis of mid-continent gravity high. Bulletin of the American Association Of
Petroleum Geologists, 51, 2381-2399.

Cottafava,G. and le Moli, G. (1969). Automatic Contour Maps. CACM, 12, 386-391.

Court, A. (1970). Map comparisons. Economic Geography, 46,435-438.

Crane, C.M. (1972), Contour plotting for functions specified at nodal points of an
irregular mesh based on an arbitrary two parameter co-ordinate system. The
Computer Journal, 15, 382-384.

Dacey, M.F. (1964). Measures of contiguity for 2-color maps. Technical Report,
Northweston University.

Davies, J, (1973). Statistics and Data Analysis in Geology, (2nd edition), John Wiley and
Sons Inc.

Dayhoff, M.C. (1963). A contour map program for x-ray crystallography. CACM, 6,
620-622.

Diaz, H.F. and Quayle, R.G. (1980). The climate of the United States since 1895: Spatial
and temporal changes. Monthly Weather Review, 108, 249-266.

de Boor, C,, (1962). Bicubic spline interpolation. Journal of Mathematics and Physics,
41, 212-218.
 244

Dionian, J.}Muirhead, C.R., Wan, S,L. and Wrixon, A.D. (1986). The risks of leukaemia
and other cancers in Thurso from radiation exposure. Report NRPB-
R196.London, Her Mayesty's Stationery Office.

Dobson, M.W. (1973). Choropleth maps without class intervals? A comment. Geog.
Anal., 5, 358-360.

Efron, B. (1979). Bootstrap Methods: Another look at the jacknife. The Annals of
Statistics 7,1-26.

Erlich, R., Baxter Pharr, R„ Healy-Williams, H. (1983). Comments on the validity of

Fourier descriptors in systematics: A reply to Bookstein e ta i Syst. Zool., 32(2).
202.

Evans, I. (1976). The selection of class intervals. Transactions, Institute of British

Geographers, 2(2), 98-124.

Evans, I.S., Catterall, J.W. and Rhind, D.W. (1975). Specific transformations are
necessary. Census Res. Unit Working, Paper 4, University of Durham.

Faraway, J.J, and Jhun, M. (1990). Bootstrap choice of bandwidths for density
estimation. Journal of the American Statistical Association, 85(412), 1119-1122.

Feller, W. (1966). An introduction to probability theory and its applications. Volume 11,
New York Wiley.

Fisher, R.A. (1935). The design of experiments. Oliver and Boyd, Edinburgh.

Fix, E. and Hodges, J.L. (1951). Discriminatory analysis, non-parametric estimation:

consistency properties. Report No. 4, Project no, 21-49-004, USAF School of
Aviation Medicine, Randolph Field, Texas.

Folland, C.K., Karl, T.R. and Vinnikov K.Y. (1990). Observed climatic variations and
change. In Climate change: the IPCC Scientific Assessment, Houghton, J.T.,
Jenkins, GJ. and Ephraums, JJ. (eds), Cambridge University Press.
 245

Fiyer, MJ. (1977). A review of some non-parametric methods of density estimation.

Journal Inst Maths Applications, 201, 335-354.

Fukunaga, K. (1972). Introduction to statistical pattern recognition. New York,

Academic Press.

Gail, M. (1978). The analysis of heterogeneity for indirect standardised mortality ratios.
J.R.Statistical Society A, 141, 224-234,

Geary R.C. (1930). The frequency distribution of the quotient of two normal variances,
J.R. Statistical Society, 93,442.

Gesler, W.M., Todd, C„ Evans, C„ Casella, G. Pittam, J. and Andrews I.I. (1980).
Spatial variations in morbidity and their relationship with community
characteristics in Central Harlem District. Society Science Medicine, 14D, 387-
396.

Gibbs, J.P. (1961). A method for comparing the spatial shapes of urban units. In Gibbs
J.P., editor, Urban Research Methods. Princeton, New Jersey. Van Nostrand Co.,
Inc, 99-106.

Gini, G. (1913-1914). Sulla misura dell concentraziono e della variabilitia dei caraterri.
Atti del Reale Institute Veneto di Scienze, Lettre ed Arti, 53, 2.

Gold, C.M., Charters, T.D. and Ramsden, J. (1977). Automated contour mapping using
triangular element data structures and an interpolant over each irregular triangular
domain. Computer Graphics, 11(2), 170-175.

