You are on page 1of 4

SL 3.

5
Use of seismic attributes as predictors of reservoir properties -- potential risks
Author Cynthia T. Kalkomey*, Mobil E&P Technical Center

Summary
Suppose we have only ten data points available to estimate
Application of seismic attributes to predicting reservoir true, butunknown, correlation coefficient, p.
properties typically relies on first identifying an attribute
that is significantly correlated with the reservoir property
being modeled. The dense seismic data is then used to
guide the interpolation between sparse well data using (1)
geostatistical, regression or neural network techniques. A
common practice is to choose the seismic attributes based
on the strength of their observed correlations with reservoir
properties measured at the wells. Often these correlations
must be estimated by a small number of sample points. where E is the expected value operator
Under these conditions, the probability of selecting a is the expected value or true mean of the
seismic attribute that is actually uncorrelated with the reservoir property (X)
reservoir property being modeled can be quite large. is the expected value of the seismic
attribute (Y).
This study quantifies the probability of observing
spuriously high correlations between the well and seismic
data; that is, the probability of observing a significant To estimate p, we calculate the sample correlation, r .
sample correlation when the seismic attributes are actually
uncorrelated with the reservoir property. The probability of
observing a spurious correlation increases as the sample .
size (number of well data points) decreases, or the number
of attributes investigated increases. It is strongly
recommended, when the sample size is small, that only
those seismic attributes which have a physically justifiable ( 2 )
relationship with the reservoir property be considered as
predictors.
where n = sample size or number of locations with
If the attribute selected is actually uncorrelated with the measurements of both the reservoir
property, but appears correlated based on data from a small property and the seismic attribute
number of wells, then not only will the estimated reservoir = reservoir property at location i
property map be biased, but its associated variance will be = seismic attribute at location i
underestimated. This could lead to confident, but
inaccurate predictions and ultimately poor decisions. This Suppose our sample correlation from the 10 data points
risk is greater, at least in this author’s opinion, than the risk is calculated to be 0.70. A 95% confidence interval for the
associated with discarding an attribute that is actually true correlation (p) would be [0.12, 0.91]. The only
correlated with the reservoir property. statement we can make with confidence is that the true
correlation between the reservoir property and seismic
Introduction attribute is between 0.12 and 0.91. If instead of ten, we
have only five data points to estimated the correlation, the
Many applications of geostatistics rely on first identifying 95% confidence interval widens to [-0.39, 0.96]. Since
an attribute from the 3D seismic volume which is strongly zero is contained in this interval, we cannot say with
correlated with the reservoir property being modeled -- for confidence that there is any correlation between the
example, identifying amplitude as being strongly correlated reservoir property and the seismic attribute. This illustrates
with porosity. This is true when applying geostatistical the problem with choosing a seismic attribute as a predictor
estimation or simulation based on kriging with external solely on the basis of the sample correlation. As we
drift, cokriging or collocated cokriging [1] where dense increase the number of attributes that we are testing for
seismic data is used to guide the spatial distribution of correlation with the reservoir property, the problem is
reservoir properties between sparse well data. Selection of exacerbated.
one or more seismic attributes is also a first step for other
deterministic estimation techniques, such as multiple linear The following sections will quantify the probability of
or nonlinear regression or neural networks. observing spurious sample correlations between seismic
attributes and a reservoir property; that is, the probability of
The problem is that the correlation between the seismic observing a value of r that appears to be significant when
attribute(s) and the reservoir property is often estimated by there is actually no correlation. This is followed by a
a small number of sample points -- data at well locations. discussion of the risks associated with selecting a seismic
As with any statistical measure, as the sample size predictor.
decreases, the width of an interval estimate of the true
parameter, for a given confidence level, increases.

1789
Use of seismic attributes as predictors
Probability of observing spurious sample correlations

Assume that we calculate a set of k independent seismic


attributes from a 3-D seismic volume. There are numerous
attributes that are considered in geostatistical or seismic
attribute analysis studies -- functions of peak and trough
amplitudes, interval travel times, acoustic impedance,
instantaneous phase, frequency, or amplitude and many
others. Also, suppose we have n locations where the
measured reservoir property can be tested for correlation
with the k seismic attributes. A tempting approach is to
calculate the sample correlation (Eq. 2) of each of the k
seismic attributes with the reservoir property and select any
that have an absolute value that exceeds some threshold,
say 0.7. If the true value of the correlation were indeed 0.7,
we would then know that 49% of the variance of the
reservoir property could be accounted for by the variation
of the seismic attribute.

