Spatial data in R

Handling spatial data in R: GRASS interface

Handling spatial data in R: methods
Worked examples in spatial statistics

Applied Spatial Data Analysis with R

Roger Bivand1

12 June 2014

1
Department of Economics, Norwegian School of Economics, Helleveien 30,
N-5045 Bergen, Norway; Roger.Bivand@nhh.no
Roger Bivand Applied Spatial Data Analysis with R
 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Overview
1 Spatial data in R
Study area
Representing spatial data
Coordinate reference systems
Raster data
Exporting data
2 Handling spatial data in R: GRASS interface
3 Handling spatial data in R: methods
Topology predicates and operations
Overlay
4 Worked examples in spatial statistics
Disease mapping
Spatial autocorrelation
Regression
Bayesian spatial regression with INLA
Roger Bivand Applied Spatial Data Analysis with R
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

Many objects of interest in data analysis can be expressed as

lists of numbers
R sees the world this way too, and almost everything is
expressed as vectors or lists of one kind or another
R at its simplest behaves like an overgrown calculator, so that:
> 2 + 2
[1] 4
> 2 * pi * 10
[1] 62.83185

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

Examples are made up of the R command line prompt >

followed by the instruction in italics, and then the output
The output of a vector printed to the screen is preceeded by
the number in sequence of the vector value in square brackets;
here the vector only has a length of one
R lets us assign values to symbolic variables, so we can also
do:
> a <- 2
> a
[1] 2
> a + a
[1] 4

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

Assignment does not print any output (usually), but we can

see what is inside a symbolic variable (object) by entering its
name
R is built up using functions, some of which we can use on our
symbolic variable (or object) a
str() shows the structure of a in summary, it is numeric —
here num, and has the value 2; length() tells us how long it is
> str(a)
num 2
> length(a)
[1] 1

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

class() tells us what class it belongs to

The values returned by functions can be assigned to new
symbolic variables, here ca, which is a character vector
> ca <- class(a)
> ca
[1] "numeric"
> str(ca)
chr "numeric"

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

R can express arithmetic and other operations directly, or through

functions, which most often are vectorised. Let’s generate some
data:
> a <- c(1, 1, 1, 1, 1, 2, 2, 2, 2, 2)
> b <- c(1, 2, 1, 2, 1, 2, 1, 2, 1, 2)
> z <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

where c() is a function used to join objects of the same kind

together into a vector.

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

Doing the addition element for element is tiresome, so the sum()

function helps — there are functions for most things in R
> a[1] + a[2] + a[3] + a[4] + a[5] + a[6] + a[7] + a[8] + a[9] +
+ a[10]
[1] 15
> sum(a)
[1] 15
> sum(a^2)
[1] 25
> sum(a * b)
[1] 23

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher
When we talk of arguments to a function, we mean the
(possibly named) arguments passed to the function for
evaluation
We can access the arguments using the function help pages,
or briefly using args():
> args(sum)
function (..., na.rm = FALSE)
NULL

Now we will look at why data-analytical software should be

used to analyse data
Problems with the data are very common, respondents who
do not answer a particular question, equipment that fails
occasionally, rain smudging field data collection sheets, etc.,
and the software needs to be able to handle them
Roger Bivand Applied Spatial Data Analysis with R
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

R refresher

As we see, sum() has an argument called na.rm, which is set

to FALSE by default
This means that NA — not available — values will not be
removed by default, and the summation will fail, returning an
NA, and alerting the analyst to problems with the data
> a[3] <- NA
> a
[1] 1 1 NA 1 1 2 2 2 2 2
> sum(a)
[1] NA
> sum(a, na.rm = TRUE)
[1] 14

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Olinda Hansen’s Disease data set

The data are discussed in Lara, T. et al. (2001) Leprosy

surveillance in Olinda, Brazil, using spatial analysis techniques.
Cadernos de Saúde Pública, 17(5): 1153–1162.
They use 243 1990 census districts, and include the
1991-1996 counts of new cases (CASES), 1993 year-end
population figures (POP) and a deprivation index (DEPRIV)
The data have been made available in cooperation with
Marilia Sa Carvalho, and their use here is partly based on
teaching notes that we wrote together some years ago
The data now take the form of a shapefile, but were initially
assembled from a Mapinfo file of census district borders and
text files of variable values

Roger Bivand Applied Spatial Data Analysis with R

Olinda Hansen’s Disease data set
So that we’ve got something to go on, let’s read in the data set, using an
interface function in the rgdal package (packages are kept in libraries and
extend R’s basic functions and classes):
> library(rgdal)

The function as we will see later takes at least two arguments, a data
source name, here the current directory, and a layer name, here the name
of the shapefile without extension. Once we have checked the names in
the imported object, we can use the spplot method in sp to make a
map:
> olinda <- readOGR(".", "setor1")

OGR data source with driver: ESRI Shapefile

Source: ".", layer: "setor1"
with 243 features and 10 fields
Feature type: wkbPolygon with 2 dimensions

> names(olinda)

[1] "AREA" "PERIMETER" "SETOR_" "SETOR_ID" "VAR5" "DENS_DEMO"

[7] "SET" "CASES" "POP" "DEPRIV"

> spplot(olinda, "DEPRIV", col.regions = grey.colors(20, 0.9, 0.3))

Map of deprivation variable

0.8

0.6

0.4

0.2

0.0
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Data frames

Very often data are stored in a data frame with rows

representing observations, here 1990 census districts in Olinda
In data frames, the columns can be of different classes — in a
matrix, all the columns have to have the same class
The columns are accessed using the $ operator; the columns
are list elements of the data frame
We can also use square brackets to access data frame values,
like look-up in a data base

Roger Bivand Applied Spatial Data Analysis with R

Olinda data

> str(as(olinda, "data.frame"))

'data.frame': 243 obs. of 10 variables:
$ AREA : num 79139 151153 265037 137696 121873 ...
$ PERIMETER: num 1270 2189 2818 2098 3353 ...
$ SETOR_ : num 2 3 4 5 6 7 8 9 10 11 ...
$ SETOR_ID : num 1 2 3 4 5 6 9 10 11 12 ...
$ VAR5 : int 242 224 223 229 228 222 218 225 227 221 ...
$ DENS_DEMO: num 4380 10563 6642 13174 13812 ...
$ SET : num 242 224 223 229 228 222 218 225 227 221 ...
$ CASES : num 1 6 5 1 1 4 6 2 1 6 ...
$ POP : num 337 1550 1711 1767 1638 ...
$ DEPRIV : num 0.412 0.168 0.192 0.472 0.302 0.097 0.141 0.199 0.228 0.223 .
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Object framework

To begin with, all contributed packages for handling spatial

data in R had different representations of the data.
This made it difficult to exchange data both within R between
packages, and between R and external file formats and
applications.
The result has been an attempt to develop shared classes to
represent spatial data in R, allowing some shared methods and
many-to-one, one-to-many conversions.
We chose to use new-style classes to represent spatial data,
and are confident that this choice was justified.

