See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/253461672

Use of multiattribute transforms to predict log properties from seismic data

Article in Geophysics · January 2001

DOI: 10.1190/1.1444899

CITATIONS READS
587 2,316

3 authors, including:

Daniel P. Hampson
CGG
38 PUBLICATIONS 2,218 CITATIONS

SEE PROFILE

All content following this page was uploaded by Daniel P. Hampson on 31 October 2018.

The user has requested enhancement of the downloaded file.

:
Use of Multi-Attribute Transforms to Predict Log Properties from Seismic Data

Dan Hampson1, Jim Schuelke2, John Quirein3

Abstract

In this paper, we describe a new method for predicting well log properties from seismic
data. The analysis data consists of a series of target logs from wells which tie a 3-D
seismic volume. The target logs may be theoretically of any type; however, the greatest
success to date has been in predicting porosity logs. From the 3-D seismic volume a
series of sample-based attributes is calculated. The objective is to derive a multi-attribute
transform, which is a linear or non-linear transform between a subset of the attributes and
the target log values. The selected subset is determined by a process of forward step-wise
regression, which derives increasingly larger subsets of attributes. An extension of
conventional cross-plotting involves the use of a convolutional operator to resolve
frequency differences between the target logs and the seismic data.

In the linear mode, the transform consists of a series of weights, which are derived by
least-squares minimization. In the non-linear mode, a neural network is trained, using the
selected attributes as inputs. Two types of neural networks have been evaluated: the
multi-layer feedforward network (MLFN), and the probabilistic neural network (PNN).
Because of its mathematical simplicity, the probabilistic neural network appears to be the
network of choice.

To estimate the reliability of the derived multi-attribute transform, cross-validation is

used. In this process, each well is systematically removed from the training set, and the
transform is re-derived from the remaining wells. The prediction error for the hidden
well is then calculated. The validation error, which is the average error for all hidden
wells, is used as a measure of the likely prediction error when the transform is applied to
the seismic volume.

The method is applied to two real data sets. In each case, we see a continuous
improvement in predictive power as we progress from single-attribute regression to linear
multi-attribute prediction to neural network prediction. This improvement is evident not
only on the training data, but more importantly, on the validation data. In addition, the
neural network shows a significant improvement in resolution over that from linear
regression.

1
Hampson-Russell Software Services Ltd, 510-715 5th Ave, SW, Calgary, Canada; E-mail: dan@hampson-
russell.com
2
Formerly Mobil Technology Company, Dallas, Texas, USA; presently ExxonMobil Upstream Research
Company, Houston, Texas, USA; E-mail: [email protected]
3
Formerly Mobil Technology Company, Dallas, Texas, USA; presently Halliburton Energy Services,
Houston, Texas, USA; E-mail: [email protected]

1
 Introduction

The integration of well log data and seismic data has been a consistent aim of
geoscientists. This has become increasingly important (and successful) in recent years
due to the shift from exploration to development of existing fields, with large numbers of
wells penetrating them. One type of integration is the forward modeling of synthetic
seismic data from the logs. A second type of integration is the inverse modeling of the
logs from the seismic data. This is called seismic inversion, and has been described by
numerous authors (e.g., Lindseth (1979), Cooke et al (1983), Oldenburg et al (1983), Chi
et al (1984)).

In this paper, we attempt to go beyond the limits of conventional seismic inversion in

several ways. First, we will directly predict log properties other than acoustic impedance,
such as porosity. This differs from previous authors who have usually modeled porosity
from the impedance derived from the inversion (Anderson (1996)). A second difference
is that we will use attributes derived from the seismic data rather than the conventional
post-stack data itself. This allows us to include pre-stack information, as well as non-
linear transforms of the post-stack data. Thirdly, instead of assuming a particular model
relating the logs and the seismic, a statistical relationship will be derived by analyzing a
set of training data at well locations. This relationship will be either linear (multivariate
regression) or non-linear (neural network). Finally, we will use the concept of cross-
validation to estimate the reliability of the derived relationship.

After describing the theoretical basis of the method, two real data examples, which
emphasize the enhanced resolution obtainable from the method, will be shown.

Multi-Attribute Linear Regression

Seismic Attributes

In this methodology, our aim is to find an operator, possibly non-linear, which can
predict well logs from neighboring seismic data. In fact, we choose to analyze not the
seismic data itself, but attributes of the seismic data. One reason why we expect this to
be more beneficial than the raw seismic data is that many of these attributes will be non-
linear, thus increasing the predictive power of the technique. A second reason is that there
is often benefit in breaking down the input data into component parts. This process is
called pre-processing or feature extraction and it can often greatly improve the
performance of a pattern recognition system by reducing the dimensionality of the data
before using it to train the system. Pre-processing can also provide a means of adding
prior knowledge into the design of the pattern recognition system.

We define a seismic attribute generally as any mathematical transform of the seismic

trace data. This, for example, includes simple attributes such as trace envelope,
instantaneous phase, instantaneous frequency, etc., and complicated attributes such as
seismic trace inversion and AVO attributes. The transform may or may not incorporate

2
other data sources. Note, for example, that trace inversion assumes other data sources,
such as the seismic wavelet, the initial guess model, constraints, etc. However, for this
analysis, we still consider the inversion result to be an attribute of the seismic trace.

As pointed out by Chen (1997), seismic attributes may be divided into two categories:

- Horizon-based attributes : These are average properties of the seismic trace between
two boundaries, generally defined by picked horizons.

- Sample-based attributes : These are transforms of the input trace in such a way as to
produce another output trace with the same number of samples as the input.

In this paper, we consider only sample-based attributes. In theory, it should be possible

to transform the Horizon-based attribute into a Sample-based attribute by repeating the
average property as many times as there are output samples between the bounding
horizons. However, we do not address this possibility in this paper.

