You are on page 1of 13

13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geophysical Prospecting, 2012, 60, 400–412 doi: 10.1111/j.1365-2478.2011.00992.x

A practical review of geostatistical processing applied to geophysical


data: methods and applications
Matthieu Bourges1∗ , Jean-Luc Mari2 and Nicolas Jeannée1
1 Geovariances, 49 bis Avenue Franklin Roosevelt, 77210 Avon, France, and 2 IFP Energies nouvelles, IFP School, 1-4 Avenue de Bois-Préau,

92852 Rueil-Malmaison, France

Received January 2010, revision accepted June 2011

ABSTRACT
Nowadays, geostatistics is commonly applied for numerous gridding or modelling
tasks. However, it is still under used and unknown for classical geophysical applica-
tions. This paper highlights the main geostatistical methods relevant for geophysical
issues, for instance to improve the quality of seismic data such as velocity cubes or
interpreted horizons. These methods are then illustrated through four examples. The
first example, based on a gravity survey presents how a geostatistical interpolation
can be used to filter out a global trend, in order to better define real anomalies. In the
second case study, dedicated to refraction surveying, geostatistical filtering is used to
filter out acquisition artefacts and identify the main geological structures. The third
one is an example of porosity being integrated geostatistically with a seismic acoustic
impedance map. The last example deals with geostatistical time to depth conversion;
the interest of performing geostatistical simulations is finally discussed.

Key words: Acquisition, Anisotropy, Data processing, Gravity, Modelling.

r time to depth conversion, by means of kriging with exter-


INTRODUCTION
nal drift (for example, two-way time as external drift), or
Geostatistics are proved to be reliable and well adapted meth- collocated cokriging.
ods when dealing with gridding tasks and risk analysis in the r property modelling. Multivariate geostatistical methods
oil and gas industry (Dubrule 2003). Compared to classical (such as cokriging) are used, for instance, to generate con-
statistics, geostatistical methods take into account the spatial sistent porosity models, based on the correlation between
variability of the target parameter, in order to provide real- acoustic impedance and porosity.
istic spatial estimates together with a quantification of the The goal of this paper is to introduce non-specialists to the
associated uncertainty. use of geostatistics for geophysical applications. For this pur-
Therefore geostatistical methods are relevant at most steps pose, the first part presents the main geostatistical concepts.
of reservoir characterization, in particular for seismic process- These concepts are afterwards illustrated through four ex-
ing applications: amples. The two first examples correspond to near-surface
r data quality control and filtering of both random acquisi- applications where the use of geostatistics is not classical
tion noise and footprints. (Chilès and Guillen 1984; Peinado 1996; Vesnaver 2006; Bri-
r mapping of structures corresponding to different wave- dle 2008). The two last examples are more common applica-
lengths, by means of factorial kriging analysis. A variogram tions to property modelling and time to depth conversion.
structure is associated to each class of wavelength (for ex- Geostatistical methods applied for geological modelling are
ample, small, medium and large). not developed hereafter, this paper being focused on geophysi-
cal applications. Therefore, subjects such as facies simulations,
∗ E-mail: bourges@geovariances.com petrophysical simulations, uncertainties and risk analysis will

400 
C 2012 European Association of Geoscientists & Engineers
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 401

not be presented. Relevant references for the latter include r in case of several variables, investigating their correlation
Dubrule (2003) or Doyen (2007). through the display of a scatter diagram.

GEOSTATISTICAL WORKFLOWS Variography

Any geostatistical process begins with data quality control and Most geostatistical models rely on a variogram model, which
analysis, thus allowing an understanding of the data prior to quantifies the spatial variability of the target parameter
any further step. The variographic analysis is then performed (depth, velocity, porosity, etc) (Chilès and Delfiner 1999).
in order to measure the spatial variability of the data (contin- Variograms are used to describe the geological spatial con-
uous, discontinuous behaviour, stationarity, non-stationarity tinuity of properties. Through the measure of the spatial vari-
etc) leading to a variogram computation. Afterwards, geosta- ability, variograms are best suited for describing the continuity
tistical modelling such as kriging/cokriging can be performed of geological structures.
using this spatial information (variogram). Variogram inter- Firstly, the experimental variogram describes how the spa-
pretation can also be used to perform geostatistical filtering tial variability between data values increases with the distance
(for example, to remove artefacts). between the data. The experimental variogram γ (h) is defined
by:

1 
N(h)
Exploratory data analysis and quality control
γ (h) = (Z(xi + h) − Z(xi ))2 , (1)
2N(h) i=1
An in-depth exploratory data analysis allows quality control
of the data set in order to ensure its consistency by detecting where Z(x) denotes the target variable measured at location
and correcting/removing anomalies. x and N(h) corresponds to the number of pairs of points
This analysis usually relies on several tools allowing a full separated by a distance of h.
understanding of the data organization: In practice, the experimental variogram is computed as
r displaying a basemap of the data, in order to validate the follows: for each pair of data points and associated to
data location. the distance h between the two points, a value equal to
r computing the histogram, to check the homogeneity of the 0.5 ∗ (Z(x + h) − Z(x))2 is computed. The set of points corre-
statistical distribution. sponding to these values is called the variogram cloud. Then
r calculating the statistical characteristics: the mean (or aver- the points of the variogram cloud are averaged into classes of
age tendency) and the variance that measures the dispersion distances to obtain the experimental variogram (Fig. 1a). The
of the values around the mean. distance between each class is called the lag value.

20.

15.
Variogram

10.

5.

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

a) b)
Figure 1 Variographic analysis. a) Variogram cloud, classes of distances and experimental variogram; b) variogram map.


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
402 M. Bourges, J.-L. Mari and N. Jeannée

During the variographic analysis, it is important to check The shape of the variogram (Fig. 2a) reflects the behaviour
for the presence of potential anisotropies. Indeed the target of the variable (Yarus and Chambers 2006):
variable might present a more continuous behaviour in a given r at small distances, the presence of a discontinuity at the
direction, for instance, variograms along a channel and per- origin, called the nugget effect, corresponds to high values
pendicular to this channel. For this purpose directional var- nearby small values.
iograms are computed. When data are densely sampled, a r at large distances: if the variogram reaches a sill, the
variogram map can be obtained by automatically calculating variable is stationary (its mean and variance are con-
directional variograms in every direction (Fig. 1b). stant whatever the location in the space). If the variogram
The experimental variogram is not sufficient for a geosta- keeps increasing, the variable is non-stationary (the vari-
tistical process because there is only one value per class of able presents a trend, for instance its mean varies regarding
distance. This is one of the reasons why an analytical mathe- the location in the space).
matical function is fitted on the experimental variogram: the Several types of variogram basic models are available to
variogram model (Fig. 2a). describe the variability of data between the origin and range.
A variogram model ensures that the spatial variability is de- The most common mathematical models are detailed below
fined for every distance and azimuth. Also it allows to smooth and illustrated in Fig. 2(b).
possible statistical fluctuations. A variogram model is usually r On the one hand, spherical or exponential models, which
defined by three parameters: the type of basic mathematical correspond to continuous but not differentiable variables:
model, its sill and its range. Usually, the variability increases the variability increases rapidly even for small distances.
and becomes stabilized at a given variability level called the The corresponding mathematical formulae are:
sill. The range (zone of influence) is the distance at which r spherical model:
the variogram reaches the sill value. For distances superior
⎧   
to the range, there is no spatial correlation. For distances be- ⎪ 3
⎨γ (h) = C 3 h


1 h
for h < a,
tween the origin and the range, there is still spatial correlation. 2 a 2 a (2)
The spatial variability increase is expressed by a mathematical ⎪


model that is described below. γ (h) = C for h ≥ a,

Figure 2 Variogram modelling. a) Fitting of a variogram model (black curve) to an experimental variogram (red curve) and an illustration of
the main features; b) spherical (blue), exponential (orange), cubic (black) and Gaussian (red) models with a range of 30 m and a sill of 1.


