ARTICLE

QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR

RESERVOIR CHARACTERIZATION
Richard L. Chambers* and Jeffrey M. Yarus**
Quantitative Geosciences, Inc.
Broken Arrow, Ok, U.S.A* Houston, Texas, U.S.A**
Introduction using a minimum phase wavelet and a gain to enhance structural
surfaces. However, the quantitative use of seismic attributes
As industry consultants involved with reservoir studies, either dictates a different processing philosophy; what we term "strati-
directly, or typically as mentors, we are the benefactors, not the graphic" processing. Processing seismic data for stratigraphic and
generators, of many data types. Often a client desires to use seismic rock and fluid properties requires zero-phase, true amplitude, and
attributes in a reservoir characterization study, or we find that migrated data, which is more costly and time consuming, but
seismic attributes are specified as general deliverables in projects necessary if most attribute studies are to succeed. Perhaps geomet-
up for competitive bid. Many times such requests are included as rical attributes describing spatial and temporal continuity do not
a project deliverable simply because it is fashionable, or there is a require such rigorous processing.
perception that "every one else is doing it."
If for example, the purpose is acoustic impedance inversion, the
Since the introduction of Complex Trace Attributes in the 1970s, data must be zero-phase, with true-amplitude recovery; otherwise
literally hundreds of new seismic attributes have emerged from a the resulting impedance cube is meaningless for quantitative inter-
variety of computational methods. Attributes may measure only pretation. AVO analysis examines the basics of amplitude varia-
one quality, termed "Primitive" attributes, or these primitive attrib- tions with offset which result from contrasts in elastic rock
utes may be combined through some statistical, neural network, or properties. AVO analysis begins with unstacked data, generating
mathematical manipulation to form "Hybrid" attributes (Taner, huge volumes of data, but may be combined with post-stack inver-
2001). The impetus behind the computation of so many attributes sion techniques to infer rock properties. Again, success depends on
is the desire to use them as predictive variables in reservoir char- zero-phase, true-amplitude seismic data.
acterization projects. Most typical studies use attributes qualita-
tively, such as in seismic stratigraphic interpretations depicting What Seismic Measures and What We Require for
internal bedding geometries and terminations, or to reveal spatial Reservoir Characterization
patterns related to depositional environments, faults or factures.
However, the trend is towards the quantitative use of single or The parameters measurable from the seismic data are (1) travel
combined attributes to predict lithology, facies, porosity, or fluid time, (2) amplitude, (3) the character of events, and (4) the patterns
type, for example. of events. From this information we often compute (Sheriff, 1992):
Depth maps of important horizons from traveltimes (and
This myriad of attributes poses a problem as few direct rela- velocity information);
tionships are established between all the attributes and physical or Velocity from differences in traveltime between source and
geological characteristics. As pointed out by Barnes (2001), because receiver;
many of the experts don't know what to make of all these attrib- Contrast in rock properties from measurements of reflection
utes, it is no wonder that the rest of us are confused. amplitude;
Locations of faults and stratigraphic changes from discontinu-
The purpose of this article is to review issues related to selecting ities in reflection patterns.
and using seismic attributes quantitatively in reservoir characteri- Dip and discontinuities from differences in traveltimes along
zation projects, rather than to describe or classify them. These latter a surface
tasks were lucidly presented in two recent CSEG Recorder articles
by Barnes (2001) and Taner (2001). From this information it is often possible to infer:
from velocity lithology, fluid content, abnormal pressure, or
Does the Seismic Data Warrant the Use of Attributes? temperature;
from lateral amplitude changes hydrocarbon locations,
Before dashing headlong into the computation of numerous changes in porosity, lithology, or thickness;
attributes, step back and look at quality of the data, determine the from seismic data patterns depositional environments, or
processing workflow, and ask what problems need to be solved. faults and fractures;
Too often we have seen that the data simply does not warrant use from changes in measurement direction velocity anisotropy,
beyond a basic structural interpretation because of poor signal or fracture orientation;
quality, low frequency content at the reservoir level, and improper from time-lapse measurements (4D seismic) locations of
processing. Data can be processed for structural interpretation changes in hydrocarbon distribution or bypassed pay.
Continued on Page 15

14 CSEG Recorder June, 2002

 ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION

Your
Continued from Page 14

For the flow simulation model the reservoir engineer needs:

The amount and spatial (vertical and horizontal) distribution of porosity;
Permeability

The nature of the fluids and their saturations;

Pressure
Temperature
Locations of barriers to flow (sealing and non-sealing faults, stratigraphic barriers, etc.)
Locations of thief zones (high permeability layers)
Data
Elastic Rock Properties

The following review is not addressed to the experienced AVO and rock property inter-
preter, but to those geoscientists and reservoir engineers who still have not been exposed to FIELD
some of the basic theory on elastic rock properties. We need to know what information is
contained in the seismic wavelet and how to extract it, before trying to use it quantitatively in
reservoir characterization studies.

