Geostatistical Methods

Applied to Seismic Processing

Warning: this document is not a textbook and only aims at providing attendees
the main notions about geostatistics and its application for seismic processing.
Extensive use is made to relevant bibliographic references, that will provide
exhaustive details about geostatistics theory.

Table of contents

1 Introduction ............................................................................................................ 3
2 Reminder on basic statistics ................................................................................... 3
3 Variography ........................................................................................................... 4
3.1 Experimental variogram................................................................................. 4
3.2 Variogram model ........................................................................................... 5
4 Kriging ................................................................................................................... 8
4.1 Best linear unbiased interpolator ................................................................... 8
4.2 Kriging systems ............................................................................................. 9
4.3 Kriging neighborhood .................................................................................... 9
4.4 Kriging Properties ........................................................................................ 10
4.5 Kriging variance........................................................................................... 10
4.6 Kriging variants ........................................................................................... 10
5 Geostatistical filtering techniques ........................................................................ 11
5.1 Principles...................................................................................................... 12
5.2 First example: Refraction surveying ............................................................ 12
5.3 Second example: Bouguer anomaly............................................................. 14
6 Data integration .................................................................................................... 15
6.1 Cokriging ..................................................................................................... 15
6.2 Collocated cokriging .................................................................................... 16
6.3 Kriging with external drift ........................................................................... 17
7 Bibliography ........................................................................................................ 19

Geostatistical Methods Applied to Seismic Processing Page 1 / 20

 Table of illustrations
Figure 1: Histogram with 2 modes................................................................................. 3
Figure 2: Histogram of porosity (left); Histogram of porosity with an outlier (right). .. 4
Figure 3: Variogram cloud, classes and experimental variogram. (left) The lag value is
equal to 1 kilometer and the tolerance is equal to 0.5. ........................................... 5
Figure 4: Variogram map. .............................................................................................. 5
Figure 5: Four basic structures. ...................................................................................... 6
Figure 6: Fitting of a variogram model (black curve) on an experimental variogram
(red points). ............................................................................................................ 7
Figure 7: Fitting of a variogram models (red and purple curve) on an experimental
variogram (dotted line). Two different models have been fitted: one with a
spherical and a nugget (purple), an other with an exponential without nugget
(red). ....................................................................................................................... 7
Figure 8: Geometric anisotropy (left) : there is more spatial correlation along the
direction 2 due to a longer range; zonal anisotropy(right) : there is more spatial
correlation along direction 1, because there is a lower sill. ................................... 8
Figure 9: a moving neighborhood with anisotropic ranges of the search ellipsoid and 4
angular sectors. ...................................................................................................... 9
Figure 10: On the left hand-side configuration, the declustering effect on the weights
is clearly visible. If the standard-deviation of the kriging error is computed, the
same values will be practically got for the two configurations. This can be
explained by the fact that adding an extra data point close to an other does not
bring an extra information. .................................................................................. 10
Figure 11: nested variogram with spherical (short range), periodic and linear
components. Each structure corresponds to a given scale of variability. ............ 11
Figure 12: Refraction survey : detection of long wavelength anomaly. Top left: raw
depth map (5m x 15m) of the weathering zone. Top right: WZ depth map
obtained by factorial kriging on the raw dataset (5m x 5m). Bottom left:
Experimental directional variograms and modeled omnidirectional variogram,
after the removal of an erroneous line. Bottom right: weathering depth map once
filtered the small scale structures, orientation of the geological structures. ........ 13
Figure 13: Detection of a furrow on a Bouguer anomaly dataset. Top left: raw data
locations (black dots) and interpolated map (universal kriging). Top right:
experimental (thin line) and modeled (bold) variogram. Bottom left: regional
Bouguer anomaly obtained by kriging of the global trend. Bottom right: residual
Bouguer anomaly after the removal of the global trend. ..................................... 14
Figure 14: simple variogram (top left, bottom right) and cross-variogram (bottom left)
.............................................................................................................................. 16
Figure 15: Displays of cokriging of porosity with Norm AI (Top left); collocated
cokriging porosity (Top right) and Norm AI (Bottom left). The collocated
cokriging is more accurate due to integration of the information brought by the
Norm AI at target location. .................................................................................. 17
Figure 16: basemap (top left),histogram (top right) and experimental variogram
(bottom left) of depth at wells .............................................................................. 18
Figure 17: a TWT map (top left) ; correlation between TWT and TVD (top right) ;
Kriging with external drift of TVD using TWT as external drift (bottom left). .. 19

Geostatistical Methods Applied to Seismic Processing Page 2 / 20

1 Introduction
Geostatistics are proved to be relevant and well adapted methods when dealing
with spatial prediction and risk analysis in the Oil&Gas industry (Dubrule, 2003;
Yarus & Chambers, 2005)
Geostatistical methods take into account the spatial variability of the target
parameter, in order to provide realistic spatial estimates together with a
quantification of the associated uncertainty.
This document introduces the basic steps of a geostatistical analysis, from
exploratory data analysis to variographic analysis and (co)kriging techniques, with a
particular look at seismic processing application.