Goodman, L. A. and Kruskal, W.H, (1954) Measures of association for cross­

classification Part 1. Journal Atnerican Statistical Association, 49,732-764.

Gotshtasby, A.', Stockman, G.C. and Page, C.V. (1986). A region based approach to
digital image registration with sub-pixel accuracy. IEEE Transactions on
Geoscience and Remote Sensing. Vol GE 24(3) 390-399.
 246

Grant, F.A. (1957). A problem in the analysis of geophysical data. Geophysics, 22,
309-344

Gray, J.M., (1972). Trend surface analysis: trends through clusters. Area, 4,102-103.
Haining, R. (1987). Trend surface models with regional and local scales of variation with
an application to aerial survey data. Technometrics, 29(4), 461-469.

Hall, P., Diciccio, T.J. and Romano, J.P. (1989). On smoothing and the bootstrap. The
Annals of Statistics, 17(2), 692-704.

Hailey, E. (1686). A historical account of the trade winds and monsoons, observable in
the seas between and near the tropics: with an attempt to assign the physical cause
of said winds. Philosophical Transactions, 153-158.

Hamming, R.W. (1983). Digital Filters, Englewood Cliffs NJ. Prentice Hall (Signal
processing Series) 2nc*edition,

Handcock, M.S. and Wallis, J.R. (1990). An approach to statistical spatial-temporal

modeling of meteorological fields. Copenhagen General Assembly of the
European Geophysical Society.

Hansen, J., Lacis, A., Rind, D., Russell, G„ Stone, P., Fung, I., Ruedy, R. and Lemer, J.
(1984). Climate sensitivity analysis of feedback mechanisms in climate processes
and climate sensitivity. Geophysics, Monogr., Ser., Vol29 ed. Hansen J. E. and
Takahashi T. 130-163, AGU, Washington D.C.

Heap, B.R. and Pink, M.G. (1969). Three contouring algorithms. DNAM Report 81,
National Physical Laboratory, Teddington.

Henshaw, D.L., Eatough, J.P. and Richardson, R.B. (1990). Radon as a causative factor
in induction of myeloid leukaemia and other cancers. The Lancet. 335,1008-1012.

Hickey, R.J., Bowers, E.J., Spence, D.E., Zemel, B.S., Clelland, A.B. and Clelland, A.B.
(1983). Radiation hormesis, public health and public policy: a commentary.
Health Physics. 44, 207-219.
 247

Hills, M. and Alexander, F, (1989), Statistical methods used in assessing the risk of
disease near a source of possible environmental pollution: a review. J.R. Statistical
Society A, 152, 353-363.

Horton, R.E, (1932). Drainage basin characteristics. Transactions of the American

Geophysical Union, 13, 350-361.

Hsu, M.L. and Robinson, A.H. (1970). The fidelity of isopleth maps, an experimental
study. University of Minnesota Press, Minneapolis.

Hugg, L. (1979). A map comparison of work disability and poverty status in the United
States. Soc. Set Med 13D, 237-240.

Jenks, G.F. and Caspall, F.C. (1971), Error on choroplethic maps: definition,
measurement, reduction. Annals Association American Geographer, 61, 217-244.

Jenks, G.F, and Coulson, M.R.C. (1963), Class intervals for statistical maps.
International Yearbook of Cartography, 3,119-134.

Jones, P.D., Raper, S. S. B., Bradley, R.S., Diaz, H. F., Kelly, P.M. and Wigley, T.M.
(1986), Northern hemisphere surface air temperature variations 1851-1984.
Journal of Climate and Applied Meteorology 25,161-179.

Journal, A.G. (1969). Rapport d'etude sur l'estimation d'une variable regionalisee.
Internal Report No. N-156 CGMM.

Journal, A.G. and Huijbregts, G.J. (1978). Mining Geostatistics. Academic Press,
London.

Karl, T.R., Diaz, H.F. and Kuklan, G. (1988). Urbanisation: Its detection and effect in
the United States climate record. Journal of Climate, 1,1099-1123.

Karl, T.R. Heim, R.R. and Quayle, R.G. (1991). The greenhouse effect in Central North
America: If not now, when? Science, 251,1058-1061.
 248

Karl, T.R. Livezey, R.E. and Epstein, E.S. (1984). Recent unusual mean winter
temperatures across the contiguous United States. Bull. Amer. Meteor. Soc., 65,
1302-1309.

Karl,T.R. and Williams (Jr.) C.N. (1987). An approach to adjusting climatological time
series for discontinuous homogeneities. Journal of Climate and Applied
Meteorology,26,1744-1763.

