P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK

Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Mathematical Geology, Vol. 33, No. 4, 2001

Teacher’s Aide
Variogram Interpretation and Modeling1
Emmanuel Gringarten2 and Clayton V. Deutsch3

The variogram is a critical input to geostatistical studies: (1) it is a tool to investigate and quantify
the spatial variability of the phenomenon under study, and (2) most geostatistical estimation or sim-
ulation algorithms require an analytical variogram model, which they will reproduce with statistical
fluctuations. In the construction of numerical models, the variogram reflects some of our understanding
of the geometry and continuity of the variable, and can have a very important impact on predictions
from such numerical models. The principles of variogram modeling are developed and illustrated with
a number of practical examples. A three-dimensional interpretation of the variogram is necessary to
fully describe geologic continuity. Directional continuity must be described simultaneously to be con-
sistent with principles of geological deposition and for a legitimate measure of spatial variability for
geostatistical modeling algorithms. Interpretation principles are discussed in detail. Variograms are
modeled with particular functions for reasons of mathematical consistency. Used correctly, such vari-
ogram models account for the experimental data, geological interpretation, and analogue information.
The steps in this essential data integration exercise are described in detail through the introduction of
a rigorous methodology.

KEY WORDS: kriging, stochastic simulation, covariance, zonal and geometric anisotropy.

INTRODUCTION

The variogram has been used widely to quantify the spatial variability of spatial
phenomena for many years; however, calculation and interpretation principles
have advanced slowly. This is particularly true in the petroleum industry due to
the limited number of well data. The preliminary steps of variogram calculation,
interpretation, and modeling are often performed hastily or even skipped altogether.
This practice should be reversed and much more attention devoted to establishing

1Received 2 August 1999; accepted 11 April 2000.

2Landmark Graphics Corp., Austin, Texas, USA, presently with T-Surf Corp., Austin, Texas, USA.
e-mail: [email protected]
3Department of Civil & Environmental Engineering, University of Alberta, Edmonton, Alberta, Canada.

e-mail: [email protected]

507
0882-8121/01/0500-0507$19.50/1 °
C 2001 International Association for Mathematical Geology
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

508 Gringarten and Deutsch

a robust model of spatial variability (variogram) before proceeding with building

numerical geological models.
Geostatistical model-building algorithms such as sequential Gaussian simu-
lation, sequential indicator simulation, and truncated Gaussian simulation take an
input variogram model and create a three-dimensional (3D) model constrained to
local data and the variogram model. The variogram has an extremely important
role to play in the appearance and behavior of the resulting 3D models.
Thorough variogram interpretation and modeling are important prerequisites
to 3D model building. The practice of variogram modeling and the principle of the
Linear Model of Regionalization have been covered in many texts (e.g., Journel
and Huijbregts, 1978; Armstrong, 1984; Cressie, 1993; Olea, 1995; Goovaerts,
1997). However, none have presented a strict and rigorous methodology to easily
and systematically produce a licit and consistent 3D variogram model. We present
a methodology of variogram interpretation and modeling whereby the variance is
divided into a number of components and explained over different length scales
in different directions.

THE VARIOGRAM

The variogram has been defined in many books and technical papers. For
completeness, however, we recall the definition of the variogram and related statis-
tics. Consider a stationary random function Y with known mean m and variance
¾ 2 . The mean and variance are independent of location, that is, m(u) = m and
¾ 2 (u) = ¾ 2 for all locations u in the study area. Often there are areal and vertical
trends in the mean m, which are handled by a deterministic modeling of the mean
and working with a residual from the locally variable mean. The variogram is
defined as
© 2ª
2° (h) = Var[Y (u) − Y (u + h)] = E [Y (u) − Y (u + h)] (1)

In words, the expected squared difference between two data values separated by
a distance vector h is the variogram. The semivariogram ° (h) is one half of the
variogram 2 ° (h). To avoid excessive jargon we simply refer to the variogram,
except where mathematical rigor requires a precise definition. The variogram is
a measure of variability; it increases as samples become more dissimilar. The
covariance is a statistical measure that is used to measure correlation (it is a measure
of similarity):

C(h) = E{[Y (u) · Y (u + h)]} − m 2 (2)

By definition, the covariance at h = 0, C(0), is the variance ¾ 2 . The covariance

C(h) is 0.0 when the values h-apart are not linearly correlated.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 509

Expanding the square in Equation (1) leads to the following relation between
the semivariogram and covariance:

° (h) = C(0) − C(h) or C(h) = C(0) − ° (h) (3)

This relation depends on the model decision that the mean and variance are constant
and independent of location. These relations are the foundation for variogram
interpretation. That is, (1) the “sill” of the variogram is the variance, which is the
variogram value that corresponds to zero correlation; (2) the correlation between
Y (u) and Y (u + h) is positive when the variogram value is less than the sill; and
(3) the correlation between Y (u) and Y (u + h) is negative when the variogram
exceeds the sill. This is illustrated by Figure 1, which shows three h-scatterplots
corresponding to three lags on a typical semivariogram. Geostatistical modeling
generally uses the variogram instead of the covariance for mainly historical reasons.
The flexibility of the variogram to handle cases of infinite variance is of little
practical consequence.
A single variogram point ° (h) for a particular distance and direction h is
straightforward to interpret and understand. Practical difficulties arise from the
fact that we must simultaneously consider many lag vectors h—that is, many
distances and directions. The variogram is a measure of “geological variability”
vs. distance. The “geologic variability” is quite different in different directions;
for example, in sedimentary formations there is typically much greater spatial
correlation in the horizontal plane.