A rock's physical properties, such as rigidity (shear modulus, ), incompressibility (bulk

modulus,(), porosity, pore fluid, clay content, gas saturation, and lithology affect how
seismic waves travel through rocks.

Rocks have elastic properties and elasticity theory provides the expressions for the velocity STACK
of seismic P-waves and S-waves in terms of elastic rock constants for simple cases. Elasticity
deals with deformation that vanishes completely upon removal of the stress which caused the
deformation, such as from the passage of a seismic wave (Sheriff, 1973). The elastic media is
determined from velocity and density () measurements. For isotropic media (Sheriff, 1992;
Hilterman, 2001),

BASIC

where is the Poisson ratio. Lam's constant , an elastic parameter sensitive to fluid
content, is related to and by = - 2/3. These physical properties are related to the ability SURVEY
of rocks to propagate seismic waves. Our interest in P-wave and S-waves is that they travel
through rocks differently depending on the fluid content and physical rock properties.

Further theory leads to the equations of Gassmann (1951) and Biot (1956) which relates
seismic velocity to porosity and the rock and fluid properties. Seismic velocity depends
strongly on porosity and often a decrease in velocity with an increase in porosity is the prin-
ciple controlling factor of velocity. We often use the time-average equation (Wyllie et al. 1956)
to compute porosity from velocity through the following expression:
SECTIONS

where is the porosity, Vf is the velocity of the interstitial fluid, and Vm is the velocity of
the rock matrix.
Continued on Page 16

June, 2002 CSEG Recorder 15

CALENDAR ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR
RESERVOIR CHARACTERIZATION
Continued from Page 15

When using attributes for the quantitative prediction of rock

JUNE 2002 properties, it is important to remember the relations in the
previous equations because they formulate the physical relation-
3rd BALKAN GEOPHYSICAL CONGRESS AND ship between the seismic attribute and rock and fluid properties.
EXHIBITION BGS/EAGE/SEG
June 24-28, 2002 Sofia, Bulgaria Seismic Attributes Property Predictors or False
www.BalkanGeophySoc.org Prophets?

AUGUST 2002 For our purposes we consider a seismic attribute as any seismi-
cally derived parameter computed from prestack or poststack
SEG/SPE DEVELOPMENT AND PRODUCTION data, before or after migration. Amplitude, phase, and frequency
FORUMIMPROVED PREDICTION, are fundamental parameters of the seismic wavelet and from these
PRODUCTIVITY AND PROFITABILITY USING few all other attributes are derived, either singly or in combina-
GEOPHYSICAL TOOLS. tions, and many of the new attributes duplicate each other because
August 25-30, 2002 of the nature of the computations. For example, bi-variate scatter
Snowmass, Colorado, USA plots of amplitude variance, average energy, RMS amplitude,
reflection strength, and average absolute amplitude show either a
www.spe.org/meetings
linear or parabolic relationship, but all these attributes contain the
same information (Barnes, 2001)
SEPTEMBER/OCTOBER 2002
With the proliferation of new attributes in the 1980s and into
2002 DOODLEBUG 50TH ANNIVERSARY
the 1990s, new methods arose to make sense of the many attributes
September 5-7, 2002 lacking geological significance. There was a curious notation that
Jasper Park Lodge, Alberta, Canada if attributes didn't make sense individually, perhaps they might
www.doodlebug.ca make sense in combination. One principle should be kept in mind
when using attributes; the physical basis for the correlation with
4TH INTERNATIONAL AZERBAIJAN properties measured at the wells. For example, a high negative
GEOPHYSICAL CONFERENCE (Azeri SEG correlation between porosity and acoustic impedance has a phys-
Section/SEG) ical basis, because velocity has an inverse relationship to porosity;
as velocity increases, the porosity typically decreases.
September 24-26, 2002 Baku
Unfortunately, there is a common practice of selecting attributes
geophiz@azdata.net based solely on the strength of their observed correlations with
properties measured at the wells, but with little thought given to
2002 SPE ANNUAL TECHNICAL CONFERENCE the validity of the correlation, except that it looks good.
AND EXHIBITION
September 29-October 2, 2002 The Problem
San Antonio, Texas, USA
Our problem is to identify and select which attributes to use,
www.spe.org
then to select a method to integrate them with properties meas-
ured at the wells. The three most common prediction methods
SEG INTL EXPOSITION AND 72ND ANNUAL used in our industry today are regression, geostatistics and neural
MEETING networks. Each method requires making an inference (prediction)
October 6-11, 2002 Salt Lake City, Utah, USA from the seismic attribute(s) based on its relationship to much
sparser information measured at well locations. We also make the
NOVEMBER 2002 assumption that the sample population (well data) is representa-
tive of the larger parent population (the reservoir). The basic data
11TH VENEZUELAN GEOPHYSICAL CONGRESS integration process is accomplished in five steps:
(SEG)
November 17-20, 2002 Caracas Hilton Convention 1. Calibration. Well data provide high-resolution, depth related
Centre local information, whereas 3D seismic data provide spatially
www.sovg.org dense, but vertically lower resolution, time related information.
Calibration is the first and most critical step in the process as the
data must be calibrated both vertically and areally.