Karl, T.R,, Williams (Jr.), C.N., Young PJ. and Wendland W.M. (1986). A model to
estimate the time of observation bias associated with monthly mean maximum,
minimum and mean temperature for the United States. Journal of Climate and
Applied Meteorology,25,145-160.

Kendall, M.G. and Stuart, A. (1977). The Advanced Theory of Statistics, Vol 1,
Distribution Theory, (4th edition), C. Griffin and Co., Ltd., London.

Kneale, G.W. and Stewart, A.M, (1987). Childhood cancers in the United Kingdom and
their relation to background radiation. In Jones R.R., Southwood R. (eds.). Radiation
and Health: The Biological Effect of Low Level Exposure to Ionising Radiation,
Chichester: John Wiley, 203-220.

Krige, D.G. (1978). Lognormal de Wijsian Geostatistics For Ore Evaluation. South
African Inst Min. Metall. Monograph Series: Geostatistics Vol. 1.

Krumbein, W.C. (1956). Regional and local components in facies maps. Bulletin
American Association Petroleum Geologists, 40, 2162-2194. ,

Krumbein, W.C. (1959). Trend surface analysis of contour-type maps with irregular
control point spacing. Journal of Geophysical Research, 64, 823-834.

Leukaemia and Lymphoma. An area atlas of distribution within areas of England and
Wales 1984-1988. (1990), Compiled by the Leukaemia Research Fund Centre for
Clinical Epidemiology at the University of Leeds.

Link, R.F. and Koch, (Jr.)G.S. (1975). Some consequences of applying lognormal
theory to pseudo-lognormal distributions. Mathematical Geology, 17(2), 117-128.
 249

Loftsgaarden and Quesenbury, C.P. (1965). A non-parametric estimate of a multivariate

probability density function. Annals Math. Statist., 1049-1051.

Lorenz, M.C. (1905). Methods of measuring the concentration of wealth. Publications of

the American Statistical Association, 9, new series, 209-219.

McCullagh, M.J. (1983). Transformation of contour strings to a regular grid based

digital elevation model. Euro-Carto, 18pp.

McCullagh, MJ. and Ross, C.G. (1980). Delaunay triangulation of a random data set for
isarithmic mapping. Cartographic Journal, 17(2), 93-99.

McGlashan, N.D. (1972). Geographical evidence on medical hypothesis. Medical

Geography: Techniques and Field Studies. Methuen, London.

McLain, D.H, (1974). Drawing contours from arbitrary data points. Computer Journal,
17, 318-324.

Mackay, J.R. (1951). Some problems and techniques in isopieth mappping. Economic
Geography, 21,1-9.

Mackay, J.R, (1953). The alternative choice in isopieth interpolation. Professional

Geographer, 5, 2-4.

Maniya, G.M. (1961). Soobshch. Ak&d. Nauk Gruzin. SSR 27, 385-390.

Matem, B. (1960). Spatial Variation Medd. Statens. Skogsforskningsinstitut.

Merriam, D.F. and Sneath, P.H.A. (1966). Quantitative Comparison of Contour Maps.
Journal of Geophysical Research, 7(4), 1105-1115.

Miesch, A.T. and Conner, JJ. (1967). Stepwise regression and non-polynomial model in
trend analysis. University of Kansas, State Geological Survey, Computer
Contribution, No.27.
 250

Miller, R.L. (1956). Trend surfaces : Their application to analysis and description of
environment of sedimentation. Journal of Geology, 64,425-446.

Miller, V.C. (1953). A quantitative geomorphic study of drainage basin characteristics in

the Clinch Mountain Area, Virginia and Tennessee. Technical Report No. £.,
Department of Geology, Columbia University, New York.

Mills, F.C. (1955). Statistical Methods (3r^ edition). Holt Rinehart and Winston, New
York.

Minnick, R.F. (1964). A method for the measurement of areal correspondance. Papers of
The Michigan Academy of Science, Arts and Letters, Vol XL1X, 333-342.

Mirchink, M. F. and Bukhartsev, V. P. (1959). The possibility of statistical studies of

structural relations. Doklady of the Academy of Sciences ofU.S.S.R., 126(5), 495-
497.

Mitchell, J.F.B. (1989). The ’greenhouse' effect and climate changes. Reviews of
Geophysics, 27,115-139.