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 403

r exponential model: Kriging estimation



−3h Kriging (Fig. 3b) provides an estimate of the target variable at
γ (h) = C 1 − exp . (3)
a a given location using the variogram model previously fitted
The exponential model reaches the sill only asymptotically. (Fig. 3a).
By convention, the practical sill is reached for the distance The estimation of the variable Z at a target x0 is derived
h = a (with a, the practical range). from a linear combination of the weighted data values at lo-
r On the other hand, smooth models such as Gaussian or cu- cations xα (in practice, data values may be chosen within a
bic ones correspond to differentiable variables (continuous moving neighbourhood centred at x0 ):
behaviour). 
Z∗ (x0 ) = λα Z(xα ). (6)
The corresponding mathematical formulae are:
r Gaussian model: The kriging weights λα are computed from the minimization
  
3h 2 of the kriging error:
γ (h) = C 1 − exp − . (4)
a    
Var Z0∗ − Z0 = − λα λβ γαβ + 2 λα γα0 . (7)
As for the exponential model, the Gaussian model reaches α β α

the sill only asymptotically. By convention, the practical sill With Z0∗ the estimated value at the target and Z0 the real
is reached for a distance h = a (with a, the practical range). unknown target value. α and β correspond to a couple of
r Cubic model:
data points. γαβ is the variogram corresponding to the couple
⎧     

⎪ h 2 35 h 3 7 h 5 3 h 7 of points (α, β) separated by a given distance xα − xβ . γα0 is

⎪ γ (h) = C 7 − + −

⎨ a 4 a 2 a 4 a the variogram between the target point x0 and the data point
⎪ for h < a, xα .



⎪ By drifting the variance and looking for the 0 values of
⎩ γ (h) = C for h ≥ a.
the derivatives we obtain the ordinary kriging system (with
(5) σ 2 , the kriging variance). The variable’s mean is unknown.
The variogram will be used for the geostatistical interpolation A particular constraint is applied on kriging weights to en-

method, called kriging. sure unbiasedness: the universality condition ( α λα = 1). To

2.5

2.0
Variogram : Porosity

1.5

1.0

0.5

0.0
0 1000 2000 3000
Distance (m)

a) b)
Figure 3 Kriging example (porosity). a) Experimental variogram and fitted model (spherical, range 3000 ft); b) kriging estimate, with data
points displayed (+).


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
404 M. Bourges, J.-L. Mari and N. Jeannée

account for this condition, a Lagrange parameter μ is added: where the f l (x) are usually monomials of the coordinates
 and the al are unknown fixed (or at least smoothly varying)

λβ γαβ + μ = γα0 arbitrary α, coefficients. Note that in general: i) there is no optimal way
β to separate m(x) and the residuals and ii) the decomposition

λα = 1, (8) is relevant to a particular scale of observation.
α Universal kriging provides an estimate of Z for this model,