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial objects

The foundation object is the Spatial class, with just two slots
(new-style class objects have pre-defined components called
slots)
The first is a bounding box, and is mostly used for setting up
plots
The second is a CRS class object defining the coordinate
reference system, and may be set to CRS(as.character(NA)),
its default value.
Operations on Spatial* objects should update or copy these
values to the new Spatial* objects being created

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Coordinate reference systems

The EPSG list among other sources is used in the workhorse

PROJ.4 library, which as implemented by Frank Warmerdam
handles transformation of spatial positions between different
CRS
This library is interfaced with R in the rgdal package, and the
CRS class is defined partly in sp, partly in rgdal
A CRS object is defined as a character NA string or a valid
PROJ.4 CRS definition
The validity of the definition can only be checked if rgdal is
loaded

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial points

The most basic spatial data object is a point, which may have
2 or 3 dimensions
A single coordinate, or a set of such coordinates, may be used
to define a SpatialPoints object; coordinates should be of
mode double and will be promoted if not
The points in a SpatialPoints object may be associated with
a row of attributes to create a SpatialPointsDataFrame object
The coordinates and attributes may, but do not have to be
keyed to each other using ID values

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial points classes and their slots

SpatialPointsDataFrame Spatial
SpatialPoints bbox
coords.nrs proj4string
data
SpatialPoints
data.frame coords
Spatial

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial*DataFrames

Spatial*DataFrames are constructed

as Janus-like objects, looking to mapping
and GIS people as “their” kind of object,
as a collection of geometric freatures,
with data associated with each feature.
But to data analysts, the object “is” a
data frame, because it behaves like one.

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial lines and polygons

A Line object is just a spaghetti collection of 2D coordinates; a
Polygon object is a Line object with equal first and last coordinates
A Lines object is a list of Line objects, such as all the contours at
a single elevation; the same relationship holds between a Polygons
object and a list of Polygon objects, such as islands belonging to
the same county
A comment is added to each Polygons object to indicate which
interior ring belongs to which exterior ring (SFS)
SpatialLines and SpatialPolygons objects are made using lists
of Lines or Polygons objects respectively
SpatialLinesDataFrame and SpatialPolygonsDataFrame
objects are defined using SpatialLines and SpatialPolygons
objects and standard data frames, and the ID fields are here
required to match the data frame row names
Roger Bivand Applied Spatial Data Analysis with R
Spatial Polygons classes and slots

SpatialLines Lines Line

lines Lines coords
Spatial ID

SpatialPolygons Polygons Polygon

polygons Polygons labpt
plotOrder plotOrder area
Spatial labpt hole
ID ringDir
area coords
Spatial comment
bbox
proj4string
Back to Olinda
So we can treat olinda as a data frame, for example adding an expected
cases variable. The object also works within model fitting functions, like
glm.
> olinda$Expected <- olinda$POP * sum(olinda$CASES, na.rm = TRUE)/sum(olinda$POP,
+ na.rm = TRUE)
> str(olinda, max.level = 2)

Formal class 'SpatialPolygonsDataFrame' [package "sp"] with 5 slots

..@ data :'data.frame': 243 obs. of 11 variables:
..@ polygons :List of 243
.. .. [list output truncated]
..@ plotOrder : int [1:243] 243 56 39 231 224 60 58 232 145 66 ...
..@ bbox : num [1:2, 1:2] 288955 9111044 298446 9120528
.. ..- attr(*, "dimnames")=List of 2
..@ proj4string:Formal class 'CRS' [package "sp"] with 1 slots

> str(as(olinda, "data.frame"))

'data.frame': 243 obs. of 11 variables:

$ AREA : num 79139 151153 265037 137696 121873 ...
$ PERIMETER: num 1270 2189 2818 2098 3353 ...
$ SETOR_ : num 2 3 4 5 6 7 8 9 10 11 ...
$ SETOR_ID : num 1 2 3 4 5 6 9 10 11 12 ...
$ VAR5 : int 242 224 223 229 228 222 218 225 227 221 ...
$ DENS_DEMO: num 4380 10563 6642 13174 13812 ...
$ SET : num 242 224 223 229 228 222 218 225 227 221 ...
$ CASES : num 1 6 5 1 1 4 6 2 1 6 ...
$ POP : num 337 1550 1711 1767 1638 ...
$ DEPRIV : num 0.412 0.168 0.192 0.472 0.302 0.097 0.141 0.199 0.228 0.223 ...
$ Expected : num 1.4 6.43 7.1 7.33 6.8 ...
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Back to Olinda

The object also works within model fitting functions, like glm; note the
number of rows.
> str(model.frame(CASES ~ DEPRIV + offset(log(POP)), olinda), give.attr = FALSE)

'data.frame': 241 obs. of 3 variables:

$ CASES : num 1 6 5 1 1 4 6 2 1 6 ...
$ DEPRIV : num 0.412 0.168 0.192 0.472 0.302 0.097 0.141 0.199 0.228 0.223 ...
$ offset(log(POP)): num 5.82 7.35 7.44 7.48 7.4 ...