Monmonier, M.S. (1973). Eigenvalues and principal components. A method for

detecting natural breaks for choroplethic maps. American Congress Survs, Mapp.
Proc. Fall Convention, 252-264.

Morgenstem, H. (1982). Uses of ecologic analysis in epidemiologic research. American

Journal Public Health, 72,1336-1344.

Morrison, J.L. (1971). Method-produced error in isarithmic mapping. American

Congress on Surveying and Mapping, Technical Monograph No. CA-5.

Mosteller, F. and Wallace, D.L. (1963). Inference in an authorship problem. Journal

American Statistical Association, 58,275-309.
Nambi, K.S. V. and Soman, S.D. (1990). Further observations on environmental radiation
and cancer in India. Health Physics, 59(3), 339-344.
 252

Richardson, L.F. (1961). The problem of contiguity. An Appendix t o 'Statistics of

Deadly Quarrels', Wright, Q. and Lenau, C.C. (editors).

Ripley, B.D. (1981). Spatial Statistics. J. Wiley and Sons, U.S.A.

Robinson, A.H. (1961). The cartographic representation of the statistical surface.

International Yearbook of Cartography, 1, 53-63,

Robinson, G. (1972). Trials on trends through clusters of cirques. Area, 4,102-103.

Robinson, A.H. and Bryson, R. A. (1959). A method for describing quantitatively the
correspondance of geographical distributions. Annals Association American
Geographer, 47, 379-391,

Robinson, A.H. and Sale, R.D. (1969). Elements of cartography (3rd edition). John
Wiley and Sons, New York.

Rosenblatt, M. (1956). Remarks on some non-parametric estimates of a density function.

Annals Mathematics Statistics, 27, 832-837.

Rothwell, M.A. (1971). A computer program for the construction of pole figures.
Journal Applied Cryst., 4,494.

Rudemo, M, (1982). Empirical Choice of Histograms and kemal density estimators.

Scand. Journal Statist, 9, 65-78.

Sanderson, D.C.W., Allyson, J.D., Martin, E., Tyler, A.N., and Scott, E.M. (1990). An
airborne gamma ray survey of three Ayrshire Districts. Scottish Universities
Research and Reactor Centre, East Kilbride.

Sanderson, D.C.W., Martin, E., Scott, E.M., Baxter, M.S. and NiRiain, C. (1992). The
use of radiometrics for epidemiological studies of leukaemia. A preliminary
investigation in S.W. England. Scottish Universities Research and Reactor Centre,
East Kilbride.
 253

Schlesinger, M.E. and Zhao, Z. (1987). Seasonal climate changes induced by doubled
carbon dioxide as simulated by the OSU atmospheric OCMI mixed layer model.
Report 70\ 73pp. Oregon State University, Climateic Institute, Corvallis,

Schmid, C.F. and MacCannell, E.H. (1955). Basic problems, techniques and theory of
isopieth mapping. Journal American Statistical Association, 50, 220-239.

Schultz, G.M. (1961). An experiment in selecting value scales for statistical distribution
maps. Surveying and Mapping, 21,224-230.

Schuster, E.F. and Gregory, C.G. (1981). On the nonconsistency of maximum likelihood
non-parametric estimators. Computer Science and Statistics: Proceedings of The
13th Symposium On the Interface. New York: Springer-Verlag, 295-298,

Scott, D.W, (1985a). Frequency Polygons: Theory and Application. Journal American
Statistical Association, 80(390), 348-354.

Scott, D.W, (1985b). Average shifted histograms: Effective non-parametric density

estimators in several dimensions. Annals of Statistics, 13(3), 1024-1040.

Scott, D.W. and Factor, L.E. (1981). Monte Carlo study of three data-based non-
parametric density estimators. Journal American Statistical Association, 76, 9-15.

Scott, D.W. and Thompson, J.R. (1983). Probability density estimation in higher
dimensions. In Gentle J.E. (ed.), Computer Science and Statistics: Proceedings of
the Fifteenth Symposium and The Interface, Amsterdam: North Holland, 173-179.

Scriptor, M.W. (1970). Nested-means map classes for statistical maps. Annals
Association American Geographer, 60, 385-393.

Seber, G.A.F. (1977). Linear Regression Analysis. John Wiley and Sons.

Sellers, W.D. (1968). Climatology of monthly precipitation patterns in the Western

United States, 1931-1966. Monthly Weather Review?, 96, 585-595.
 254

Shephard, D.S. (1968a). A two-dimensional interpolation function for computer mapping

of irregularily spaced data. Harvard Theoretical Geography, Paper No. 15,
Laboratory for Computer Graphics, Harvard University Graduate School of
Design, Cambridge, Mass.