σ2 = λα γα0 − μ. without having to estimate the trend itself. However, when
α
estimating this trend is precisely the goal, it can be obtained
Solving the kriging system leads to the kriging weights. directly from the universal kriging system by just equalling
The kriging variance σ 2 is computed from the variogram zero its right-hand side term. The residual component can be
and the kriging weights. It corresponds to the uncertainty obtained by subtracting the estimated mean from the estimate
attached to the kriging. of the variable itself: Y∗ (x) = Z∗ (x) − m∗ (x) or by filtering
directly the drift component (see below). Both approaches are
equivalent because of the linearity of the kriging system.
Cokriging/collocated cokriging
In practice, one option is to calculate the trend representing
When several correlated variables are available (for example, the non-stationary part from the best explanatory external
porosity and acoustic impedance, see application 3.3 below), drifts (monomials). Then the residuals are computed by re-
a multivariate geostatistical estimation called cokriging can be moving the trend from the data. A variogram is fitted on the
performed: the use of one (or several) correlated variable(s) residuals (stationary part). Finally, the variable is estimated
allows one to improve the estimation of the main property. with a model combining both stationary and non-stationary
An extension of cokriging called collocated cokriging is par- parts.
ticularly appropriate for the data integration of seismic sur-
veys for time to depth conversion or property modelling. In
Geostatistical filtering
this method, the seismic variable is incorporated at each target
location. It allows bringing more information to the estima- In seismic processing, a typical decomposition can be done be-
tion. So the estimate is expected to have a better consistency tween structure(s) corresponding to the signal and structure(s)
(Chilès and Delfiner 1999). corresponding to the noise. When dealing with nested vari-
These multivariate techniques require the computation and ograms (i.e., variograms with several structures), each struc-
fitting of a variogram model that contains simple variograms ture corresponds to a given scale of variability of the main
for each variable and a cross variogram measuring the spatial variable. One interest is to map or remove the component
correlation between the variables. corresponding to a given structure: this method is called fac-
torial kriging (Matheron 1982; Chilès and Delfiner 1999).
The underlying assumption is that the variable of interest,
Geostatistical trend extraction Z(x), can be decomposed into several components:
Sometimes the experimental variogram of a variable keeps Z(x) = Y1 (x) + · · · + Yn (x) + m(x), (11)
increasing. This means that the variable has a non-stationary
behaviour. In this case, a particular geostatistical approach where Yi (x) (i = 1, . . . , n) correspond to stationary and un-
might interestingly be applied: universal kriging (Chilès and correlated components of the phenomenon and m(x) repre-
Delfiner 1999). sents the unknown mean (large-scale trend). Once a variogram
In the universal kriging case, a non-stationary variable can model is fitted, the efficiency of the filtering technique depends
be decomposed as follows: on the ability to interpret each variogram structure in terms
of the spatial component.
Z(x) = m(x) + Y(x), (9) The variogram of the variable Z(x) is decomposed into as
where m(x) corresponds to the non-stationary trend and Y(x) many structures as Yi components.
correspond to stationary residuals. In simple cases, the trend γ (h) = γ1 (h) + γ2 (h) + · · · + γn (h). (12)
m(x) may be modelled by a simple polynomial expression:
 Strictly speaking, factorial kriging consists in estimating,
m(x) = al f l (x), (10) knowing the original data Z(xi ) and the variogram model of


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 405

one particular component Yi (x) (zero-mean). From a kriging non-stationary behaviour: the universal kriging approach was
system point of view, factorial kriging only requires a slight applied.
modification of the right-hand side term, keeping only the After the removal of a global linear trend on both coordi-
variogram component corresponding to the target Yi (x). It is nates (non-stationary part), a variogram analysis (Fig. 5b) was
sometimes convenient to present things in a slightly different performed on data residuals. Then, the raw Bouguer anomaly
way, particularly if one aims at filtering out one component was interpolated (100 m × 100 m) (Fig. 5a).
(for instance, acquisition noise) instead of estimating one par- Geostatistical filtering was used to decompose the Bouguer
ticular component. In this case, the kriging system is solved anomaly into a global trend corresponding to a regional
with, in the right-hand side term, the entire variogram model anomaly (Fig. 6a) and a short scale structure correspond-
except the component to be filtered out. ing to a residual anomaly (Fig. 6b). On the residual Bouguer
anomaly map, one can clearly see a negative anomaly oriented
north west–south east. It delineates a furrow in the molassic
APPLICATIONS basement.
Gravity data

The gravity data set was gathered by the Geophysical Insti- Refraction survey
tute, University of Lausanne, under the direction of Prof. Ray- The university of Poitiers (France) has a hydrogeological ex-
mond Olivier (Atlas gravimétrique de la Suisse au 1/10000). perimental site built nearby the campus for the sole purpose
The area of Aubonne is located between Geneva and Lau- of providing facilities to develop long-term monitoring and
sanne in Switzerland. The geology of this area consists in experiments for a better understanding of flow and transfers
a Molasse basement (Chattien sandstone) dug with furrows in fractured rocks. The concerned aquifer, 20–130 m in depth,
filled by quaternary sediments. The Bouguer model is com- consists of tight karstic carbonates of Middle Jurassic age. It
puted using the 1967 Ellipsoı̈de and the model density is 2.4. lies on the borderline, named the ‘Poitou threshold’, between
This density corresponds to the Molassic basement. The to- the Paris and the Aquitaine sedimentary basins. The top of the
pographic correction was applied up to 167 km (Olivier and reservoir was initially flat and horizontal 150 million years
Simard 1981). ago but has been eroded and weathered since, during the Cre-
Firstly, the experimental variogram computed on the raw taceous and Tertiary ages. It is shaped today by hollows and
anomaly keeps increasing (Fig. 4). So this variable shows a bumps with a magnitude reaching up to 35 m.
Refraction seismic surveying, described in detail by Mari
and Porel (2007), was used to map the irregular shape of the
9
top of karstic reservoir.
Variogram : Bouguer Anomaly (mGals)

