Stockton dataset

Contents
1 variograms
1.1 variogram figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 variogram envelopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 directional variogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2
2
3
4

2 estimating parameters
2.1 estimating parameters with variofit()
2.2 likfit() . . . . . . . . . . . . . . . . .
2.3 simple kriging . . . . . . . . . . . . .
2.4 comparing to our original figure . . .

5
5
7
8
9

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3 assuming anisotropy (azimuthal angle = 135 deg)

10
3.1 anisotropic directional variogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 anisotropic log-likelihoods (with loglik.GRF()) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

*All below data are de-trended.

Created on January 21, 2014, with R version 2.14.1 (2011-12-22) on an x86_64-pc-linux-gnu machine.
1

variograms

1.1

variogram figures
classical estimator

modulus estimator
0.8

0.20

0.0000

0.0005

0.0010

0.4

0.10
0.05
0.00

0.2

0.0

semivariance

0.15

0.6

0.0015 0.0000

0.0005

0.0010

0.0015

distance

0.020

0.020

distance

0.015
0.000

0.005

0.005
0.000

0.0005

0.0010

0.0015 0.0000
10

0.0000

distance

08

06

04

02

00

modu us va og am

04
03
02

01

0.0005

0.0010
distance

00

0.010

semivariance

0.010

0.015

0.0015

1.2

variogram envelopes

to test the constant mean model

> set.seed(1)
> env1R <- variog.mc.env(s1R.geo, ob=bin1R)
variog.env: generating 99 simulations by permutating data values
variog.env: computing the empirical variogram for the 99 simulations
variog.env: computing the envelops
> plot(bin1R, env=env1R, main="de-meaned model")
> rm(bin1R,bin2R)

demeaned model

0.020

0.010
0.005
0.000

semivariance

0.015

0.0000

0.0005

0.0010
distance

0.0015

1.3

directional variogram

0.025

> vario4R <- variog4(s1R.geo, max.dist = maxD)

> plot(vario4R)

0.015
0.010
0.005
0.000

semivariance

0.020

0
45
90
135

0.0000

0.0002

0.0004

0.0006

0.0008

distance

0.0010

0.0012

0.0014

estimating parameters

2.1

estimating parameters with variofit()

> pSill <- .00665

> phi
<- .0001
> bin
<- variog(s1R.geo, uvec=seq(0,maxD,l=10),
+
max.dist=maxD)
variog: computing omnidirectional variogram
> variogk15 <- variofit(bin, cov.model='mat',
+
ini.cov.pars=c(pSill, phi),
+
kappa=1.5,
+
nugget=.014,
+
fix.nugget=FALSE)
variofit: covariance model used is matern
variofit: weights used: npairs
variofit: minimisation function used: optim
> summary(variogk15)
$pmethod
[1] "WLS (weighted least squares)"
$cov.model
[1] "matern"
$spatial.component
sigmasq
phi
7.277033e-03 9.361362e-05
$spatial.component.extra
kappa
1.5
$nugget.component
tausq
0.01512544
$fix.nugget
[1] FALSE
$fix.kappa
[1] TRUE
$practicalRange
[1] 0.0004660384
$sum.of.squares
value
0.1846355
$estimated.pars
tausq
sigmasq
phi
1.512544e-02 7.277033e-03 9.361362e-05
$weights
[1] "npairs"
$call
5

variofit(vario = bin, ini.cov.pars = c(pSill, phi), cov.model = "mat",

fix.nugget = FALSE, nugget = 0.014, kappa = 1.5)
attr(,"class")
[1] "summary.variomodel"
> rm(bin,phi,pSill)

We see here that for = 1.5, variofit() suggests the Matern correlation function
1

(u) =
2( 32 )

u
9.4 105

 32


K

u
9.4 105

with 2 = 0.01512. We can quickly plot that against the empirical variogram:

0.020

0.010

0.000

0.005

semivariance

0.015

0.0000

0.0005

0.0010
distance

Looks like a pretty good fit.