Taner et al.(1994), has provided a long list of sample-based attributes, of which our
experience has shown these to be particularly important for log prediction:

- integrated trace
- integrated absolute amplitude of the trace
- near angle stack
- AVO intercept / gradient
- frequency and absorption estimates
- seismic velocity

Conventional Cross Plotting

Given a particular attribute of the seismic data, the simplest procedure for deriving the
desired relationship between target data and seismic attribute is to cross plot the two:

Figure 1: Conventional cross plot between the target log “den-porosity’ and the seismic attribute
“Attribute”.

3
Figure 1 shows an example in which a target log property, in this case “den-porosity”, is
plotted against a seismic attribute, called “Attribute”. The assumption is that the target
log has been integrated to travel-time at the same sample rate as the seismic attribute.
Effectively, this integration reduces the target log to the same resolution as the attribute,
which is usually significantly coarser than the log property. Each point in the cross plot
consists of a pair of numbers corresponding to a particular time sample.

Assuming a linear relationship between the target log and the attribute, a straight line
may be fit by regression:

y = a + bx (1)

The coefficients a and b in this equation may be derived by minimizing the mean-squared
prediction error:

1 N
E   ( yi  a  bxi ) 2
2
(2)
N i1
where the sum is over all points in the cross plot.

The calculated prediction error, E, is a measure of the goodness-of-fit for the regression
line defined by equation (1). An alternative measure is the normalized correlation
coefficient defined by:

 xy

 x y (3)

where:

N
1
 xy 
N
 (x
i 1
i  mx )( yi  m y ) (4)

N
1
x 
N
 (x
i 1
i  mx ) 2 (5)

N
1
y 
N
(y
i 1
i  my ) 2 (6)

4
 N
1
mx 
N
x
i 1
i (7)

N
1
my 
N
y
i 1
i (8)

Note that the linear requirement may be relaxed somewhat by applying a non-linear
transform to either the target data or the attribute data or both, as shown in Figure 2:

Figure 2: Applying non-linear transforms to both the target and the attribute data improves the correlation
between the two.

Extension of Cross Plotting to Multiple Attributes

The extension of the conventional linear analysis to multiple attributes (multivariate

linear regression) is straightforward. Assume, for simplicity, that we have three attributes
as shown in Figure 3:

Figure 3: Assuming the case of three seismic attributes, each target log sample is modeled as a linear
combination of attribute samples at the same time.

5
At each time sample, the target log is modeled by the linear equation:

L(t) = w0+w1A1 (t)+w2 A2 (t)+w3 A3 (t) (9)

The weights in this equation may be derived by minimizing the mean-squared prediction
error, as extended from equation (2):

N
1
E 2

N
 (L  w
i 1
i 0  w1 A1i  w2 A2i  w3 A3i )2 (10)

As shown in the Appendix, the solution for the four weights produces the standard
normal equations:

1
 w0   N A A  A    L 
1i 2i 3i i
w   A
 1     1i  A  A A  A A   A L 
2
1i 1i 2i 1i 3i 1i i

 w2   A2i  A A  A  A A   A L  2 (11)

  
1i 2i 2i 2i 3i 2i i

 w3   A3i  A A  A A  A   A L  2

1i 3i 2i 3i 3i 3i i

Just as in the single-attribute case, the mean-squared error (10) calculated using the
derived weights constitutes a goodness-of-fit measure for the transform, as does the
normalized correlation, defined in (3), where the x-coordinate is now the predicted log
value and the y-coordinate is the real log value.

Use of the Convolutional Operator

The derivation of the multi-attribute regression assumes a single weight for each
attribute. The problem with this approach is illustrated in Figure (4):

Figure 4: A comparison between the target log on the left with the seismic attribute on the right emphasizes
the difference in frequency content. This observation suggests the use of a convolutional operator to
resolve the difference.

6
This figure shows that the frequency content of the target log is typically much higher
than that of the seismic attribute. Consequently, correlating the log with the attributes on
a sample-by-sample basis may not be optimal. The alternative is to assume that each
sample of the target log is related to a group of neighboring samples on the seismic
attribute as shown in Figure 5:

Figure 5: Using a five-point convolutional operator to relate the seismic attributes to the target log.

The use of the convolutional operator is also suggested by the classic convolutional
model in geophysics. If the well log, for example, happens to be acoustic impedance,
then the five-point operator shown in Figure (5) is closely related to the seismic wavelet.
In general, for any other log property, we can expect the wavelet to “smear” the effects of
each log sample over a range of contiguous seismic samples.

The extension of equation (9) to include the convolutional operator is:

L = w0 + w1*A1 + w2*A2 + w3*A3 (12)

where * represents convolution, and the wi are operators of a specified length. Note that
the number of coefficients has now increased to:

(number of attributes times operator length) + 1

Once again , the operator coefficients may be derived by minimizing the mean-squared
prediction error:

1 N
E   ( Li  w0  w1 * A1i  w2 * A2i  w3 * A3i ) 2
2
(13)
N i1

7
As shown in the Appendix, this is equivalent to introducing a series of new attributes,
which are time-shifted versions of the original attributes.

Determining Attributes by Step-Wise Regression

In the previous sections, we have derived equations which allow us to determine optimal
operators for any given set of attributes. These operators are optimal in the sense that the
mean-squared prediction error between the actual target logs and the predicted target logs
is minimized. The next issue to be addressed is how to select the attributes.

One possible procedure could be Exhaustive Search. Let’s assume, for example, that we
wish to find the best M attributes out of a total list of N attributes, for a given operator
length, L. One obvious procedure is to try all combinations of M attributes. For each
combination, the optimal weights are derived using equation 11 above. That combination
with the lowest prediction error is then selected.

The problem with Exhaustive Search is that the computation time can very quickly
become excessive. Suppose, for example, that we have a total of N = 25 attributes, and
we wish to derive the best combination of M=5 attributes for an operator of length L=9.
In this case, there are 25*24*…*21 = 6,375,600 combinations of 5 attributes to be
checked. Each of these combinations requires the solution of a linear system with 5*9+1
= 46 unknowns.