Understanding Variogram Behavior

The link between geological variations and observed variogram behavior must
be understood for reliable variogram interpretation and modeling. Figure 2(A–C)
shows three geologic images and corresponding semivariograms in the vertical
and horizontal directions for each image. In practice, we do not have an exhaustive
image of the variable and the variogram behavior must be interpreted and related
to geological principals from directional variograms. The primary variogram be-
haviors are as follows:
1. Randomness or lack of spatial correlation: Certain geological variations
appear to have no spatial correlation. These random variations are the
result of deterministic geological processes. At some scales, however,
the processes are highly nonlinear and chaotic leading to variations that
have no spatial correlation structure. Typically, only a small portion of the
variability is explained by random behavior. For historical reasons, this
type of variogram behavior is called the nugget effect. In early mining
geostatistics, the presence of gold nuggets in drillhole samples would lead
to apparently random variations—hence, nugget effect.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

510 Gringarten and Deutsch

Figure 1. Semivariogram with the h-scatterplots corresponding to three different lag distances. Note that
the correlation on the h-scatterplot is positive when the semivariogram value is below the sill, zero when the
semivariogram is at the sill, and negative when the semivariogram is above the sill.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 511

2. Decreasing spatial correlation with distance: Most depositional processes

impart spatial correlation to petrophysical properties. The magnitude of
spatial correlation decreases with separation distance until a distance at
which no spatial correlation exists, the range of correlation. In all real
depositional cases the length scale or range of correlation depends on
direction—that is, the vertical range of correlation is much less than the
horizontal range due to the larger lateral distance of deposition. Although
the correlation range depends on distance, the nature of the decrease in
correlation is often the same in different directions. The reasons for this
similarity are the same reasons that underlie Walther’s Law, which is a
principle of sedimentary geology that tells us horizontal variability is en-
countered in the vertical direction; just at different length scales. This type
of variogram behavior is called geometric anisotropy.
3. Geologic trends: Virtually all geological processes impart a trend in the
petrophysical property distribution—for example, fining or coarsening up-
ward or the systematic decrease in reservoir quality from proximal to distal
portions of the depositional system. Such trends can cause the variogram
to show a negative correlation at large distances. In a fining upward sed-
imentary package, the high porosity at the base of the unit is negatively
correlated with low porosity at the top. The large-scale negative correla-
tion indicative of a geologic trend show up as a variogram that increases
beyond the sill variance ¾ 2 . As we will see later, it may be appropriate to
remove systematic trends prior to geostatistical modeling.
4. Areal trends: These have an influence on the vertical variogram—that
is, the vertical variogram will not encounter the full variability of the
petrophysical property. There will be positive correlation (variogram ° (h)
below the sill variance ¾ 2 ) for large distances in the vertical direction.
This type of behavior is called zonal anisotropy. A schematic illustration
of this is given in Figure 3.
5. Stratigraphic layering: There are often stratigraphic layer-like features or
vertical trends that persist over the entire areal extent of the study area.
These features lead to positive correlation (variogram ° (h) below the sill
variance ¾ 2 ) for large horizontal distances. Although large-scale geologic
layers are handled explicitly in the modeling, there can exist layering
and features at a smaller scale that cannot be handled conveniently by
deterministic interpretation. This type of variogram behavior is also called
zonal anisotropy because it is manifested in a directional variogram that
does not reach the expected sill variance.
6. Geologic cyclicity: Geological phenomenon often occur repetitively over
geologic time leading to repetitive or cyclic variations in the facies and
petrophysical properties. This imparts a cyclic behavior to the variogram—
that is, the variogram will show positive correlation going to negative
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

512 Gringarten and Deutsch

 P1: Vendor/GDX/GIR/GAY
Mathematical Geology [mg]
PP103-297990
P2: GDX/LZX/FGL

Variogram Interpretation and Modeling

Figure 2. A, image showing geometric anisotropy with corresponding vertical (red) and horizontal (black) variograms. The image
March 29, 2001
Tally: FTK

is 100 pixels by 50 pixels and the variograms were calculated to a total distance of 35 pixels in the both directions. B, image showing
cyclicity in the vertical direction (red variogram) and zonal anisotropy in the horizontal direction (black variogram). The image is
100 pixels by 50 pixels and the variograms were calculated to a total distance of 35 pixels in both directions. C, image showing a trend
in the vertical direction (red variogram) and zonal anisotropy in the horizontal direction (black variogram). The image is 100 pixels
10:48

by 50 pixels and the variograms were calculated to a total distance of 35 pixels in both directions.
513
Style file version June 30, 1999
 514
P1: Vendor/GDX/GIR/GAY
Mathematical Geology [mg]
PP103-297990
P2: GDX/LZX/FGL
March 29, 2001
Tally: FTK
10:48