8 Due to the limitations of the area, the length of the seismic


line could not exceed 250 m in the in-line direction. In the
7 cross-line direction, the extension of the area does not exceed
300 m. As a result, 20 receiver lines were implemented, with
6
a 15 m distance between the adjacent lines. Each line was
5 composed of 48 geophones. A 5 m distance between the two
adjacent geophones was selected. An explosive source (25 g)
4 was detonated. Such a source makes it easy to identify and pick
first arrivals. A direct shot and a reverse shot were recorded
3
per receiver line. The weathering velocity that is in the area
2 around 850 m/s is given by the slope of the direct wave. To
obtain the velocity of the refractor (top of the reservoir) and its
1 depth, Hagedoorn’s plus–minus method was applied on each
line independently and the weathering zone’s depth values
0
0 1000 2000 3000 4000 5000 were interpolated on a 5 m × 5 m grid (Fig. 8a) (Hagedoorn
Distance (m)
1959). In the area, the velocity of the refractor is of the order
Figure 4 Experimental variogram of the raw Bouguer anomaly. of 3350 m/s.


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
406 M. Bourges, J.-L. Mari and N. Jeannée

0.05
156000

Variogram : Bouguer Anomaly Residuals


155000

154000 0.04

153000 (mGals)
152000 -47 0.03
-48
Y (m)

-49
151000 -50
-51
-52
150000 -53
-54 0.02
-55
149000 -56
-57
-58
148000 -59
-60
-61 0.01
147000 -62

146000
515000 517500 520000 522500 525000
0.00
X (m) 0 1000 2000 3000
Distance (m)

Figure 5 Gravimetry survey data analysis. a) Raw data locations (black dots) and interpolated map (universal kriging); b) experimental (thin
line) and fitted variograms (bold).

Figure 6 Detection of a furrow on a Bouguer anomaly data set. a) Regional Bouguer anomaly obtained by kriging of the global trend; b) residual
Bouguer anomaly after the removal of the global trend. A furrow with an S-SE orientation is visible.

The velocity model obtained by the plus–minus method for anisotropic spatial behaviour (see Fig. 7a). The variogram
after geostatistical processing and the depth map were used as along the in-line direction shows a higher continuity than the
a static model for tomographic inversion (Mari and Mendes variogram along the cross-line direction. These variograms are
2010). fitted by a model comprising five components (see Fig. 7a):
Geostatistical methods were used to study the spatial vari-
γ (h) = C0 + γaa (h) + γgh (h) + γg f 1 (h) + γg f 2 (h), (13)
ability of the seismic maps and to detect some footprints due
to the acquisition procedure. Directional variograms are com- with
puted on the weathering zone’s depth variable along the two r C0 a nugget effect, interpreted as random acquisition noise
main directions (in-line (IL) and cross-line (XL)) to search (even in the in-line direction);


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 407

25
552 31676
598
644 3045030354
29840
20 690 15
25070
Variogram : depth

Variogram : depth
736 24960
XL
1197
1764 1575 1386 20608
15 782
1953 10
828
2142 13526

2331
10 874

2520 IL
8160
5
920
5
2709
After removal of an
945 All data 900 anomalic line
0 0
0 50 100 150 0 25 50 75 100 125
Distance (m) Distance (m)

a) b)
Figure 7 Variograms of the depth of the weathering zone. a) Anisotropic variogram model with added variability along the cross-line direction
(XL); b) omnidirectional variogram after the removal of an anomalic line.

