Mathematical Geology, Vol. 21, No.

7, 1989

Factorial Kriging Analysis Applied to Geological

Data from Petroleum Exploration

O. Jaquet 2

A regionalized variable, thickness of the reservoir layer, from a gasfieM is decomposed by factorial
kriging analysis. Maps of the obtained components may be associated with depositional environ-
ments that are favorable for petroleum exploration.

KEY WORDS: geostatistic8, factorial kriging analysis, spectral analysis, petroleum exploration.

INTRODUCTION
Factorial kriging analysis is a geostatistical method, developed by Matheron
(1982), allowing the decomposition of a regionalized variable into different
components that may be mapped separately for analysis.
From the experimental variogram of a regionalized variable, a nested var-
iogram model composed of several basic models is fitted. The range of each
basic model is associated with a certain spatial scale.
Factorial kriging analysis allows different components corresponding to
these spatial scales or "frequency bands" to be estimated, by cokriging. The
estimate of each component is carried out at the nodes of a regular grid which
is contoured, in a second step, to yield a map.
Galli and Sandjivy (1985) theoretically compared factorial kriging analysis
(spatial domain) and spectral analysis (frequency domain) demonstrating their
formal equivalence. Factorial kriging analysis seems to offer larger operational
possibilities than spectral analysis, as spectral analysis is not directly applicable
to irregularly spaced data.
Factorial kriging analysis--used in geophysics (Chil6s and Guillen, 1984;
Galli et al., 1984), geochemistry (Sandjivy, 1984) and remote sensing (Ma and
Royer, 1987)--is applied, in this case study, to geological data from a gas field.

~Manuscript received 31 March 1988; accepted 15 February 1989.

2COLENCO AG, Parkstrasse 27, 5401 Baden, Switzerland.

683
0882-8121/89/1000-0683$06.00/1 © 1989InternationalAssociationfor MathematicalGeology
684 Jaquet

THEORETICAL OVERVIEW: UNIVARIATE FACTORIAL

KRIGING ANALYSIS IN THE STATIONARY CASE
Matheron (1982) has developed the general theory of factorial kriging anal-
ysis. In the univariate case, the regionalized variable z(x), known at experi-
mental points x~, is interpreted as a realization of a stationary second-order
random function Z(x), admitting the decomposition
U

Z(x) = E yU(x)
u=l

YU (x) are mutually orthogonal stationary random functions. This decomposi-

tion is related directly to the nested variogram model G (h)
U

G ( h ) = E G u(h)
u=l

To each Y~ (x) corresponds a variogram model G" (h) associated with a certain
spatial scale. The estimate of components yu (x) is made by cokriging, using
a linear combination of the experimental Z~ with weights ~'
N

Y~*(x) = E X~Z~
c~=l

where YU*(x) is the estimator of Y" (x). The expectation o f Z (x) is an unknown
value denoted by the symbol m
E[Z (x)] = E E[Y ~ (x)] = m
U

Estimator Y"* (x) is required to be unbiased whatever the (unknown) value of

m

E[Y~*(x) - Y~ (x)] = 0

eIZx:zo- = o

m E X'~ - E[Y u (x)] = 0

Thus, weights k~ must satisfy the universality condition

N

E x~=o
a=l

which imposes
E [ r " (x)] = o
Factorial Kriging Analysis 685

and the model becomes

Z ( x ) = Z Y~(x) + m
u

Matheron (1982) has shown that the same relation, in terms of the estimators,
gives the coherence condition

Z* (x) = Z r"* (x) + m*

u

Minimization of the estimation variance

Var [Y"*(x) - Y" (x)] = - Z Z X°~ X~ G (x,~, xe) + 2 ~ X~ G" (x,~, x)

c~ B ce

under the constraint

= o
o/

leads to a cokriging system for each Y" (x) with N + 1 equations

Z X~ a (x,, x~) = G" (x~, x) - ~,,

o/

Zx2 = 0

where #u is a Lagrange multiplier.

