Revision 328f151e07864e976578f9134025cbfdb72eb2f3 authored by Doug Nychka on 19 July 2013, 23:45:29 UTC, committed by cran-robot on 19 July 2013, 23:45:29 UTC
1 parent a968cb7
Krig.Rd
% fields, Tools for spatial data
% Copyright 2004-2011, Institute for Mathematics Applied Geosciences
% University Corporation for Atmospheric Research
% Licensed under the GPL -- www.gpl.org/licenses/gpl.html
\name{Krig}
\alias{Krig}
\alias{resid.Krig}
\alias{fitted.Krig}
\alias{coef.Krig}
\title{
Kriging surface estimate
}
\description{
Fits a surface to irregularly spaced data. The Kriging model assumes
that the unknown function is a realization of a Gaussian
random spatial processes. The assumed model is additive
Y = P(x) + Z(X) + e, where P is a low order polynomial and Z is a
mean zero,
Gaussian stochastic process with a
covariance that is unknown up to a scale constant. The main advantages of
this function are the flexibility in specifying the covariance as an
R language function and also the supporting functions plot, predict,
predict.se, surface for
subsequent analysis. Krig also supports a correlation model where the mean
and marginal variances are supplied.
}
\usage{
Krig(
x, Y, cov.function = "stationary.cov", lambda = NA, df
= NA, GCV=FALSE, Z = NULL, cost = 1, knots = NA, weights = NULL,
m = 2, nstep.cv = 200, scale.type = "user", x.center =
rep(0, ncol(x)), x.scale = rep(1, ncol(x)), rho = NA,
sigma2 = NA, method = "GCV", verbose = FALSE, mean.obj
= NA, sd.obj = NA, null.function =
"Krig.null.function", wght.function = NULL, offset =
0, outputcall = NULL, na.rm = TRUE, cov.args = NULL,
chol.args = NULL, null.args = NULL, wght.args = NULL,
W = NULL, give.warnings = TRUE, ...)
\method{fitted}{Krig}(object,...)
\method{coef}{Krig}(object,...)
resid.Krig(object,...)
}
\arguments{
\item{x}{
Matrix of independent variables. These could the locations for spatial
data or the indepedent variables in a regression.
}
\item{Y}{
Vector of dependent variables. These are the values of the surface
(perhaps with measurement error) at the locations or the dependent
response in a regression.
}
\item{cov.function}{
Covariance function for data in the form of an R function (see
Exp.simple.cov as an example).
Default assumes that correlation is an exponential function of distance.
See also \code{stationary.cov} for more general choice of covariance
shapes. \code{exponential.cov} will be faster if only the exponential
covariance form is needed.
}
\item{Z}{
A vector of matrix of covariates to be include in the fixed part of the
model. If NULL (default) no addtional covariates are included.}
\item{lambda}{
Smoothing parameter that is the ratio of the error variance (sigma**2)
to the scale parameter of the
covariance function (rho). If omitted this is estimated by GCV ( see
method below).
}
\item{df}{
The effective number of parameters for the fitted surface. Conversely,
N- df, where N is the total number of observations is the degrees of
freedom associated with the residuals.
This is an alternative to specifying lambda and much more interpretable.
NOTE: GCV argument defaults to TRUE if this argument is used.
}
\item{GCV}{ If TRUE matrix decompositions are done to allow estimating
lambda by GCV or REML and specifying smoothness by the effective degrees of
freedom. So the GCV switch does more than just supply a GCV estimate. Also if
lambda or df are passed the estimate will be evaluated at those values,
not at the GCV/REML estimates of lambda.
If FALSE Kriging estimate is found under a fixed lambda model. }
\item{cost}{
Cost value used in GCV criterion. Corresponds to a penalty for
increased number of parameters. The default is 1.0 and corresponds to the
usual GCV function.
}
\item{knots}{
A matrix of locations similar to x. These can define an alternative set of
basis functions for representing the estimate. One choice may be a
space-filling subset of the original x locations, thinning out the
design where locations cluster. The
default is to put a "knot" at all unique locations. (See details.)
}
\item{weights}{
Weights are proportional to the reciprocal variance of the measurement
error. The default is equal weighting i.e. vector of unit weights.
}
\item{m}{
A polynomial function of degree (m-1) will be
included in the model as the drift (or spatial trend) component.