A much faster, although less optimal, procedure is called Step-Wise Regression (Draper
and Smith, 1966). The assumption in this procedure is that if the best combination of M
attributes is already known, then the best combination of M+1 attributes includes the
previous M attributes as members. Of course, the previously calculated coefficients must
be re-derived. The process is illustrated in the series of steps:

Step 1: Find the single best attribute by Exhaustive Search. For each attribute in the list,

Amplitude Weighted Phase,

Average Frequency,
Apparent Polarity,
etc,

solve for the optimal coefficients and calculate the prediction error. The best attribute is
the one with the lowest prediction error. Call this attribute1.

Step 2: Find the best pair of attributes, assuming that the first member is attribute1. For
each other attribute in the list, form all pairs,

(attribute1, Amplitude Weighted Phase),

(attribute1, Average Frequency),
etc.

8
For each pair, solve for the optimal coefficients and calculate the prediction error. The
best pair is the one with the lowest prediction error. Call this second attribute from the
best pair attribute2.

Step 3: Find the best triplet of attributes, assuming that the first two members are
attribute1 and attribute2 . For each other attribute in the list, form all triplets,

(attribute1, attribute2, Amplitude Weighted Phase),

(attribute1, attribute2, Average Frequency),
etc.

For each triplet, solve for the optimal coefficients and calculate the prediction error. The
best triplet is the one with the lowest prediction error. Call this third attribute from the
best triplet attribute3.

Carry on this process as long as desired.

The first thing to note is that the computation time for this process will be much shorter
than for Exhaustive Search. For the example above, the number of combinations to
check is now 25+24+…+21 = 115, instead of 6,375,600. In addition, the size of the
linear system to be solved starts at 9+1 = 10 for the first 25 combinations and increases
linearly to 5*9+1 = 46 for the last 21 combinations.

The problem with Step-Wise Regression is that we cannot be sure of deriving the optimal
solution. In other words, the combination of five attributes found may not, in fact, be the
best five which would be found by Exhaustive Search. However, it can be shown that
each additional attribute found has a prediction error less than or equal to the previous
smaller combination. This can be proven by contradiction: if the new prediction error is
greater, then simply set all the weights to zero for this new attribute, and the prediction
error will be equal to the previous set.

One advantage of Step-Wise Regression is that it relieves us from the need to worry about
whether the attributes in the total list are linearly independent. This is because Step-Wise
Regression automatically chooses the next attribute whose contribution in a direction
orthogonal to the previous attributes is greatest. Assume, for example, that two of the
attributes, say Ai and Aj are scaled versions of each other : Aj = a + b*Ai . This would
represent the extreme case of linear dependence. As the Step-Wise Regression proceeds,
one or the other of them will be chosen first, say Ai. From then on, the other attribute, Aj,
will never be chosen. This is because once Ai is included, the improvement by adding the
other attribute, Aj, is precisely zero. In summary, because we are using Step-Wise
Regression, we may have any arbitrary total attribute list, and the only penalty incurred
by using linearly dependent attributes is computation time.

At this point, we can define the general term Multi-Attribute Transform:

9
 A Multi-Attribute Transform is a set of attribute types along with rules for
transforming the attributes into the desired output log.

In the analysis so far, the transformations are linear weights applied to either the
attributes themselves or to non-linear transforms of the attributes. The next section
extends this analysis to include Neural Networks.

Neural Networks

The analysis so far has been linear. The limitation which this imposes can be understood
by examining Figure 6:

Figure 6: Cross plot of target log against seismic attribute. The regression line has been fit by minimizing
Equation 2.

This figure shows a target log, called “P-wave”, cross plotted against a single seismic
attribute. As before, the regression line has been calculated by minimizing the mean-
squared prediction error. Visually, we might guess that a higher-order curve would fit the
points better. A number of options exist for calculating this curve. One option, which
has been discussed above, is to apply a non-linear transform to either or both the
variables, and fit a straight line to the transformed data. A second option is to fit a higher
order polynomial. In this section, we examine a third option, which is to use a Neural
Network to derive the relationship.

10
Multi-Layer Feedforward Neural Network

Neural Networks have been used for some years in geophysics (McCormack, M.D.,
1991; Schultz, et al., 1994; Schuelke, Quirein, and Sarg, 1997). Recently Liu (1998)
described the use of a multi-layer feedforward Neural Network (MLFN) to predict log
properties directly from seismic data. The MLFN is the traditional network shown in
Figure 7:

Figure 7: Multi-Layer Feedforward Neural Network architecture

The properties of the MLFN have been described in numerous textbooks (eg, Masters,
1994). The network consists of an input layer, an output layer, and one or more hidden
layers. Each layer consists of nodes, and the nodes are connected with weights. The
weights determine the result from the output layer. In our implementation, the input
layer has as many input nodes as there are attributes. If a convolutional operator is being
used, the number of effective attributes is increased by the operator length. For example,
for an operator length of 3, each attribute is repeated 3 times, corresponding to a time
sample shift of –1, 0, and +1. The output layer has one node, since we are predicting a
single log property. We use a single hidden layer, with the number of nodes set by
experimentation.

The training process consists of finding the optimum weights between the nodes. The
training is performed by presenting training “examples” to the network. Each example
consists of data for a single time sample:

{ A1, A2, A3, L}

where Ai are the attributes and L is the measured target log value. There are as many
training examples as there are cumulative seismic samples within the analysis windows
from all the wells available.

The problem of estimating the weights can be considered a non-linear optimization

problem, where the objective is to minimize the mean-squared error between the actual

11
target log values and the predicted target log values. This problem has traditionally been
solved by back-propagation, which is a form of gradient descent. Modern methods now
use conjugate-gradient and simulated annealing to speed convergence and avoid local
minima. (Masters, 1994).