Figure 3. In presence of areal trends (illustrated at the left) each well will not “see” the full range of variability, that is, wells in the
higher valued areas (e.g., well A) encounter mostly high values whereas wells in the lower valued areas (e.g., well B) encounter mostly
low values. The vertical variogram in this case does not reach the total variability, that is, it shows a zonal anisotropy.
Style file version June 30, 1999

Gringarten and Deutsch

P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 515

correlation at the length scale of the geologic cycles going to positive

correlation and so on. These cyclic variations often dampen out over large
distances as the size or length scale of the geologic cycles is not perfectly
regular. For historical reasons, this is sometimes referred to as a hole
effect. In early mining geostatistics, cyclicity was observed in “down-hole”
variograms; hence the name hole effect.

Real variograms almost always reflect a combination of these different variogram

behaviors. Considering the three images and their associated variograms presented
in Figure 4, we see evidence of all the behaviors mentioned above: nugget effect
most pronounced on the top image, geometric anisotropy and zonal anisotropy on
all, a trend on the middle one, and cyclicity most pronounced on the bottom image.
The top image is an example of migrating ripples in a man-made eolian sandstone (a
modified 2 foot by 5 foot image from the U.S. Wind Tunnel Laboratory), the central
image is a 5 inch by 10 inch example of convoluted and deformed laminations
from the Brazos River (heavily modified from image on page 131 of Sandstone
Depositional Environments, Scholle and Spearing, 1982), and the bottom image
is a modified 75 foot by 125 foot photograph of wedge and tabular cross strata
from near Moab, Utah. These are only illustrations to show real variograms from
near-exhaustive data.
The intent of this paper is to present a systematic procedure for variogram
interpretation and modeling of real geological features.

Requirement for a 3D Variogram Model

All directional variograms must be considered simultaneously to understand-

ing the variogram behavior. The experimental variogram points are not used di-
rectly in subsequent geostatistical steps; a parametric variogram model is fitted to
the experimental points. A detailed methodology for this fitting is a central theme
of this paper. There are a number of reasons why experimental variograms must
be modeled:

1. The variogram function ° (h) is required for all distance and direction
vectors h within the search neighborhood of subsequent geostatistical cal-
culations; however, we only calculate the variogram for specific distance
lags and directions (often, only in the principal directions of continuity).
There is a need to interpolate the variogram function for h values where
too few experimental data pairs are available. In particular, the variogram
is often calculated in the horizontal and vertical directions, but geostatis-
tical simulation programs require the variogram in off diagonal directions
where the distance vector simultaneously contains contributions from the
horizontal and vertical directions.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

516 Gringarten and Deutsch

 P1: Vendor/GDX/GIR/GAY
Mathematical Geology [mg]
PP103-297990
P2: GDX/LZX/FGL

Variogram Interpretation and Modeling

Figure 4. Three different geologic images with the corresponding directional variograms. Note the cyclicity (the undulations in both
March 29, 2001
Tally: FTK

directional variograms on the bottom example), trends (particularly noticeable in the vertical direction of the middle variogram), geometric
anisotropy (the shorter range of correlation in the vertical direction on all variograms), and zonal anisotropy (the horizontal variogram
not raching the sill, which is most noticeable on the top variogram).
10:48

517
Style file version June 30, 1999
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

518 Gringarten and Deutsch

2. There is also a need to introduce geological information regarding ani-

sotropy, trends, sampling errors, and so on in the model of spatial correla-
tion. As much as possible, we need to filter artifacts of data spacing and data
collection practices and make the variogram represent the true geological
variability.
3. Finally, the covariance counterpart of the variogram measure ° (h), for
all distance and direction vectors h, must have mathematical property
of positive definiteness, that is, we must be able to use the variogram, or
its covariance counterpart, in kriging and stochastic simulation. A positive
definite model ensures that the kriging equations can be solved and that the
kriging variance is positive—in other words, a positive definite variogram
is a legitimate measure of distance.

For these reasons, geostatisticians have fit variograms with specific known positive
definite functions like the spherical, exponential, Gaussian, and hole effect vari-
ogram models (Journel and Huijbregts, 1978; Cressie, 1993; Christakos, 1984).
It should be mentioned that any positive definite variogram function can be used,
including tabulated variogram or covariance values. The use of any arbitrary func-
tion or nonparametric table of variogram values would require a check to ensure
positive definiteness (D. E. Myers, 1991). In general, the result will not be positive
definite and some iterative procedure would be required to adjust the values until
the requirement for positive definiteness is met.
With a “correct” variogram interpretation, the use of traditional paramet-
ric models is adequate to achieve a good fit. In fact, the traditional parametric
models permit all geological information to be accounted for and realistic vari-
ogram behavior to be fit. Moreover, the use of traditional variogram models allows
straightforward transfer to existing geostatistical simulation codes (Deutsch and
Journel, 1997).