6

0.0015

2.2

likfit()

> mlfit <- likfit(s1R.geo, cov.model='mat',

+
ini=c(7.277033e-03, 9.361362e-05),
+
kap=1.5, nug=1.512544e-02)
--------------------------------------------------------------likfit: likelihood maximisation using the function optim.
likfit: Use control() to pass additional
arguments for the maximisation function.
For further details see documentation for optim.
likfit: It is highly advisable to run this function several
times with different initial values for the parameters.
likfit: WARNING: This step can be time demanding!
--------------------------------------------------------------likfit: end of numerical maximisation.
> mlfit
likfit: estimated model parameters:
beta
tausq sigmasq
phi
"0.0060" "0.0148" "0.0071" "0.0001"
Practical Range with cor=0.05 for asymptotic range: 0.0004660384
likfit: maximised log-likelihood = 774.7
> #summary(mlfit)

2.3
>
+
>
>
>

simple kriging

grid

<- expand.grid(seq(min(s1R$x),max(s1R$x),l=100),
seq(min(s1R$y),max(s1R$y),l=100))
kc
<- krige.control(obj.model = mlfit)
set.seed(1)
pred
<- krige.conv(s1R.geo, loc=grid, krige=kc)

krige.conv: model with constant mean

krige.conv: Kriging performed using global neighbourhood

0.7180
0.7175
0.7170

Y Coord

0.7185

0.7190

> image1 <-image(pred, loc = grid, col = gray(seq(1, 0, l=21)))

1e03

5e04

0e+00
X Coord

5e04

1e03

2.4
>
>
>
>
+
>
+
+
+
>

comparing to our original figure

require(ggmap)
paste0 <- function(..., collapse = NULL) paste(..., sep = "", collapse = collapse)
cLat=(37.95564 + 38.05752)/2; cLon=(-121.3681 -121.2502)/2
stocktonMap_toner <- qmap(c(cLon,cLat), zoom=12, source = "stamen",
maptype = "toner")
stocktonMap_toner +
geom_point(aes(x=long, y=lat, colour=residual, size=residual), data=s) +
scale_colour_gradient(low = "aquamarine1", high = "navyblue") +
guides(colour = guide_legend())
rm(s)

residual

0.25

0.00

0.25

0.50

3
3.1

assuming anisotropy (azimuthal angle = 135 deg)

anisotropic directional variogram
3
4

radians), and

0.020

0.025

0.030

0
45
90
135

0.000

0.000

0.005

0.005

0.010

0.010

semivariance

0.015

0.020

0
45
90
135

0.015

0.025

0.035

coords.aniso() takes two arguments: anisotropy angle psiA (which here is 135 degrees =
anisotropy ratio psiR which Ive varied in the below: 1 (no change), 1.5, 2, 5:

0.0012

0.0000

0.025

0.030

distance

0.0004
0
45
90
135

0.00

0.000

0.005

0.01

0.010

0.02

semivariance

0.03

0.04

0
45
90
135

0.0008

0.020

0.0004

0.015

0.0000

3.2

anisotropic log-likelihoods (with loglik.GRF())

Looking at the log likelihoods for each of the above:

> # log likelihood for basic model
> loglik.GRF(s1R.geo, cov.model='mat',
+
cov.pars=c(7.277033e-03, 9.361362e-05),
+
kap=1.5, nug=1.512544e-02,
+
psiA=0, psiR=1)
[1] 774.5323
> # log likelihood for model, psiA=(3*pi)/4, psiR=1.5
> loglik.GRF(s1R.geo, cov.model='mat',

10

0.0008
distance

0.0012

+
+
+

cov.pars=c(7.277033e-03, 9.361362e-05),
kap=1.5, nug=1.512544e-02,
psiA=(3*pi)/4, psiR=1.5)

[1] 766.9429
> # log likelihood for model, psiA=(3*pi)/4, psiR=2
> loglik.GRF(s1R.geo, cov.model='mat',
+
cov.pars=c(7.277033e-03, 9.361362e-05),
+
kap=1.5, nug=1.512544e-02,
+
psiA=(3*pi)/4, psiR=2)
[1] 758.4905
> # log likelihood for model, psiA=(3*pi)/4, psiR=5
> loglik.GRF(s1R.geo, cov.model='mat',
+
cov.pars=c(7.277033e-03, 9.361362e-05),
+
kap=1.5, nug=1.512544e-02,
+
psiA=(3*pi)/4, psiR=5)
[1] 724.8255
All these numbers seem very similar.
Given the similarity, I wanted to check with you both before chopping the points up into quadrants and rerunning. (I also ran out of time.) :)

11