In this study, estimate of I1" (x) is made with a moving neighborhood.
This choice is justified by the geological context. Thus, the model is expressed

Z* (x) = ~ Y~* (x) + m* (x)

u

where m* (x) is a local mean dependent on the size of the neighborhood, m (x)
is considered as a low frequency function at the neighborhood scale. The esti-
mate of m (x) is added to the large-range ( = "low frequency") component Y"*
(x). For mapping, the large-range component Y~,* (x) is represented by merg-
ing Y"* (x) and m* (x)

Y~* (x) = yu, (x) + m* (x)

DATA
The data are derived from 214 irregularly spaced wells from a gas field
(surface area of over 10,000 km 2) in northern Europe (Fig. 1). Measured vari-
able E4 is thickness of a reservoir (Fig. 2) in carbonates.
686 Jaquet

10 P.o 3o 4o 5o 60 70 80
I I I I,,, 1 I 1 ,,, I

90- -90

~B0- -BO

70- -70

G0-

50- -50
t
40- ,N,i,O •
30- -aO

20- -~0

10- -19

" I
"'~ n ~ ~ ~ ~ A ~
Fig. 1. Distribution of wells; symbol size proportional
to the thickness of the reservoir (coordinates expressed
in kin).

EXPERIMENTAL VARIOGRAMS
Experimental variograms of regionalized variable E4 were calculated in
four directions: E-W, NE-SW, N-S, and NW-SE. The mean variogram is
superimposed on the directional variograms (Fig. 3).

0.115

0.I0

t 0.06

q
mte~
Fig. 2. Histogram of reservoir thickness.
Factorial Kriging Analysis 687

II00c " "~""

t ~

I I i -,, I ...... I ~ I -..L

Fig. 3. Directional experimental vafiograms of E4 (dashed lines). Mean var-

iogram represented by solid line.

FIT OF THE VARIOGRAM MODEL

Directional variograms o f E 4 exhibit only slight anisotropies which can be

neglected. The mean experimental variogram is fitted up to 50 km with a model
consisting o f two nested spherical models (structures), (Fig. 4), whose param-
eters fire:
Structure 1: sill CI = 2,600 m 2
range al = 10 km
Structure 2: sill C~ = 2,900 m 2
range a2 = 30 k m

The model was selected using the cross-validation criterion (Jaquet, 1985).
The fitted theoretical variogrfim matches the mean experimental variogram
fairly well to a distance o f 50 kin; hence thickness may be considered as a
stationary random function in any neighborhood with a diameter up to 50 kin.

/"k

N _%-

]" iII
N

I ..... t,, I I ,.I 1 ~

10 l0 SO 40 ll0 II~

Fig. 4. Variogram with two nested spherical models fitted to the mean experi-
mental variogram.
688 Jaquet

FACTORIAL KRIGING ANALYSIS

Components yl(x) and YZ(x) associated, respectively, with the 10 km and
30 km range are estimated by cokriging at nodes of a regular grid (mesh di-
mensions: 3 x 3 kin) for contouring. The local mean re(x) is estimated by
kriging, and then is added to the estimated component Y2*(x).
The size of the moving neighborhood used is 8 points. This size is suffi-
cient for estimating the large-range component Y2(x).
The map of the 10 km range component Y~* (x) (Fig. 5) shows the "high
frequency" relative structural characteristics of E4. The map of the 30 km
range component Y~* (x) (Fig. 6) shows the "low frequency" relative struc-
tural characteristics of E4.
The sum of components y1, (x) and Y~* (x) of the small and large range,
respectively, is equal to estimator E4*(x) obtained by kriging

E4*(x) = yl, (x) + Y~* (x)

Variable E4* (x) is mapped (Fig. 7). Decomposition of E4 is shown clearly

by comparision (Fig. 7 with Figs. 5 and 6).

10 20 30 4O 5O 6O 70 80
; I I I 1 I I ,,,I.... f
, , [
110 -4 1tt

too-t
9O-4

6O-4
I

I
40-t
I

'°1 /-t-

,o4 I .....
F .4-o i
- Lo ~o 3', ,~ ,,o 6, 7,
Fig. 5. Map of the lO-kmrange componentYI s (x).
Factorial Kriging Analysis 689

Fig. 6. Map of the 30-km range component Y~*(x).