r γaa (h) an anisotropic spherical short range (5 m) component along the cross-line direction was due to this anomalic line.
in the cross-line direction due to acquisition artefacts; Therefore an omnidirectional variogram is computed, most
r γgh (h) a mid-range (50 m) cubic structure corresponding to interest being in the small-scale variability, up to 50–75 m.
geological heterogeneities; The variogram model is constituted by a nugget effect, a cu-
r γg f 1 (h) and γg f 2 (h) two long-range structures correspond- bic structure with a range of 57 m and a long scale spherical
ing to long wavelength geological features modelled with structure with a range of 145 m (see Fig. 7b). Finally a krig-
spherical structures with a range of 100–400 m. ing with this model is performed to obtain the filtered depth
The additional variability source in the cross-line direction, map of the weathering zone (Fig. 8d). Factorial kriging is
attributable to acquisition artefacts is modelled by a spherical used to filter out both the nugget and the small-scale struc-
basic structure with a range equal to 5 m in the cross-line ture (cubic with a range of 57 m). This map enables one to
direction. This range is chosen in consistence with the expected identify the geological structures at the scale of the map. The
mapping resolution (5 m × 5 m). Also the large-scale spherical azimuth of the main structure is N90; it corresponds to a frac-
model has a smaller range (275 m compared to 400 m) in the tured area. The reference is true north, indicated by the arrow
cross-line direction. N0.
The nugget effect and the small-scale spherical model (5 m
in the cross-line direction), which have no physical meaning
Porosity modelling
and are due to acquisition artefacts are filtered through fac-
torial kriging. The filtered depth map is shown in Fig. 8(b). The data set is constituted of a set of wells with porosity val-
By substracting the filtered map to the raw one, an anomalic ues and a seismic grid of normalized acoustic impedance (see
line is detected around Y = 150 m (Fig. 8c). The residual map Fig. 9b). To model porosities, one first option is to use the
shows that the depth of the weathering zone at the location porosity known at the wells location and interpolate between
of the anomalic line is underestimated. This is due to a lo- the wells by means of kriging (Fig. 10a). Partly due to the small
cal under estimation of the velocity in the weathering zone number of wells, this outcome is really smooth and does not
for that line. The variographic analysis is performed again seem geologically consistent. More dense information can be
after the removal of the anomalic line. The added variability integrated in order to improve the estimation of porosity. As


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
408 M. Bourges, J.-L. Mari and N. Jeannée

Figure 8 Geostatistical processing applied to refraction data. a) Depth map after linear interpolation in the cross-line direction (5 m × 5 m); b)
weathering zone’s depth map after filtering by factorial kriging; c) detection of an anomalic line (in-line 150); d) weathering zone’s depth map
after filtering of small-scale structures.

porosity is linked to acoustic impedance, it will be relevant to This is the case: the correlation at wells between porosity and
use the dense seismic acoustic impedance information. So a acoustic impedance is 0.83 (see Fig. 10b).
collocated cokriging of porosity integrating the seismic infor- For this geostatistical interpolation, the collocated cok-
mation is performed. riging takes into account the spatial correlation through a
For a collocated cokriging, the integration of seismic in- bivariate variogram model between porosity and acoustic
formation will be more important if the correlation is high. impedance. The collocated cokriging of porosity is displayed


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 409

12
rho=0.83

11

10
Porosity (%)

-1.0 -0.5 0.0


Norm Ai (Pa.s/m)

a) b)
Figure 9 Data analysis. a) Correlation between porosity and normalized acoustic impedance; b) 2D map of normalized acoustic impedance.

Figure 10 Porosity modelling using acoustic impedance. a) Map of porosity kriging; b) map of porosity collocated cokriging using normalized
acoustic impedance as the secondary variable.

in Fig. 10(b). This outcome presents more variability than the 3D velocity modelling and time to depth conversion
kriging outcome. This added variability comes from the seis-
The data set is composed of:
mic normalized acoustic impedance integrated at the target lo- r a two-way time (TWT) map of the top reservoir horizon.
cation through the collocated cokriging process (see Fig. 9b). r a 3D seismic survey of the root-mean-squared (RMS) ve-
By construction, this estimate still honours porosity values at
locity already gridded.
data locations.