Shephard, D.S. (1968b.). A two-dimensional interpolation function for irregularily

spaced data (revised). Proceedings of the 23rd National Conference of the
Association for Computing Machinery, Brandon/Systems Press, Inc., Princeton,
N.J.

Siegel, S. (1956). Non -Parametric Statistics For The Behavioral Sciences. McGraw -
Hill, New York.

Siegel, A.F. and Benson, R.H. (1982), A robust comparison of biological shapes.
Biometrics, 38,341-350.

Silverman, B.W. (1978). Choosing the widow width when estimating a density.
Biometrika, 65,1-11.

Silverman, B.W. (1986). Density estimationfor statistics and data analysis. Chapman
and Hall.

Skaggs, R.H. (1975). Drought in the United States 1931-1940. Annals Association
American Geographer, 65, 391-402.

Sneath, P.H.A. (1967). Trend surface analysis of transformation grids. Journal Zoology.,
London 151, 65-122.

Steffensen, J.F. (1927). Interpolation. Williams and Wilkins,

Stevens, S.S. (1946). On the theory of scales of measurement. Science, 103, 677-680.

Stockman, G„ Kapstein, S. and Bennett, S. (1982). Matching images to models for

registration and object detection via clustering, IEEE Trans Pattern Anal. Machine
Intelligence. Voi PAMI-4 no.3, 229-241.
 255

Stoddart, D.R. (1965). The shape of atolls. Marine Geology, 3, 369-383.

Stone, CJ. (1984). An asymptotically optimal window selection rule for kernel density
estimates. Ann. Statist, 12,1285-1297.

Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions (with

discussion).//?. Statistical Society B, 36, 111 -147.

Student, I. (1907). On the error of counting with a haemacytometer. Biometrika, 5, 351-

360.

Switzer, P. (1973). Applications of random process models to the description of spatial

distributions of qualitative geologic variables. 24ft1International Geological
Congress, Montreal Irving, E. (edj. 1-11.

Switzer, P. (1975). Estimation of the accuracy of qualitative maps. Display and Analysis
of Spatial Data, J.C. Davies and MJ. McCullagh (eds.), Wiley, New York, 1-13.

Switzer, P., Mohn, C.M., and Heitman, R.E. (1964). Statistical analysis of ocean terrain
and contour plotting. Procedures, Project Trident Report No. 1440464, Arthur D.
Little, Inc., Cambridge, Mass.

Thompson W. D'Arcy (1917,1961). On growth andform. Bonner J.T. (ed) Cambridge

University Press, Cambridge.

Tipper, J.C. (1979). Surface Modelling Techniques. Kansas Geological Survey Series on
Spatial Analysis, No. 4.

Tobler W.R. and Lau, J. (1970). Isopieth mapping using histoplanes. Geographical
Analysis, 10, 273-271.

Tsonis, A. A. and Eisner, J.B. (1989). Testing the global warming hypothesis.
Geophysical Research Letters. 16, 795 - 797.
 256

Tukey, P. A. and Tukey, J.W. (1981). Geographical display of data sets in three or more
dimensions. In Barnett, V. (ed.) InterpretingMultivariate Data. Chichester: Wiley,
189-275.

Ujeno, Y. (1983). Relation between cancer incidence or mortality and external natural
background radiation in Japan. In Biological Effects of Low Level
Radiation, Vienna, I.A.E.A. 253-262.

Unwin, D.J. (1981). Introductory Spatial Statistics, Methuen and Co., Ltd. London.

Walter, S.D. and Bimie, S.E. (1991). Mapping mortality and morbidity patterns: an
international comparison. Int. J. Epidemiology.

Wamtz, W. (1959). Toward a geography of price. Philadelphia.

Washington, W. M. and Meehl, G.A. (1984). A seasonal cycle experiment on the climate
sensitivity due to a doubling of carbon dioxide with an atmospheric general
circulation model coupled to a simple mixed layer ocean model. J. Geophys. Res.,
89, 9475-9503,

Watson, G. (1971). Trend Surface Analysis. Journal Mathematical Geology. 3(3), 215-
226.

Weatherald, R.T. and Manabe, S. (1986). An investigation of cloud cover change in

response to thermal forcing. Climatic Change, 8, 5-24.

