Scribd
Upload
31 views

Gebhardt+Gebhardt-Bayesian Methods in Geostatistics

This document discusses using Bayesian methods to incorporate prior knowledge about trend parameters into kriging estimation and monitoring network design in R. Specifically, it describes implementing Bayesian kriging which accounts for prior distributions of trend parameters through modified covariance terms. It also allows for optimal monitoring network design by minimizing prediction variance from kriging estimates using criteria like G-optimality and I-optimality. An example demonstrating the R libraries for these Bayesian geostatistical methods is provided.

Uploaded by

Kenneth Ugalde
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
31 views

Gebhardt+Gebhardt-Bayesian Methods in Geostatistics

This document discusses using Bayesian methods to incorporate prior knowledge about trend parameters into kriging estimation and monitoring network design in R. Specifically, it describes implementing Bayesian kriging which accounts for prior distributions of trend parameters through modified covariance terms. It also allows for optimal monitoring network design by minimizing prediction variance from kriging estimates using criteria like G-optimality and I-optimality. An example demonstrating the R libraries for these Bayesian geostatistical methods is provided.

Uploaded by

Kenneth Ugalde
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Bayesian Methods in Geostatistics:

Using prior knowledge about trend parameters for


Kriging and Design of Experiments
Albrecht Gebhardt∗ Claudia Gebhardt†
[email protected] [email protected]
Submitted to useR! 2004

Abstract
This paper covers an approach to incorporate prior knowledge about trend
parameters into Kriging estimation. The implementation in R covers predic-
tion as well as monitoring network design.

1 Introduction
Kriging, the central method for spatial prediction of a regionalised variable Z(x), x ∈
Rd , is based on some prerequisites: translation invariant covariance structure (sec-
ond order stationarity)

Cov (Z(x), Z(x + h)) = C(h), h ∈ Rd (1)

and knowledge about an underlying trend. Depending on the existence or absence


of a trend function one has to choose between Ordinary or Universal Kriging. Ad-
ditional prior knowledge about the trend parameter can be incorporated into the
prediction using a Bayesian approach. This approach has been implemented in R
for prediction and monitoring network design purposes.

2 Bayesian Kriging
Given a data set Z = (Z(xi ))i=1,...,n and using a linear trend model

Z(x) = f (x)> θ + ε(x), θ ∈ Θ ⊆ Rr (2)

Universal Kriging takes a weighted average Ẑ(x0 ) = w> Z of the data as estimator
at location x0 . It assumes unbiasedness E Ẑ(x0 ) = E Z(x0 ) which leads to the so
called universality constraint F> w = f 0 . Choosing the weights w which minimise
the variance of the prediction yields finally the Universal Kriging estimator

Ẑ(x0 ) = c>
0C
−1
(Z − Fµ̂) + f >
0
θ̂ (3)

using the notations f 0 = f (x0 ), F = (f (x1 ), f (x2 ), . . . , f (xn ))> , C ij = C(xi −xj )


and (c0 )i = C(xi −x0 ) i, j = 1, . . . , n where θ̂ = (F> C−1 F)−1 F> C−1 Z corresponds
to the generalised least squares estimator of θ.
∗ University of Klagenfurt, Institute of Mathematics, Austria
† University of Klagenfurt, University Library, Austria

1
Bayesian Kriging requires knowledge about the prior distribution of the trend
parameter θ. This approach (see Omre (1987)) makes only use of the first two
moments E θ = µ and Cov θ = Φ. In contrast to Universal Kriging the condition
of unbiasedness gets now thrown away in favour of a modified Bayesian Kriging
estimator involving a bias component w0

Ẑ(x0 ) = w> Z + w0 . (4)

Again the weights are chosen to minimise the prediction variance. The solution of
the corresponding Lagrange system is

Ẑ(x0 ) = c̃>
0 C̃
−1
(Z − Fµ) + f >
0
µ (5)

which involves the following modified covariance terms

C̃0 = C0 + f >
0
Φ f0 c̃0 = c0 + F Φ f 0 C̃ = C + F Φ F> . (6)

3 Monitoring network design


Monitoring networks evolve over time. After starting with an initial set of measure-
ment sites it is a common task to expand (or shrink) this network in an optimal way.
Kriging does not only return estimated values but also the associated prediction
variance. As this prediction variance only depends on the (planned) measurement
locations it is possible to use G- and I-optimality criteria minimising maximum or
mean prediction variance to choose an optimal subset fo a given set of candidate
measurement points.

4 Implementation Details and Example


The implementation (see Gebhardt (2003)) is primarily based on the sgeostat
library. In a first step a library rgeostat implementing universal kriging estimation
in Fortran code based on LAPACK subroutines using the same data structures as in
sgeostat has been written. Another library baykrig implements the above shown
Bayesian approach based on empirical prior data (see Pilz (1991)). Finally library
kdesign puts all things together and provides functions for optimal network design
based on complete enumeration among a finite set of candidate measurement points.
The usage of these libraries will be demonstrated in a short example.
Preliminary versions of these libraries are available at ftp://ftp.uni-klu.ac.
at/pub/R.

References
C. Gebhardt. Bayessche Methoden in der geostatistischen Versuchsplanung. PhD
thesis, University of Klagenfurt, 2003.
H. Omre. Bayesian Kriging - merging observations and qualified guesses in kriging.
Mathematical Geology, 19:25–39, 1987.
J. Pilz. Ausnutzung von a-priori-Kenntnissen in geostatistischen Modellen. In
G. Peschel, editor, Beiträge zur Mathematischen Geologie und Geoinformatik,
volume 3, pages 74–79, Köln, 1991. Verlag Sven von Loga.

You might also like