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
410 M. Bourges, J.-L. Mari and N. Jeannée

r a set of wells with depth and two-way time variables known variograms of the model are proportional to the simple
at the intercepts with key horizons. variogram of the secondary ‘dense’ seismic variable. A high
The goals are to obtain a 3D cube of average velocity and correlation between the two variables is required.
to perform the time to depth conversion of the top reservoir r the variogram modelling is performed on the seismic vari-
horizon. One possible solution to perform the time to depth able (see Fig. 11b). The vertical variogram keeps increasing
conversion is to model the average velocity on the top reser- due to vertical non-stationarity.
voir horizon and then calculate the depth through the usual r the 3D modelling of average velocity using the RMS veloc-
relationship between depth, time and velocity. ity as a secondary variable is performed. A 3D model of
First of all, the average velocity is calculated at the wells (us- average velocity is obtained.
ing two-way time and depth). Then, due to the relationship r average velocity values are extracted along the two-way
between average velocity and RMS velocity, it is interesting time seismic horizon (see Fig. 12a).
to use the 3D seismic survey as secondary correlated informa- r finally, depth values are calculated on the top reser-
tion. The purpose is to obtain a 3D model of average veloc- voir using two-way time and average velocity (see
ity, extract the average velocity values corresponding to the Fig. 12b).
top reservoir horizon and then calculate the horizon’s depth There are several implications in performing such a 3D
values. geostatistical time to depth conversion. Firstly, a reliable
We proceed as follows: 3D velocity model is obtained, taking into account of
r RMS velocity values are migrated at the wells location: the the spatial variability and integrating the seismic informa-
correlation is high between the average velocity and RMS tion. Therefore, reliable depth horizons can be extracted.
velocity (∼0.86) (see Fig. 11a). Collocated cokriging can be Secondly, the uncertainties attached to the velocity model
used with the RMS velocity as a secondary and collocated can be quantified through geostatistical simulations (sev-
variable. As there are only eight wells, collocated cokriging eral equiprobable realizations of the velocity models are ob-
with the multivariate model fitted at the wells cannot be tained). Several corresponding realizations of depth horizons
performed, due to the lack of wells information. Therefore a may be extracted, thus allowing assessment of the uncer-
collocated cokriging with the assumption of Markov-Bayes tainties for the structural modelling and for the volumetrics
is performed: we consider that all the simple and cross computation.

rho=0.87 50000
4100
Variogram : Vrms on grid

4000 40000 D-90


Vavg (m/s)

3900
30000
N17
3800
20000

3700
N107
10000

3600

0
3000 3250 3500 3750 4000 4250 0 1 2 3 4 5 6 7
Vrms (m/s) Distance (km)
a) b)
Figure 11 Data analysis. a) Correlation cloud at wells between average velocity (Vavg) and RMS velocity (Vrms); b) variogram models computed
on the RMS velocity seismic variable (Vrms).


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Geostatistical processing applied to geophysical data 411

Figure 12 Time to depth conversion. a) Average velocity extracted from the time of the top reservoir horizon; b) depth of the top reservoir
horizon.