Webb 111, T., Bartlein, PJ. and Kutzbach, J.E, (1987). Climatic change in eastern North
America during the past 18000 years: Comparisons of pollen data with model
results. In Ruddiman W.F. and Wright H.E. Jr. (Eds.) North America and adjacent
oceans during the last deglaciation: Boulder, Colorado, Geological Society of
America, The Geology of North America, V K-3.

Werrity, A. (1969). On the form of drainage basins. Papers in Geography, The

Pennsylvania State University, Pennsylvania, Papers in Geography.
 257

White, R.R. (1972). Probability maps of leukaemia mortalities in England and Wales. In
Medical Geography: Techniques and Field Studies. McGlashen, N.D. (ed.),
Methuen, London, 173-185.

Whitten, E.H.T. (1957). Composite trends in granite: Modal variation and ghost
stratigraphy in part of the Donegal Granite, Eire. Journal of Geophysical
Research, 64, 835-848.

Whitten, E.H.T, (1970). Orthogonal polynomial trends for irregularily spaced data.
Journal of Mathematical Geology,2 ,141-152.

Wilson, C.A. and Mitchell, J.F.B. (1987). A doubled carbon dioxide climate sensitivity
experiment with a GCM including a simple ocean. J. Geophy. Res, 92,13315-
13343.

Woodward, W.A. and Gray, H.L. (1992). 'Global Warming' and the problem of testing
for trend in time series data. Department of Statistical Science, Southern Methodist
University, 19 pages.

Young, I.T., Walker, J.E. and Bowie, J.E. (1974). An analysing technique for biological
shape, Medinfo. 74. M.T.T 843.
 258

APPENDIX

Year Season Temp. Area Perime Orient­ Centroid displacement

level -ter ation
X y
1930 Spring 40°C 728.5 127.2 1,827 97.95 39.07
(a) (232.0) (51.39)
(b ) 19.23 19.55 0.525 115.9 115.9
(2.607) (4.753)

45°C 627.1 142.9 1.828 97.40 38.73

(a) (184.4) (37.89)
(b) 38.83 24.50 0.6215 116.1 39.33
(5.260) (7.663)
(c) 8.604 10.77 6.229 107.9 35.59
(1.787) (1.196)

50<>C 337.7 89.08 1.669 88.24 37.83

(a) (95.42) (29.16)
(b ) 28.59 33.83 2.867 118.5 44.37
(3.894) (13.95)

55°C 231.0 88.63 1.757 86.68 37.56

(a) (64.71) (16.38)
(b) 8,062 11.84 1.040 89.38 49.36
(1.384) (0.068)
(c) 4.915 11.53 1.009 118.8 46.33
(2.045) (0.432)

60°C 30.87 24.16 3,065 96.22 38.11

(a) (5.223) (6.820)
(b ) 135.5 72.73 2.093 82.79 37.74
(13.91) (30.98)
 259

1980 Spring 40°C 728.5 127.2 1.826 97.94 39.06

(a) (229.1) (51.39)
(b) 17.91 19.02 0.561 115.9 39.07
(2.527) (4.666)

45°C 613.3 159,1 1.829 97.13 38.80

(a) (171.2) (34.57)
(b) 39.93 25.05 0,559 116,1 39.34
(5.203) (8.487)

50°C 318.4 86.00 1.998 87.87 37.75

(a) (89.40) (28.50)
(b) 23.27 32.74 2.875 118.6 44.63
(3.437) (13.26)
(c) 24.29 12.985 2.985 110.8 40.72
(2.302) (7.074)

55°C 248.5 80.80 1.716 86.98 37.74

(a) (73.51) (19,08)
(b) 6.656 6.979 1.014 89.35 49.36
(1.280) (0.068)
(c) 9.507 13.87 2.314 118.6 46.23
(2.851) (1.172)
(d) 4.847 8.224 2.792 110.9 40.00
(0.810) (0.940)

60<>C 29.41 23.07 3.206 96.33 37.98

(3.779) (6.584)
(b) 92.50 59.20 1.812 81.18 38.37
(23.11) (23.11)
(c) 14.48 14.26 2.796 89.60 32.56
(2.410) (2.530)
 260

1930 Summer 65°C 805.0 128.4 1.8388 98.38 39.40

(a) (280.1) (55.40)
(b) 9.583 11.84 0.069 116.1 38.10
(1.489) (1.730)

70°C 576.5 189.4 1.483 87.03 38.61

(a) (216.8) (15.05)