Roger Bivand Applied Spatial Data Analysis with R

 The GridTopology class

The GridTopology class is the key > library(sp)

> getClass("GridTopology")
constituent of raster representations,
Class "GridTopology" [package "sp"]
giving the coordinates of the south-west
Slots:
raster cell, the two cell sizes in the
Name: cellcentre.offset
metric of the coordinates, giving the step Class: numeric

to successive centres, and the numbers Name: cellsize

Class: numeric
of cells in each dimension. We’ll return
Name: cells.dim
to examples after importing some data: Class: integer
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Spatial grids and pixels

There are two representations for data on regular rectangular

grids (oriented N-S, E-W): SpatialPixels and SpatialGrid
SpatialPixels are like SpatialPoints objects, but the
coordinates have to be regularly spaced; the coordinates are
stored, as are grid indices
SpatialPixelsDataFrame objects only store attribute data
where it is present, but need to store the coordinates and grid
indices of those grid cells
SpatialGridDataFrame objects do not need to store
coordinates, because they fill the entire defined grid, but they
need to store NA values where attribute values are missing

Roger Bivand Applied Spatial Data Analysis with R

Spatial grid and pixels classes and their slots

SpatialGridDataFrame Spatial
SpatialGrid SpatialGrid bbox
data Spatial proj4string
grid

GridTopology
cellcentre.offset
SpatialPixelsDataFrame cellsize
SpatialPixels SpatialPixels cells.dim
data grid
grid.index SpatialPoints
SpatialPoints coords
Spatial
data.frame
A SpatialGridDataFrame object
The space shuttle flew a radar topography mission in 2000, giving 90m
resolution elevation data for most of the world. The data here have been
warped to a UTM projection, but for the WGS84 datum - we’ll see below
how to project and if need be datum transform Spatial objects:
> DEM <- readGDAL("UTM_dem.tif")

UTM_dem.tif has GDAL driver GTiff

and has 122 rows and 121 columns

> summary(DEM)

Object of class SpatialGridDataFrame

Coordinates:
min max
x 288307.3 299442.7
y 9109607.8 9120835.2
Is projected: TRUE
proj4string :
[+proj=utm +zone=25 +south +ellps=WGS84 +units=m +no_defs]
Grid attributes:
cellcentre.offset cellsize cells.dim
x 288353.4 92.02797 121
y 9109653.8 92.02797 122
Data attributes:
Min. 1st Qu. Median Mean 3rd Qu. Max.
-1.00 3.00 11.00 19.45 31.00 88.00

> is.na(DEM$band1) <- DEM$band1 < 1

> image(DEM, "band1", axes = TRUE, col = terrain.colors(20))

Elevation at Olinda

9120000
9118000
9116000
9114000
9112000
9110000

288000 290000 292000 294000 296000 298000

The original data set CRS
Very often, a good deal of work is needed to find the actual coordinate
reference system of a data set. Typically, researchers use what is to hand,
without thinking that incomplete CRS specifications limit their ability to
integrate other data. Here we were told that the map was from the 1990
census, that it was most likely UTM zone 25 south, but not its datum.
Investigation showed that until very recently, the most used datum in
Brazil has been Corrego Alegre:
> proj4string(olinda)

[1] NA

> EPSG <- make_EPSG()

> EPSG[grep("Corrego Alegre", EPSG$note), 1:2]

code note
154 4225 # Corrego Alegre 1970-72
460 5524 # Corrego Alegre 1961
2741 5536 # Corrego Alegre 1961 / UTM zone 21S
2742 5537 # Corrego Alegre 1961 / UTM zone 22S
2743 5538 # Corrego Alegre 1961 / UTM zone 23S
2744 5539 # Corrego Alegre 1961 / UTM zone 24S
3140 22521 # Corrego Alegre 1970-72 / UTM zone 21S
3141 22522 # Corrego Alegre 1970-72 / UTM zone 22S
3142 22523 # Corrego Alegre 1970-72 / UTM zone 23S
3143 22524 # Corrego Alegre 1970-72 / UTM zone 24S
3144 22525 # Corrego Alegre 1970-72 / UTM zone 25S
The original data set CRS

In order to insert the parameters for the transformation to the WGS84

datum, we have to override the initial imported values:
> set_ReplCRS_warn(FALSE)

[1] FALSE

> proj4string(olinda) <- CRS("+init=epsg:22525")

> proj4string(olinda)

[1] "+init=epsg:22525 +proj=utm +zone=25 +south +ellps=intl +towgs84=-206,172,-6,0,0,0,0 +units=m +no_defs

> proj4string(DEM)

[1] "+proj=utm +zone=25 +south +ellps=WGS84 +units=m +no_defs"

> proj4string(DEM) <- CRS("+init=epsg:31985")

> proj4string(DEM)

[1] "+init=epsg:31985 +proj=utm +zone=25 +south +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs"

> set_ReplCRS_warn(TRUE)

[1] TRUE
Getting olinda to WGS84

As we see, although both olinda and DEM are UTM zone 25 south, they
differ in their ellipsoid and datum. Using spTransform methods in rgdal
we can undertake a datum shift for olinda, making it possible to
overplot in register:
> olinda1 <- spTransform(olinda, CRS(proj4string(DEM)))

> image(DEM, "band1", axes = TRUE, col = terrain.colors(20))

> plot(olinda1, add = TRUE, lwd = 0.5)
Getting olinda to WGS84

9120000
9118000
9116000
9114000
9112000
9110000

288000 290000 292000 294000 296000 298000

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Reading rasters

There are very many raster and image formats; some allow
only one band of data, others think data bands are RGB,
while yet others are flexible
There is a simple readAsciiGrid function in maptools that
reads ESRI Arc ASCII grids into SpatialGridDataFrame
objects; it does not handle CRS and has a single band
Much more support is available in rgdal in the readGDAL
function, which — like readOGR — finds a usable driver if
available and proceeds from there
Using arguments to readGDAL, subregions or bands may be
selected, which helps handle large rasters

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

But let’s keep things local

Using Landsat 7 panchromatic imagery,

we can get a similar contextual effect,
here with histogram equalised 15m
resolution:
> pan <- readGDAL("L8s.tif")

L8s.tif has GDAL driver GTiff

and has 800 rows and 800 columns

> bb0 <- set_ReplCRS_warn(FALSE)

> proj4string(pan) <- CRS("+init=epsg:31985")
> bb0 <- set_ReplCRS_warn(TRUE)
> brks <- quantile(pan$band1,
+ seq(0, 1, 1/255))
> pan$lut <- findInterval(pan$band1,
+ brks, all.inside = TRUE)

> image(pan, "lut", col = grey.colors(20))

