gstat.Rnw

%% Document started, Sat Jul  3 19:30:52 CEST 2004, my 37th birthday,
%% while being stuck for 24 hours at Philadelphia airport, on my way 
%% back from the joint TIES/Accuracy 2004 symposium in Portland, ME.
%% Continued, Oct 28, during the Apmosphere mid-term review. Oh, shame.

% \VignetteIndexEntry{ The meuse data set: a tutorial for the gstat R package }

\documentclass[a4paper]{article}
\usepackage{hyperref}

\newcommand{\code}[1]{{\tt #1}}

\SweaveOpts{echo=TRUE}

\title{The meuse data set: a brief tutorial\\
for the {\tt gstat} R package }
\author{\href{mailto:edzer.pebesma@uni-muenster.de}{Edzer Pebesma}}
\date{\today}

\begin{document}
\maketitle

\section{Introduction}
The \code{meuse} data set provided by package \code{sp} is a
data set comprising of four heavy metals measured in the top soil
in a flood plain along the river Meuse, along with a handful of
covariates. The process governing heavy metal distribution seems
that polluted sediment is carried by the river, and mostly deposited
close to the river bank, and areas with low elevation. This document
shows a geostatistical analysis of this data set. The data set was
introduced by Burrough and McDonnell, 1998.

This tutorial introduced the functionality of the R package \code{gstat},
used in conjunction with package \code{sp}. Package \code{gstat} provides
a wide range of univariable and multivariable geostatistical modelling,
prediction and simulation functions, where package \code{sp} provides
general purpose classes and methods for defining, importing/exporting
and visualizing spatial data.

\section{R geostatistics packages}
Package \code{gstat} (Pebesma, 2004) is an R package that provides basic
functionality for univariable and multivariable geostatistical analysis,
including 
\begin{itemize}
\item variogram modelling, residual variogram modelling, and cross variogram
modelling using fitting of parametric models to sample variograms
\item geometric anisotropy specfied for each partial variogram model
\item restricted maximum likelihood fitting of partial sills 
\item variogram and cross variogram maps
\item simple, ordinary, universal and external drift (co)kriging
\item (sequential) Gaussian (co)simulation equivalents for each of
the kriging varieties
\item indicator (co)kriging and sequential indicator (co)simulation
\item kriging in a local or global neighbourhood 
\item block (co)kriging or simulation for each of the varieties, for
rectangular or irregular blocks
\end{itemize}
Other geostatistical packages for R usually lack part of these options
(e.g. block kriging, local kriging, or cokriging) but provide others:
e.g. package \code{geoR} and \code{geoRglm} (by Paulo Ribeiro and Ole
Christensen) provide the model-based geostatistics framework described
in Diggle et al. (1998), package \code{fields} (Doug Nychka and others)
provides thin plate spline interpolation, covariance functions for
spherical coordinates (unprojected data), and routines for spatial
sampling design optimization.

\section{Spatial data frames}
As an example, we will look at the meuse data set, which is a
regular data frame that comes with package \code{gstat} (remove the
88 from the colour strings to make a plot without alpha transparency
on windows or X11):

<<fig=FALSE>>=
library(sp)
data(meuse)
class(meuse)
names(meuse)
coordinates(meuse) = ~x+y
class(meuse)
summary(meuse)
coordinates(meuse)[1:5,]
bubble(meuse, "zinc", 
	col=c("#00ff0088", "#00ff0088"), main = "zinc concentrations (ppm)")
@

% the following is needed because lattice plots (bubble wraps xyplot) do not
% show without an explicit print; in order not to confuse users, we hide this:
<<fig=TRUE,echo=FALSE>>=
print(bubble(meuse, "zinc", 
	col=c("#00ff0088", "#00ff0088"), main = "zinc concentrations (ppm)")
)
@

and note the following: 
\begin{enumerate}
\item the function \code{coordinates}, when assigned (i.e. on
the left-hand side of an \verb|=| or \verb|<-| sign), promotes the
\code{data.frame} meuse into a \code{SpatialPointsDataFrame}, which knows about
its spatial coordinates; coordinates may be specified by a formula,
a character vector, or a numeric matrix or data frame with the actual
coordinates
\item the function \code{coordinates}, when not assigned, {\em retrieves}
the spatial coordinates from a \code{SpatialPointsDataFrame}.
\item the two plotting functions used, \code{plot} and \code{bubble}
assume that the $x$- and $y$-axis are the spatial coordinates.
\end{enumerate}

\section{Spatial data on a regular grid}

<<fig=TRUE,echo=TRUE>>=
data(meuse.grid)
summary(meuse.grid)
class(meuse.grid)
coordinates(meuse.grid) = ~x+y
class(meuse.grid)
gridded(meuse.grid) = TRUE
class(meuse.grid)
image(meuse.grid["dist"])
title("distance to river (red = 0)")
library(gstat)
zinc.idw = idw(zinc~1, meuse, meuse.grid)
class(zinc.idw)
spplot(zinc.idw["var1.pred"], main = "zinc inverse distance weighted interpolations")
@