75°C 339.5 86.40 1.654 88.39 38.41

(a) (94.01) (29.60)
(b) 17.28 30.53 0.2782 119.4 45.59
(4.040) (12.87)
(c) 6.156 16.25 0.144 111.5 40.32
(2.276) (2.276)

80°C 96.00 62.56 1.793 86.52 37.26

(a) (63.98) (7.490)
(b) 13.43 23.58 2.685 96.64 38.96
(2.811) (5.841)
(c) 1.531 3.368 0.001 90.26 31.94
(0.314) (0.298)
(d) 0.2812 4.163 2.356 82.63 37.75
(0.013) (0.023)

85°C 6.844 9.748 2.469 75.62 40.89

(a) (1.959) (1.029)
(b) 1.727 4.866 1.213 97.41 41.00
(0.435) (0.192)

1980 Summer 65°C 782.0 128.3 1.833 98.33 39.38

(a) (254.1) (54.98)
(b) 9.583 11.84 0.069 116.1 38.10
(1.489) (1.730)
 261

70°C 522.0 146.2 1.895 93.21 38.65

(163.2) (30.71)
(b) 28.37 41.28 0.042 105.3 38.97
(0.613) (0.523)
(c) 0.646 0.537 0.075 106.4 38.23
(0.078) (0.078)

75°C 339.5 86.40 1.654 88.39 38.41

(a) (94.01) (29.60)
(b) 19.81 31.65 0.284 119.4 45.05
(3.910) (12.07)
(c) 6.156 16.25 0.1444 111.5 40.32
(2.276) (3.861)
(d) 2.969 5.240 0.001 97.91 30.64
(0.590) (0.511)

80°C 216.7 96.63 1.725 87.05 38.26

(a) (60.31) (15.80)
(b) 1.793 5.079 6.242 111.6 40.14
(0,356) (0.371)
(c) 0.2320 0.273 0,012 93.16 40.10
(0.238) (0.264)
(d) 0.2812 4.164 2.356 82.63 37.75
(0.012) (0.123)

85°C 15.09 14.60 2.404 75.67 40.90

(a) (2.918) (2.280)
(b) 3.875 11.78 0.001 86.76 39.55
(2.731) (1.064)
(c) 2.125 9.064 0.003 97.21 36.15
(0.049) (0.536)
(d) 0.2186 4.853 1.198 97.35 40.84
(0.312) (0.296)
(e) 1.250 3.014 0.0012 81.47 34.52
(2.731) (1.075)
 262

1930 Autumn 45°C 678.1 141.7 4.975 97.58 38.86

(a) (192.3) (41.19)
(b) 30.21 21.90 0,5008 116.0 39.50
(4.134) (6.215)
(c) 3.133 6.522 0.9559 107.9 35.39
(0.7244) (0.4139)

50°C 387.0 100.2 1.633 89.10 38.03

(a) (112.9) (31.23)
(b) 37,87 39.59 3.277 118.5 43.21
(3.091) (21.93)
(c) 46.36 34.52 3,070 110.9 39.49
(3.142) (17.54)

55°C 281.7 79.32 5.140 87.27 37.91

(a) (79.77) (23.22)
(b) 29.03 12.99 2.914 118.9 44.82
(2.655) (13.37)
(c) 5.654 8.921 2.873 110.9 39.96
(0.9479) (1.095)

60°C 195.5 93.95 1.732 86.10 37,97

(a) (55.25) (14.25)
(b) 45.32 24.98 3.105 95.99 37.91
(5.989) (6.831)

65<>C 96.82 9.718 2.513 89.71 31.14

(a) (0,580) (0.536)
(b) 14.39 13.83 2,174 34.60 34.60
(1.830) (1.830)
(c) 7.249 10.93 2.310 95.99 37.27
(1.976) (1.063)
(d) 3.312 6.075 0,813 97.23 40.57
(0.6398) (0.328)
 263

(e) 23.76 18.97 1.778 86.37 39.46

(6.723) (2.280)
(f) 28.38 20.71 5.443 76.02 40.44
(4.985) (4.111)

1980 Autumn 45°C 658.9 142.1 4.969 97.57 38.82

(a) (190.0) (41.09)
(b) 33.48 22.64 0.4665 116.1 39.42
(4.605) (6.844)
(c) 6.386 9.308 6.273 108.1 35.28
(1.246) (1.015)