The "m" notation is from thin-plate splines where m is the
derivative in the penalty function. With m=2 as the default a linear
model in the locations will be fit a fixed part of the model.
}
\item{nstep.cv}{
Number of grid points for the coarse grid search to
minimize the GCV RMLE and other related criteria for finding lambda,
the smoothing parameter.
Default is 200, fairly large to avoid some cases of closely spaced local
minima. Evaluations of the GCV and related objective functions are cheap
given the matrix decompositions described below.
}
\item{scale.type}{
This is a character string among: "range", "unit.sd", "user", "unscaled".
The independent variables and knots are scaled to the specified scale.type.
By default no scaling is done. This usuall makes sense for spatial locations.
Scale type of "range" scales the data to the interval (0,1) by forming
(x-min(x))/range(x) for each x. Scale type of "unit.sd"
Scale type of "user" allows specification of an x.center and x.scale by the
user. The default for "user" is mean 0 and standard deviation 1. Scale
type of "unscaled" does not scale the data.
}
\item{x.center}{
Centering values to be subtracted from each column of the x matrix.
}
\item{x.scale}{
Scale values that are divided into each column after centering.
}
\item{rho}{
Scale factor for covariance.
}
\item{sigma2}{
Variance of the errors, often called the nugget variance. If weights are
specified then the error variance is sigma2 divided by weights.
Note that lambda is defined as the ratio sigma2/rho.
}
\item{method}{
Determines what "smoothing" parameter should be used. The default
is to estimate standard GCV
Other choices are: GCV.model, GCV.one, RMSE, pure error and REML. The
differences are explained below.
}
\item{verbose}{
If true will print out all kinds of intermediate stuff. Default is false,
of course as this is used mainly for debugging.
}
\item{mean.obj}{
Object to predict the mean of the spatial process. This used in when
fitting a correlation model with varying spatial means and varying
marginal variances. (See details.)
}
\item{sd.obj}{
Object to predict the marginal standard deviation of the spatial process.
}
\item{null.function}{
An R function that creates the matrices for the null space model.
The default is fields.mkpoly, an R function that creates a polynomial
regression matrix with all terms up to degree m-1. (See Details)
}
\item{wght.function}{
An R function that creates a weights matrix to the observations.
This is only needed if the weight matirx has off diagonal elements.
The default is NULL indicating that the weight matrix is a diagonal, based
on the weights argument. (See details)
}
\item{offset}{
The offset to be used in the GCV criterion. Default is 0. This would be
used when Krig is part of a backfitting algorithm and the offset is
other model degrees of freedom from other regression components.
}
\item{outputcall}{
If NULL the output object will have a \$call argument based on this call.
If no NULL the output call will have whatever is passed. This is kludge
for the Tps function so that it return a Krig object but have the right
call argument. Sorry no one promised that fields would be pretty.
}
\item{cov.args}{
A list with the arguments to call the covariance function. (in addition to the locations)
}
\item{na.rm}{If TRUE NAs will be removed from the \code{y} vector and the
corresponding rows of \code{x} -- with a warning.
If FALSE Krig will just stop with a message. Once NAs are removed all
subsequent analysis in fields does not use those data. }
\item{chol.args}{
Arguments to be passed to the cholesky decomposition in Krig.engine.fixed.
The default if NULL, assigned at the top level of this function, is
list( pivot=FALSE). This argument is useful when working with
the sparse matrix package. }
\item{wght.args}{
Optional arguments to be passed to the weight function (wght.function)
used to create the observation weight matrix.}
\item{W}{The explicit observatoin weight matrix.}
\item{null.args}{ Extra arguments for the null space function
\code{null.function}.
If \code{fields.mkpoly} is passed as
\code{null.function} then this is set to a list with the
value of \code{m}. So the default is use a polynomial of degree m-1 for the
null space (fixed part) of the model.
}
\item{give.warnings}{ If TRUE warnings are given in gcv grid search limits.
If FALSE warnings are not given. Best to leave this TRUE!}
\item{\dots}{
Optional arguments that appear are assumed to be additional arguments
to the covariance function. Or are included in methods functions (resid,
fitted, coef) as a
required argument.}
\item{object}{ A Krig object}
}
\value{
A object of class Krig. This includes the predicted values in
fitted.values and the residuals in residuals. The results of the grid
search to minimize the generalized cross validation function are
returned in gcv.grid.