Continued on Page 17

16 CSEG Recorder June, 2002

 ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION

clean
Continued from Page 16

2. Choice of the seismic attribute(s). The primary objective is to identify the attribute(s) that
works best as a predictor for the reservoir property of interest. However, care must be
taken when choosing the seismic attribute, because it is not unusual to find spurious or
false correlations that do not reflect any physical basis for the relationship. The probability
of finding a false correlation increases with the number of seismic attributes considered
and is inversely proportional to the number of data control points. This concept is
discussed below.
and
3. Prediction. The areal distribution of the variable of interest is mapped by integrating the
well data and the seismic attribute. This prediction step is typically done by either linear
simple
or non-linear regression models, neural networks, or using a geostatistical method like
Colocated CoKriging.

4. Cross-validation. This is the systematic removal of wells, one-by-one, and re-estimating

their values based on the model selected. Cross-validation is not always performed, but it NON-PROPRIETARY
does provide a means to validate the contribution of the secondary information to improve
the prediction.
FORMATS
5. Management Decisions. Depending on the project objectives, management decisions
based on the prediction may include location of in-fill wells, developing a depletion
strategy, designing a secondary or tertiary recovery program. Perhaps the decision is more
basic are there enough in-place hydrocarbons to justify developing the prospect? SECURE
The basic process is simple, but the key is which attributes to use and how to perform the WEB ACCESS
calibration.

Probability of observing a false correlation

Most of us are familiar with the correlation coefficient, r, which is a measure of the strength
of the relationship between one or more variables. But, we must remember that for a given
correlation coefficient, we should make some estimate of its validity. For example, is a corre- OPEN
lation of 0.83 good or bad? Most practitioners are aware that the smaller the number of
samples, the greater the uncertainty about the true value of the correlation. However, few of MEDIA BACKUP
us know about what statisticians call experiment-wise error rates. As we generate more
seismic attributes, there is a greater chance of observing at least one large spurious (false)
correlation value; a large correlation computed purely by chance. Kalkomey (1997) discusses
the potential risks when using seismic attributes as predictors of reservoir properties and
illustrates the impact of spurious correlations. The article is highly recommended reading, but
for those lacking the time, we have taken the liberty of summarizing some of her work in the
following paragraphs. NO SOFTWARE
If we consider only one seismic attribute, acoustic impedance as a predictor of porosity for INSTALLATION
example, then the probability of observing the absolute value of the sample correlation coef-
ficient, r, greater than some constant R, given the true correlation () is zero can be found from
the following expression:

where n is the sample size or the number of locations (wells) where the reservoir property DISC FILE
and the seismic attribute are measured, and t is distributed as a Student's t- critical value, with
n-2 degrees of freedom. STORAGE
The probability of a spurious correlation depends solely on R (the magnitude of the
spurious sample correlation) and n, the number of well measurements, based on the assump-
tion of random sampling. It is interesting that when data are spatially correlated, as is the case

Continued on Page 18
June, 2002 CSEG Recorder 17
ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 17