As an example of the behavior of MLFN, Figure 8 shows the prediction curve for the
same data as Figure 6 using the MLFN with 5 nodes in the hidden layer. In this case,
since there is only one attribute, there is a single node in the input layer.

Figure 8: Prediction curve derived by Multi-Layer Feedforward Neural Network with 5 nodes in the hidden
layer. The data is the same as that for Figure 6.

Figure 8 demonstrates two aspects of the behavior of the MLFN. The positive aspect is
that the data values over most of the attribute range are modeled more accurately than is
the case with linear regression. The negative aspect is the instability apparent at the low
attribute values as the network attempts to model this data too closely. This is an
example of a condition known as “over-training”, which we will discuss in more detail
below.

Probabilistic Neural Network

An alternative type of Neural Network is the Probabilistic Neural Network (Masters,

1995, Specht, 1990, 1991). The Probabilistic Neural Network (PNN) is actually a
mathematical interpolation scheme, which happens to use a neural network architecture
for its implementation. This is a potential advantage, since by studying the mathematical
formulation, we can often understand its behavior much better than the MLFN (which
tends to be a black box).

The data used by the PNN is the same training data used by MLFN. It consists of a series
of training “examples”, one for each seismic sample in the analysis windows from all the
wells.

{ A11, A21, A31, L1}

12
 { A12, A22, A32, L2}
{ A13, A23, A33, L3}
.
.
.
{ A1n, A2n, A3n, Ln}

where there are n training examples and three attributes. The values Li are the measured
target log values for each of the examples.

Given the training data, the PNN assumes that each new output log value can be written
as a linear combination of the log values in the training data. For a new data sample with
attribute values:

x = {A1j, A2j, A3j}

the new log value is estimated as:

n

 L exp( D( x, x ))
i i
Lˆ ( x)  i 1
n (14)
 exp( D( x, x ))
i 1
i

where

2
x x 3
D( x, xi )    j ij 
  
j 1 
(15)
j 
The quantity D(x,xi) is the “distance” between the input point and each of the training
points xi. This distance is measured in the multi-dimensional space spanned by the
attributes, and is scaled by the quantity  j , which may be different for each of the
attributes.

Equations 14 and 15 describe the application of the PNN network. The training of the
network consists of determining the optimal set of smoothing parameters,  j . The
criterion for determining these parameters is that the resulting network should have the
lowest validation error.

Define the validation result for the mth target sample as:

13
  L exp(  D( x
i m , xi ))
Lˆ m ( xm )  im
n (16)
 exp( D( xm , xi ))
im

This is the predicted value of the mth target sample when that sample is left out of the
training data. Since we know the value of this sample, we can calculated the prediction
error for that sample. Repeating this process for each of the training samples, we can
define the total prediction error for the training data as:

n
EV ( 1 , 2 , 3 )   ( Li  Lˆi ) 2 (17)
i 1

Note that the prediction error depends on the choice of the parameters,  j . This quantity
is minimized using a non-linear conjugate gradient algorithm described in Masters
(1995). The resulting network has the property that the validation error is minimized.

The performance of the PNN on the simple cross plot data is shown in Figure 9:

Figure 9: Prediction curve derived by the Probabilistic Neural Network. The data is the same as that for
Figure 6.

From this figure, we can see that the PNN has the desirable characteristic of following the
data as closely as the MLFN, but does not have the same instability at the limits of the
attribute range. The biggest problem with the PNN is that because it carries around all its
training data and compares each output sample with each training sample, the application
time can be slow.

14
 Validation

In this section, we examine the question of how to determine the correct number of
attributes to use. As discussed previously, we can show that a multi-attribute transform
with N+1 attributes must always have a prediction error less than or equal to the
transform with N attributes. As more attributes are added, we can expect an
asymptotically declining prediction error, as shown in Figure 10:

Figure 10: Plot of prediction error against number of attributes used in the transform. Mathematically, the
curve must continue to decline asymptotically.

Of course, while the additional attributes always improve the fit to the training data, they
may be useless or worse when applied to new data, not in the training set. This is
sometimes called “over-training” and has been very well described by Kalkomey (1997).
Effectively, using higher numbers of attributes is analogous to fitting a cross plot with
increasingly higher order polynomials.

A number of statistical techniques have been derived to measure the reliability of the
higher order attribute fits (e.g., Draper and Smith, 1966). Unfortunately, most of these
techniques apply to linear regression, and are not immediately applicable to non-linear
prediction using neural networks. For this reason, we have chosen the process of Cross-
Validation, which can be applied to any type of prediction.

Cross-Validation consists of dividing the entire training data set into two subsets, the
Training Data Set and the Validation Data Set. The Training Data Set is used to derive
the transform, while the Validation Data Set is used to measure its final prediction error.
The assumption is that over-training on the Training Data Set will result in a poorer fit to
the Validation Data Set. This is illustrated in Figure 11:

15
Figure 11: Illustration of cross-validation. Two curves are used to fit the data points. The solid curve is a
low order polynomial. The dashed curve is a higher order polynomial. The dashed curve fits the training
data set better, but shows a poorer fit when compared with the validation data set.

In our analysis, the natural sub-division of data is by well. In other words, the Training
Data Set consists of training samples from all wells, except some specified hidden well.
The Validation Data Set consists of samples from that hidden well. In the process of
Cross-Validation, the analysis is repeated as many times as there are wells, each time
leaving out a different well. The total validation error is the root mean square average of
the individual errors:

N
1
  eVi
2 2
EV (18)
N i 1

where: EV is the Total Validation Error,

eVi is the validation error for well i, and
N is the number of wells in the analysis

Figure 12 shows the same plot as Figure 10, except that now the Total Validation Error
has been added. As expected, the Validation Error for any particular number of attributes
is always higher than the training error. This is because removing a well from the
training set will always result in a decrease in predictive power. Also note that the
Validation Error curve does not decline monotonically. In fact, it exhibits a broad local
minimum around four attributes, and then gradually increases. We interpret this to mean
that all additional attributes after the fourth are over-training the system. Generally, if a
Validation Error curve exhibits a distinct minimum, we assume that the number of
attributes at that point is optimum. If the Validation Error curve shows a broad
minimum, such as Figure 12, or shows a series of local minima, we select the point at

16
which the curve stops declining convincingly. This would correspond to the first two
attributes in Figure 12.