> plot(olinda1, add = TRUE, border = "brown",
+ lwd = 0.5)
> title(main = "Landsat panchromatic channel 15m")

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Constructing and using NDVI

Landsat ETM also lets us construct a

30m resolution normalised difference
vegetation index (of more relevance for
other diseases), and plot it using a
ColorBrewer palette:
> letm <- readGDAL("L_ETMps.tif")

L_ETMps.tif has GDAL driver GTiff

and has 400 rows and 400 columns

> bb0 <- set_ReplCRS_warn(FALSE)

> proj4string(letm) <- CRS("+init=epsg:31985")
> bb0 <- set_ReplCRS_warn(TRUE)
> letm$ndvi <- (letm$band4 - letm$band3)/(letm$band4 +
+ letm$band3)
> library(RColorBrewer)
> mypal <- brewer.pal(5, "Greens")
> greens <- colorRampPalette(mypal)

> image(letm, "ndvi", col = greens(20))

> plot(olinda1, add = TRUE)

Roger Bivand Applied Spatial Data Analysis with R

Spatial classes provided by sp

This table summarises the classes provided by sp, and shows how
they build up to the objects of most practical use, the
Spatial*DataFrame family objects:

data type class attributes extends

points SpatialPoints none Spatial
points SpatialPointsDataFrame data.frame SpatialPoints
pixels SpatialPixels none SpatialPoints
pixels SpatialPixelsDataFrame data.frame SpatialPixels
SpatialPointsDataFrame
full grid SpatialGrid none SpatialPixels
full grid SpatialGridDataFrame data.frame SpatialGrid
line Line none
lines Lines none Line list
lines SpatialLines none Spatial, Lines list
lines SpatialLinesDataFrame data.frame SpatialLines
polygon Polygon none Line
polygons Polygons none Polygon list
polygons SpatialPolygons none Spatial, Polygons list
polygons SpatialPolygonsDataFrame data.frame SpatialPolygons
Methods provided by sp

This table summarises the methods provided by sp:

method what it does

[ select spatial items (points, lines, polygons, or
rows/cols from a grid) and/or attributes variables
$, $<-, [[, [[<- retrieve, set or add attribute table columns
spsample sample points from a set of polygons, on a set of
lines or from a gridded area
bbox get the bounding box
proj4string get or set the projection (coordinate reference sys-
tem)
coordinates set or retrieve coordinates
coerce convert from one class to another
over combine two different spatial objects
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Using Spatial family objects

Very often, the user never has to manipulate Spatial family
objects directly, as we have been doing here, because methods
to create them from external data are also provided
Because the Spatial*DataFrame family objects behave in
most cases like data frames, most of what we are used to
doing with standard data frames just works — like [ or $
(but no merge, etc., yet)
These objects are very similar to typical representations of the
same kinds of objects in geographical information systems, so
they do not suit spatial data that is not geographical (like
medical imaging) as such
They provide a standard base for analysis packages on the one
hand, and import and export of data on the other, as well as
shared methods, like those for visualisation
Roger Bivand Applied Spatial Data Analysis with R
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Writing objects
In rgdal, writeGDAL can write for example multi-band
GeoTiffs, but there are fewer write than read drivers; in
general CRS and geogreferencing are supported — see
gdalDrivers
The rgdal function writeOGR can be used to write vector files,
including those formats supported by drivers, including now
KML — see ogrDrivers
External software (including different versions) tolerate output
objects in varying degrees, quite often needing tricks - see
mailing list archives
In maptools, there are functions for writing sp objects to
shapefiles — writeSpatialShape, etc., as Arc ASCII grids —
writeAsciiGrid, and for using the R PNG graphics device for
outputting image overlays for Google Earth
Roger Bivand Applied Spatial Data Analysis with R
 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Using the direct KML method

One of the nice things about linking to OSGeo software is that when
someone contributes a driver, it is there for other software too. With the
increasing availability of software like Google Earth (or Google Maps), the
ability to display data or results in a context that is familiar to the user
can be helpful, but we need geographical coordinates:
> olinda_ll <- spTransform(olinda1, CRS("+proj=longlat +datum=WGS84"))
> writeOGR(olinda_ll, dsn = "olinda_ll.kml", layer = "olinda_ll", driver = "KML",
+ overwrite_layer = TRUE)

Roger Bivand Applied Spatial Data Analysis with R

 Study area
Spatial data in R
Representing spatial data
Handling spatial data in R: GRASS interface
Coordinate reference systems
Handling spatial data in R: methods
Raster data
Worked examples in spatial statistics
Exporting data

Olinda census district boundaries in GE

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

GIS interfaces

Because GIS can be used as databases, and their tools can be

better suited to some analyses and operations, it may be
sensible to use one in addition to data analysis software
There is an extra effort required when using linked software
systems, because they may not integrate easily
Since R is open source, and R spatial packages use open
source components, staying with open source GIS means that
many of the underlying software components are shared
This certainly applies to R and GRASS, and through GRASS,
also to R and QGIS — QGIS seemes to be more file-based
than GRASS, which has an underlying data storage
specification

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

R — GIS interfaces

GIS interfaces can be as simple as just reading and writing files —

loose coupling, once the file formats have been worked out, that is
Loose coupling is less of a burden than it was with smaller, slower
machines, which is why the GRASS 5 interface was tight-coupled,
with R functions reading from and writing to the GRASS database
directly
The GRASS 6 interface spgrass6 on CRAN also runs R within
GRASS, but uses intermediate temporary files; the package is under
development but is quite usable
Use has been made of COM and Python interfaces to ArcGIS;
typical use is by loose coupling except in highly customised work
situations

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Layering of shells

The interface between R and GRASS

uses the fact that GRASS modules can
be run as command line programs at the
shell prompt. The shell has certain
R interactive session
environment variables for GRASS set, for system("g.region −p")
system("psql midwest")

example saying where the data is stored, GRASS environment and location

but is otherwise a regular shell, from System shell and environment (csh, ksh, bash)

Underlying operating system

which R can be started. This instance of
R inherits the environment variables set
by GRASS

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Which software component faces the user?

