matern: Evaluate the Matern correlation function

View source: R/matern.R

maternR Documentation

Evaluate the Matern correlation function

Description

Returns the Matern covariance for the distances supplied.

Usage

	matern( x, param=c(range=1, variance=1, shape=1),
	type=c('variance','cholesky','precision', 'inverseCholesky'),
	y=NULL)
	## S3 method for class 'SpatVector'
matern(x,  param,
	type=c('variance','cholesky','precision', 'inverseCholesky'),
	y=NULL)
	## Default S3 method:
matern( x, param,
	type=c('variance','cholesky','precision', 'inverseCholesky'),
	y=NULL)
	## S3 method for class 'dist'
matern( x, param,
	type=c('variance','cholesky','precision', 'inverseCholesky'),
	y=NULL)
	## S3 method for class 'SpatRaster'
matern( x,  param,
	type=c('variance','cholesky','precision', 'inverseCholesky'),
	y=NULL)
	fillParam(param)

Arguments

x

A vector or matrix of distances, or SpatRaster or SpatVector of locations, see Details below.

param

A vector of named model parameters with, at a minimum names range and shape (see Details), and optionally variance (defaults to 1) and nugget (defaults to zero). For Geometric Anisotropy add anisoRatio and either anisoAngleDegrees or anisoAngleRadians

type

specifies if the variance matrix, the Cholesky decomposition of the variance matrix, the precision matrix, or the inverse of the Cholesky L matrix is returned.

y

Covariance is calculated for the distance between locations in x and y. If y=NULL, covariance of x with itself is produced. However, if x is a matrix or vector it is assumed to be a set of distances and y is ignored.

Details

The formula for the Matern correlation function is

M(x) = \frac{variance}{\Gamma(shape)} 2^{1-shape} \left( \frac{ x \sqrt{8 shape} }{range} \right)^{shape} besselK(x \sqrt{8 shape}/ range, shape)

The range argument is sqrt(8*shape)*phi.geoR, sqrt(8*shape)*scale.whittle.RandomFields, and 2*scale.matern.RandomFields.

Geometric anisotropy is only available when x is a SpatRaster or SpatVector. The parameter 'anisoAngle' refers to rotation of the coordinates anti-clockwise by the specified amount prior to calculating distances, which has the effect that the contours of the correlation function are rotated clockwise by this amount. anisoRatio is the amount the Y coordinates are divided by by post rotation prior to calculating distances. A large value of anisoRatio makes the Y coordinates smaller and increases the correlation in the Y direction.

When x or y are rasters, cells are indexed row-wise starting at the top left.

Value

When x is a vector or matrix or object of class dist, a vector or matrix of covariances is returned. With x being SpatVector , y must also be SpatVector and a matrix of correlations between x and y is returned. When x is a Raster, and y is a single location a Raster of covariances between each pixel centre of x and y is returned.

Examples

param=c(shape=2.5,range=1,variance=1)
u=seq(0,4,len=200)
uscale = sqrt(8*param['shape'])* u / param['range']

theMaterns = cbind(
	dist=u, 
	manual=	param['variance']* 2^(1- param['shape']) * 
			( 1/gamma(param['shape'])  )  * 
			uscale^param['shape'] * besselK(uscale , param['shape']),
	geostatsp=geostatsp::matern(u, param=param)
)
head(theMaterns)
matplot(theMaterns[,'dist'], 
	theMaterns[,c('manual','geostatsp')],
	col=c('red','blue'), type='l', 
	xlab='dist', ylab='var')
legend('topright', fill=c('red','blue'),
	legend=c('manual','geostatsp'))



# example with raster
myraster = rast(nrows=40,ncols=60,extent=ext(-3, 3,-2,2))
param = c(range=2, shape=2,	anisoRatio=2, 
	anisoAngleDegrees=-25,variance=20)

# plot correlation of each cell with the origin
myMatern = matern(myraster, y=c(0,0), param=param)


plot(myMatern, main="anisortopic matern")


# correlation matrix for all cells with each other
myraster = rast(nrows=4,ncols=6,extent = ext(-3, 3, -2, 2))
myMatern = matern(myraster, param=c(range=2, shape=2))
dim(myMatern)

# plot the cell ID's
values(myraster) = seq(1, ncell(myraster))
mydf = as.data.frame(myraster, xy=TRUE)
plot(mydf$x, mydf$y, type='n', main="cell ID's")
text(mydf$x, mydf$y, mydf$lyr.1)
# correlation between bottom-right cell and top right cell is
myMatern[6,24]

# example with points
mypoints = vect(
	cbind(runif(8), runif(8))
	)
# variance matrix from points
m1=matern(mypoints, param=c(range=2,shape=1.4,variance=4,nugget=1))
# cholesky of variance from distances
c2=matern(dist(crds(mypoints)), param=c(range=2,shape=1.4,variance=4,nugget=1),type='cholesky')

# check it's correct
quantile(as.vector(m1- tcrossprod(c2)))

# example with vector of distances
range=3
distVec = seq(0, 2*range, len=100)
shapeSeq = c(0.5, 1, 2,20)
theCov = NULL
for(D in shapeSeq) {
	theCov = cbind(theCov, matern(distVec, param=c(range=range, shape=D)))
}
matplot(distVec, theCov, type='l', lty=1, xlab='distance', ylab='correlation',
	main="matern correlations")
legend("right", fill=1:length(shapeSeq), legend=shapeSeq,title='shape')
# exponential

distVec2 = seq(0, max(distVec), len=20)
points(distVec2, exp(-2*(distVec2/range)),cex=1.5, pch=5)
# gaussian
points(distVec2, exp(-2*(distVec2/range)^2), col='blue',cex=1.5, pch=5)
legend("bottomleft", pch=5, col=c('black','blue'), legend=c('exp','gau'))

# comparing to geoR and RandomFields

if (requireNamespace("RandomFields", quietly = TRUE) &
requireNamespace("geoR", quietly = TRUE) 
) { 

covGeoR = covRandomFields = NULL

for(D in shapeSeq) {
	covGeoR = cbind(covGeoR, 
		geoR::matern(distVec, phi=range/sqrt(8*D), kappa=D))
	covRandomFields = cbind(covRandomFields,
		RandomFields::RFcov(x=distVec, 
		model=RandomFields::RMmatern(nu=D, var=1,
				scale=range/2) ))
}



matpoints(distVec, covGeoR, cex=0.5, pch=1)
matpoints(distVec, covRandomFields, cex=0.5, pch=2)

legend("topright", lty=c(1,NA,NA), pch=c(NA, 1, 2), 
	legend=c("geostatsp","geoR","RandomFields"))
}
 

geostatsp documentation built on March 19, 2024, 3:08 a.m.