2 Reminder on basic statistics

An in-depth exploratory data analysis allows a quality control of the data set and
to ensure its consistency: no data sampling issue, no erroneous values.
Displaying a basemap, the computation of the histogram, mean and standard
deviation allow a better understanding of the data.
For instance, the histogram allows to check the homogeneity of the population
(see Figure 1). If there are different modes, that is to say 2 sub-populations, it can
be interesting to see what these populations corresponds to. Sometimes it can be
relevant to perform a separate geostatistical analysis on each sub-population.

Figure 1: Histogram with 2 modes.

It is common practice to summarize the statistical distribution by:

- the average (mean) value, which characterizes the central tendency of
the distribution:
1 n
m= zi
n i =1
where n is the count of observations in the sample.

- the variance, which characterizes the spread of the distribution around

the mean:
1 n
s2 = ( z i - m )2
n i =1
where s is the standard deviation.

Geostatistical Methods Applied to Seismic Processing Page 3 / 20

 Outliers can have a significant impact on such basic statistics. It is therefore
important to identify them and check their correctness, as they will play a large
role in the subsequent (geo)statistical analysis.
The example presented on Figure 2 illustrates this: starting from a porosity
dataset comprising 55 values between 6% and 12 %, basic statistics are computed:
Mean = 8.2%, Standard Deviation =1.4%. Then, adding a 56th value equal to 25%, the
basic statistics are computed again and change to: Mean = 8.5%, Standard Deviation
= 2.6%. It can therefore be seen that, due to just one outlier value, the standard
deviation has nearly increased twice. It is commonly said that such statistics are
not robust against outliers/atypical values.

0.15 Nb Samples: 56
Nb Samples: 55
0.4 Minimum: 6.1
Minimum: 6.1
Maximum: 11.8 Maximum: 25.0
Mean: 8.2 Mean: 8.5
Std. Dev.: 1.4 Std. Dev.: 2.6
0.3
0.10

0.2

0.05

0.1

0.00 0.0
6 7 8 9 10 11 12 5 10 15 20 25
Porosity Porosity
Figure 2: Histogram of porosity (left); Histogram of porosity with an outlier (right).

3 Variography
Most geostatistical models rely on variogram model, which quantifies the spatial
variability of the target parameter (depth, velocity, porosity). This variogram
model is fitted on an experimental variogram derived from the data values.
Hereafter are presented the different steps of a variographic analysis:
- 1: compute the experimental variogram,
- 2: fit a variogram model on the experimental one. The variogram model
will allow to derive the most relevant spatial estimate for the target
parameter.

3.1 Experimental variogram

The experimental variogram g (h) is defined by:
N
g ( h) =
1
( Z ( x + h ) - Z ( x ) )2
2 * N ( h) 1
where Z (x) denotes the target variable measured at location x and N (h)
corresponds to the number of pairs of points separated by a distance of h.
The variogram cloud is usually computed first. (See Figure 3) Each point of the
variogram cloud corresponds to a given pair of points separated by a distance h
(abscissa) and has the corresponding value 0.5 * (Z ( x + h) - Z ( x ) ) (ordinate).
2

By means of the variogram cloud, it is possible to look at the local outliers.

Those are outliers that are not seen on the histogram, but that create high values
in the variogram cloud: they could arise from a large value surrounded by small
neighbors.

Geostatistical Methods Applied to Seismic Processing Page 4 / 20

 The experimental variogram synthesizes the spatial variability information
provided by the variogram cloud. To obtain it, the information contained in the
variogram cloud is averaged by classes of distance. A distance called the lag is
defined, together with a tolerance usually equal to 50% of the lag and a given
number of lags to be calculated (multiples of the initial lag value). For each
calculation lag, the experimental variogram point is obtained by averaging all the
variogram cloud points contained in the corresponding class (distance comprised
between the lag +- the tolerance).
A initial lag value can be the averaged distance between the data points.
It is usually relevant to compute the experimental variogram up to half the field
size. At larger distances, the confidence in the experimental variogram decreases
with the number of calculation pairs.
Furthermore, it is important to check for anisotropies. Indeed, the target
variable might present a more continuous behavior in a given direction, as for
instance variograms along a channel and perpendicular to this channel. In that
purpose directional variograms are computed. When data are densely sampled, a
variogram map can be calculated : it is calculating automatically directional
variograms in every direction (See Figure 4).
Distance (Kilometer)
0. 1. 2. 3. 4.
20. 20.

15. 15.
Variogram
Variogram

10. 10.

5. 5.