Finally, although for research purposes

it may be prefered to have data analysis
external

software, such as R, facing the user, it is databases

possible to try to “embed” this

component in the workflow, so that the external
GIS functionality
R/GIS
interfaces
GIS/application
GUI

end user does not need so much training

— but then an “expert” has to codify the external
files
steps needed in advance

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

The GRASS location

GRASS uses a structure of locations, with fixed projection and

variable region (window) and resolution
The PERMANENT mapset contains data that is not supposed
to change, and users should create their own mapsets, where
work-in-progress can be kept
Each mapset can contain many directories, including
directories for raster and vector data components, projection
metadata, and various windows used in parts of the location
Raster layers with different resolutions are resampled to the
current region when used

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Getting GRASS running

The stable version of GRASS, 6.4.*, has a native Windows

build, so GRASS modules and the GRASS shell are now much
easier to install
GRASS has very mature raster analysis modules, including a
complete map calculator, as well as many other useful tools
On the vector side, GRASS 6 uses a topological vector model,
cleaning input data for oddities; keeping the geometries
associated with the data can be challenging

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Installing GRASS

The easiest way is using the GRASS download, which is large,

but contains all the dependencies on Windows.
Type the full path to Rgui.exe at the shell prompt, for me
"C:/Program Files/R/R-3.1.0/bin/x64/Rgui.exe"
You may need to set the path to R_LIBS in user environment
variables, or to add it to the command line in the GRASS shell
prompt, or use .libPaths() within R

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Using GRASS inside R

There are two ways of using GRASS

inside R, by starting R inside a running
GRASS session, or by initiating GRASS
> library(spgrass6)
inside R, here using a throw-away > myGRASS <- "/home/rsb/topics/grass/g643/grass-6.4.3"
> initGRASS(myGRASS, tempdir(),
location to show how the streams were + SG = DEM, override = TRUE)
generated, set to the extent and
resolution of the elevation data from
SRTM:

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Moving data and running GRASS commands

Functions are provided in spgrass6 for

moving raster and vector data both
ways. In addition, all GRASS commands > writeRAST6(DEM, "dem", flags = "o")
> execGRASS("g.region", rast = "dem")
replying to an interface description > execGRASS("r.resamp.rst", input = "dem",
+ ew_res = 14.25, ns_res = 14.25,
request can be run programmatically + elev = "DEM_resamp", slope = "slope",
+ aspect = "aspect")
from within R using execGRASS. Here > execGRASS("g.region", rast = "DEM_resamp")
> DEM_out <- readRAST6(c("DEM_resamp",
we resample the DEM to the resolution + "slope", "aspect"))

of the panchromatic image, and reset

the region to the output raster:

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Creating the stream network

This is where the work gets done, with

GRASS command r.watershed
analysing the resampled DEM to detect
> execGRASS("r.watershed", elevation = "DEM_resamp",
depressions through which streams may + stream = "stream", threshold = 1000L,
+ convergence = 5L, memory = 300L)
flow; it can also output the boundaries > execGRASS("r.thin", input = "stream",
+ output = "stream1", iterations = 200L)
of watersheds, but these are not needed
here. The output is in raster form, and
may also be thinned:

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Returning results

9120000
9118000
First we vectorise the raster stream

9116000
network, and read it into the R session
for display:

9114000
> execGRASS("r.to.vect", input = "stream1",
+ output = "stream", feature = "line")

9112000
> stream1 <- readVECT6("stream")
9110000

288000 290000 292000 294000 296000 298000

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface
Handling spatial data in R: methods
Worked examples in spatial statistics

Using Jasiewicz/Metz r.stream.extract

Using the same resampled input raster

DEM, we can also use the
r.stream.extract GRASS add-on.
Add-ons are not distributed with GRASS > execGRASS("r.stream.extract",
+ elevation = "DEM_resamp",
itself, but can be compiled and installed + stream_vect = "v_stream",
+ threshold = 1000)
for if GRASS is installed on the system, > stream1 <- readVECT6("v_stream")

and the same build tools. We will be

using this computed stream line version
here:

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Operating on geometries

The rgeos package interfaces the

GEOS/JTS topology suite providing
> library(rgeos)
predicates and operations for geometries.
> getScale()
The simplest are measures for planar
[1] 1e+08
geometries — only these are supported. A
> pols <- as(olinda1, "SpatialPolygons")
> Area <- gArea(pols, byid = TRUE) complication is that computational
> olinda1$area <- sapply(slot(olinda1,
+ "polygons"), slot, "area") geometry may represent objects using
> all.equal(unname(Area), olinda1$area)
different scaling factors, leading to
[1] TRUE
geometries becoming “unclean”; GEOS uses
a fixed precision representation.

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Intersections - channels and tracts

> lns <- as(stream1, "SpatialLines")

> system.time(GI <- gIntersects(lns,
+ pols, byid = TRUE))
When many geometries are involved in
user system elapsed
0.031 0.000 0.031 topological operations, like finding the
> unname(which(GI[2, ])) drainage channels in each tract and
[1] 50 measuring their length, it is often prudent
> res <- numeric(length = nrow(GI)) to run the equivalent predicate first — here
> system.time(for (i in seq(along = res)) {
+ if (any(GI[i, ])) { gIntersects, and then only operate on
+ resi <- gIntersection(lns[GI[i,
+ ]], pols[i]) combinations of objects which meet that
+ res[i] <- gLength(resi)
+ } condition. We assign the drainage channel
+ })
lengths to the tracts object.
user system elapsed
0.305 0.000 0.307

> olinda1$stream_len <- res

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Intersections - channels and tracts

> res0 <- numeric(length = length(pols))

> length(res0)

[1] 243

> system.time(lnsInt <- gIntersection(lns, We can see that using gIntersection

+ pols, byid = c(FALSE, TRUE)))
directly gives the same sums of lengths, but
user system elapsed
5.261 0.000 5.274 at a greater cost in time, because each tract
> length(lnsInt) is intersected with all streams in turn, not
[1] 142 just the ones that do intersect each other.
> res0[as.integer(row.names(lnsInt)) +
+ 1] <- gLength(lnsInt, byid = TRUE)
> all.equal(res, res0)

[1] TRUE

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Intersections - channels and tracts