50°C 394.7 100.7 5.114 89.32 38.01

(a) (115.4) (31.96)
(b) 41.70 41.88 3.347 118.4 42.91
(3.241) (23.36)
(c) 56.28 37.13 6.232 111.0 39.54
(4.196) (19.04)

55°C 272.6 85.08 4.848 87.26 37.79

(a) (74.57) (20.30)
(b) 8.659 10.91 2.898 110.8 39.94
(1.399) (1.542)
(c) 27.91 11.62 2.905 118.9 45.14
(2.956) (11,29)

60°C 139.5 74.70 5.262 82.90 37.54

(a) (30.95) (14.32)
(b) 46.26 25.53 3.051 96.47 38.30
(6.639) (7.975)

65°C 6,634 9.555 2.492 89.82 32.16

(a) (1.222) (1.073)
(b) 15.55 14.42 2.203 81.23 34.62
(3.371) (2.006)
 264

(c) 8.956 11.46 2.236 96.58 36.31

(2.111) (1.771)
(d) 6.548 9.306 0.9040 96.72 40.39(0.
(1.379) (0.760)
(e) 17.77 15.68 1.846 86,02 39.24
(4.566) (1.857)
(f) 22.43 18.58 5.440 75.88 40.41
(4.174) (3.317)

1930 Winter 40°C 658.4 134.5 4.958 98.13 38.25

(a) (213.8) (37.96)
(b) 37.16 27.15 5.968 115.5 40.45
(4.866) (8.513)
(c) 5.616 8.710 0.2913 107.7 35.73
(1.153) (0.870)

45°C 379.3 96.55 1.666 89.19 37.61

(a) (118.0) (28.56)
(b) 86.02 69.51 3.690 115.5 39,33
(13.70) (21.51)
(c) 5.216 8.764 0.7221 110.4 45.38
(1.215) (0.8052)

50°C 283.7 84.33 4.883 87.52 37.29

(a) (87.12) (23.37)
(b) 20.43 32.62 3.156 119.1 45.57
(2.292) (15.18)
(c) 0.1778 1.819 6.107 111.5 34.07
(0.022) (0.912)

55°C 160.0 57.80 5.225 83.48 36.88

(a) (41.96) (17.84)
(b) 3.156 8.067 3.478 120.0 46.27
(0.3220) (1.040)
 265

(c) 7.222 2.259 3.335 120.2 38.63

(0.1364) (1.035)
(d) 50.02 26.58 6.236 96.24 37.93
(7.335) (7.335)

60°C 54.20 35.24 5.383 79.01 37.13

(a) (13.16) (11.30)
(b) 44.91 35.70 2.626 87.63 36.70
(6.893) (12.66)
(c) 27.96 23.17 3.286 96.14 37.60
(3.642) (5.783)
(d) 3.156 3.038 2.975 120.5 38.70
(0.004) (0.2783)
(e) 0.4062 2.584 3.359 120.5 46.38
(0.0096) (0,1920)

1980 40°C 643.1 134.5 4.958 97.79 38.54

(a) (199.8) (38.02)
(b) 24.65 20.34 0.5066 115.9 39.55
(3.383) (5.273)
(c) 0.1904 1.700 0.070 108.0 35.68
(0.036) (0.041)

45°C 475.9 182.5 1.643 95.51 37.74

(a) (184.4) (25.01)

50°C 265.9 84.7 4.932 87.68 36.83

(a) (87.49) (20.80)
(b) 24.72 33.55 6.268 118.9 43.82
(2.755) (15.74)
(c) 10.26 12,18 2.939 110.9 39.54
(1.410) (2.235)

55°C 147.6 56.58 5.213 83.53 36.45

(a) (42.22) (16.59)
 266

(b) 9.623 14.97 2.526 118.9 46.15

(2,611) (1.569)
(c) 8.141 2.919 3.393 120.1 38.64
(0.1518) (1.879)
(d) 50.16 26.03 3.214 96.61 37.48
(8.097) (7,802)

60°C 18,61 15.81 1.728 86.68 38.95

(a) (4.465) (2.095)
(b) 49.40 33.60 2.930 79.29 36.68
(12.55) (10.64)
(c) 32.19 22.84 3.215 96.58 37.18
(4.573) (5.916)
(d) 5.218 4.593 3.242 120.4 38.69
(0.031) (0.529)
(e) 4.469 6.824 2.832 120.1 46.19
(0.1937) (0.8290)
(f) 20.38 16.79 5.988 89.52 32.64
(3.464) (3.317)

Gl a s g o w '" !
uNn'rp^nY s
i