0. 0.
0. 1. 2. 3. 4.
Distance (Kilometer)

Figure 3: Variogram cloud, classes and Figure 4: Variogram map.

experimental variogram. (left) The lag
value is equal to 1 kilometer and the
tolerance is equal to 0.5.

But the experimental variogram is not sufficient because the values are known
only at the lag and multiples of the lag distances. So it is compulsory to fit a
variogram model on the experimental in order to have values for all distances and
azimuths.

3.2 Variogram model

Fitting a model on the experimental variogram is compulsory for several reasons
(see Figure 6):
- The spatial variability has to be defined at every distances.
- The model will smoothe the possible statistical fluctuations.
- The model ensures a positive kriging variance.
A variogram model is usually defined by three parameters: the type of basic
structure, its sill and its range. Usually, the variability increases and get stabilized
at a given variability level called the sill. The range is the distance at which the

Geostatistical Methods Applied to Seismic Processing Page 5 / 20

variogram reaches the sill value. For distances superior to the range, there is no
spatial correlation anymore.
Several types of variogram basic structures (See Figure 5) are available to
describe how the variability is increasing. The most common structures are:
- on one hand, spherical or exponential structures, which correspond to
continuous but not differentiable variables: the variability increases
rapidly even for small distances;
- on the other hand, smooth structures such as the gaussian or the cubic
ones, which correspond to differentiable variables.

1.5 1.5
Spherical Exponential

1.0 1.0
Variogram : V

0.5 Variogram : V 0.5

0.0 0.0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Distance (m) Distance (m)

1.5
1.5
Cubic
Gaussian

1.0
Variogram : V

1.0
Variogram : V

0.5 0.5

0.0 0.0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Distance (m) Distance (m)
Figure 5: Four basic structures.

When the variogram reaches a sill, it is called a bounded or stationary variogram.

In the stationary case, there is the following relationship between the variogram g
(h) and the spatial covariance C(h) :

g (h) = C(0) C(h)

with C(h) = E((Z(x+h)- m(Z)) (Z(x) m(Z))), the spatial covariance between 2
points separated by a distance h.
It is possible to have a variogram that keeps increasing. (without a sill and a
range). In that case, the relationship between covariance and variogram can no
more be applied. The stationarity assumption being weaker for the variogram
calculation than for the covariance, the variogram is usually computed in order to
model the spatial variability.

Geostatistical Methods Applied to Seismic Processing Page 6 / 20

 The other important point is the behavior at the origin. Sometimes, there is a
nugget effect that corresponds to a discontinuity at the origin(see Figure 6).

Figure 6: Fitting of a variogram model (black curve) on an experimental variogram (red

points).

For the same experimental variogram several variogram models are sometimes
possible(See Figure 7). On the same experimental variogram of porosity, we can fit
2 different models. Of course, different models will give different estimates.
Differences at small distances will have an important impact, wheras structures at
large distances will give consistent estimates. Then, close variogram models will
give similar kriging estimates. The choice might also depend on a priori information
the user has about the spatial behavior of the target parameters.
In order to choose the most consistent model, some methods as the cross-
validation are available.

2.5

2.0
Variogram : Porosity

1.5

1.0

0.5

0.0
0 2500 5000 7500 10000
Distance (ft)

Figure 7: Fitting of a variogram models (red and purple curve) on an experimental

variogram (dotted line). Two different models have been fitted: one with a spherical
and a nugget (purple), an other with an exponential without nugget (red).

Two different anisotropy models exist (see Figure 8) :

- geometric anisotropy : The model reaches the same sill in different
directions but at different ranges.

Geostatistical Methods Applied to Seismic Processing Page 7 / 20

 - zonal anisotropy : The model reaches different sills in different
directions.

Figure 8: Geometric anisotropy (left) : there is more spatial correlation along the
direction 2 due to a longer range; zonal anisotropy(right) : there is more spatial
correlation along direction 1, because there is a lower sill.

Furthermore, a variogram can be modeled with several nested structures, each

one corresponding to a different scale of variability.

4 Kriging

4.1 Best linear unbiased interpolator

Kriging is a linear interpolation. The estimation of the variable Z at the target is
derived from a linear combination of the weighted data values(There is x0 target
location; xa data locations):
Z * ( x0 ) = la * Z ( xa )

with a, data point.

The unknown values are the kriging weights la.
In order to perform a kriging estimation, an unbiasdness condition must be
fulfilled:
[ ]
E Z 0* - Z 0 = 0
So we have :

m(la - 1) = 0
Two options are resulting :
- The mean is known (simple kriging), there is no condition on the weights.
- The mean is unknown (ordinary kriging). We must have (universality
condition):
la =1
The kriging weights are obtained by minimizing the variance of the kriging error
( )
Var Z 0* - Z 0 . So we get:
( )
Var Z 0* - Z 0 = la l b Cab - 2la Ca 0 + C 00

a,b correspond to a couple of given data points. So Cab is the spatial

covariance corresponding to the couple of points (a,b) separated by a given
distance h.