for the variables we work with in the petroleum industry, Eq. 5 Hypothesis Testing
gives a conservative estimate of the probability of a spurious corre-
lation. This is because the effective sample size is smaller than the The most frequent use of statistics is to test some hypothesis.
actual sample size; and as n decreases, the probability of a spurious Our hypothesis is that the reservoir property measured at the wells
correlation increases (Kalkomey, 1997). Table 1 shows the proba- correlates with the seismic attribute. This hypothesis, call the null
bility of observing a spurious correlation for different levels of the hypothesis (H0) is the hypothesis under test. It is called the null
sample correlation, R, and different sample sizes, n. This is the base hypothesis because we assume that the correlation between the
case when only one seismic attribute is considered. seismic attribute and the well measurement is true. The alternative
hypothesis (Ha) is that the correlation is not true. We may reject H0
This table is used to assess the chance that the sample correla- when, in fact it is true, or we may accept H0 when it is false and
tion, r, is actually false or uncorrelated with the reservoir property some alternative hypothesis is true. These are Type I and Type II
being predicted. For example, given 5 wells, and an r = 0.7, there is errors, respectively. Table 3 illustrates a Decision Table and the
a 19% probability that the correlation is false. Perhaps we are goodness of a statistical test of a hypothesis is measured by proba-
willing to accept this risk, however there is another aspect of the bilities of making a Type I or Type II error (Sokal and Rohlf, 1969)
correlation coefficient that should be considered the confidence
limits of the true correlation coefficient. For this example, the 95% Property and Property and
confidence limits are from a minimum r of -0.48 (P97.5) and a Decision attribute are attribute are
maximum r of 0.98 (P2.5). Because the minimum r is negative, we uncorrelated correlated
cannot say with confidence that there is any correlation and we
should reject this attribute as a predictor. Considering one seismic Keep seismic
attribute and a sample correlation of 0.7, we need 9 samples before attribute as a Type I Error Correct Decision
the minimum r is positive, but its value is only 0.07, with a 4% predictor (no error)
chance that the correlation is false.
Reject seismic Correct Decision
The problem of spurious correlations is only enhanced when attribute as a (no error) Type II Error
considering more than one attribute at a time. Because many attrib- predictor
utes contain the same information it is important to select inde-
pendent attributes for multivariate analysis. Table 2 is similar to Table 3. A Decision Table for the Null Hypothesis
Table 1, except in this case 5 independent attributes are considered
as the predictors (weighted, linear combinations) of a reservoir The Decision Table illustrates the possible outcomes one must
property. Returning to the example above for 5 samples and a consider to assess the risk in selecting seismic attributes as predic-
correlation of 0.7, we see (Table 2) that there is a 65% chance that at tors of reservoir properties.
least one of the 5 attributes is falsely corrected with the reservoir Type I error: We commit this error when the attribute is selected
property measured at the wells. The confidence limits on the corre- but the correlation is false. The prediction is less accurate, but more
lation coefficient remain the same as before. precise and the cost is an inaccurate predication with confidence.
Type II error: This type of error occurs when a true correlation
exists, but we fail to use the seismic attribute as a predictor. Now
we have less accurate and less precise prediction and the cost is

Sample size Samples

R 5 10 20 50 100 R 5 10 20 50 100

0.1 0.87 0.78 0.67 0.49 0.32 0.1 1.00 1.00 1.00 0.97 0.86

0.3 0.62 0.40 0.20 0.03 0.00 0.3 0.99 0.92 0.67 0.16 0.01

0.5 0.39 0.14 0.02 0.00 0.00 0.5 0.92 0.53 0.12 0.00 0.00

0.7 0.19 0.02 0.00 0.00 0.00 0.7 0.65 0.12 0.00 0.00 0.00

0.9 0.04 0.00 0.00 0.00 0.00 0.9 0.17 0.00 0.00 0.00 0.00

Table 1. Probability of observing a spurious correlation between a reservoir prop- Table 2. Probability of observing a spurious correlation between a reservoir property
erty and a single seismic attribute (modified from Kalkomey, 1997, p 248) and five independent seismic attributes (modified from Kalkomey, 1997, p 248)

Continued on Page 20
18 CSEG Recorder June, 2002
ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 18

more uncertainty than justified. and that the well locations sampled the reservoir in such a manner
creating a false, high correlation, thus we commit a Type I error.
Kalkomey (1997) believes that for most cases the economic
consequences of making highly confident, but inaccurate predic- Data integration was by the Markov-Bayes approach to
tions (Type I error) are more severe than the consequence of a Type Colocated Cokriging. With only seven wells, computation of a
II error. variogram is not possible; therefore the variogram model (Fig. 3)

An Example

The previous discussion focused on aspects of the selection of a

seismic attribute and the possibility that the correlation may be
false. The following example illustrates some of these principles.
The seismic attribute is acoustic impedance derived from zero-
phase, true amplitude, and 58-fold 3D seismic data. Inversion to
acoustic impedance followed the method of Lindseth (1979), using
sonic and density logs from three wells to calibrate the inversion.
Frequency content at the reservoir depth is 7-110 Hz between 1220-
1370-m. The low frequency (0-7 Hz) contribution to impedance,
due to density, was derived from the conversion of seismic
stacking velocities to interval velocities, using a Gardner's relation-
ship. The primary target was a siliciclastic interval 16-28-m thick,
Figure 2. Map of seismic acoustic impedance and porosity percent at 7 well loca-
averaging 22-m (see Chambers, et al. 1994, for a detailed discussion
tions.
of this example). The quarter-wavelet tuning thickness is about 13-
m using 70 Hz as the dominant frequency and an average velocity was derived from the seismic acoustic impedance data (33,800
of 3,650 m/s. The objective of this example is to use seismic samples. Recall that in the Markov-Bayes method, the sill of the
acoustic impedance to predict porosity away from the wells. primary variogram is scaled based on the ratio of variances
(seismic AI/porosity), using the same scale lengths and anisotropy
Figure 1 shows a scatter plot of seismic acoustic impedance and as the seismic AI model. The cross-covariance model has a "sill"
porosity measured at 7 well locations. The sample correlation coef- based on the magnitude of the correlation coefficient.
ficient is -0.95, with only a 0.1% chance of a false correlation. The
95% confidence limits are -0.69 (P97.5) and -0.99 (P2.5). The nega- Figure 4 illustrates the colocated cokriging results using the
tive correlation coefficient is expected because acoustic impedance minimum correlation (-0.69, Fig. 4A), the sample correlation (-0.95,
(AI) varies inversely with the magnitude of porosity. Even with
these high correlations, we assume that the pattern displayed by
the seismic AI (Fig. 2) is related to the true distribution of porosity
and that the high correlation isn't serendipitous. That is, there still
may be a chance that the distribution pattern of the true porosity is
quite different from the pattern displayed by the seismic attribute