CONCLUSIONS allow one to quantify the uncertainties attached to the mod-


elling of any variable and are of the upmost importance for
The aim of this paper was to introduce the main geostatistical
reservoir risk analysis.
workflows valuable for geophysical processing. Topics such
as data analysis (statistics, variogram), geostatistical interpo-
lation (kriging), data integration (collocated cokriging) and ACKNOWLEDGEMENTS
geostatistical filtering were covered and illustrated through
We thank the University of Poitiers and IFP for their per-
case studies.
mission to use the refraction data recorded on the Poitiers
The first example, based on a gravity data set, shows how
hydrogeological site. We thank Raymond Olivier and Insti-
to apply geostatistics when dealing with a non-stationary vari-
tut Geophysique de Lausanne for their permission to present
able. Regional trend and small-scale structures (residuals) of
the Aubonne gravity data set. We are grateful to Dominique
a Bouguer anomaly variable were extracted. The residuals
Chapellier for providing useful information and help on
present a negative anomaly that was interpreted as a furrow
gravity processing. Geostatistical results were obtained with
in the bedrock.
ISATIS software1 . We thank the anonymous reviewers for
The second example deals with refraction data. Geosta-
their fruitful comments regarding this paper.
tistical filtering allowed detecting and suppressing acquisi-
tion footprints. Afterwards, long wavelength structures corre-
sponding to a fractured area were extracted through factorial REFERENCES
kriging.
The third case study is a typical illustration of data integra- Bridle R. 2008. Gaining a geostatistical advantage in near-surface
modelling. 78th SEG meeting, Las Vegas, Nevada, USA, Expanded
tion for petrophysical modelling. Seismic acoustic impedance
Abstracts, 1218–1222.
and wells with porosity were used. To perform the integra- Chilès J.P. and Delfiner P. 1999. Geostatistics Modelling Spatial Un-
tion, a method called collocated cokriging was applied. In certainty. Wiley.
this case, due to the strong relationship between porosity and Chilès J.P. and Guillen A. 1984. Variogrammes et krigeages pour la
acoustic impedance, variographic analysis was performed on gravimétrie et le magnétisme. Sciences de la Terre, Série Informa-
the seismic attribute. tique Géologique 20, 455–468.
Deraisme et al. 2009. ISATIS Case Studies. Geovariances & Ecole des
The last example is a time to depth conversion using a 3D
Mines de Paris, version 9, 724p.
seismic RMS velocity cube, average velocity at wells and a Doyen P.M. 2007. Seismic Reservoir Characterization: An Earth
TWT map at a key horizon. As for the petrophysical mod- Modelling Perspective. EAGE.
elling case study, collocated cokriging was performed. The Dubrule O. 2003. Geostatistics for Seismic Data Integration on Earth
interests of performing stochastic simulations for quantifying Models, Distinguished Instructor Series, N◦ 6. SEG/EAGE.
Hagedoorn G.J. 1959. The plus-minus method of interpreting seismic
uncertainties was discussed.
refraction sections. Geophysical Prospecting 7, 158–182.
Finally, this paper focused on interpolation methods.
Stochastic approaches such as geostatistical simulations were
only mentioned. But it is important to keep in mind that in- 1ISATIS Case Studies: Geovariances & Ecole des Mines de Paris,
terpolations and simulations are complementary. Simulations version 9 (2009).


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412
13652478, 2012, 3, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1365-2478.2011.00992.x by Universidad Nacional Autonoma De Mexico, Wiley Online Library on [11/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
412 M. Bourges, J.-L. Mari and N. Jeannée

Mari J.L. and Mendes M. 2010. 3D Near surface imaging thanks model for terraincorrection in rugged topography. Geophysics 46,
to refraction tomography and delay method. 72nd EAGE meeting, 1054–1056.
Barcelona, Spain, Expanded Abstracts, M034. Peinado A. 1996. 3D refraction statics: A geostatistical approach.
Mari J.L. and Porel G. 2007. 3D seismic imaging of a near-surface 66th SEG meeting, Denver, Colorado, USA, Expanded Abstracts,
heterogeneous aquifer: A case study. Oil and Gas Science and Tech- 1654–1657.
nology 63, 179–201. doi:10.2516/ogst/2007077 Vesnaver A., Bridle R., Henry B., Ley II R., Rowe W. and Wyllie A.
Matheron G. 1982. Pour une analyse krigeante des données 2006. Geostatistical integration of near-surface geophysical data.
régionalisées. Internal report N-732, Centre de Géostatistique Geophysical Prospecting 54, 667–680.
et de Morphologie Mathématique, Ecole des Mines de Yarus J.M. and Chambers R.L. 2006. Practical geostatistics – An
Paris. armchair overview for petroleum reservoir engineers. Journal of
Olivier R.J. and Simard R.G. 1981. Improvement of the conic prism Petroleum Technology 58, 78–86.


C 2012 European Association of Geoscientists & Engineers, Geophysical Prospecting, 60, 400–412

You might also like