Geostatistical Methods Applied to Seismic Processing Page 8 / 20

 By drifting the variance and looking for the 0 values of the derivative we get the
kriging system. Solving the kriging system leads to the kriging weights, that can be
replaced in the initial linear combination.

4.2 Kriging systems

First, for the simple kriging (known mean), the kriging variance is drifted (by la):

l C =C b
Whatever a
ab a0

s 2 = C00 - la Ca 0
Solving the kriging system leads to the kriging weights.

Secondly, for the ordinary kriging (unknown mean), due to the universality
condition, we must add a constrain called the Lagrange parameter. It is necessary
to minimize the quantity:
j = la lb Cab - 2 la Ca 0 + C00 + 2m ( la -1)
This quantity is drifted by la and m and we get the following kriging system:

b
l b Cab + m = Ca 0 Whatever a

a
la = 1

s 2 = C 00 - la Ca 0 + m
Solving the kriging system leads to the kriging weights.
The kriging system can be written in matrix form. Solving this system requires to
invert the matrix of the Cab. This is the most time consuming thing and inverting
large matrices lead to numerical instabilities. This explains why a moving
neighborhood is usually used for large datasets.

4.3 Kriging neighborhood

Two neighborhood types are used: the unique and the moving neighborhoods.
(see Figure 9)
With a unique neighborhood, all the data points are used in order to estimate
the target value. On the contrary a moving neighborhood uses only a subset of the
data nearby the target location.

10000

-0.480
-0.004
Y (ft)

0.35918.672 -0.829
8.711
5000
-0.498
48.23620.214

-0.173 3.5612.459

-0.228

0 5000 10000
X (ft)

Figure 9: a moving neighborhood with anisotropic ranges of the search ellipsoid and 4
angular sectors.

Geostatistical Methods Applied to Seismic Processing Page 9 / 20

 To create a moving neighborhood, the following parameters have to be defined
are:
- the ranges of the ellipsoid
- the number of angular sectors
- the optimum number of data points per angular
It is preferable to have look on the variogram, on the data sampling when
defining a neighborhood.
A moving neighborhood is compulsory when :
- the covariance matrix is too large: In order to solve the kriging system, it
is necessary to invert the matrix of the covariance between the data.
Inverting large matrices leads to numeric instabilities. A moving
neighborhood becomes compulsory in order to decrease the number of
data points, therefore the size of the matrix.
- When dealing with unbounded variograms ( non stationary behavior), it
can be good to use an moving neighborhood, to have a local stationarity.

4.4 Kriging Properties

One characteristic of the kriging estimation is called the declustering effect. In
case of declustered data, the kriging estimator will give less weight to data points
that are closed to each other (they contain redundant information in case of spatial
continuity and more weight to lonely data points.(See Figure 10 as example)

Figure 10: On the left hand-side configuration, the declustering effect on the weights is
clearly visible. If the standard-deviation of the kriging error is computed, the same
values will be practically got for the two configurations. This can be explained by the
fact that adding an extra data point close to an other does not bring an extra
information.

4.5 Kriging variance

The kriging variance describes the uncertainty attached to the kriging
estimation: the knowledge of the uncertainties is very useful: It is helpful when
implementing new wells. Nevertheless the kriging variance depends only of the
variogram information not on the data values.

4.6 Kriging variants

Other Kriging techniques are available : Kriging with faults, kriging with
inequalities, kriging with variance of measurement error. See Chils & Delfiner
(1999) or Geovariances (2008) for details.

Geostatistical Methods Applied to Seismic Processing Page 10 / 20

5 Geostatistical filtering techniques
When dealing with nested variograms, (i.e a variogram with several structures)
each structure corresponds to a given scale of variability of the main variable (see
Figure 11). This method is called the factorial kriging analysis.

0.3

Variogram : Resistivity 0.2

0.1

0.0
0 1 2 3 4 5 6 7 8 9 10
Distance (m)

Figure 11: nested variogram with spherical (short range), periodic and linear
components. Each structure corresponds to a given scale of variability.

Assume that the variable of interest, Z(x), can be decomposed in several

components :
Z ( x) = Y1 ( x) + ... + Yn ( x) + m( x)

where the Yi ( x) (i = 1, K , n) correspond to stationary and uncorrelated

components of the phenomenon and m(x) represents the unknown mean (large
scale trend). Once a variogram model is fitted, the efficiency of the filtering
technique depends on the ability to interpret each variogram structure in terms of
spatial component.
For instance, in the refraction example (see paragraph 5.2), the variogram will
be fitted by a model constituted of 5 components:
g (h) = c 0 + g aa (h) + g gh (h) + g gf 1 (h) + g gf 2 (h)
with:
- c0 a nugget effect, interpreted as random acquisition noise,