> system.time(tree <- gBinarySTRtreeQuery(lns, It is also possible to build and query a
+ pols))
Sort-Tile-Recursive tree to check envelopes
user system elapsed
0.007 0.000 0.007 of objects for intersection. The advantages
> tree[[2]] of doing this increase with the number of
[1] 46 47 48 50 objects, as the predicate output matrix is
> res1 <- numeric(length = length(tree)) returned dense, although many elements are
> system.time(for (i in seq(along = res1)) {
+ if (!is.null(tree[[i]])) { FALSE. The list returned by STR tree query
+ gi <- gIntersection(lns[tree[[i]]],
+ pols[i]) is sparse, and tree query does make a big
+ res1[i] <- ifelse(is.null(gi),
+ 0, gLength(gi)) difference for big data sets. Here, the cost
+ }
+ }) of taking intersections for tracts and
user system elapsed streams with intersecting envelopes but
0.465 0.000 0.467
which do not intersect makes this approach
> all.equal(res, res1)
slower than using gIntersects.
[1] TRUE

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Intersections - buffered channels and tracts

> buf50m <- gBuffer(lns, width = 50)

> GI1 <- gIntersects(buf50m, pols, Just knowing the lengths by tract may be
+ byid = TRUE)
> res <- numeric(length = nrow(GI1)) less useful than knowing the proportion of
> system.time(for (i in seq(along = res)) {
+ if (any(GI1[i, ])) { areas within 50m of drainage channels. We
+ out <- gIntersection(buf50m[GI1[i,
+ ]], pols[i]) can use gBuffer to create a single
+ res[i] <- gArea(out)
+ } SpatialPolygons object, which we can
+ })
intersect with the tracts — using an STR
user system elapsed
2.999 0.000 3.008 tree here will not lead to savings because
> olinda1$buf_area <- res we only have one buffer object.
> prop <- olinda1$buf_area/olinda1$area
> olinda1$prop_50m <- prop

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Intersections - buffered channels and tracts

> system.time(bufInt <- gIntersection(buf50m,

+ pols, byid = c(FALSE, TRUE)))
As we can see, because the buffer is
user system elapsed
3.339 0.000 3.347 defined as a single entity, running
> length(bufInt) gIntersection directly takes about the
[1] 166 same amount of time as subsetting; we still
> res0 <- numeric(length = length(pols)) have to put the shorter output into the
> res0[as.integer(row.names(bufInt)) +
+ 1] <- gArea(bufInt, byid = TRUE) correct rows of the polygons object.
> all.equal(res, res0)

[1] TRUE

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Drainage channels in Olinda

If we merge the tracts within the city
limits, we can make a cookie-cutter, and Prop. of area < 50m of streams
use it to clip/crop other objects to match.
So now we can display the proportions of 0.8

the areas of tracts within 50m of drainage

channels in the city, omitting channels 0.6

outside the city bounds:

> bounds <- gUnaryUnion(pols) 0.4
> stream_inside <- gIntersection(lns,
+ bounds)

> bl <- colorRampPalette(brewer.pal(5, 0.2

+ "Blues"))
> tt <- "Prop. of area < 50m of streams"
> spplot(olinda1, "prop_50m", 0.0
+ col.regions = bl(20), col = "transparent",
+ sp.layout = list("sp.lines",
+ stream_inside), main = tt)

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R
Handling spatial data in R: GRASS interface Topology predicates and operations
Handling spatial data in R: methods Overlay
Worked examples in spatial statistics

Overlay operations

The basic approach is to query one spatial

data object using a second spatial data Normalized difference vegetation index

object of the same or of a different class.

0.3
The query result will typically be another
0.2
spatial data object or an object pointing to
one of the input objects. We also remove 0.1

two census tracts with missing population:

0.0
> o_ndvi <- over(olinda1, letm[,
+ , "ndvi"], fn = median)
> o_alt <- over(olinda1, DEM_out[, −0.1
+ , "DEM_resamp"], fn = median)
> library(maptools)
−0.2
> olinda1A <- spCbind(olinda1,
+ cbind(o_ndvi, o_alt))
> olinda2 <- olinda1A[!is.na(olinda1A$POP), −0.3
+ ]

Roger Bivand Applied Spatial Data Analysis with R

Extract operations

It is possible to use extract methods from the raster package to

aggregate queried values; note that the arguments are reversed in order
compared to over methods in sp. We convert our
SpatialPixelsDataFrame object to a RasterStack object, and obtain
when querying with a SpatialPolygons object, a list of numeric
matrices with the values of all variables in rasters in the RasterStack
object, for all the cells with centres falling in each Polygons object. As
we can see by comparison, extract methods from the raster package
yield the same results as over methods in sp for 28.5m NDVI medians.
> library(raster)

> TMrs <- stack(letm)

> e1 <- extract(TMrs, as(olinda2, "SpatialPolygons"))
> all.equal(sapply(e1, function(x) median(x[, "ndvi"])), olinda2$ndvi)

[1] TRUE
 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Probability maps and smoothing

Choynowski proposed a method for making probability maps

in 1959, still described by contemporary authors — it splits
between upper and lower tails
It is also possible to say just ppois(Observed, Expected) to
get a similar result. If the observations are distributed as
assumed, this will indicate which regions appear unusual
Probability maps typically mirror the spatial pattern of relative
risks; using Empirical Bayes smoothing will use the population
at risk to adjust relative risks
Empirical Bayes smoothing is implemented for the
method-of-moments global and local cases in spdep and for
the ML global case in DCluster

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Choynowski probability map

Choynowski probability map

low
N/S
high

The legacy Choynowski approach is

implemented in spdep
> library(spdep)
> ch <- choynowski(olinda2$CASES,
+ olinda2$POP)

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Poisson probability map

Poisson probability map

[0,0.05) (53)
[0.05,0.1) (24)
[0.1,0.2) (17)
[0.2,0.8) (82)
[0.8,0.9) (8)
[0.9,0.95) (8)
The Poisson probability map in one [0.95,1] (49)

tail is identical with the Choynowski

approach, but shows both tails rather
than folding them together
> pm <- probmap(olinda2$CASES,
+ olinda2$POP)
> names(pm)

[1] "raw" "expCount" "relRisk"

[4] "pmap"

> olinda2$SMR <- pm$relRisk

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Weaknesses of probability maps

Poisson probability map

80
If the underlying distribution of the
data does not agree with our

60
assumption, we may get several
possible processes mixed up,

Frequency

40
overdispersion with spatial
dependence:

20
> table(findInterval(pm$pmap,
+ seq(0, 1, 1/10)))

1 2 3 4 5 6 7 8 9 10
77 17 26 9 12 16 8 11 8 57
0

0.0 0.2 0.4 0.6 0.8 1.0

pm$pmap

Roger Bivand Applied Spatial Data Analysis with R

 Some Empirical Bayes rates?