<<fig=TRUE,echo=FALSE>>=
print(spplot(zinc.idw["var1.pred"], main = "zinc inverse distance weighted interpolations"))
@

If you compare the bubble plot of zinc measurements with the map with
distances to the river, it becomes evident that the larger concentrations
are measured at locations close to the river. This relationship can be
linearized by log-transforming the zinc concentrations, and taking the
square root of distance to the river:

<<fig=TRUE,echo=TRUE>>=
plot(log(zinc)~sqrt(dist), meuse)
abline(lm(log(zinc)~sqrt(dist), meuse))
@

\section{Variograms }

Variograms are calculated using the function \code{variogram}, which
takes a formula as its first argument: \verb|log(zinc)~1| means that we
assume a constant trend for the variable log(zinc).

<<fig=FALSE>>=
lzn.vgm = variogram(log(zinc)~1, meuse)
lzn.vgm
lzn.fit = fit.variogram(lzn.vgm, model = vgm(1, "Sph", 900, 1))
lzn.fit
plot(lzn.vgm, lzn.fit)
@

<<fig=TRUE,echo=FALSE>>=
print(plot(lzn.vgm, lzn.fit))
@

Instead of the constant mean, denoted by \verb|~1|, we can specify a
mean function, e.g. using \verb|~sqrt(dist)| as a predictor variable:

<<fig=FALSE>>=
lznr.vgm = variogram(log(zinc)~sqrt(dist), meuse)
lznr.fit = fit.variogram(lznr.vgm, model = vgm(1, "Exp", 300, 1))
lznr.fit
plot(lznr.vgm, lznr.fit)
@

<<fig=TRUE,echo=FALSE>>=
print(plot(lznr.vgm, lznr.fit))
@

In this case, the variogram of residuals with respect to a fitted mean
function are shown. Residuals were calculated using ordinary least
squares.

\section{Kriging}

<<fig=FALSE>>=
lzn.kriged = krige(log(zinc)~1, meuse, meuse.grid, model = lzn.fit)
spplot(lzn.kriged["var1.pred"])
@

<<fig=TRUE,echo=FALSE>>=
print(spplot(lzn.kriged["var1.pred"]))
@

\section{Conditional simulation}
<<fig=FALSE>>=
lzn.condsim = krige(log(zinc)~1, meuse, meuse.grid, model = lzn.fit, 
    nmax = 30, nsim = 4)
spplot(lzn.condsim, main = "four conditional simulations")
@

<<fig=TRUE,echo=FALSE>>=
print(spplot(lzn.condsim, main = "four conditional simulations"))
@

For UK/residuals:

<<fig=FALSE>>=
lzn.condsim2 = krige(log(zinc)~sqrt(dist), meuse, meuse.grid, model = lznr.fit, 
    nmax = 30, nsim = 4)
spplot(lzn.condsim2, main = "four UK conditional simulations")
@

<<fig=TRUE,echo=FALSE>>=
print(spplot(lzn.condsim2, main = "four UK conditional simulations"))
@

\section{Directional variograms}
The following command calculates a directional sample variogram, where
directions are binned by direction angle alone. For two point pairs,
$Z(s)$ and $Z(s+h)$, the separation vector is $h$, and it has a direction.
Here, we will classify directions into four direction intervals:

<<fig=FALSE>>=
lzn.dir = variogram(log(zinc)~1, meuse, alpha = c(0, 45, 90, 135))
lzndir.fit = vgm(.59, "Sph", 1200, .05, anis = c(45, .4))
plot(lzn.dir, lzndir.fit, as.table = TRUE)
@

<<fig=TRUE,echo=FALSE>>=
print(plot(lzn.dir, lzndir.fit, as.table = TRUE))
@

Looking at directions between 180 and 360 degrees will repeat the
image shown above, because the variogram is a symmetric measure:
$(Z(s)-Z(s+h))^2=(Z(s+h)-Z(s))^2$.

The first plot gives the variogram in the zero direction, which is
North; 90 degrees is East. By default, point pairs are assigned to the
directional variorgram panel with their nearest direction, so North
contains everything between -22.5 and 22.5 degrees (North-West to
North-East). After classifying by direction, point pairs are binned by
separation distance class, as is done in the usual omnidirectional case.

In the figure, the partial sill, nugget and model type of the model are
equal to those of the omnidirectional model fitted above; the range
is that in the direction with the largest range (45$^o$), and the
anisotropy ratio, the range in the 135 direction and the range in the
45 direction, estimated ``by eye'' by comparing the 45 and 135 degrees
sample variograms. Gstat does not fit anisotropy parameters automatically.

We do not claim that the model fitted here is ``best'' in some way; in
order to get to a better model we may want to look at more directions,
other directions (e.g. try {\tt alpha = c(22, 67, 112, 157) }), and to
variogram maps (see below). More elaborate approaches may use directions
in three dimensions, and want to further control the direction tolerance
(which may be set such that direction intervals overlap).

