Multivariate Geostatistics

Mg. Sc. Marco A. Cotrina

mcotrinat@[Link]
Motivation
Situations where multivariate geostatistics could be used
• Existence of several variables of interest:
• Environmental science: contaminant concentrations
• Mining: grades in polymetallic deposits; geo-metallurgical attributes
• Hydrogeology: petrophysical properties
• Existence of several sources of information:
• remote sensing, digital elevation model, geophysics
• data with different supports or different quality.
Definitions
• Coregionalization: set of regionalized variables {z1,... zN} defined in the same
region of space (field).

• Isotopic sampling: when the variables

are known at the same locations

• Heterotopic sampling (partial or total)

Exploratory tools

Maps

Histograms
Exploratory tools

Quantile-quantile plots: to compare the

distributions of two variables

Scatter plot:
• to detect a dependence between variables
• to detect outlying data
• to choose a regression model
Exploratory tools
----------------------------------------------------
| Cd | Co | Cr | Cu | Ni | Pb |
Zn
Cd | 1.00| 0.26| 0.58| 0.15| 0.49| 0.22| 0.62
----------------------------------------------------
Correlation coefficient / correlation Co | 0.26| 1.00| 0.48| 0.19| 0.74| 0.16| 0.44
Cr | 0.58| 0.48| 1.00| 0.21| 0.71| 0.26| 0.61
Cu | 0.15| 0.19| 0.21| 1.00| 0.22| 0.82| 0.66
Ni | 0.49| 0.74| 0.71| 0.22| 1.00| 0.27| 0.59
matrix: to summarize the linear Pb | 0.22| 0.16| 0.26| 0.82| 0.27| 1.00| 0.67
Zn | 0.62| 0.44| 0.61| 0.66| 0.59| 0.67| 1.00
---------------------------------------------------
relationships between variables

Principal components analysis: to construct

uncorrelated variables (“factors”) that decompose
the multivariate information.
Exploratory tools
Multivariate classification: to model the
relationship between known categories and
dependent variables, in order to predict to
which category belongs a data for which
only the dependent variables are known.
Multivariate clustering: to combine data
into groups such that data within the same
group are similar, and data from different
groups are dissimilar
Variogram analysis
The objective is to describe and to model the spatial dependence
relationships between a set of coregionalized variables. The following
tools will be used:
• “Direct” covariance functions or variograms, i.E., Covariance
functions or variograms of each variable
 To describe the spatial continuity of each variable
• “Cross” covariance functions or variograms
 To measure the spatial dependence between two variables
Variogram analysis
Let Zi and Zj be two jointly stationary random functions, and h a lag separation
vector. Their cross covariance function is defined as follows:

while their cross variogram is defined by:

Variogram analysis
One alternative to the cross variogram is the so-called “pseudo cross variogram”,
defined as:

However, this tool is difficult to interpret when the variables are not measured in the
same units (the difference between the two variables does not have a physical
meaning)
hardly ever used in applications.
Variogram analysis
Mathematical properties of a cross covariance function
• Symmetry: in general, for i  j, cij(h)  cji(h) and cij(h)  cij(–h), but one always
has cij(h) = cji(–h)

• Sign: the cross covariance function between two variables may take on negative
values, even at the origin, which indicates that the variables are negatively
correlated.
• Extreme values: the maximum value of a cross covariance function may occur for
a non-zero lag separation vector (delay effect, frequent with variables that
depend on time and evolve asynchronously).
Variogram analysis
Mathematical properties of a cross variogram
• Symmetry with respect to the indices: ij(h) = ji(h)
• Parity: ij(h)  ij(h)
• The cross variogram vanishes at the origin:: ij(0)  0
• The cross variogram may be negative, which occurs with negatively
correlated variables
• Relationship with the cross covariance function:
Variogram analysis
Inference: sample cross variogram
The cross variogram for a given lag separation vector h can be estimated by using
the pairs of data whose separation coincides with h:

with N(h) = {(,) such that x – x  h} (with tolerances)

|N(h)| = cardinal of N(h) (number of data pairs)

 The sample cross variogram cannot be calculated in the case of a totally heterotopic
sampling.
Variogram analysis
The calculation of the sample cross
variogram requires defining the same N

parameters as in the univariate case: p

 hH

• Distances of interest (lags, number of lags) Paso


Paso 1 Paso Paso Paso
E

0 2 3 4

• Direction of interest: azimuth, dip

• Tolerance on the distance
• Tolerance on the azimuth and dip
• Band width and band height
Variogram analysis
Example: standardize data of cobalt and nickel concentrations
Variogram analysis
Link between direct and cross variograms
• Cauchy-Schwarz inequality