Figure 3. Anisotropic variogram model of seismic acoustic impedance. Dashed lines

are the experimental points and the solid lines are the nested model. Vertical units
are variance and the horizontal units are meters.

Fig 4B), and the maximum correlation (-0.99, Fig 4C). Because of the
high correlation and only 7 wells, the footprint of the seismic
attribute is very pronounced. The difference between Figures 4B and
Figure 1. Scatterplot of porosity percent and seismic acoustic impedance at 7 well 4C are imperceptible, whereas differences can be seen when these
locations. figures are compared to the image in Figure 4A. The zone of intense

Continued on Page 21
20 CSEG Recorder June, 2002
 ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 20

Figure 6. A map of porosity based on kriging 55 values. The wells indicated by the
black asterisk are the locations of the 7 wells shown in figures 2, 4, and 5.

Figure 7 illustrates the value of integrating a seismic attribute

when mapping porosity using only 7 porosity values. For this
study we had the advantage of knowing the porosity values at 48
other well locations, thus we can test the accuracy of using the
seismic attribute as a predictive variable. Of the 48 'new' locations,
Figure 4 illustrates the colocated cokriging results using the minimum correlation only 9 were misclassified using the criterion of finding porosity * 9
(-0. 69, Fig. 4A), the sample correlation (-0.95, Fig 4B) and the maximum correla- percent. The wells shown as open circles are the values for the 7
tion (-0.99, Fig 4C). wells and are predicted perfectly, because colocated cokriging is an
red is not as broad in Figure 4A as it is in the other two images. From exact interpolator, like kriging (the red line shows the perfect line
past experiences, suppose the operator knows that wells encoun- of correlation).
tering * 9% porosity, within this reservoir interval, are good
producers. If the purpose of the study was to locate in-fill drilling
locations, we could use any of the maps shown in Figure 4 to make
such a decision, assuming that we did not make a Type I error (using
an attribute when the correlation is false). Before selecting a final
location, it may be prudent to run conditional simulations with each
correlation coefficient and create a risk map showing probability that
porosity is * 9%. Figure 5 illustrates a risk map based on 100 condi-
tional simulations with an r = -0.95. The only wells satisfying the *
9% condition are the three shown with a white asterisk (*) with
porosity values of 12.1, 13.5, and 17.4 percent (see Fig. 2)

Figure 7. Scatter plot of measure porosity (Y-axis) versus predicted porosity (X-
axis). Nine wells were misclassified out of the 48 'new' wells.

The additional 48 wells were drilled in a typical west Texas 40-

acre 5-spot pattern. If seismic data had been available when
designing the original in-fill program, we can see that many wells
Figure 5. A risk map based on 100 conditional simulations with an r = -0.95
should not have been drilled using the porosity cutoff criteria.
showing the probability that porosity is * 9%.

The 55 well locations, shown in Figure 6, were the true well What about using Linear Regression?
density (40-acre, 5-spot) at the time of this study. The map is based
on kriging 55 porosity values using an anisotropic variogram. Reviewing the scatterplot shown in Figure 1 and recalling that
Compare this map to the maps shown in Figure 4 based on 7 wells the correlation coefficient is -0.95, it would seem logical to simply
and seismic acoustic impedance used as the predictor variable. The use linear regression to predict porosity from seismic acoustic
large scale features shown in the map, based on kriging porosity impedance. The map can be made using the following regression
data only, are similar in appearance to those in Figure 4, but the equation: Y= c-bX.
map lacks the higher frequency heterogeneity added by the use of
the seismic attribute. Continued on Page 22

June, 2002 CSEG Recorder 21

ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 21

Porosity = 62.13 - (0.00157 * AI) (Eq. 7) whereas solid circles are locations of over-estimates of porosity. The
desired pattern would be a random distribution of open and solid
where c is the intercept and b is the slope. circles, indicating no bias in the estimation error. Comparing the two
results, Figure 11 B shows much more clustering of the two symbols
when using linear regression _ the southwestern quadrant is almost
always over-estimated, and there is more under-estimation of the
measured porosity in the northwestern quadrant, which is an unde-

Figure 8. The map of porosity based on a regression relation.