Importance of the Variogram in Geostatistics

The variogram is used by most geostatistical mapping and modeling algo-

rithms. Object-based facies models and certain iterative algorithms, such as simu-
lated annealing, do not use variograms. Not only is the variogram used extensively,
it has a great effect on predictions. Occasionally there are enough data to control
the appearance and behavior of the numerical models; however, these cases are
infrequent and of lesser importance than the common case of sparse data control.
The available data are too widely spaced to provide effective control on the nu-
merical model. The variogram provides the only effective control on the resulting
numerical models.
The lack of data, which makes the variogram important, also makes it difficult
to calculate, interpret, and model a reliable variogram (Cressie and Hawkins, 1980;
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 519

Genton, 1998a). Practitioners have been aware of this problem for some time with
no satisfactory solution. Variogram modeling is important and the “details” often
have a crucial impact on prediction. In particular, the treatment of zonal anisotropy
and systematic vertical or horizontal trends is critical.

VARIOGRAM INTERPRETATION AND MODELING

Establishing the Correct Variable

Variogram calculation is preceded by selection of the variable of interest. It is

rare in modern geostatistics to consider untransformed data. The use of Gaussian
techniques requires a prior Gaussian transform of the data and the variogram model
of these transformed data. Indicator techniques require an indicator coding of the
facies or of the continuous variable at a series of thresholds. Also, systematic areal
or vertical trends should be removed from the variable prior to transformation and
variogram calculation; see the next section.
When data is skewed or has extreme high or low values; estimated variograms
often exhibit erratic behaviors. Various robust alternatives to the traditional vari-
ogram have been proposed in the literature (Cressie and Hawkins, 1980; Genton,
1998a). These include madograms, rodograms, general and pairwise relative vari-
ograms. They are generally used to determine ranges and anisotropy, which cannot
be detected with the traditional variogram. However, these measures of spatial
variability should not be modeled, as they cannot serve as input for subsequent
estimation or simulation algorithms. Instead, it is recommended to transform the
data to Normal space before performing variogram calculations, e.g., with the
normal score transform (Deutsch and Journel, 1997). This has some important
advantages: (1) the difference between extreme values is dampened and (2) the
theoretical sill is known to be 1. Also, some algorithms (e.g., p-fields) may require
that variogram calculations be performed on a Uniform score transformation of
the data. However, the Uniform score and Normal score variograms are generally
so similar that the latter can be used most of the time.

Removing the Trend

As mentioned above, the first important step in all geostatistical modeling

exercises is to establish the correct property to model and to make sure (inasmuch
as it is possible) that this property is stationary over the domain of the study. Indeed,
if the data shows a systematic trend, this trend must be modeled and removed
before variogram modeling and geostatistical simulation. Variogram analysis and
all subsequent estimations or simulations are performed on the residuals. The
trend is added back to estimated or simulated values at the end of the study.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

520 Gringarten and Deutsch

There are problems associated with defining a reasonable trend model and
removing the deterministic portion of the trend; however, it is essential to consider
deterministic features such as trends deterministically. The presence of a significant
trend makes the variable nonstationary, that is, it is unreasonable to expect the mean
value to be independent of location. Residuals from some simple trend model
are easier to consider stationary. Removing a trend by estimaiton from the data
themselves can introduce a bias; however, this bias is considered less significant
than the errors introduced by leaving the trend alone.
Trends in the data can be identified from the experimental variogram, which
keeps increasing above the theoretical sill, see earlier discussion. In simple terms,
this means that as distances between data pairs increase the differences between
data values also systematically increase.
To illustrate the above, consider the porosity data shown in Figure 5, which
clearly exhibits a trend in the porosity profile along the well. Porosity increases with
depth due to a fining-upward of the sand sequence. The (normal-score) variogram
corresponding to this porosity data is shown in Figure 6. It shows a systematic
increase well above the theoretical sill of 1. One could fit a “power” or “fractal”
variogram model to the experimental variogram, however, since these models do
not have a sill value (it is infinite), they cannot be used in simulation algorithms

Figure 5. Porosity profile along a vertical well with a clear vertical trend. A linear
trend model is fitted to the data (solid line).
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 521

Figure 6. Variogram of the normal score transform of the porosity values shown in Figure 5.
The vertical trend in porosity reveals itself as a continuous increase in the variogram above
the sill value of 1.0.

such as sequential Gaussian simulation. But above all, they are not representative
of the property of interest.
A linear trend was fitted to the porosity profile (see Figure 5) and then removed
from the data. The resulting residuals constitute the new property of interest and
their profile is shown in Figure 7. The (normal-score) variogram of the residuals is
shown in Figure 8, which now exhibits a clearer structure reaching the theoretical
sill of 1 at about 7 distance units.