- g aa (h) an anisotropic spherical short range component in the crossline

direction due to acquisition artifacts,
- g gh (h) a mid-range cubic structure corresponding to geological
heterogeneities,
- g gf 1 (h) and g gf 2 (h) two long-range structures corresponding to long
wavelength geological features.
Strictly speaking, factorial kriging consists in estimating, knowing the original
data Z(xi) and the variogram model, one particular component Yi (x) (zero-mean).
From a kriging system point of view, factorial kriging only requires a slight
modification of the right-hand side term, keeping only the variogram component
corresponding to the target Yi (x) . It is sometimes convenient to present things in a
slightly different way, particularly if one aims at filtering out one component (for

Geostatistical Methods Applied to Seismic Processing Page 11 / 20

instance acquisition noise) instead of estimating one particular component. In this
case, the kriging system is solved with, in the right-hand side term, the entire
variogram model except the component to be filtered out.
Geostatistical filtering techniques are commonly applied to improve the quality
of seismic data such as velocity cubes or interpreted horizons.
One noticeable application of factorial kriging to gravimetry is presented in
Chils and Guillen (1984). In particular, the authors compare the factorial kriging
approach with a spectral analysis. Whereas the latter requires a preliminary
interpolation on a regular grid, the factorial kriging approach has the advantage of
performing both the filtering and the smoothing in one step.

5.1 Principles
For a non-stationary variable, a extent drift estimation may be performed. A
variable can be defined as following:
Z ( x ) = m( x ) + Y ( x )
m(x) corresponds to the trend and can be either stationary or not. Y(x)
correspond to stationary residuals (with a zero mean).
The trend m(x) may sometimes be modeled by a simple polynomial expression:
m( x ) = a l f l ( x )
l

where the f l (x) are usually monomials of the coordinates and the al are unknown
fixed (or at least smoothly varying) coefficients. Note that in general: (i) there is
no optimal way to separate m(x) and the residuals Y ( x ) = Z ( x ) - m( x ) , and (ii) the
decomposition is relevant for a particular scale of observation. Universal kriging
provides an estimate of Z for this model, without having to estimate the trend
itself. However, when estimating this trend is precisely the goal, the latter can be
obtained directly from the universal kriging system (written in covariances) by just
equating to zero its right-hand side term. The residual component Y (x ) can be
obtained by subtracting the estimated mean from the estimate of the variable
itself: Y * ( x) = Z * ( x ) - m * ( x ) or by filtering directly the drift component; both
approaches are equivalent because of the linearity of the kriging system. This
universal kriging approach will be illustrated on the gravimetry example.

One noticeable application of factorial kriging to gravimetry is presented in Chils

and Guillen (1984). In particular, the authors compare the factorial kriging
approach with a spectral analysis. Whereas the latter requires a preliminary
interpolation on a regular grid, the factorial kriging approach has the advantage of
performing both the filtering and the smoothing in one step.

5.2 First example: Refraction surveying

The dataset comes from the Experimental Hydrogeological Site (EHS), located
near Poitiers city (France). The concerned aquifer consists of tight karstic Jurassic
carbonates of about 100 m in thickness beneath a weathering zone of Tertiary
clays.
Refraction seismic surveying, described in detail by Mari and Porel (2007), has
been used to map the irregular shape of the top of karstic reservoir (Figure 12, top
left).
The experimental variogram, computed in the main directions of continuity (in
line and cross line directions) has been modeled by several nested structures: a
small nugget effect (random noise), a cubic structure with a range equal to 50m,
consistent with the size of the geological heterogeneities, two spherical structures
with ranges equal to 100m and 400m. The cross-line variogram presents an

Geostatistical Methods Applied to Seismic Processing Page 12 / 20

additional variability source, attributable to acquisition artifacts. This component
is modeled by a spherical basic structure with a range equal to 5m. The nugget
effect and the small scale spherical model (5m in the crossline direction), which
have no physical meaning and are due to acquisition artifacts, are filtered out
during the interpolation process by factorial kriging (Figure 12, top right).
By comparison with the original dataset, the interpolated map allowed to
identify an erroneous acquisition line, responsible for the additional variability
component in the cross line direction. After the removal of this line, an
omnidirectional variogram is computed on the updated dataset (Figure 12, bottom
left), being mostly interested by the small scale variability, up to 50-75m. The
resulting variogram model is constituted by a nugget effect, a cubic structure with
range 55m and a long scale spherical structure with range 145m.The final WZ map
is obtained using kriging, with filtering of the nugget effect (random acquisition
noise) and small scale structure (cubic 55m) in order to obtain a map which makes
appear the large scale structures. Figure 12 (bottom right) shows the resulting
weathering depth map. The map enables to identify the directions of the geological
structures at the scale of the map: the azimuth of the main structure is N90.