Figure 12: The same plot as Figure 10, except that the Total Validation Error is now shown as the upper
curve. Note that attributes past the second contribute little improvement to the Validation Error, and, in
fact, gradually cause an increase in prediction error.

Establishing the statistical significance of a prediction has been addressed by many

authors in the past. Specht (1991) points out that the PNN provides a “consistent
estimator”, asymptotically (i.e., given enough training samples) converging to the
underlying joint probability density function of the attributes and prediction variables.
Leonard et al (1992) show how cross-validation (the validation error) can be used to
calculate confidence intervals for the predictions. Leonard shows for a well, that the
confidence interval is directly proportional to the well validation error and inversely
proportional to the square root of the number of wells.

The extrapolation uncertainty must also be addressed, i.e., the behavior of the prediction
between and away from the training wells. Leonard suggests an approach to determine if
there is sufficient training data by estimating local training data density in the
neighborhood of the input sample to be used for the prediction. We plan on
implementing a variation of this approach in future research. Currently, extrapolation
accuracy is assessed by visual observation of the seismic predictions between wells and
comparison with the actual seismic data.

Example 1

The first example comes from the Blackfoot area of Western Canada. This data has been
supplied by the University of Calgary Crewes Consortium, and consists of a 3-D seismic
volume which ties 13 wells. The primary target is the Glauconitic member of the

17
Mannville group. The reservoir occurs at a depth of around 1550m (1060ms), where
Glauconitic sand and shale fill valleys incised into the regional Mannville stratigraphy.
The objectives of the survey are to delineate the channel and distinguish between sand-
fill and shale-fill. Each of the wells contains a porosity log, which is used as the target
for this example. The seismic volume has been processed through a model-based
inversion algorithm to produce an acoustic impedance volume. This volume is used as an
attribute in the process. Figure 13 shows a single inline from each of the seismic
volumes.

Figure 13: A single inline from the input 3-D volumes. The upper plot shows the post-stack seismic data.
The lower plot shows the seismic inversion result, which is used as an attribute for this analysis. The color
scale is acoustic impedance.

18
For each of the 13 wells, a single composite trace has been extracted from the 3-D
volumes by averaging the 9 nearest traces around the borehole. The training data for one
of the wells is shown in Figure 14.

Figure 14: Training data for a single well. The curve on the left is the target porosity log from the well.
The center curve is the composite seismic trace from the 3-D volume at the well location. The right curve
is the composite acoustic impedance from the seismic inversion. The horizontal red lines show the analysis
window.

Note that the porosity log has been converted from depth to time and sampled at the same
2 ms sample rate as the seismic data. Because we will be correlating over a time
window, the quality of the depth to time conversion is critical for this process. The
analysis window is indicated by the horizontal red lines and is less than 100 ms. Figure
15 shows a cross plot of the target porosity values against the Seismic Inversion attribute,
using points from the analysis windows of all 13 wells. The normalized correlation
coefficient is 0.41, indicating a fairly poor correlation using this attribute alone.

Figure 15: Cross plot of porosity against acoustic impedance from the seismic inversion, using points
within the analysis windows from all 13 wells. The normalized correlation is .41.

19
The attributes which are available for the multi-attribute analysis consist of 17 attributes
derived from the seismic trace plus the external attribute Seismic Inversion. The
complete list is:

Amplitude Envelope
Amplitude Weighted Cosine Phase
Amplitude Weighted Frequency
Amplitude Weighted Phase
Average Frequency
Apparent Polarity
Cosine Instantaneous Phase
Derivative
Derivative Instantaneous Amplitude
Dominant Frequency
Instantaneous Frequency
Instantaneous Phase
Integrate
Integrated Absolute Amplitude
Second Derivative
Time
Seismic Inversion

The list is further increased by applying the following non-linear transforms to each of
the previous attributes:

Natural Log
Exponential
Square
Inverse
Square Root

The analysis consists of applying the step-wise regression described above. A seven-
point convolutional operator was used. The tabular results are shown in Figure 16. Each
line of this table shows a multi-attribute transform with an increasing number of
attributes. The first line, for example, shows that the single best attribute is 1/(Seismic
Inversion). Using this attribute with a seven-point convolutional operator gives a
prediction error of 5.5%. (This error is the absolute prediction error in the units of
Porosity, which is %). The second line shows that the best pair of attributes is 1/(Seismic
Inversion) and Integrate. Using this pair gives a prediction error of 5.08%. The table
shows combinations up to 10 attributes.

20
Figure 16: The results of step-wise regression, applied to the porosity prediction problem. Each line shows
a different multi-attribute transform with the number of attributes listed in the first column. The multi-
attribute transform for each line includes all attributes above it. The prediction error for that transform is
shown in the last column in the units of the target log (i.e., % porosity).

The same information is displayed graphically in Figure 17, which also shows the
validation results. The lower curve is the prediction error when all wells are used in the
analysis. As expected, this curve decreases as attributes are added. The upper curve
shows the average validation error, as defined above. We interpret this curve to mean
that adding attributes after the sixth causes over-training of the system.

21
Figure 17: The results of step-wise regression in graphical form. The lower curve shows the prediction
error when all wells are used in the analysis. The upper curve shows the validation error, as defined
previously.

Figure 18 shows the result of applying the derived multi-attribute transform with six
attributes. In this figure, the original porosity logs are shown in black with the predicted
logs in red. The normalized correlation for all the wells is now 0.69, compared with the
cross plot correlation of 0.41, derived from a single attribute alone.