Variance for Variogram Interpretation

There is often confusion about the correct variance to use for variogram
interpretation. It is important to have the variance ¾ 2 , or C(0) value, to correctly
interpret positive and negative correlation. Recall that a semivariogram value ° (h)
above the sill variance implies negative correlation between Y (u) and Y (u + h),
whereas a semivariogram value ° (h) below the sill implies positive correlation.
There has been some discussion in the literature about the correct variance to use for
the sill variance and for variogram interpretation. This discussion was summarized
by Goovaerts (1997, page 103), who references Journel and Huijbregts (1978,
page 67) and the article by Barnes (1991). There are three issues that must be
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

522 Gringarten and Deutsch

Figure 7. Vertical profile of the residual porosity values (after removal of the
linear trend).

Figure 8. Variogram of the normal score transform of the residual of porosity values shown
above. The variogram reaches the expected sill value of 1.0.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 523

discussed before making a recommendation regarding the correct variance to use

for variogram interpretation and modeling: (1) the dispersion variance, which
accounts for the difference between our finite domain and the infinite stationary
variance; (2) declustering weights, which account for the fact that our data, and
all summary statistics, such as the sample variance ¾ˆ 2 , are not representative of
the entire domain, and (3) outlier values, which can cause erratic and unstable
estimates of the variance.
The main point being made by Journel and Barnes is that the sample variance
¾ˆ 2 is not an estimator of the stationary variance ¾ 2 , •, it is an estimator of the
dispersion variance of samples of point support • within the area of interest A,
denoted D 2 (•, A) in conventional geostatistical notation. Only when the sample
area A approaches an infinite domain does the sample variance ¾ˆ 2 approach the
stationary variance ¾ 2 . We should note that the dispersion variance D 2 (•, A) is
the average variogram value °̄ (A, A), which can only be calculated knowing the
variogram. A recursive approach could be considered to identify the correct sta-
tionary variance ¾ 2 and the variogram ° (h); however, this is unnecessary. The
data used to estimate the variogram represent the area of interest A and not an
infinite domain. Thus, the point where Y (u) and Y (u + h) are uncorrelated is the
dispersion variance D 2 (•, A). In other words, we should use the sample variance
as the sill of the sample semivariogram and acknowledge that, in all rigor, it is a
particular dispersion variance.
The second issue relates to the use of the naı̈ve sample variance or the sample
variance accounting for declustering weights. The use of declustering weights
is very important to ascertain a reliable histogram, mean, variance, and other
summary statistics. Although the use of declustering weights is important, they
are not used in the calculation of the variogram, that is, there are more experimental
pairs in areas of greater sampling density. Therefore, the sill (or semivariogram
value corresponding to zero correlation) is reached at the naı̈ve sample variance.
Efforts have been made (Omre, 1984) to incorporate declustering weights into
the variogram calculation; however, they provide no better variogram and are
difficult to implement in practice. As a result, declustering weights should not
be used in data transformation or variance calculation for variogram calculation,
interpretation, and modeling.
The last issue that must be addressed is the influence of outlier sample values.
It is well known in statistics that the variance, being a squared statistic, is sensitive
to outlier values. For this reason, the sample variance may be unreliable. We should
note, however, that this is not a problem with transformed data; the Gaussian and
indicator transform remove the sensitivity to outlier data values. The resulting
Gaussian distribution has no outliers by construction and there are only 0’s and 1’s
after indicator transformation.
The correct variance for variogram interpretation is the naı̈ ve equal-weighted
variance.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

524 Gringarten and Deutsch

METHODOLOGY FOR VARIOGRAM INTERPRETATION

AND MODELING

The methodology advocated in this paper is classical in that it assumes the

regionalized variable is made up of a sum of independent random variables. Each
constituent random variable has its own variogram structure. The component var-
iogram structures may be added arithmetically to create a complete 3D variogram
model to be used for geostatistical algorithms. Since each variogram structure
corresponds to a specific underlying geological phenomenon, the actual modeling
phase (traditional curve fitting exercise) is preceded by a necessary interpretation
stage.
In an approach similar to the well-established model identification part of
well test interpretation (Gringarten, 1986), where the pressure response is par-
titioned into different time regions, the total variance of the phenomenon un-
der study is divided into variance regions. The behavior of the variogram in
each region is then identified as being of a specific structure type. This is anal-
ogous to the model recognition step of well test analysis where the pressure-
derivative signature of each time region is associated to a specific flow regime.
Well test analysis defines three time regions: (1) early time corresponding to
near-wellbore effects (e.g., wellbore storage, hydraulic fractures); (2) middle time
characterized by the basic reservoir behavior (e.g., single or double porosity, com-
posite); and (3) late time accounting for reservoir boundaries (e.g., faults, no-
flow, or constant pressure boundaries). Similarly, three major variance regions
can be defined for variogram analysis: (1) short-scale variance (nugget effect);
(2) intermediate-scale variance (geometric anisotropy—there may exist more than
one such structure); and (3) large-scale variance (zonal anisotropy, hole effect).
The methodology proceeds sequentially, identifying first the short-scale variance,
then intermediate-scale structures, and finally large-scale features. For the var-
iogram interpretation to be consistent in 3D, all structures contributing to the
total variance must exist in all directions and their variance contributions must be
equal.
The steps of the methodology are as follows:

1. Compute and plot experimental variograms in what are believed to be

the principal directions of continuity based on a priori geological knowl-
edge (variogram calculation is not covered in this paper; see Deutsch and
Journel, 1997). If geological information is ambiguous, one can use 2D
variogram maps to determine major horizontal directions of continuity [A
variogram map is a plot of experimental variogram values in a coordinate
system (h x , h y ) with the center of the map corresponding to the variogram
at lag 0.0 (Goovaerts, 1997)]. Consider the horizontal and vertical exper-
imental variograms shown in Figure 9A.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Figure 9. A, experimental horizontal and vertical variograms; B, structure one repre-

sents a nugget effect; C, structure two represents geometric anisotropy; D, structure three
represents zonal anisotropy.