Figures 1 and 2 reveal the factors that influence the


probability of observing spurious correlations. Figure 1
shows the probability of observing at least one sample
correlation, between the property measured at the well and
a seismic attribute, greater than or equal to 0.4 when the
actual correlation is zero. Figure 2 presents the same
results for an observed sample correlation of 0.6. These
figures illustrate that as the number of well data points Probability tables and their derivation
decreases or the number of attributes considered increases,
the likelihood of observing at least one spurious sample The probability of observing the absolute value of a sample
correlation increases. Comparing Figures 1 and 2, one can correlation (r) greater than or equal to R, given 0, can
also see that the probability of a spurious correlation
be found from the relationship [2]
decreases, for a given sample size and number of seismic
attributes considered, as the magnitude of the observed
sample correlation increases. = = (3)

where n is the sample size or number of lo&ions


with measurements of both the reservoir
property and the seismic attribute
and t is distributed as a Student’s t with
n-2 degrees of freedom.

It can be readily shown that if is the probability of


observing a spurious correlation when considering a single
attribute, then the probability of observing at least one
spurious correlation when considering a set of k attributes is

= (4)

From Eq. 4 we can see that the penalty for increasing the
number of attributes considered from k - 1 to k is

(5)

Tables 1-5 give the probability of observing a spurious


sample correlation for one or more seismic attributes out of
a set of k = 5, 10, 20, 40, and 60 independent attributes.
Each table gives the probability of observing a sample
correlation greater than or equal to 0.1, 0.2, 0.3, 0.4, 0.5,
0.6,0.7,0.8, and 0.9 when the reservoir property is actually
uncorrelated with the attributes, given a sample size of 5,
10, 15,20,25,35, 50,75, or 100.

1790
Use of seismic attributes as predictors

Table 1 Table 5
Probability of observing absolute sample correlation R Probability of observing absolute sample correlation R
for at least one of a set of k = 5 seismic attributes for at least one of a set of k = 60 seismic attributes
when well property and attributes actually uncorrelated when well property and attributes actually uncorrelated
Sample Size Sample Size
R 5 10 15 20 25 35 50 75 100 R 5 10 15 20 25 35 50 75 100
0.10 1.00 1.00 1.00 1.00 0.99 0.98 0.97 0.92 0.86 0.10 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
0.20 1.00 0.99 0.96 0.92 0.87 0.76 0.59 0.36 0.21 0.20 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.94
0.30 0.99 0.92 0.80 0.67 0.54 0.34 0.16 0.04 0.01 0.30 1.00 1.00 1.00 1.00 1.00 0.99 0.88 0.42 0.14
0.40 1.00 1.00 1.00 0.99 0.95 0.65 0.21 0.02 0.00
0.40 0.97 0.77 0.53 0.34 0.22 0.08 0.02 0.00 0.00 0.50 1.00 1.00 0.97 0.78 0.48 0.13 0.01 0.00 0.00
0.50 0.92 0.53 0.26 0.12 0.05 0.01 0.00 0.00 0.00 0.60 1.00 0.98 0.66 0.27 0.09 0.01 0.00 0.00 0.00
0.60 0.81 0.29 0.09 0.03 0.01 0.00 0.00 0.00 0.00 0.70 1.00 0.77 0.20 0.03 0.01 0.00 0.00 0.00 0.00
0.70 0.65 0.12 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.80 1.00 0.28 0.02 0.00 0.00 0.00 0.00 0.00 0.00
0.80 0.42 0.03 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.90 0.90 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.90 0.17 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 These tables can be used to assess the risk of selecting a
seismic attribute that is actually uncorrelated with the
Table 2 reservoir property being modeled. For example, look at the
Probability of observing absolute sample correlation column headed 10 and row labeled 0.60 of Table 3. We see
for at least one of a set of k = 10 seismic attributes that there is a 75% chance of observing a sample
when well property and attributes actually uncorrelated correlation coefficient 0.60 for at least one of the 20
Sample Size attributes considered, given we have only 10 well
R 5 10 15 20 25 35 50 75 100 measurements, when actually the true correlation is zero.
0.10 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 0.98
0.20 1.00 1.00 1.00 0.99 0.98 0.94 0.83 0.59 0.38 Risks associated with selection of seismic attribute
0.30 1.00 0.99 0.96 0.89 0.79 0.57 0.29 0.09 0.02
0.40 1.00 0.95 0.78 0.57 0.39 0.16 0.04 0.00 0.00 There are two types of error in the process of selecting a
0.50 0.99 0.78 0.45 0.22 0.10 0.02 0.00 0.00 0.00 seismic attribute to use as a predictor of a reservoir
0.60 0.96 0.50 0.17 0.05 0.02 0.00 0.00 0.00 0.00 property. Figure 3 identifies the possible events. To
0.70 0.88 0.22 0.04 0.01 0.00 0.00 0.00 0.00 0.00 rigorously assess the risk associated with each of the errors,
0.80 0.67 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 their probability of occurrence and cost must be quantified.
0.90 0.32 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Table 3 Well property and Well property and
seismic attribute seismic attribute
Probability of observing absolute sample correlation R are actually are actually
for at least one of a set of k = 20 seismic attributes
when well property and attributes actually uncorrelated uncorrefated correlated
Select seismic
Sample Size attribute for use in
R 5 10 15 20 25 50 75 100 predicting
0.10 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 reservoir property
0.20 1.00 1.00 1.00 1.00 1.00 1.00 0.97 0.83 0.61
0.30 1.00 1.00 1.00 0.99 0.96 0.81 0.50 0.16 0.05 Reject seismic
0.40 1.00 1.00 0.95 0.81 0.62 0.29 0.08 0.01 0.00 attribute for use in Correct decision
0.50 1.00 0.95 0.70 0.39 0.20 0.04 0.00 0.00 0.00 predicting (no error occurs) Type II Error
0.60 1.00 0.75 0.31 0.10 0.03 0.00 0.00 0.00 0.00 reservoir property
0.70 0.98 0.39 0.07 0.01 0.00 0.00 0.00 0.00 0.00
0.80 0.89 0.10 0.01 0.00 0.00 0.00 0.00 0.00 0.00 Figure 3. Possible outcomes when selecting a seismic attribute as a
0.90 0.53 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 predictor of a reservoir property.
Table 4
Probability of observing absolute sample correlation R A Type I error would occur if no relationship truly existed
for at least one of a set of k = 40 seismic attributes between the seismic attribute and the reservoir property of
when well property and attributes actually uncorrelated interest, yet we select the seismic attribute as a predictor
because it has a strong sample correlation with the reservoir
Sample Size property being modeled. A Type II error would occur if a
R 5 10 15 20 25 35 50 75 100 physical relationship did exist between the seismic attribute
0.10 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
0.20 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.97 0.85 and the reservoir property of interest, but we reject the
0.30 1.00 1.00 1.00 1.00 1.00 0.96 0.75 0.30 0.09 seismic attribute for use as a predictor (even though the
0.40 1.00 1.00 1.00 0.97 0.86 0.50 0.15 0.01 0.00 observed sample correlation is strong) because the physical
0.50 1.00 1.00 0.91 0.63 0.36 0.09 0.01 0.00 0.00 relationship is not understood.
0.60 1.00 0.94 0.52 0.19 0.06 0.01 0.00 0.00 0.00
0.70 1.00 0.62 0.14 0.02 0.00 0.00 0.00 0.00 0.00 There is a cost associated with each of these types of errors.
0.80 0.99 0.20 0.01 0.00 0.00 0.00 0.00 0.00 0.00 The cost of a Type I error is predictions that are (a) less
0.90 0.78 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 accurate (i.e., the estimated value is farther away from the