GEOLOGICAL INTERPRETATION
Variability of the thickness of reservoir consisting of carbonate rocks is
the result of different geological phenomena (sea level changes, faults, etc.)
which were active during sedimentation. Two main environments are distin-
guished, a strong-energy and a weak-energy environment.
The strong-energy environment is associated with locally agitated condi-
tions (currents, waves, etc.) in limited areas. The weak-energy environment is
distributed more broadly. The measured thickness is the sum of different facies
associated with these depositional conditions.
An interpretation is proposed which associates the small range component
Yl(x) with the strong-energy environment, and the large range component y2(x)
with the weak-energy environment.
The map of I11" (x) (Fig. 5) shows the relative structural characteristics of
the "strong energy facies" of the reservoir, whereas the map of Y~* (x) (Fig.
6) represents the relative structural characteristics of the "weak energy facies"
of the reservoir.
This hypothesis allows the geological model to be refined from a sedimen-
tological and physical point of view, and moreover has an economic application
690 Jaquet

-U0

-TD

-601

-40

'111

Fig. 7. Map of the estimator E4*(x).

because sediments deposited in the strong-energy environment generally are

more porous.

CONCLUSIONS
From a strictly geostatistical point of view, results of the decomposition
of a regionalized variable, herein, thickness of the reservoir, obtained by fac-
torial kriging analysis are conclusive.
In order for this decomposition to have physical meaning, the different
variogram parameters must be related to the geology.
Compared to factorial kriging analysis, the analog method of spectral anal-
ysis can be applied only to data defined with a regular mesh; otherwise, the
data must first be interpolated at grid nodes. On the other hand, factorial kriging
analysis presents the advantage of working directly with the data, because the
variogram calculation does not require regularly spaced data.
Furthermore, factorial kriging analysis, used with a moving neighborhood,
allows a local approach. This methodology is particularly well suited for study
of heterogeneous geological models, because only a hypothesis of local station-
arity is required.
In this study, factorial kriging analysis has been applied only in the uni-
Factorial Kriging Analysis 691

vafiate case. The technique has additional potential, however, as it can be ap-
plied in the multivariate case (Matheron, 1982).

ACKNOWLEDGMENTS
I would like to thank particularly Beb Erdgas and Erdol GmbH, Hannover,
Germany, for making available data for this study. I also wish to thank Hans
Wackemagel and Alain Galli for their help and advice during this study, which
was carried out at the Centre de G6ostatistique de Fontainebleau.
BLUEPACK-3D was used for computations and drawing of the figures.

References
Chiles, J. P., and Guillen A., 1984, Variogrammes et Krigeages pour ta Gravim6trie et le Mag-
n6tisme: Sciences de la Terre, S6rie Informatique G6ologique, V. 21, p. 455-468.
Galli, A., Gerdil-Neuillet, F., and Dadou, C., 1984, Factorial Kriging Analysis: A Substitute to
Spectral Analysis of Magnetic Data, in Verly, G, et al. (Eds.), Geostatistics for Natural Re-
sources Characterization; Reidel, Dordrecht, The Netherlands, p. 543-557.
Galli, A., and Sandjivy, L., 1985, Analyse Krigeante et Analyse Spectrale: Sciences de la Terre,
S6rie Informatique G6ologique, V. 21, p. 115-124.
Galli, A., and Wackemagel, H., Multivariate Geostatistical Methods for Spatial Data Analysis, in
Diday, E., et al. (Eds.), Data Analysis and Information, V. 5, North Holland, in press.
Jaquet, O., 1985, Multivariate Geostatistical Study of a Gas Field: S-171, Centre de G6ostatistique,
Fontainebleau, France, 33 p.
Ma, Y. Z., and Royer, J. J., 1987, Le Comportement des Pond6rateurs en Analyse Krigeante.
Comparaison avec les Filtres Classiques Utilis6s en Traitement d'Images: S6minaire CFSG,
15-16 June 1987, Centre de G6ostatistique, Fontainebleau, France, 24 p.
Matheron, G., 1982, Pour une Analyse Krigeante de Donn6es R6gionalis~es: Note interne, N-732,
Centre de G6ostatistique, Fontainebleau, France, 22 p.
Sandjivy, L., 1984, Analyse Krigeante des Donn6es G6ochimiques (6tude d'un cas Monovariable
darts le ModUle Stationnaire): Sciences de la Terre, S6rie Informatique G6ologique No 18, p.
141-172.