525
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

526 Gringarten and Deutsch

2. Place a horizontal line representing the theoretical sill. Use the value of
the experimental (stationary) variance for continuous variables, 1 if the
data has been transformed to normal score, and p(1 − p) for categorical
variables where p is the global proportion of the category of interest. A
feature of our proposed methodology is that the variograms are system-
atically fitted to the theoretical sill and the whole variance below the sill
must be explained in the following steps.
3. If the experimental variogram clearly rises above the theoretical sill, then
it is very likely that there exists a trend in the data. The trend should
be removed as detailed in the above section Removing the Trend, before
proceeding to interpretation of the experimental variogram.
4. Interpretation
• Short-scale variance: The Nugget effect is a discontinuity in the vari-
ogram at the origin corresponding to short scale variability. On the ex-
perimental variogram, it can be due to measurement errors or geological
structures with correlation ranges shorter than the sampling resolution.
It must be chosen as to be equal in all directions. It is picked from the
directional experimental variogram exhibiting the smallest nugget. It
is the interpreter’s decision to possibly lower it or even set it to 0.0.
Structure one of the example in Figure 9B corresponds to the nugget
effect.
• Intermediate-scale variance: Geometric anisotropy corresponds to a
phenomenon with different correlation ranges in different directions.
Each direction encounters the total variability of the structure. There may
exist more than one such variance structure. Structure two in Figure 9C
represents geometric anisotropy with longest correlation range in the
horizontal direction.
• Large-scale variance: (1) Zonal anisotropy is characterized by direc-
tional variograms reaching a plateau at a variance lower than the theo-
retical sill, i.e., the whole variability of the phenomenon is not visible
in those directions. Structure three in Figure 9D corresponds to zonal
anisotropy, only the vertical direction contributes to the total variability
of the phenomenon at that scale; (2) hole effect is representative of a
“periodic” phenomenon (cyclicity) and characterized by undulations on
the variogram. The hole effect does not actually contribute to the total
variance of the phenomena; however, its amplitude and frequency must
be identified during the interpretation procedure, also, it can only exist
in one direction.
5. Modeling
Once all the variance regions have been explained and each structure has
been related to a known geological process, one may proceed to vari-
ogram modeling by selecting a licit model type (spherical, exponential,
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 527

Gaussian, etc.) and correlation ranges for each structure. This step can be
referred to as the parameter estimation part of variogram analysis. Con-
straining the variogram model by a prior interpretation step with identifi-
cation of structure types can help fit the experimental variograms (Genton,
1998b).
In the presence of sparse horizontal data, variance structures visible
on the vertical variogram must be forced onto often nonexistent experi-
mental horizontal variograms. Horizontal ranges corresponding to these
structures are then “borrowed” from ancillary data (analogue outcrops,
densely drilled fields, depositional models, seismic), as shown in section
Incorporating Analogue Data.
Practitioners have refrained from rigorous variogram interpretation
and modeling due to sparse data and inadequate software. Increasingly,
data from horizontal wells and analogue fields or outcrops is becoming
available. Software can also be designed to aid in 3D variogram inter-
pretation and modeling rather than promote bad practice, which includes
misinterpretation of trends, zonal anisotropy, and not linking vertical and
horizontal variograms.

SOME EXAMPLES

Figure 10 shows horizontal and vertical variograms from a Canadian reservoir.

The variogram was fitted with two exponential models (short-scale structure and
long range structure) and a dampened hole-effect model in the vertical direction.
Note that the good horizontal variogram is due to the availability of horizontal
wells.
Five variance regions were used for the horizontal and vertical variogram of
Figure 10 (Table 1). The first small component is an isotropic nugget effect. The
next three are exponential variogram structures with different range parameters.
Three exponential structures are required to capture the inflection point at a variance
value of about 0.3 on the vertical variogram, and the long range structure in the
vertical variogram in the variance region 0.8 to 1.0. The fifth dampened hole
effect variogram structure only applies in the vertical direction and adds no net
contribution to the variance. Note that the dampening factor is five times the range.
Figure 11 shows horizontal and vertical variograms for the Amoco data, which
was made available to the Stanford Center for Reservoir Forecasting for testing
geostatistical algorithms (Journel, 1991). Note the zonal anisotropy in the vertical
directions due to systematic areal variations in the average porosity.
Three variance regions were for the horizontal and vertical variogram of
Figure 11 (Table 2). The first two components are anisotropic spherical variogram
structures. The last variogram captures the zonal anisotropy in the vertical direction.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

528 Gringarten and Deutsch

Figure 10. Horizontal and vertical variogram fitted with a combination of ex-
ponential and dampened hole effect variograms. Calculated from a Canadian
reservoir.