• Matrix positive definiteness

 The direct and cross variograms cannot be modeled independently

Variogram analysis
Variogram fitting
There exist many modeling strategy to fit the direct and cross variograms :
• Intrinsic correlation model
• Markov model
• Linear coregionalization model
• Bilinear coregionalization model
• Models for variables linked by functional or differential equations
Variogram analysis
Intrinsic correlation model
All the direct and cross variograms {ij(h), i  1... N, j  1... N} are proportional:

where:
• g(h) is a basic variogram model (e.g., nugget effect, spherical, exponential...)
with given range and with unit sill
• C is a symmetric, positive semi-definite matrix (i.e., a symmetric matrix with
non-negative eigenvalues) called the sill matrix
Variogram analysis
Markov model
Knowing the value of Zi at location x, the values of the other variables at
the same location do no longer depend on the values of Zi at locations
different from x
The cross variograms {ij(h), j  1... N} are proportional to the direct
variogram ii(h)
Variogram analysis
Example: the intrinsic correlation and Markov models are ill-suited to the
modeling of the following sample variograms, insofar as the direct and
the cross variogram do not have the same shape
Variogram analysis
Linear coregionalization model
This model extends the concept of nested structures to the multivariate
context. It consists in modeling the direct and cross variograms with the
same set of basic nested structures.
A sufficient condition to obtain a mathematically consistent model is
that, for each nested structure, the sill matrix is symmetric positive
semi-definite, i.e., that its eigenvalues are non-negative.
Variogram analysis
The linear coregionalization model can be written in the following
fashion:

where:
• g1(h),... gS(h) are basic variogram models (e.g., nugget effect, spherical,
exponential...) with given ranges and with unit sills
• C1,... CS are symmetric sill matrices, with non-negative eigenvalues
Variogram analysis
In the case of two variables, one has

so that the positive semi-definiteness condition is fulfilled if

In the case of three or more variables, the positive semi-definiteness condition is

more difficult to check (one has to calculate the eigenvalues of each sill matrix in
order to verify that none of them is negative). There exist automatic sill fitting
algorithms, for which the user has to propose the set of basic variogram model, in
particular, their number (S), types and ranges .
Variogram analysis
Complement: algorithm for automatic sill fitting

Let us consider a linear coregionalization model:

It is of interest to minimize the following weighted sum of squares, which
involves a set of K lag separation vectors {hk / k = 1,… K}:

where
• wk is a positive weight
• (hk ) is the matrix of direct and cross variogram models
Variogram analysis
Initialization
An initial set of sill matrices (C1,… CS) is chosen, for instance, null
matrices.
Iteration
Calculate WSS for the current set of sill matrices
• For s = 1, … S
─ Calculate:
Variogram analysis
Perform the eigen-decomposition of

~
C s  Q s  s Qt with Q s Qt s  I (identity matrix) and Λs diagonal matrix
s

+
─ Set to zero the negative values of Λs to obtain a diagonal matrix Λs .

 t
─ Update the sill matrix C s , by setting C s Q  Q
s s s

• Calculate WSS for the set of updated sill matrices. If the difference with the WSS
value calculated at the beginning of the iteration is small, the algorithm is stopped.
Otherwise, one goes back to the iteration step.
Variogram analysis
Example 1: variograms of cobalt and nickel concentrations
Variogram analysis
Example 2: direct variograms of copper, silver, arsenic and antimony
grades
Variogram analysis
Example 2 (cont.): cross variograms
Spatial estimation: cokriging
Cokriging aims at performing the joint estimation of several
coregionalized variables, taking into account their spatial dependences.
This technique generalizes kriging: the estimator is constructed as a
linear combination of the available data, with no bias and with the
smallest possible variance for the estimation error.
Cokriging is advantageous when the target variable (primary variable) is
under-sampled with respect to other cross-correlated variables
(covariates).
Spatial estimation: cokriging
As for kriging, their exist several variants of cokriging:
• Simple cokriging: with known means
• Ordinary cokriging: with unknown means
• Block cokriging
• Specific multivariate variants:
• Mixed cokriging: some means are known, but others are unknown
• Collocated cokriging: this variant uses the data of the covariates only at the
location targeted for estimation (collocated cokriging) and also at the
locations where the primary variable is known (multi-collocated cokriging).
It requires that all the covariates are exhaustively known.
Spatial estimation: cokriging
Cokriging neighborhood
The selection of relevant data to perform
estimation is more complex than in the
univariate
case:
• search for the closest data locations,
irrespective of which variables are known at
these locations?
• two-part search? (one search for the closest
data of the primary variable; another search
for the closest data of the covariates).
Simple cokriging
Assumptions
• The coregionalized variables have known mean values{mi, i  1... N}.
In general, these means are assumed to be constant in space and
equal to the (declustered) sample means of the data of each
variable.
• The set of direct and cross covariance functions Cij(h), i, j  1... N} is
also known.
Simple cokriging
Notations
• For the sake of simplicity, let us consider the case when all the variables are
known at the same locations {x,  = 1… n} (isotopic sampling).
• It is convenient to use vector and matrix notations in order to establish the
cokriging equations. Let us denote by:
Z the N × 1 vector whose generic term is zi
m the N × 1 vector whose generic term is mi
C(x – x) the N × N matrix whose generic term is Cij(x – x)
Simple cokriging
Simple cokriging estimator
The estimator of the entire set of variables at location x0 can be written as:

where a is a N × 1 vector and {,  = 1… n} are N × N matrices defined by

Simple cokriging
The previous system of equations can be solved by
• Matrix inversion
• Gauss elimination
• LU decomposition of the covariance matrix.
Note that a single system has to be solved in order to simultaneously estimate all
the variables. In the case of a heterotopic sampling, the system of equations is
obtained by removing the rows and columns corresponding to the missing data.
Simple cokriging
Simple cokriging errors
The variance-covariance matrix of cokriging errors is:

The diagonal terms of this matrix are the variances of cokriging errors (cokriging
variances).
Ordinary cokriging
Assumptions
• The coregionalized variables have unknown mean values
• The set of direct and cross variograms {ij(h), i, j  1... N} is known.
By considering that the mean values are unknown, one can extend the estimator to
the cases when these means are constant only at a local scale, but not at the global
scale: the mean values can vary in space, provided that they are constant at the
scale of the cokriging neighborhood.
In the following, let us denote by (x – x) the N × N matrix whose generic term is
ij(x – x)
Ordinary cokriging
Ordinary cokriging estimator
The estimator of the entire set of variables at location x0 can be written as:

where {,  = 1… n} are N × N matrices (weight matrices) defined by

( is a matrix of Lagrange multipliers, I the identity matrix, 0 the zero matrix)

Ordinary cokriging
The previous system of equations can be solved by
• Matrix inversion
• Gauss elimination
Note that a single system has to be solved in order to simultaneously
estimate all the variables. In the case of a heterotopic sampling, the
system of equations is obtained by removing the rows and columns
corresponding to the missing data.
Ordinary cokriging
Ordinary cokriging errors
The variance-covariance matrix of cokriging errors is:

The diagonal terms of this matrix are the variances of cokriging errors (cokriging
variances).
Ordinary cokriging
Standardized ordinary cokriging
When the unknown means are the same, the equations giving the cokriging
weights, Lagrange multipliers and variance-covariance matrix of cokriging errors are
modified as follows:

( is a 1 × N vector of Lagrange multipliers and 1 is a N × 1 vector of ones)

Remarks on the cokriging system
The cokriging weights and the variances of cokriging errors depend on
geometric information (data and target locations) and on variogram
information (nugget effect, anisotropy…), but they do not depend on the
data values.
The cokriging system is singular if
• There are duplicated data
• The coregionalized variables are linearly dependent
Properties of cokriging
• Exact interpolation: at the data locations, the estimation coincides with the data
values
• Smoothing effect: the estimated values are less dispersed than the true values.
Cokriging tends to underestimate high-value areas and to overestimate low-value
areas.
• Additivity: the estimation of an average value is the same as the average of the
point-support estimations
• Unbiasedness: the average of estimation errors over a region tends to zero when
the region becomes infinitely large.
The alternative between kriging and cokriging
From a theoretical standpoint, cokriging always provides more accurate
estimations than kriging:
The error variance is smaller
• Linear relationships between variables are reproduced.
From a practical standpoint, cokriging may yield to results similar to that
of kriging, but it requires more CPU time and greater modeling efforts
(variogram analysis).
The alternative between kriging and cokriging
Cokriging coincides with kriging in the following two situations:
• The variables do not have spatial cross-correlations (the cross-
variograms are equal to zero)
• The direct and cross variograms are proportional (intrinsic correlation
model) and the sampling pattern is isotopic
 the variables have the same spatial continuity and turn out to be
redundant
The alternative between kriging and cokriging
In contrast, cokriging improves the results of kriging when the
primary variable is under-sampled with respect to other cross-
correlated variables (heterotopic sampling)
 advantageous when covariates are more accessible or less
expensive to sample
The alternative between kriging and cokriging
Pros of cokriging
• Accounts for correlations between variables
• Reproduces linear relationships (e.g.: compositional variables)
• Improves the accuracy of the estimation, in particular when the covariates are
highly correlated and when the sampling pattern is heterotopic
Cons
• It provides smooth models
• It does not reproduce non-linear relationships (e.g.: inequalities…)
• The implementation requires a coregionalization model (model of the direct and
cross variograms).
The alternative between kriging and cokriging
Pros of cokriging
• Accounts for correlations between variables
• Reproduces linear relationships (e.g.: compositional variables)
• Improves the accuracy of the estimation, in particular when the covariates are
highly correlated and when the sampling pattern is heterotopic
Cons
• It provides smooth models
• It does not reproduce non-linear relationships (e.g.: inequalities…)
• The implementation requires a coregionalization model (model of the direct and
cross variograms).
Gracias
 Multivariate Geostatistics
Mg. Sc. Marco A. Cotrina

mcotrinat@[Link]