22
Figure 18: Applying the multi-attribute transform using six attributes and a seven-point operator. Only the
first three wells are shown. The original porosity log is shown in black. The predicted log is shown in red.
The normalized correlation coefficient for all the wells is 0.69.

Figure 19 shows a similar plot, but in this case the predicted log for each well has been
calculated with a different transform. For each well, the same six attributes were used,
but the weights were re-calculated, using reduced training data which did not include that
target log. Effectively this simulates the result of drilling that well after the analysis.

23
Figure 19: The validation result for the multi-attribute analysis. This is the same as Figure 18 except that
the multi-attribute transform for each well has been re-derived with that well data removed from the
analysis. This simulates the effect of drilling the well after the prediction. The normalized correlation for
all the wells is 0.60.

Figure 20 shows the distribution of prediction errors over the 13 wells. While the
distribution is highly variable, the difference between the two curves is fairly consistent
at about 0.4% in porosity units.

Figure 20: The prediction errors for each of the 13 wells. The lower curve shows the prediction error when
the specified well is used in the analysis. The upper curve shows the validation error, when the well is not
used in the analysis.

24
Using the same six attributes, the Probabilistic Neural Network was trained, as outlined
in the previous section. Figure 21 shows the prediction results for the PNN. Because the
PNN contains a copy of all the target data within its operator, prediction results are
always higher than is the case with linear regression. Mathematically this is analogous to
kriging in which derived maps will always honor the input well information. As with the
kriging technique (Deutsch and Journel, 1992), the real measure of performance is cross-
validation, and this is shown in Figure 22. The average normalized correlation for all the
wells is 0.62. While this is only marginally better than the validation result for
multivariate linear regression (0.60), the high frequency enhancement for thin layers can
be observed at specific locations (eg, between 1050ms and 1075ms).

Figure 21: Applying the Probabilistic Neural Network using six attributes and a seven-point operator. Only
the first three wells are shown. The original porosity log is shown in black. The predicted log is shown in
red. The normalized correlation coefficient for all the wells is 0.95.

25
Figure 22: The validation result for the Probabilistic Neural Network. This is the same as Figure 21 except
that PNN for each well has been re-derived with that well data removed from the analysis. The normalized
correlation for all the wells is 0.62.

Ultimately the value of multi-attribute analysis has to be measured in terms of its

improvement over the basic cross plotting of a single attribute. Figure 23 shows the
continuous improvement in prediction ability as we move from a cross plot using the
single best attribute (Acoustic Impedance) to multivariate linear regression with the seven
point operator and six attributes to the PNN. In each case, we are showing the validation
result, which is the expected performance on this well in a blind test. Note, in particular,
the enhancement of the thin bed resolution.

26
 Acoustic Impedance Multivariate Linear Neural Network
Regression

Figure 23: Comparison of validation results for a single well using three prediction methods. Each panel is
a blind test, i.e., this well was not used in the operator derivation. The left panel shows linear regression
applied to the Acoustic Impedance volume. The middle panel shows multivariate linear regression using six
attributes and a seven-point operator. The right panel shows the result of using Probabilistic Neural
Network with the same six attributes.

Each of the three derived transforms was then applied to the 3-D seismic and inversion
volumes. The result in each case is a volume of estimated porosity. Figure 24 shows a
single inline through each of the volumes. The anomaly at about 1090 ms is a known
sand channel. Note the high resolution result achieved with the Probabilistic Neural
Network.

27
Figure 24: Application of the derived transforms to the 3-D volumes. The upper panel shows the
regression curve applied to the acoustic impedance volume. The middle panel shows the multivariate
linear transform with six attributes and a seven-point operator. The lower panel shows the Probabilistic
Neural Network with six attributes and a seven-point operator. The inserted log is the target porosity log at
this location. The color scale is in the units of % porosity.

28
Finally, Figure 25 shows a data slice through the PNN volume. The data slice tracks 12
ms below the high porosity marker at 1070 ms and averages the estimated porosity for
the next 6 ms. We can see clearly the high porosity channel trending north-south through
the center of the volume.

Figure 25: A data slice through the porosity volume estimated using the Probabilistic Neural Network. The
color scale is in the units of % porosity.

Example 2

The second example comes from the Pegasus Field in West Texas. The Pegasus Field is
located 25 miles south of Midland, Texas in the southwestern portion of what is now the
Midland basin. Hydrocarbons are structurally trapped at Pegasus in a northeast to
southwest trending faulted anticline, 7 miles long (N-S) by 4 miles wide (E-W). The
Devonian reservoir, at a depth of 11,500 feet, is one of six producing intervals that range
in depth from 5,500 to 13,000 feet. A more detailed description of the field, the geology,
and the reservoir is given by Schuelke, Quirein and Sarg (1997). The purpose of this
example is to show the improvement in resolution both vertically and horizontally using
the PNN neural network vs. the multivariate linear regression.

The same validation procedure followed in Example 1 was used to determine which and
how many seismic attributes to use in this porosity prediction exercise (see Schuelke,

29
Quirein and Sarg, 1997). Ten wells provided the calibration between the well porosity
and the seismic attributes. The target region was restricted to the Devonian interval
approximately 100 ms, or 500 feet thick. The case study area was a small 2 by 3 mile
subset of the full 3-D survey area. Figure 26 shows a seismic inline from the 3-D survey
that intersects one of the ten calibration wells. The red curve is the porosity log for the
Peg 21-06 well. Deflection to the right shows increasing porosity. The Top Devonian is
at 1710 ms, and the Base Devonian is at 1820 ms for this well location. There is one very
thin high porosity (24%) zone in this well at approximately 1765 ms. The intervals above
and below this high porosity interval in the Devonian are tight, less than 5% porosity.
The intervals above and below the Devonian interval are shales. The porosity log shows
false porosity in these zones.

Figure 26. An example seismic line from the 3D survey through the Peg 21-06 well. The porosity log is
show as the red curve with increasing porosity to the right.