The map of porosity based on a regression relation is shown in

Figure 8. At first glance this porosity map looks very similar to a
map made using colocated cokriging (Figure 4). However, the
range of porosity in Figure 5 is greater than the predicted porosi- Figure 9. Scatter plot of measure porosity (Y-axis) versus predicted porosity (X-
ties in Figure 4; there is nothing wrong with this, except that the axis) using a linear regression model (Eqn. 7).
regression method predicted negative porosities. Compare Figures
7 and 9 illustrating scatter plots between the measured and
predicted porosity from colocated cokriging and regression. The
most obvious difference is the inability of the regression method to
predict the porosity values used to compute the regression equa-
tion (the open circles). The number of underestimations is the same
as colocated cokriging, but there are 4 additional locations that
over estimate porosity. The fundamental problem with such an
application is that we just, inappropriately, turned a point estima-
tion method into a spatial estimator. Traditional regression
methods use only data available at the target location and fail to
use existing spatial correlations from additional secondary data
control points and the primary attribute to be estimated (Journel,
1989). The fact that regression uses only one data point during the Figure 10. Scatter plot of measured porosity versus residual porosity. Residual
estimation is valid, because traditional regression assumes data porosity is measured porosity minus porosity predicted by linear regression at the
55 well locations.
independence. Although the results seem fine, the b term (slope of
the function) imparts a spatial linear bias
(trend) in the estimates during the
mapping process. The bias becomes
apparent in an analysis of the residuals
shown (Figure 10) as a scatter plot of
measured porosity versus residual
porosity, where residual porosity is meas-
ured porosity minus porosity predicted
by linear regression at the 55 well loca-
tions. The correlation coefficient is 0.69.

The displays in Figure 11 illustrate the

spatial distribution of the estimation
errors based on a cross-validation using
colocated cokriging (Fig. 11 A) and linear
regression (Fig. 11 B). The open circles Figure 11. The spatial distribution of the residual errors based on a cross-validation using colocated cokriging (A) and linear
represent under-estimates of porosity, regression (B).

Continued on Page 23
22 CSEG Recorder June, 2002
 ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 22

sirable attribute in an estimator algorithm. The colocated cokriging Figure 12 illustrates the integration of geological features using
error distribution is more random, honoring the unbiasedness trait multiple-point statistics to infer spatial patterns. The image on the
of the system of kriging algorithms. left represents a time slice through a 3D seismic amplitude cube. The
image shows the presence of channel-like features, which are to be
Integrating Geometrical Attributes included in the reservoir model. The seismic data image is sampled,
and the resampled data are used in the data integration step. The
Geometrical attributes depict spatial and temporal patterns center image depicts a conceptual model of the channels, in this case,
related to bedding geometries, discontinuities such as faults and a meandering channel system, with a user specified wavelength,
fracture swarms, bedding similarity, bedding dips, event character- sinuosity, etc. The right image is one of many possible pixel-based
istics, and can be used to quantify features related to depositional
patterns, and related lithology (Taner, 2001). Continued on Page 24

Traditional geostatistics uses the vari-

ogram model to capture spatial continuity by
measuring the degree of variability or
conversely correlation/continuity between Theres no getting round it,
any two locations in space. Because the vari-
ogram is based only on two-point statistics, it
cannot easily model curvi-linear features,
the Earth is
such as channels, nor can it model strongly
contiguous patterns, such as fractures. anisotropic
The incorporation of geometrical features,
like channels, deltas, reefs, etc. into reservoir
models is typically accomplished using
object-based (Boolean) geostatistical methods.
Reservoir models built using objects is
appealing to the geologists because they look
realistic, that is, channels look like channels.
However object-based methods have their
limitations in that the algorithms require
numerous input parameters describing the
geometrical dimensions, such that object
modeling becomes nearly deterministic, espe-
cially when many wells are used as condi-
tioning information. However, when many
wells are used, object modeling often fails
because the parametric conditions imposed CGG is exploiting the natural
by the modeler cannot honor the conditioning phenomenon of anisotropy
well data.
instead of trying to get round it.
With A + processing, CGG A + is for top marks in time,
A new method, called multiple-point
delivers a better, more focused depth and multi-component
statistics has been introduced that allows
geological patterns integration through the image of your reservoir and processing. Our mission is to
Digital Imagery copyright 2000 PhotoDisc Inc

framework of pixel-based modeling (Journel, provides additional knowledge give you added insight, added
1997, 2002; Caers, 2000; Strebelle, 2000). about the Earths properties. confidence and added value.
Training images rather than variograms
models are used to depict the prior geological
conceptual model, then a sequential-based
Turn the Earths anisotropy to your advantage
simulation algorithm is used to generate
multiple realizations, with each realization
honoring the well data and the multiple-point CGG Canada Services Ltd.
Suite 700, 404-6th Avenue S.W. Calgary, Alberta T2P 0R9
statistics (patterns) derived from the training (403) 266-1011 rvesely@ca.cgg.com
images. www.cgg.com