The coef.Krig function only returns the coefficients, "d", associated with the
fixed part of the model (also known as the null space or spatial drift).
\item{call}{
Call to the function
}
\item{y}{
Vector of dependent variables.
}
\item{x}{
Matrix of independent variables.
}
\item{weights}{
Vector of weights.
}
\item{knots}{
Locations used to define the basis functions.
}
\item{transform}{
List of components used in centering and scaling data.
}
\item{np}{
Total number of parameters in the model.
}
\item{nt}{
Number of parameters in the null space.
}
\item{matrices}{
List of matrices from the decompositions (D, G, u, X, qr.T).
}
\item{gcv.grid}{
Matrix of values from the GCV grid search. The first column
is the grid of lambda values used in the search, the second column
is the trace of the A matrix, the third column is the GCV values and
the fourth column is the estimated value of sigma conditional on the vlaue
of lambda.
}
\item{lambda.est}{
A table of estimated smoothing parameters with corresponding degrees
of freedom and estimates of sigma found by different methods.
}
\item{cost}{
Cost value used in GCV criterion.
}
\item{m}{
Order of the polynomial space: highest degree polynomial is (m-1).
This is a fixed part of the surface often referred to as the drift
or spatial trend.
}
\item{eff.df}{
Effective degrees of freedom of the model.
}
\item{fitted.values}{
Predicted values from the fit.
}
\item{residuals}{
Residuals from the fit.
}
\item{lambda}{
Value of the smoothing parameter used in the fit.
Lambda is defined as sigma**2/rho. See discussion in details.
}
\item{yname}{
Name of the response.
}
\item{cov.function}{
Covariance function of the model.
}
\item{beta}{
Estimated coefficients in the ridge regression format
}
\item{d}{
Estimated coefficients for the polynomial basis functions that span the
null space
}
\item{fitted.values.null}{
Fitted values for just the polynomial part of the estimate
}
\item{trace}{
Effective number of parameters in model.
}
\item{c}{
Estimated coefficients for the basis functions derived from the
covariance.
}
\item{coefficients}{
Same as the beta vector.
}
\item{just.solve}{
Logical describing if the data has been interpolated using the basis
functions.
}
\item{shat}{
Estimated standard deviation of the measurement error (nugget effect).
}
\item{sigma2}{
Estimated variance of the measurement error (shat**2).
}
\item{rho}{
Scale factor for covariance. COV(h(x),h(x\code{)) = rho*cov.function(x,x})
If the covariance is actually a
correlation function then rho is also the "sill".
}
\item{mean.var}{
Normalization of the covariance function used to find rho.
}
\item{best.model}{
Vector containing the value of lambda, the estimated variance of the
measurement error and the scale factor for covariance used in the fit.
}
}
\details{
This function produces a object of class Krig. With this object it is
easy to subsequently predict with this fitted surface, find standard
errors, alter the y data ( but not x), etc.
The Kriging model is: Y.k= f(x.k) = P(x.k) + Z(x.k) + e.k
where ".k" means subscripted by k, Y is the dependent variable observed
at location x.k, P is a low order polynomial, Z is a mean zero, Gaussian
field with covariance function K and e is assumed to be independent
normal errors. The estimated surface is the best linear unbiased
estimate (BLUE) of f(x)= P(x) + Z(x) given the observed data. For this
estimate K, is taken to be rho*cov.function and the errors have variance
sigma**2. In more conventional geostatistical terms rho is the "sill" if
the covariance function is actually a correlation function and sigma**2
is the nugget variance or measure error variance (the two are confounded
in this model.) If the weights are given then the variance of e.k is
sigma**2/ weights.k . In the case that the weights are specified as a
matrix, W, using the wght.function option then the assumed covariance
matrix for the errors is sigma**2 Wi, where Wi is the inverse of W. It
is straightforward to show that the estimate of f only depends on sigma
and rho through the ratio lambda = sigma**2/ rho. This parameter, termed
the smoothing parameter plays a central role in the statistical
computations within \code{Krig}. See also the help for thin plate
splines, (\code{Tps}) to get another perspective on the smoothing
parameter.
This function also supports a modest extension of the Generalized
Kriging model to include other covariates as fixed regression type
components. In matrix form Y = Zb + F + E where Z is a matrix of
covariates and b a fixed parameter vector, F the vector of function values at
the observations and E a vector of errors. The The \code{Z} argument in
the function is the way to specify this additional component.