Figure 12 shows facies (presence of limestone coded as 1, dolomite and

anhydrite coded as zero) variograms calculated from a major Saudi Arabian reser-
voir; see Benkendorfer and others (1995) for the original variograms. Note the
interpretable horizontal variograms and the consistent vertical and horizontal
variograms.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 529

Table 1. Parameters for Five Variance Regions of Variogram on Figure 10a

Variance contribution Type of variogram Horizontal range, m Vertical range, m

0.05 Nugget
0.29 Exponential 100.0 0.015
0.46 Exponential 175.0 0.450
0.20 Exponential 100.0 0.500
0.20 Dampened Hole Effect 0.060

a The variance contribution, a variogram type, a range in the horizontal direction, and a vertical range

specify each variance region. No horizontal anisotropy is considered.

Two variance regions were identified for the facies variogram on Figure 12
(Table 3). Note that the sill in this case is 0.24 (related to the relative proportion
of limestone to dolomite). Both are anisotropic exponential variograms.

SOFTWARE IMPLICATIONS

The methodology and examples presented above have a number of implica-

tions on software design for variogram calculation, interpretation, and modeling.
Firstly, it is evident that multiple directions must be considered simultaneously.
It is poor practice to consider each direction independently and attempt to merge
one-dimensional variograms after modeling. Variance contributions identified in
one direction must automatically persist in all directions.
One can imagine a semiautomatic fitting procedure whereby once the total
variance has been divided into different nested structures, the parameters of the
nested structures (i.e., model type and ranges) are estimated with some type of
constrained optimization procedure. Provided the contribution of all structures is
consistent in all three major directions, this is all that is required to construct a
licit 3D-variogram model. A completely automatic variogram-fitting algorithm,
even if it generates consistent 3D models, can potentially lead to nongeological
models. In addition, it would not allow the incorporation of external information
in the presence of sparse data.

Table 2. Parameters for Three Variance Regions of Variogram on Figure 11a

Variance contribution Type of variogram Horizontal range, ft Vertical range, ft

0.50 Spherical 750.0 6.0

0.40 Spherical 2000.0 50.0
0.10 Spherical 7000.0 ∞

a The variance contribution, a variogram type, a range in the horizontal direction, and a vertical

range specify each variance region. No horizontal anisotropy is considered.

P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

530 Gringarten and Deutsch

Figure 11. Horizontal and vertical variogram fitted with a combination three
spherical variograms. These variograms were calculated from the “Amoco” data,
made available for testing geostatistical algorithms. Note the zonal anisotropy
evident in the vertical direction and the trend in the horizontal variogram.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 531

Figure 12. Horizontal and vertical facies variogram for a major Arabian carbonate
reservoir (SPE 29869). There are two horizontal variograms on the upper figure; the
dashed line is in the NE–SW direction and the solid line is in the NW–SE direction.

Kriging-based softwares that make use of the variogram models proposed

here should be able to isolate or filter any particular nested structure. This could
be of some advantage when considering, for example, seismic data, which is a
low-pass filter; high frequency or short-scale components can be filtered out in
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

532 Gringarten and Deutsch

Table 3. Parameters for Two Variance Regions of Variogram on Figure 12a

Variance Type of Horz. range, Horz. range, Vertical

contribution variogram NW-SE, ft NE-SW, ft range, ft

0.10 Exponential 150.0 400.0 0.8

0.14 Exponential 2500.0 4000.0 1.2

a The variance contribution, a variogram type, a range in the horizontal NW-SE

direction, a range in the horizontal NE–SW direction, and a vertical range specify
each variance region.

the measurement. A systematic interpretation procedure can put some rigor in the
often arbitrary partitioning of the variogram for Factorial Kriging.
The evident consequence on stochastic simulation is that a licit variogram
model can be used in a variety of stochastic simulation methods. Another implica-
tion is that the variable of interest can be constructed as a sum of different random
variables, each corresponding to a nested structure. An advantage is that there
are methods for fast simulation of one particular nested structure, e.g., moving
window methods for the spherical, exponential, or Gaussian variogram structures
(note that moving window type methods are especially attractive with parallel pro-
cessing computers. The CPU speed advantage on conventional single processor
computers is questionable.).
Most importantly, the implementation of a systematic variogram interpre-
tation and modeling procedure removes the mystery and art that have generally
surrounded variogram analysis. It allows users to infer, in a straightforward and reli-
able way, licit variogram models while at the same time acquiring an understanding
of the spatial continuity/variability of the phenomenon under study. Furthermore,
a rigorous methodology would imply that different users would arrive at similar
variogram models.