6
EB_ml

We can compare the relative rate and

the EB relative rate: 4

> library(DCluster)
> olinda2$RR <- olinda2$CASES/olinda2$Expected
> olinda2$EB_ml <- empbaysmooth(olinda2$CASES,
+ olinda2$Expected)$smthrr 3
RR
> spplot(olinda2, c("RR", "EB_ml"),
+ col.regions = brewer.pal(10,
+ "RdBu"), at = c(seq(0,
2
+ 1, 0.25), seq(1, 6,
+ 1)))

0
 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Neighbours

●
●
● ●
●●

In order to investigate spatial ●

●
●
● ●
● ●● ●●
● ● ●●

● ●
●
●
● ●
●
●
●

●
●

● ● ●● ● ●
●

dependence, we need a list of ●

●
●
●
●
● ●

●
●
● ●● ● ●
● ●
● ●

neighbours. We take Queen ●

●
●

●
●
●

● ●
●
●
●
●
● ● ●

●
● ● ●
contiguities of the tract polygons: ●
●●
●
●

● ●
● ●
●

●
●
●●
●
● ●

●
●
●
●
● ●
● ●
●
●

● ● ● ● ●
●
● ● ● ●● ● ● ● ● ●
●
● ● ●
> nb <- poly2nb(olinda2) ●
●
● ● ● ●
●
● ●
● ● ●
●
●
●
● ● ● ● ● ● ● ● ●
● ●
> nb ● ●● ●
● ● ●
● ● ● ● ●
● ● ● ● ● ● ●
● ●
● ● ●
● ● ● ●
● ● ● ●
Neighbour list object: ●
● ●
● ●
●● ● ●
●●
●
●
●
●
● ● ●
Number of regions: 241 ●
● ●
●
●
●
●
●
● ● ●
●
●
●● ● ● ● ● ● ● ●
Number of nonzero links: 1324 ● ● ●
● ●
● ●
Percentage nonzero weights: 2.279575 ●
●
● ●
Average number of links: 5.493776 ●
●
● ●
●

● ● ●
●
●

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Local Empirical Bayes smoothing

6
EB_mm_local

5
If instead of shrinking to a global
rate, we shrink to a local rate, we 4
may be able to take unobserved
heterogeneity into account; here we 3
RR EB_ml
use the list of neighbours:
> olinda2$Observed <- olinda2$CASES 2
> eb2 <- EBlocal(olinda2$Observed,
+ olinda2$Expected, nb)
> olinda2$EB_mm_local <- eb2$est
1

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Moran’s I

Histogram of t

DCluster provides a permutation ●

10

0.15
●
●
●●
●
●
●

bootstrap test for spatial ●

●
●
●
●
●
●

0.10
●
●●
●
●

8
●
●
●
●
●
●

autocorrelation of the difference ●

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

0.05
●●
●
●
●

between observed and expected

6
●
●●
●
●

Density
●
●●
●
●
●●
●
●●
●
●
●
●

t*
●●
●
●
●
●
●
●
●●
●
●

counts: ●
●
●

0.00
●
●●
●
●●
●
●
●●
●

4
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
> lw <- nb2listw(nb) ●●
●

−0.05
●
●●
●
●
●
●●
●●
●
●
●●
●
●
> set.seed(130709) ●●

2
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
> moran.boot <- boot(as(olinda2, ●
●
●
●
●
●

−0.10
●●●
+ "data.frame"), statistic = moranI.boot,
0 ●
+ R = 999, listw = lw, n = length(nb),
+ S0 = Szero(lw)) −0.1 0.1 0.2 0.3 0.4 −3 −2 −1 0 1 2 3

t* Quantiles of Standard Normal

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Moran’s I

Histogram of t

0.15
●

10
●●

0.10
●

It also provides parametric ●

●
●
●
●
●
●●

8
●●
●
●
●
●
●
●
●
●
●
●
●●

bootstraps for variants, including the ●

●
●
●
●

0.05
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●

6
●
●
●
●

Density
●

Negative Binomial: ●
●
●
●
●
●
●
●
●
●
●

t*
●
●
●
●
●

0.00
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●

4
●
●
> moran.pgboot <- boot(as(olinda2, ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
+ "data.frame"), statistic = moranI.pboot, ●
●●

−0.05
●
●●
●
●
●●
●
●●
●
●
●
+ sim = "parametric", ran.gen = negbin.sim, ●
●
●
●
●
●
●
●
●
●
●
●

2
●
●
●●
●
●●
+ R = 999, listw = lw, n = length(nb), ●
●
●
●
●
●
●
●
●
●
●
●●

−0.10
●●
●
+ S0 = Szero(lw)) ●
●●
●
●
●

0 ●

−0.1 0.1 0.2 0.3 0.4 −3 −2 −1 0 1 2 3

t* Quantiles of Standard Normal

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Assunção and Reis’ correction

The Assunção and Reis’ correction to Moran’s I is implemented in

spdep:
> EBImoran.mc(olinda2$CASES, olinda2$Expected, lw, nsim = 999)

Monte-Carlo simulation of Empirical Bayes Index

data: cases: olinda2$CASES, risk population: olinda2$Expected

weights: lw
number of simulations + 1: 1000

statistic = 0.3941, observed rank = 1000, p-value = 0.001

alternative hypothesis: greater

Roger Bivand Applied Spatial Data Analysis with R

Maybe some modelling?
To show how modelling and cartography can be inter-woven, let us fit
some quasipoisson models, after seeing that — once again — spatial
dependence and overdispersion seem to be entwined:
> m0p <- glm(CASES ~ 1 + offset(log(Expected)), data = olinda2, family = poisson)
> DeanB(m0p)

Dean's P_B test for overdispersion

data: m0p
P_B = 50.2478, p-value < 2.2e-16
alternative hypothesis: greater

> m0 <- glm(CASES ~ 1 + offset(log(Expected)), data = olinda2, family = quasipoisson)