1791
Use of seismic attributes as predictors

true value) but at the same time, (b) more precise than they [2] F. A. Graybill. Theory and Application of the Linear
would have been had we not used the seismic attribute. Model. Duxbury Press, Belmont, CA, 1975.
This results in interval estimates of the reservoir property
being interpolated that are too narrow, and likely not to [3] J. Aitchison and I. R. Dunsmore Statistical Prediction
contain the true value. The cost of a Type II error is Analysis. Cambridge University Press, Cambridge,
predictions that are (a) less accurate and (b) less precise 1975.
than they would have been had we used the related seismic
attribute. This results in confidence intervals on the true
value of the property being predicted that are too wide, but
likely to contain the true value.

If the attribute selected is actually uncorrelated with the


property, but appears significantly correlated based on a
small number of wells, the cost will be that associated with
making inaccurate predictions with confidence. The other
side of this coin is the risk of discarding an attribute, which
is truly correlated with the reservoir property, because the
correlation cannot be explained. The probability of this
occurring cannot be quantified, but the cost is that
associated with having more uncertainty in the predicted
property than would have been the case had the attribute
been used.

Conclusions

It has been shown that the probability of observing


spuriously high sample correlations between well and
seismic data can be quite large when either the number of
wells is small and/or the number of seismic attributes
considered is large. Selection of a seismic attribute for use
as a covariate in geostatistical simulation, regression or
neural network predictions based solely on the strength of
the sample correlation coefficient can be risky. This is
because such a selection could lead to confident, but
inaccurate predictions and ultimately poor decisions.
Therefore, it is strongly recommended, especially when the
number of wells is small, that only those seismic attributes
that have a physically justifiable relationship with the
reservoir property be considered as candidates for
predictors.
If one decides to use an attribute which has a significant
likelihood that its correlation with the reservoir property is
spurious, then it is recommended that the influence of the
uncertainty in the magnitude of the correlation be
incorporated. This uncertainty can be included in most
prediction techniques using a Bayesian [3] approach.

Acknowledgments

I would like to thank Mobil Technology Corporation for


permission to publish this paper. I am also grateful to Sean
Boemer, Stan Cullick and Naji Saad for their careful
review of this paper.

References

[1] C. Deutsch and A. Joumel. GSLIB: Geostatistical


Software Library. Oxford University Press, New
York, NY, 1992.

1792

You might also like