June, 2002 CSEG Recorder 23

ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 23

simulations. The simulation honors the statistics of the channel as colocated cokriging and colocated cosimulations, offer attractive
morphology shown in the center image, and reproduces the shapes, means to integrate seismic attribute and well information, without
locations, and orientations of the channel-like features in the seismic an estimation bias, and account for the scale (support) differences
amplitude data (left image). The multiple-point statistics works in between the two data types. Geometrical seismic attributes probably
2D and 3D. require less rigorous data processing. Integration of geometrical
information in reservoir model is not as straight forward as the "rock
The advantage of this method is the easy implementation and property" attributes, however, the new multiple-point statistics
conditioning to well data, regardless of the number of wells, unlike methods have good promise.
object-based modeling. The training image could be much more
complex with the incorporation of cross-cutting features, such as a References
fracture system normal to the channel orientation. Two immediate Barnes, A. E., 2001, Seismic attributes in your facies, pp. 41-47, September Issue
issues to consider are the choice of the training image(s) and the
Biot, M. A., 1956, Theory of propagation of elastic waves in a fluid-saturated porous
conceptual geological model, i.e., meandering or braided channels, solid, J. Acoustic. Soc. Am., 28, pp. 168-191.
for this example can make a significance difference in the connec-
Caers, J., 2000, Modeling facies distributions from seismic using neural nets,
tivity. Some current problems with the multiple-point statistics
Stanford Center for Reservoir Forecasting Annual Report, No. 13, Vol.1.
method are: 1) how to evaluate the relative contributions (weight) of
the training image and the soft data, especially when they are some- Caers, J., 2002, History matching under training-image based geological model
constraints, Stanford Center for Reservoir Forecasting Annual Report, No. 15,
what inconsistent; 2) how to select samples from the training image,
Vol.1.
i.e., randomly, or on a Cartesian grid with an equal sampling
interval, for example; and 3) it is not always possible to preserve the Chambers, R. L., M. A. Zinger and M. C. Kelly, 1994, Constraining
non-stationarity of the information in the final results, if non-station- Geostatistical Reservoir Descriptions with 3-D Seismic Data to Reduce Uncertainty,
Stochastic Modeling and Geostatistics, J. M. Yarus and R. L. Chambers, Eds.
arity is present in the seismic data, or the conceptual model.
AAPG Computer Applications in Geology, No. 3, pp. 143-58.
Gassmann, F., 1951, Elastic waves through a packing of spheres, Geophysics, 16,
pp. 673-685.
Hilterman, F. J., 2001, Seismic Amplitude Interpretation, 2001 Distinguished
Instructor Short Course Series No. 4, sponsored by the Society of Exploration
Geophysicists, European Assoc. of Geoscientists and Engineers, Houston,
TX, p. 2-24.
Journel, A. G., 1989, Fundamentals of Geostatistics in Five Lessons, Short course
in Geology, Vol. 8, American Geophysical Union, 40 p.
Journel, A. G., 1997, Deterministic geostatistics: a new visit, in Baffi, E. and
Shofield N. eds., Geostatistics-Wollongong, 1, Kluwer Academic Press,
Dordrecht, pp. 174-187.
Journel, A. G., 2002, Combining knowledge from diverse information sources: an
alternative to Bayesian analysis, in press, J. Mathematical Geology.
Kalkomey, C. T., 1997, Potential risks when using seismic attributes as predictors
of reservoir properties. The Leading Edge, March., pp. 247-251.
Figure 12. These images illustrate the integration of geological features using
multiple-point statistics to infer spatial patterns. The left image is seismic ampli- Lindseth, R. O., 1979, Synthetic sonic logs - a process for stratigraphic interpreta-
tude data showing channel-like features, the center image is the conceptual model, tion, Geophysics, 44, pp. 3-26.
and the right image is one pixel-based realization (modified from Caers, 2002).
Sheriff, R. E., 1973, Encyclopedic Dictionary of Exploration Geophysics, Society of
Exploration Geophysicists, Tulsa, OK, p. 69.
Summary
Sheriff, R. E., 1992, Basic petrophysics and geophysics, in Reservoir Geophysics,
R. E. Sheriff, ed., Investigations in Geophysics No. 7, Society of Exploration
Seismic attributes can be important qualitative and quantitative Geophysicists, Tulsa, OK, pp. 37-49.
predictors of reservoir properties and geometries when correctly
Sokal, R. R. and J. F. Rohlf, 1969, Biometry, W. H. Freeman and Co., San
used in reservoir characterization studies. It is critical when using
Francisco, pp. 155-173.
attributes as quantitative predictors of reservoir properties to
consider: 1) how the seismic data were processed; 2) the physical Strebelle, S., 2000, Sequential simulation drawing structures from training images,
Stanford Center for Reservoir Forecasting Annual Report, No. 13, Vol.1.
basis of the correlation; 2) the possibility of false correlations when
the number of well data are few, and/or when the number of attrib- Taner, M. T., 2001, Seismic attributes, CSEG Recorder, pp. 48-56, September
utes are many; 3) when using multiple attributes, the attributes must Issue.
provide independent information about the reservoir property; 4) Wyllie, M. R. J., Gregory, A. R., and Gardner, L. W., 1956, Elastic waves in
and, although, easily applied, linear regression is not recommended heterogeneous and porous media, Geophysics, 21, pp. 41-70. R
because the results will be biased. Geostatistical data methods, such
Continued on Page 25