> m1 <- update(m0, . ~ . + DEPRIV)
> m2 <- update(m1, . ~ . + ndvi)
> m3 <- update(m2, . ~ . + DEM_resamp)
> anova(m0, m1, m2, m3, test = "F")

Analysis of Deviance Table

Model 1: CASES ~ 1 + offset(log(Expected))

Model 2: CASES ~ DEPRIV + offset(log(Expected))
Model 3: CASES ~ DEPRIV + ndvi + offset(log(Expected))
Model 4: CASES ~ DEPRIV + ndvi + DEM_resamp + offset(log(Expected))
Resid. Df Resid. Dev Df Deviance F Pr(>F)
1 240 1212.5
2 239 1096.6 1 115.871 24.9914 1.12e-06 ***
3 238 1059.4 1 37.183 8.0197 0.005025 **
4 237 1038.2 1 21.281 4.5900 0.033179 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Plotting the fitted values

40
f2 f3

We can use spplot methods to plot 35

sequences of variables on the same scale 30

using lattice graphics, once more using a

25
Color Brewer palette:
> olinda2$f1 <- fitted(m1) 20
CASES f1
> olinda2$f2 <- fitted(m2)
> olinda2$f3 <- fitted(m3)
15
> spplot(olinda2, c("CASES", "f1",
+ "f2", "f3"), col.regions = brewer.pal(8, 10
+ "Reds"), at = seq(0, 40,
+ 5))
5

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Residual maps

r3 30
So we can also plot residual maps for
the three nested models, but getting the 20
palette divergence in the right place
takes a little more care: 10

> olinda2$r1 <- residuals(m1,

+ type = "response")
0
> olinda2$r2 <- residuals(m2, r1 r2
+ type = "response")
> olinda2$r3 <- residuals(m3,
+ type = "response") −10

> spplot(olinda2, c("r1", "r2",

+ "r3"), col.regions = brewer.pal(8, −20
+ "RdBu"), at = c(-30, -15,
+ -10, -5, 0, 5, 10, 15, 31))

−30

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Bayesian spatial regression

In order to model the data, we need to try to fit random

effects with the stipulated distributions
The Poisson-Gamma model (PG) fits a model without
spatially structured random effects
The Besag, York and Molliè (BYM) model includes a spatially
structured random effect
The Besag structure is a conditional spatialautoregressive
model

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

INLA — integrated nested Laplace approximation

These models are not very complicated, but many WinBUGS

models run for hours or days — making inference is impeded
by time and computation constraints
INLA — http://www.r-inla.org/ provides ways to reach
very similar inferences much faster, using integrated nested
Laplace approximation instead of Gibbs sampling in MCMC
The inla function in the INLA package (installed from the
homepage of the INLA project, not yet CRAN) uses the
familiar formula interface
It adds an extra function in the formula to define which model
is to be used

Roger Bivand Applied Spatial Data Analysis with R

 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

INLA BYM model

By using the standard White book framework of formula, data, and

family, inla is easy to use, but only offers some built-in models:
> if (!("INLA" %in% .packages(all = TRUE))) source("http://www.math.ntnu.no/inla/givemeINLA.R")
> library(INLA)
> olinda2$INLA_ID <- 1:nrow(olinda2)
> INLA_BYM <- inla(CASES ~ DEPRIV +
+ ndvi + DEM_resamp + f(INLA_ID,
+ model = "bym", graph = nb2mat(nb,
+ style = "B")) + offset(log(Expected)),
+ family = "poisson", data = as(olinda2,
+ "data.frame"), control.predictor = list(compute = TRUE))

Roger Bivand Applied Spatial Data Analysis with R

INLA BYM model

> summary(INLA_BYM)

Call:
c("inla(formula = CASES ~ DEPRIV + ndvi + DEM_resamp + f(INLA_ID, ", " model = \"bym\", graph = nb2mat

Time used:
Pre-processing Running inla Post-processing Total
0.1771 0.7633 0.0490 0.9894

Fixed effects:
mean sd 0.025quant 0.5quant 0.975quant mode kld
(Intercept) -0.2314 0.1860 -0.5995 -0.2306 0.1318 -0.2289 0
DEPRIV 0.4183 0.3438 -0.2577 0.4184 1.0934 0.4185 0
ndvi 0.6982 0.5930 -0.4660 0.6977 1.8642 0.6967 0
DEM_resamp -0.0126 0.0063 -0.0249 -0.0126 -0.0002 -0.0127 0

Random effects:
Name Model
INLA_ID BYM model

Model hyperparameters:
mean sd 0.025quant 0.5quant 0.975quant mode
Precision for INLA_ID (iid component) 5.4605 1.9042 2.7427 5.1196 10.0987 4.5074
Precision for INLA_ID (spatial component) 1.5733 0.5878 0.6928 1.4862 2.9701 1.3210

Expected number of effective parameters(std dev): 145.62(7.199)

Number of equivalent replicates : 1.655

Marginal Likelihood: -616.09

Posterior marginals for linear predictor and fitted values computed
 INLA BYM model

We can also extract the INLA BYM

4
standardised rates:
> olinda2$INLA <- exp(INLA_BYM$summary.linear.predictor[, 3
+ 1] - log(olinda2$Expected))

> spplot(olinda2, c("INLA"), col.regions = brewer.pal(10,

+ "RdBu"), at = c(seq(0, 1, 2
+ 0.25), seq(1, 6, 1)))

0
 Spatial data in R Disease mapping
Handling spatial data in R: GRASS interface Spatial autocorrelation
Handling spatial data in R: methods Regression
Worked examples in spatial statistics Bayesian spatial regression with INLA

Summary

Using spatial data (in R) permits attention to be given to the

important steps of data handling and representation —
rushing to analysis always leads to problems
Handling spatial data is never “easy”, because representational
choices lead to support consequences, which qualify or
undermine inferential outcomes
Being close to the data and close to the code may give
considerable insight and freedom, and permits co-working with
disciplines sharing the same frame of understanding
Just publishing is not going to be enough; we will need to be
able to demonstrate how our conclusions have been reached
by releasing code and data (much data is public already).

Roger Bivand Applied Spatial Data Analysis with R