WZ Depth WZ Depth FK

16556
WZ Depth Filtered
N0
20 16378
16128

14254

12408
15
Variogram : WZ

14141
13679 13498
14536
12404

9732
11061 N90
12140

10953
10
6304

6462

5270

5
3913

855

0
0 50 100 15
Distance (m)

Figure 12: Refraction survey : detection of long wavelength anomaly. Top left: raw
depth map (5m x 15m) of the weathering zone. Top right: WZ depth map obtained by
factorial kriging on the raw dataset (5m x 5m). Bottom left: Experimental directional
variograms and modeled omnidirectional variogram, after the removal of an erroneous

Geostatistical Methods Applied to Seismic Processing Page 13 / 20

line. Bottom right: weathering depth map once filtered the small scale structures,
orientation of the geological structures.

5.3 Second example: Bouguer anomaly

The gravimetric data were gathered by the Geophysical Institute, University of
Lausanne, under the direction of Professor Raymond Olivier (Atlas gravimtrique de
la Suisse au 1/10000). The area of Aubonne is located between Geneva and
Lausanne in Switzerland. The geology of this area consists in a Molasse basement
(Chattien sandstone) dug with furrows filled by quaternary sediments. The Bouguer
model is computed using the 1967 Ellipsode and the model density is 2.4. This
density corresponds to the Molassic basement. The topographic correction has been
applied up to 167 km (Olivier and Simard, 1981).
Firstly, the raw Bouguer anomaly has been interpolated (100m x 100m) using
universal kriging (Figure 13 top left). Variogram analysis (Figure 13 top right) has
been performed on data residuals after the removal of a global linear trend on both
coordinates. Geostatistical filtering has been used to decompose the Bouguer
anomaly into a long wavelength structure (regional anomaly, corresponding to the
global trend, Figure 13 bottom left) and a short wavelength structure (residual
anomaly, Figure 13 bottom right). On the residual Bouguer anomaly map, one can
clearly see a negative anomaly oriented North West South East. It points out a
furrow in the bed rock. A borehole, drilled at 522.3/150.4, has given a depth of
146m.
156 0.05
Variogram: Bouger Anomaly - Residuals

Anom Bouger KU
155

154 0.04

153 (mGals)
152 -47
-48 0.03
Y (km)

-49
151 -50
-51
-52
150 -53
-54 0.02
-55
149 -56
-57
-58
148 -59
-60 0.01
-61
147 -62

146
515.0 517.5 520.0 522.5 525.0 0.00
0 1 2 3
X (km) Distance (km)
156
Anom Bouger Trend 156
155 Anom Bouger residuals
155
154
154
153 (mGal
153 (mGals)
152 -47
-48 152
Y (km)

1.0
-49
Y (km)

151 -50
-51 151
-52 0.5
150 -53
-54 150
-55 0.0
149 -56
-57 149
-58
148 -59 -0.5
-60 148
-61
147 -62
147 -1.0

146 146
515.0 517.5 520.0 522.5 525.0 515.0 517.5 520.0 522.5 525.0
X (km) X (km)

Figure 13: Detection of a furrow on a Bouguer anomaly dataset. Top left: raw data
locations (black dots) and interpolated map (universal kriging). Top right: experimental
(thin line) and modeled (bold) variogram. Bottom left: regional Bouguer anomaly
obtained by kriging of the global trend. Bottom right: residual Bouguer anomaly after
the removal of the global trend.

To conclude, variogram analysis and geostatistical filtering can be used to detect

and filter both random acquisition noise and foot prints in near-surface geophysics.
Futhermore geostatistical filtering techniques can be used to decompose
"geophysical anomalies" and identify target geological features.

Geostatistical Methods Applied to Seismic Processing Page 14 / 20

6 Data integration
When the main variable is known only at sparse locations and that this variable is
correlated to a dense secondary variable, it is interesting to take into account of
the information brought by this secondary variable in order to improve the
estimation of the main variable. This is the aim of geostatistical multivariate
techniques. Of course, it is possible to use several variables in order to improve the
estimation of the main variable.

6.1 Cokriging
As for kriging, the cokriging estimation is a linear combination of the data values
(both of the main and secondary variables).
Considering the main variable Z1 and the secondary variable Z2, we have:
Z 1* ( x0 ) = la1 Z 1 ( xa ) + l1b Z 2 (xb )

l1 are the weights associated to the variable Z1 and l2 are the weights
associated to the variable Z2.
As there are several variables, simple variograms for each variables and cross
variograms between the variables are compulsory to describe the joint spatial
variability between Z1 and Z2 (see Figure 14). The simple and cross variograms are
linked each other through the linear model of co-regionalization; in this model,
all simple and cross variograms are linear combinations of the same basic
structures.
Hereafter is the mathematical formula of a cross variogram between two
variables Z1 & Z2:
g 1, 2 (h ) = E [(Z 1 (x + h ) - Z 1 (x ))(Z 2 (x + h ) - Z 2 ( x ) )]
with the corresponding cross covariance formula:

[( )(
C1, 2 (h ) = E Z 1 ( x ) - m Z1 Z 2 ( x + h ) - mZ 2 )]
There is a particular case where cokriging is equal to kriging: when the two
variables are known at the same data location (isotopic case) and that simple and
cross variograms are all proportional (intrinsic correlation).
As for kriging, the variance of the estimation error is minimized (in that formula,
the variable Z1 is estimated):

( )
Var Z 01 - Z 01 = C 00
*
11
+ la1 la1 Cab
11
+ la2 la2 Cab22 + 2la1 l b2 Cab
12
- 2la1 Ca110 - 2la2 Ca120

where a,b are a couple of points separated by a distance of h. 1 and 2

correspond respectively to the variables Z1 and Z2. So Ca1,,2b correspond to the spatial
cross covariance between variables Z1 and Z2 for the couple of points a,b.

For further analysis the cokriging system is available in the bibliography (see
Wackernagel).

Geostatistical Methods Applied to Seismic Processing Page 15 / 20

 0.06

Variogram : Norm Ai at wells

0.05

0.04

0.03

0.02

0.01

0.00
0 2500 5000 7500 10000
Distance (ft)
Variogram : Norm Ai at wells & Porosity

0.3 2.5

0.2
2.0

Variogram : Porosity
0.1
1.5

0.0

1.0
-0.1

-0.2 0.5

-0.3 0.0
0 2500 5000 7500 10000 0 2500 5000 7500 10000
Distance (ft) Distance (ft)

Figure 14: simple variogram (top left, bottom right) and cross-variogram (bottom left)

6.2 Collocated cokriging

When there is a dense secondary variable, known at the target location, it is
possible to apply a collocated cokriging. In that case we add an extra information
in the estimation : the value of the secondary variable at the target location. If the
correlation is high between the 2 variables, the collocated cokriging will be more
accurate than the cokriging due to the extra information.
Z 1* ( x0 ) = la1 Z 1 ( xa ) + la2 Z 2 (xb ) + l 02 Z 2 ( x0 )
Furthermore, the collocated cokriging is heterotopic, the main and secondary
variable are not known at the same location. if the secondary variable is known on
the output and there is a strong linear correlation, the assumption of Markov-
Bayes can be made : All the simple and cross variograms are all proportional to the
simple variogram of the secondary variable. This method is simpler than the linear
model of co-regionalization.
Hereafter is an example between 2 variables Z1 and Z2. (Z2 is the secondary
variable). In the case of the assumption of Markov- Bayes, we have (in covariance) :

C1, 2 (h ) = (C1, 2 (0 ) / C 2 (0)) * C 2 (h )

C1 (h ) = (C1 (0 ) / C 2 (0 )) * C 2 (h )
Where C1,2(h) is the spatial cross-covariance between variables Z1 and Z2. C2(h) and
C1(h) are the spatial covariances of respectively Z2 and Z1.

Hereafter are some displays of the different methods.

Geostatistical Methods Applied to Seismic Processing Page 16 / 20

 Porosity (%)

12.0
10000
10000 Colloc cokriging porosity
Kriging Porosity 11.5
11.0
10.5
10.0

Y (ft)
9.5
5000
Y (ft)

5000 9.0
8.5
8.0
7.5
7.0
6.5
0 0
6.0
0 5000 10000 0 5000 10000
N/A
X (ft) X (ft)

Image

0.00
10000
Norm AI -0.10

-0.20

-0.30

-0.40
Y (ft)

5000 -0.50

-0.60

-0.70

-0.80

-0.90
0
-1.00
0 5000 10000
N/A
X (ft)

Figure 15: Displays of cokriging of porosity with Norm AI (Top left); collocated cokriging
porosity (Top right) and Norm AI (Bottom left). The collocated cokriging is more
accurate due to integration of the information brought by the Norm AI at target
location.

6.3 Kriging with external drift

When a non-stationary variable has to be estimated and that this variable is
correlated to a dense secondary variable (with a good correlation), a method
called Kriging with external drift can be performed.
In the external drift method, the non-stationary part (or Trend) of the main
variable is modeled by means of the secondary variable as an external drift. The
relationship between the main variable and the secondary variable is linear (linear
regression). The trend of the variable can be modeled as following:
E[ Z ( x)] = a 0 + a1 S ( x)
and then the variable Z(x) :
Z ( x) = a 0 + a1 S1 ( x) + R ( x)
Z(x) is the main non stationary variable that is estimated, a0 and a1 are constant
coefficients, S(x) is the external drift and R(x) corresponds to the stationary
residuals. In the external drift method the trend is first modeled by means of a
linear regression with the external drift then a kriging of the stationary residuals is
performed. At the end, the kriged residuals are added to the modeled trend.
A classical application of kriging with external drift is time-to-depth conversion:
with depth known at wells location and a map of TWT. TWT is used as external
drift. There is a linear relationship between TWT and depth.