If the parameters rho and sigma2 are omitted in the call, then they are
estimated in the following way. If lambda is given, then sigma2 is
estimated from the residual sum of squares divided by the degrees of
freedom associated with the residuals. Rho is found as the difference
between the sums of squares of the predicted values having subtracted off
the polynomial part and sigma2. These estimates are the MLE's under Gaussian
assumptions on the process and errors.
If lambda is also omitted is it estimated
from the data using a variety of approaches and then the values for sigma
and rho are found in the same way from the estimated lambda.
A useful extension of a stationary correlation to a nonstationary
covariance is what we term a correlation model.
If mean and marginal standard deviation objects are included in the call.
Then the observed data is standardized based on these functions. The
spatial process is then estimated with respect to the standardized scale.
However for predictions and standard errors the mean and standard
deviation surfaces are used to produce results in the original scale of
the observations.
The GCV function has several alternative definitions when replicate
observations are present or if one uses a reduced set knots. Here are the
choices based on the method argument:
GCV: leave-one-out GCV. But if
there are replicates it is leave one group out. (Wendy and Doug prefer
this one.)
GCV.one: Really leave-one-out GCV even if there are replicate
points. This what the old tps function used in FUNFITS.
rmse: Match the estimate of sigma**2 to a external value ( called rmse)
pure error: Match the estimate of sigma**2 to the estimate based on
replicated data (pure error estimate in ANOVA language).
GCV.model:
Only considers the residual sums of squares explained by the basis
functions.
REML:
The process and errors are assumed to the Gaussian and the likelihood is
concentrated (or profiled) with respect to lambda. The MLE of lambda is
found from this criterion. Restricted means that the likelihood is formed from a linear transformation of the observations that is orthogonal to the
column space of P(x).
WARNING: The covariance functions often have a nonlinear parameter(s) that
often control the strength of the correlations as a function of separation,
usually referred to as the range parameter. This parameter must be
specified in the call to Krig and will not be estimated.
}
\section{References}{
See "Additive Models" by Hastie and Tibshirani, "Spatial Statistics" by
Cressie and the FIELDS manual.
}
\seealso{
summary.Krig, predict.Krig, predict.se.Krig, predict.surface.se,
predict.surface, plot.Krig,
surface.Krig
}
\examples{
# a 2-d example
# fitting a surface to ozone
# measurements. Exponential covariance, range parameter is 20 (in miles)
fit <- Krig(ozone$x, ozone$y, theta=20)
summary( fit) # summary of fit
set.panel( 2,2)
plot(fit) # four diagnostic plots of fit
set.panel()
surface( fit, type="C") # look at the surface
# predict at data
predict( fit)
# predict using 7.5 effective degrees of freedom:
predict( fit, df=7.5)
# predict on a grid ( grid chosen here by defaults)
out<- predict.surface( fit)
surface( out, type="C") # option "C" our favorite
# predict at arbitrary points (10,-10) and (20, 15)
xnew<- rbind( c( 10, -10), c( 20, 15))
predict( fit, xnew)
# standard errors of prediction based on covariance model.
predict.se( fit, xnew)
# surface of standard errors on a default grid
predict.surface.se( fit)-> out.p # this takes some time!
surface( out.p, type="C")
points( fit$x)
\dontrun{
# Using another stationary covariance.
# smoothness is the shape parameter for the Matern.
fit <- Krig(ozone$x, ozone$y, Covariance="Matern", theta=10, smoothness=1.0)
summary( fit)
#
# Roll your own: creating very simple user defined Gaussian covariance
#
test.cov <- function(x1,x2,theta,marginal=FALSE,C=NA){
# return marginal variance
if( marginal) { return(rep( 1, nrow( x1)))}
# find cross covariance matrix
temp<- exp(-(rdist(x1,x2)/theta)**2)
if( is.na(C[1])){
return( temp)}
else{
return( temp\%*\%C)}
}
#
# use this and put in quadratic polynomial fixed function
fit.flame<- Krig(flame$x, flame$y, cov.function="test.cov", m=3, theta=.5)