CONCLUSION

The variogram is used throughout geostatistical reservoir modeling as a mea-

sure of spatial variability. Subsequent 3D models reproduce by construction the
variogram, in a statistical sense. For this reason, it is essential that the variogram
be representative of the true heterogeneity present in study area. It represents
the modeler’s quantitative understanding of the spatial variability of the property
of interest given the data, and any related additional geological and geophysical
information that may be available.
The interpretation methods presented in this paper are reminiscent of the
revolution in well testing that came about with the pressure derivative and the
development of a rigorous analysis methodology based on the principles of model
identification and model verification. Model identification consists of partitioning
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

Variogram Interpretation and Modeling 533

the pressure response into time regions. Each time region is associated to a spe-
cific flow regime based on the pressure response signature (model recognition)
and the parameters required to model each regime are evaluated ( parameter es-
timation). In the case of variograms, the variance is divided into different regions
that correspond to different scales of geologic variability. Each variance region
(structure) is characterized by a specific geological variability behavior (nugget
effect, geometric and zonal anisotropy, and hole effect). Each behavior is modeled
analytically and requires a licit model type (spherical, exponential, Gaussian, etc.)
and a correlation range. In well test analysis, model verification is partly achieved
by comparing the resulting models to various graphical representation of the pres-
sure response other than the pressure derivative (or log–log) plot. In variogram
analysis, this step is built-in by insuring a consistent interpretation in all principle
directions.
Even though the importance and potential impact of the variogram model
is generally acknowledged, the practice of variogram analysis is often done half-
heartedly if at all. This paper presents nothing new to the expert and experienced
practitioner who understands and generally applies all the principles discussed
here. However, geostatistics is being increasingly used by practicing geologists,
geophysicists, and engineers who often find themselves at a loss when having to
infer a representative variogram model. The methodology presented here provides
a framework for understanding experimental variograms and complementing them
with ancillary information in the presence of sparse data, yielding a consistent and
licit 3D-variogram model.

REFERENCES

Armstrong, M., 1984, Improving the estimation and modeling of the variogram, in G. Verly and others,
eds., Geostatistics for Natural Resources Characterization: Reidel, Dordrecht, Holland, p. 1–20.
Barnes, R. J., 1991, The variogram sill and the sample variance: Math. Geology, v. 23, no. 4, p. 673–678.
Benkendorfer, J. P., Deutsch, C. V., LaCroix, P. D., Landis, L. H., Al-Askar, Y. A., and Cole, J.,
1995, Integrated reservoir modelling of a major Arabian carbonate reservoir, SPE Paper 29869,
in Proceedings of SPE Middle East Oil Show, Bahrain, p. 311–326.
Christakos, G., 1984, On the problem of permissible covariance and variogram models: Water Resources
Res. v. 20, no. 2, p. 251–265.
Cressie, N., 1985, Fitting variogram models by weighted least squares: Math Geology, v. 17, p. 563–586.
Cressie, N., 1993, Statistics for spatial data, Wiley, New York, 900 p.
Cressie, N., and Hawkins, D. M., 1980, Robust estimation of the variogram, I: Math. Geology, v. 12,
no. 2, p. 115–125.
Deutsch, C. V., and Journel, A. G., 1997, GSLIB: Geostatistical software library and user’s guide,
2nd Edition: Oxford University Press, New York, 369 p.
Genton, M. G., 1998a, Highly robust variogram estimation: Math Geology, v. 30, no. 2, p. 213–221.
Genton, M. G., 1998b, Variogram fitting by generalized least squares using an explicit formula for the
covariance structure: Math Geology, v. 30, no. 4, p. 323–345.
Goovaerts, P., 1997, Geostatistics for natural resources evaluation: Oxford University Press, New York,
483 p.
P1: Vendor/GDX/GIR/GAY P2: GDX/LZX/FGL Tally: FTK
Mathematical Geology [mg] PP103-297990 March 29, 2001 10:48 Style file version June 30, 1999

534 Gringarten and Deutsch

Gringarten, A. G., 1986, Computer-aided well test analysis, SPE Paper 14099, in Proceedings of the
1986 International Meeting on Petroleum Engineering, Beijing, China, March 17–20.
Journel, A. G., 1991, The Amoco data set: Exploratory data analysis, Stanford Center for Reservoir
Forecasting, Report 4.
Journel, A. G., and Huijbregts, Ch. J. 1978. Mining geostatistics: Academic Press, New York, 600 p.
Myers, D. E. 1991, Pseudo-cross variograms, positive-definiteness, and cokriging, Math Geology, v. 23,
no. 6, p. 805–816.
Olea, R. A., 1995, Fundamentals of semivariogram estimation, modeling, and usage in J. M. Yarus
and R. L. Chambers, eds., Stochastic modeling and geostatistics: Principles, methods, and case
studies, AAPG Computer Applications in Geology, no. 3, p. 27–36.
Omre, H., 1984, The Variogram and its estimation, in G. Verly and others, eds., Geostatistics for natural
resources charcterization, Reidel, Dordrecht, Holland, v. 1, p. 107–125.
Scholle, P. A., and Spearing, D., 1982, Sandstone depositional environments: American Association
of Petroleum Geologists, Tulsa, Oklahoma, 648 p.