For the residual variogram from the linear regression model using
\code{sqrt(dist)} as covariate, the directional dependence is much
less obvious; the fitted model here is the fitted isotropic model
(equal in all directions).

<<fig=FALSE>>=
lznr.dir = variogram(log(zinc)~sqrt(dist), meuse, alpha = c(0, 45, 90, 135))
plot(lznr.dir, lznr.fit, as.table = TRUE)
@

<<fig=TRUE,echo=FALSE>>=
print(plot(lznr.dir, lznr.fit, as.table = TRUE))
@

\section{Variogram maps}

Another means of looking at directional dependence in semivariograms
is obtained by looking at variogram maps. Instead of classifying
point pairs $Z(s)$ and $Z(s+h)$ by direction and distance class {\em
separately}, we can classify them {\em jointly}. If $h=\{x,y\}$ be the
two-dimentional coordinates of the separation vector, in the variogram
map the semivariance contribution of each point pair $(Z(s)-Z(s+h))^2$
is attributed to the grid cell in which $h$ lies. The map is centered
around $(0,0)$, as $h$ is geographical distance rather than geographical
location. Cutoff and width correspond to some extent to map extent
and cell size; the semivariance map is point symmetric around $(0,0)$,
as $\gamma(h)=\gamma(-h)$.

<<fig=FALSE>>=
vgm.map = variogram(log(zinc)~sqrt(dist), meuse, cutoff = 1500, width = 100, 
	map = TRUE)
plot(vgm.map, threshold = 5)
@

<<fig=TRUE,echo=FALSE>>=
print(plot(vgm.map, threshold = 5))
@

The threshold assures that only semivariogram map values based on at
least 5 point pairs are shown, removing too noisy estimation.

% The plot is plagued by one or two extreme values, corresponding to cells
% with very small number of point pairs, which should be removed.

\section{Cross variography}

Fitting a linear model of coregionalization.

<<fig=FALSE>>=
g = gstat(NULL, "log(zn)", log(zinc)~sqrt(dist), meuse)
g = gstat(g, "log(cd)", log(cadmium)~sqrt(dist), meuse)
g = gstat(g, "log(pb)", log(lead)~sqrt(dist), meuse)
g = gstat(g, "log(cu)", log(copper)~sqrt(dist), meuse)
v = variogram(g)
g = gstat(g, model = vgm(1, "Exp", 300, 1), fill.all = TRUE)
g.fit = fit.lmc(v, g)
g.fit
plot(v, g.fit)
vgm.map = variogram(g, cutoff = 1500, width = 100, map = TRUE)
plot(vgm.map, threshold = 5, col.regions = bpy.colors(), xlab = "", ylab = "")
@

<<fig=TRUE,echo=FALSE>>=
print(plot(v, g.fit))
@

<<fig=TRUE,echo=FALSE>>=
print(plot(vgm.map, threshold = 5, col.regions = bpy.colors(), ylab = "", xlab = ""))
@

\section*{References}
\begin{itemize}
% \item Abrahamsen, P., F. Espen Benth, 2001. Kriging with inequality
% constraints.  Mathematical Geology 33 (6), 719--744.

% \item Bivand, R.S., 2003. Approaches to classes for spatial data in R.
% In: K.~Hornik \& F.~Leisch (Eds.), Proceedings of the 3rd International
% Workshop on Distributed Statistical Computing (DSC 2003) March 20--22,
% Vienna, Austria. ISSN 1609-395X; available from [1].
\item Burrough, P.A., R.A. McDonnell, 1998.  Principles of Geographical
Information Systems, 2nd Edition. Oxford University Press.

\item Diggle, P.J., J.A. Tawn, R.A. Moyeed, 1998. Model-based
geostatistics. Applied Statistics 47(3), pp 299-350.

% \item Pebesma, E.J., Wesseling, C.G., 1998. Gstat, a program for
% geostatistical modelling, prediction and simulation. Computers \&
% Geosciences, 24 (1), pp. 17--31.

\item Pebesma, E.J., 2004.  Multivariable geostatistics
in S: the gstat package.  Computers \& Geosciences
\href{http://dx.doi.org/10.1016/j.cageo.2004.03.012}{30: 683-691}.

% \item Ver Hoef, J.M., Cressie, N.A.C, 1993. Multivariable Spatial
% Prediction.  Mathematical Geology, 25 (2), pp. 219--240.

\item Wackernagel, H., 1998. Multivariate Geostatistics; an introduction
with applications, $2^{\mbox{nd}}$ edn., Springer, Berlin, 291 pp.

\end{itemize}

\end{document}

# g = gstat(NULL, "log.zinc", log(zinc)~1, meuse)
# g = gstat(g, "log.zinc.res", log(zinc)~sqrt(dist), meuse)
# lplot(vgm.map[["map"]], c("log.zinc", "log.zinc.res"))

% vim:syntax=tex