Figures 27 and 28 show the porosity predictions from the seismic attributes for the same
seismic line. Figure 27 is the multivariate linear regression prediction, and Figure 28 is
the prediction using the PNN neural network. The color scale is in porosity units, with
the higher porosity colored light green and yellow. The tight to very low porosity values
are colored gray. Both predictions show the porosity development within the mid
Devonian interval. The multivariate linear regression results, however, show a smoothed,
or more averaged, result. The PNN neural network prediction retains more of the
dynamic range and high frequency porosity content as exhibited by the porosity logs at
the well locations. The highs and lows of porosity are retained, as well as the time
resolution. This is to be expected, as the neural network result is a non-linear function
that more closely follows the training or control data from the wells, while the linear
regression approach provides only an average fit to the control data. Away from the well
control the PNN results show the lateral porosity variability expected in this stratigraphic
controlled reservoir. Because the network has been trained over a relatively small time
window, and because the attributes are calculated from bandlimited seismic data, we do
not expect the general trend or low-frequency component of the predicted porosity to be

30
reliable. To some extent, this trend has been provided by the seismic inversion attribute,
but the trend in that attribute is itself, derived from the initial model for the inversion.

Figure 27. The multivariate linear regression result. Porosity is shown in color. The log porosity for Well
Peg 21-06 is displayed as a black curve. The thin high porosity zone in the middle of the Devonian is
correctly identified, but the temporal resolution is less on the prediction and the absolute value is less than
for the actual porosity log.

Figure 28. The PNN neural network result. Porosity is shown in color. The log porosity for Well Peg 21-
06 is displayed as a black curve. The thin high porosity zone in the middle of the Devonian is correctly
identified, and both the thickness and the absolute value of the prediction matches the log porosity. Away
from the well control the predictions show the lateral variability expected in this reservoir.

31
The benefits of this improved vertical and lateral resolution are evident on a time slice
view through the two porosity volumes. Figure 29 is a time slice through the multivariate
linear regression result at a time of 1760 ms. The maximum porosity value from a 10 ms
window centered at this time is displayed. The porosity color coding is the same as for
the inline displays in Figures 27 and 28. The multivariate linear regression results show
the general outline of the higher porosity zone for this time slice through the reservoir
interval. However, much of the lateral variability in porosity and the higher values of
porosity are missing. Figure 30 is the time slice through the PNN neural network
prediction at the same time interval. The PNN result shows the lateral variability of
porosity better and matches the extremes of porosity, as indicated by the log data. This
degree of resolution is required in estimating flow barriers or designing a horizontal
drilling program. Indeed, possible porosity barriers can be seen in Figure 30 as low
porosity zones (gray colors) between the high porosity zones (green color).

Figure 29. A time slice through the multivariate linear regression results. The predicted porosity is shown
in color. Producible porosity zones are in the bright colors grading from red to yellow, with the highest
porosity shown in green. Tight zones are indicated with the gray colors.

32
Figure 30. A time slice through the PNN neural network results. The predicted porosity is shown in color.
Producible porosity zones are in the bright colors grading from red to yellow, with the highest porosity
shown in green. Tight zones are indicated with the gray colors.

Conclusions

We have demonstrated the use of multiple seismic attributes for predicting well log
properties. In our analysis, seismic attributes are defined as any sample-based transform
of the seismic data. Two mathematical formulations have been used: multivariate linear
regression, and Neural Network prediction. For each of these cases, the selection of
appropriate attributes has been determined by forward step-wise regression, which builds
up groups of attributes sequentially. We have introduced a modification of conventional
regression which includes a convolutional operator applied to each of the attributes. This

33
operator is assumed to be time-invariant – hence, the process is applied to a targeted time
window. For any derived multi-attribute transform, the measure of performance has been
cross-validation, which systematically removes wells from the analysis and measures the
prediction error for these wells.

We have described two types of Neural Network, as applied to this problem: the multi-
layer feed forward network (MLFN), and the Probabilistic Neural Network (PNN). Each
of these networks uses the same attributes derived by the multivariate linear regression
analysis. In each case, we expect an increase in resolution due to the non-linear behavior
of the network.

We have demonstrated this methodology on two data sets. In each case, we have seen an
increase in predictive power and resolution as we progress from conventional cross-
plotting to multivariate linear regression to Neural Network. This improvement is
evident not only on the training data, but is supported by the validation data as well.

In summary, the methodology can be thought of as an extension of conventional seismic

inversion. Both processes deal with the same input data (seismic and logs), and both
attempt to predict a log property. The main advantages of the new algorithm over
conventional inversion are:

- It predicts other logs besides acoustic impedance (eg., porosity)

- It may use other attributes besides the conventional stack for this purpose.
- It does not rely on any particular forward model.
- It can achieve greatly enhanced resolution.
- It does not require a knowledge of the seismic wavelet.
- It uses cross-validation as a measure of success.

All of these advantages are gained, however, only if there is sufficient well control. This
means not only a large enough number of wells, but also, a distribution of log data which
spans the range of expected conditions in the earth. Our current research is aimed at
quantifying how well these conditions are satisfied in practical exploration cases.

Acknowledgements

The authors are grateful for the help of Research Geophysicist Todor Todorov in
preparing and analyzing the data in Example 1, in addition to his continuous support
during the research project. We also thank Brian Russell for providing insights into the
mathematical operation of both the multivariate analysis and the PNN network. Finally,
we are grateful to the members of the University of Calgary CREWES Consortium and to
Mobil Exploration & Producing U.S., Midland for allowing us to publish the results
derived on their data.

34
 References
Anderson, J.K., 1996, Limitations of seismic inversion for porosity and pore fluid: Lessons from chalk
reservoir characterization exploration: 66th Annual Internat. Mtg., Soc. Expl. Geophys., Expanded
Abstracts, 309-312.