24 CSEG Recorder June, 2002

 ARTICLE Contd
QUANTITATIVE USE OF SEISMIC ATTRIBUTES FOR RESERVOIR CHARACTERIZATION
Continued from Page 24

Richard L. Chambers is co-owner of Jeffrey M. Yarus is co-owner of

Quantitative Geosciences, Broken Arrow, Quantitative Geosciences, Houston, Texas.
Oklahoma, U.S.A. He earned B.A. and While attending the College of Wooster, he
M.Sc. degrees in geology from the received a scholarship to study geology for
University of Montana, and then attended one year at the University of Durham,
Michigan State University, earning a Ph.D. England, before earning his B.A. at
in geology, with an emphasis on quantita- Wooster. He earned his MS and PhD
tive sedimentology and statistical applica- degrees in geology from the University of
tions. He worked for the Great Lakes Environmental Research South Carolina where he refined his interests in computer
Laboratory, NOAA, U.S. Department of Commerce for 8 years mapping and numerical and statistical analysis. He joined
before joining Phillips Petroleum Company, where for 8 years Amoco Production Company in New Orleans, Louisiana, as a
his activities included research in high-resolution P- and S-wave production geologist for the Gulf Coast Region. He left Amoco,
seismic imaging and developing exploration plays and lease to work as an independent for 8 years on a variety of domestic
acquisition strategies as a member of a regional basin studies basins in the Rockies, mid-continent and Appalachian regions.
team. In 1988, he joined the Amoco Production Company He joined Marathon Oil Company's Petroleum Technology
Research Center where he worked for 9 years in various tech- Center, in 1988, where he was instrumental in instituting
nical and management capacities related to the application of desktop computer mapping and geostatistical technology.
geostatistical methods for exploration and production problems. Before co-founding Quantitative Geosciences, with Richard L.
In 1997, he became a consultant providing training and geosta- Chambers, he held various managerial positions with
tistical modeling expertise to the petroleum industry. His profes- GeoGraphix, Geomath, Roxar (formerly Smedvig Technologies)
sional activities include publishing, consulting, lecturing and and Knowledge Reservoir. His professional activities include
offering applied training courses in reservoir characterization publishing (in AAPG's first volume on computer applications in
and geostatistical applications. In late 2001, he co-founded, with 1992), consulting, lecturing and offering applied training courses
Jeffrey M. Yarus, Quantitative Geosciences, with offices in in reservoir characterization and geostatistical applications. He
Broken Arrow, Houston, and London. has been an adjunct professor at the University of Colorado,
Denver and a visiting professor at the Colorado School of Mines.

Yarus and Chambers are co-editors of Stochastic Modeling and Geostatistics, AAPG Computer Applications in Geology, No. 3, published
in 1994. The second volume, with a new collection of papers, should be published in the AAPG Computer Application in Geology Series
by late 2002.

Volunteer Awards Night

The CSEG held the annual volunteer appreciation and awards Barry Korchinski and George Fairs received the meritorious
night on May 16th at the 400 Club. The night serves to thank the service award, given to members who have made a prominent or
many volunteers who contribute their free time to the various lengthy contribution to the society or industry as a whole.
social and academic functions throughout the year, and to recog-
nize major contributions that individuals have made to the society Next year we hope to get even more people out when the
over their lifetime. awards will be presented just before the Icebreaker of the 2003 joint
CSEG/CSPG convention.
This year Bob Comer received the Honorary Membership,
awarded to members who have gained recognition for distin- We would also like to thank the evenings sponsors: Godfrey
guished contributions in the field of geophysics or in service to the and Godfrey accountants, Genesis Corporate Search, and Phoenix
CSEG, either nationally or internationally. Press, and all the volunteers and participants in CSEG functions.
Continued on Page 40

CSEG Website Up and Running

(www.cseg.ca)

June, 2002 CSEG Recorder 25