Geostatistical Methods Applied to Seismic Processing Page 17 / 20

 The simplest way to check the non-stationary behavior of the variable is to look
at the shape of the variogram. If the variogram keeps increasing, the variable
might be considered as non-stationary.

The following example is taken from a case study available in Isatis (Non stationary
and volumetrics). The following data are available :
- TVD and depth from seismic known at wells locations
- Depth from seismic map

depth at wells
Nb Samples: 87
20000
Minimum: 2197.00
Maximum: 2343.00
0.3
Mean: 2241.17
Std. Dev.: 42.09
17500

Frequencies
Y (m)

0.2

15000

0.1

12500

0.0
332500 335000 337500 340000 2200 2250 2300 2350
X (m) depth at wells

157

2000
Variogram : depth at wells

169

1500

188

1000

232

500 295
252
230
271
22 256
0
0 1000 2000 3000
Distance (m)
Figure 16: basemap (top left),histogram (top right) and experimental variogram (bottom
left) of depth at wells

The experimental variogram keeps increasing, so depth at wells is considered as

a non-stationary variable (see Figure 16). Therefore a non-stationary estimation has
to be performed. Due to the presence of a depth from seismic map and a good
linear correlation between depth from seismic and depth at wells (see Figure 17
below), a kriging with external drift of depth at wells using the depth from seismic
map as external drift is calculated.

Geostatistical Methods Applied to Seismic Processing Page 18 / 20

 Depth from Seismic (Background) rho=0.982
2250
20000

depth at wells (seismic)

2225
Y (m)

15000

2200

10000
2175

5000 2150

320000 325000 330000 335000 340000 345000 2200 2250 2300 2350
X (m) depth at wells

Depth (m)

2450
Depth from Wells (Ext. Drift) 2430
2410
20000 2390
2370
2350
2330
Y (m)

15000
2310
2290
2270
10000 2250
2230
2210
2190
5000

320000 325000 330000 335000 340000 345000 N/A

X (m)

Figure 17: a TWT map (top left) ; correlation between TWT and TVD (top right) ; Kriging
with external drift of TVD using TWT as external drift (bottom left).

7 Bibliography
Chils J.P. and Guillen A. [1984] Variogrammes et krigeages pour la gravimtrie et
le magntisme. Sciences de la Terre, Srie Informatique Gologique, 20, 455-
468.
Chils J.P. and Delfiner P. [1999] Geostatistics modeling spatial uncertainty. Wiley
series in probability and statistics.
Dubrule O. [2003] Geostatistics in Petroleum Geology. Distinguished Instructor
Short Course Distinguished Instructor series, N6, SEG & EAGE.
Geovariances [2008] ISATIS technical references. Geovariances & Ecole des Mines
de Paris, version 8.0, 138 p.
Mari J.L and Porel, G. [2007] 3D seismic imaging of a nearsurface heterogeneous
aquifer: a case study. Oil & Gas Science and Technology, Rev IFP, DOI:
10.2516/ogst:2007077.
Mari J.L, Porel G and Bourbiaux B. [2008] From 3D Seismic to 3D Reservoir
Deterministic Model Thanks to Logging Data: the Case Study of a Near Surface
Heterogeneous Aquifer. Oil & Gas Science and Technology, Rev. IFP, DOI:
10.2516/ogst:2008049.
Matheron G. [1982] Pour une analyse krigeante des donnes rgionalises. Internal
report N-732, Centre de Gostatistique et de Morphologie Mathmatique,
Ecole des Mines de Paris, Janvier 1982.
Olivier R.J. and Simard R.G. [1981] Improvement of the conic prism model for
terrain correction in rugged topography. Geophysics, 46(7), 1054-1056.
Wackernagel H. [1995] Multivariate Geostatistics. Springer, 203 p.

Geostatistical Methods Applied to Seismic Processing Page 19 / 20

Yarus J.M. and Chambers R.L. [2000] Petroleum Geostatistics for
Nongeostatisticians Part 1. The Leading Edge; May 2000; v. 19; no. 5; p.
474-479; DOI: 10.1190/1.1438630, Society of Exploration Geophysicists.
Yarus J.M. and Chambers R.L. [2000] Petroleum Geostatistics for
Nongeostatisticians Part 2. The Leading Edge; June 2000; v. 19; no. 6; p.
592-599; DOI: 10.1190/1.1438664, Society of Exploration Geophysicists.
Yarus J.M. and Chambers R.L. [2005] Stochastic Modeling and Geostatistics
Principles, methods and Case Studies. Ed. Amer Assn of Petroleum Geologists,
379 p.

Geostatistical Methods Applied to Seismic Processing Page 20 / 20