#
# note how range parameter is passed to Krig.
# BTW: GCV indicates an interpolating model (nugget variance is zero)
# This is the content of the warning message.
# take a look ...
surface(fit.flame, type="I")
}
#
# Thin plate spline fit to ozone data using the radial
# basis function as a generalized covariance function
#
# p=2 is the power in the radial basis function (with a log term added for
# even dimensions)
# If m is the degree of derivative in penalty then p=2m-d
# where d is the dimension of x. p must be greater than 0.
# In the example below p = 2*2 - 2 = 2
#
out<- Krig( ozone$x, ozone$y,cov.function="Rad.cov",
m=2,p=2,scale.type="range")
# See also the Fields function Tps
# out should be identical to Tps( ozone$x, ozone$y)
#
# A Knot example
data(ozone2)
y16<- ozone2$y[16,]
# there are some missing values -- remove them
good<- !is.na( y16)
y<- y16[good]
x<- ozone2$lon.lat[ good,]
#
# the knots can be arbitrary but just for fun find them with a space
# filling design. Here we select 50 from the full set of 147 points
#
xknots<- cover.design( x, 50, num.nn= 75)$design # select 50 knot points
out<- Krig( x, y, knots=xknots, cov.function="Exp.cov", theta=300)
summary( out)
# note that that trA found by GCV is around 17 so 50>17 knots may be a
# reasonable approximation to the full estimator.
#
\dontrun{
# the plot
surface( out, type="C")
US( add=TRUE)
points( x, col=2)
points( xknots, cex=2, pch="O")
}
\dontrun{
## A quick way to deal with too much data if you intend to smooth it anyway
## Discretize the locations to a grid, then apply Krig
## to the discretized locations:
##
RM.approx<- as.image(RMprecip$y, x=RMprecip$x, nx=20, ny=20)
# take a look:
image.plot( RM.approx)
# discretized data (observations averaged if in the same grid box)
# 336 locations -- down form the full 806
# convert the image format to locations, obs and weight vectors
yd<- RM.approx$z[RM.approx$ind]
weights<- RM.approx$weights[RM.approx$ind] # takes into account averaging
xd<- RM.approx$xd
obj<- Krig( xd, yd, weights=weights, theta=4)
# compare to the full fit:
# Krig( RMprecip$x, RMprecip$y, theta=4)
}
\dontrun{
# A correlation model example
# fit krig surface using a mean and sd function to standardize
# first get stats from 1987 summer Midwest O3 data set
data(ozone2)
stats.o3<- stats( ozone2$y)
mean.o3<- Tps( ozone2$lon.lat, c( stats.o3[2,]))
sd.o3<- Tps( ozone2$lon.lat, c( stats.o3[3,]))
#
# Now use these to fit particular day ( day 16)
# and use great circle distance
fit<- Krig( ozone2$lon.lat, ozone2$y[16,],
theta=350, mean.obj=mean.o3, sd.obj=sd.o3,
Covariance="Matern", Distance="rdist.earth",
smoothness=1.0,
na.rm=TRUE) #
# the finale
surface( fit, type="I")
US( add=TRUE)
points( fit$x)
title("Estimated ozone surface")
}
\dontrun{
#
#
# explore some different values for the range and lambda using REML
theta <- seq( 100,500,,40)
PLL<- matrix( NA, 40,80)
# the loop
for( k in 1:40){
# call to Krig with different ranges
# also turn off warnings for GCV search
# to avoid lots of messages. (not recommended in general!)
PLL[k,]<- Krig( ozone2$lon.lat,ozone2$y[16,],
cov.function="stationary.cov",
theta=theta[k], mean.obj=mean.o3, sd.obj=sd.o3,
Covariance="Matern",smoothness=.5,
Distance="rdist.earth", nstep.cv=80,
give.warnings=FALSE, na.rm=TRUE)$gcv.grid[,7]
#
# gcv.grid is the grid search output from
# the optimization for estimating different estimates for lambda including
# REML
# default grid is equally spaced in eff.df scale ( and should the same across theta)
# here
}
# get lambda grid from looping
k<- 1
lam<- Krig( ozone2$lon.lat,ozone2$y[16,],
cov.function="stationary.cov",
theta=theta[k], mean.obj=mean.o3, sd.obj=sd.o3,
Covariance="Matern",smoothness=.5,
Distance="rdist.earth", nstep.cv=80,
give.warnings=FALSE, na.rm=TRUE)$gcv.grid[,1]
# see the 2 column of $gcv.grid to get the effective degress of freedom.
contour( theta,log(lam) , PLL)
}
}
\keyword{spatial}
% docclass is function
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...