Chen, Q. and Sidney, S., 1997, Seismic attribute technology for reservoir forecasting and monitoring: The
Leading Edge, Vol. 16, No. 5, May, 1997, 445-456.

Chi, C. Y., Mendel, J. M. and Hampson, D., 1984, A computationally fast approach to maximum-
likelihood deconvolution : Geophysics, 49, no. 05, 550-565.

Cooke, D. A. and Schneider, W.A., 1983, Generalized lienar inversion of reflection seismic data:
Geophysics, 48, no. 06, 665-676.

Deutsch, C. V. and Journel, A. G., 1992, GSLIB Geostatistical Software Library and User’s Guide: Oxford
University Press.

Draper, N.R. and Smith, H., 1966, Applied regression analysis: John Wiley & Sons, Inc.

Kalkomey, C.T., 1997, Potential risks when using seismic attributes as predictors of reservoir properties:
The Leading Edge, Vol. 16, No. 9, 247-251.

Leonard, J.A., Kramer, M.A., and Ungar, L.H., 1992, Using radial basis functions to approximate a
function and its error bounds: IEEE Transactions on Neural Networks, Vol. 3, No. 4.

Lindseth, Roy O., 1979, Synthetic sonic logs – A process for stratigraphic interpretation: Geophysics, 44,
no. 01, 3-26.

Liu, Z., and Liu, J., 1998, Seismic-controlled nonlinear extrapolation of well parameters using neural
networks: Geophysics, Vol. 63, No. 6, 2035-2041.

Masters, T., 1994, Signal and image processing with neural networks: John Wiley & Sons, Inc.

Masters, T., 1995, Advanced algorithms for neural networks: John Wiley & Sons, Inc.

McCormack, M.D., 1991, Neural computing in geophysics: The Leading Edge, Vol. 10, No. 1, 11-15.

Oldenburg, D. W., Scheuer, T. and Levy, S., 1983, Recovery of the acoustic impedance from reflection
seismograms: Geophysics, 48, no. 10, 1318-1337.

Schuelke, J.S., Quirein, J.A., and Sarg, J.F., 1997, Reservoir architecture and porosity distribution, Pegasus
Field, West Texas -- an integrated sequence stratigraphy - seismic attribute study using neural networks:
67th Annual Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, 668-671.

Schultz, P.S., Ronen, S., Hattori, M., and Corbett, C., 1994, Seismic guided estimation of log propties,
Parts 1, 2, and 3: The Leading Edge, Vol.13, No. 5,6,&7, May, 1994, 305-310, June, 1994, 674-678, July,
1994, 770-776.

Specht, Donald, 1990, Probabilistic neural networks: Neural Networks, 3, 109-118.

Specht, Donald, 1991, A general regression neural network, IEEE Transactions on Neural Networks, 2(6),
568-576.

Taner, M.T., Schuelke, J.S., O’Doherty, R., Baysal, E., 1994, Seismic attributes revisited: 64 th Annual
Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, 94, 1104-1106.

35
 Appendix

Multi-Attribute Linear Regression

Multi-Attribute Linear Regression is an extension of simple linear regression to M

variables. That is, we will use M attributes, A1, A2, … , AM, to predict the log, L. To do
this, we must determine the M+1 weights, w0, w1, w2, …., wM, which, when multiplied
by the particular set of attribute values, give the closest result to the log in a least-squared
sense. For simplicity, assume that M=3. If we have N samples in our log, we can then
write the following set of equations:

L1 = w0 + w1A11 + w2A21 + w3A31

L2 = w0 + w1A12 + w2A22 + w3A32
. . . . .
. . . . .
. . . . .
LN = w0 + w1A1N + w2A2N + w3A3N (A-1)

where Aij is the jth sample of the ith attribute.

Notice that equations (A-1) can be written as:

 L1  1 A11 A21 A31   w0 

 L  1 A A22 A32   w1 
 2 12
(A-2)
         w2 
    
 LN  1 A1N A2 N A3 N   w3 

or:

L = AW (A-3)

where L is an (N x 1) matrix containing the known log values, A is an (N x 4) matrix

containing the attribute values, and W is a (4 x 1) matrix with the unknown weights.

This can be solved by least squares minimization to give:

W = [ATA]-1ATL (A-4)

As a detailed computation, note that:

1
 w0   N A A A
1i 2i 3i    Li 
w   A   
 1     1i A A A A A   A1i Li 
2
1i 1i 2i 1i 3i 

 w2   A2i A A A A A  A2i Li 
2  (A-5)
   2   
1i 2i 2i 2 i 3i

 w3   A3i A A A A A
1i 3i 2i 3i 3i   A3i Li 

36
Multi-Attribute Linear Regression using Convolutional Weights

Next, let us generalize the preceding equations by assuming that we have a convolutional
sum:

L = w0 + w1*A1 + w2*A2 +….+ wN*AN (A-6)

where w0 = a constant,
and wi = l-point convolutional filters.

To simplify, consider the above equation, using only two attributes and four sample
values. Also, consider the case of a 3-point operator, which we could write as:

wi = [wi(-1), wi(0), wi(+1)]

Under these circumstances equation (A-6) can be written in the following matrix form:

 L1   w1 (0) w1 (1) 0 0   A11   w2 (0) w2 (1) 0 0   A21 

L   w (1) w (0) w (1)    
0   A12  w2 (1) w2 (0) w2 (1) 0   A22 
 2  w   1 1 1

 L3  0
 0 w1 (1) w1 (0) w1 (1)  A13   0 w2 (1) w2 (0) w2 (1)  A23 
       
 L4   0 0 w1 (1) w1 (0)   A14   0 0 w2 (1) w2 (0)   A24 

(A-7)

This can then be re-arranged as:

 L1   A12   A11   0   A22   A21   0 

L  A  A  A  A  A   
 2   w  w (1)  13   w (0)  12   w (1)  11   w (1)  23   w (0)  22   w (1)  A21 
 L3  0 1
 A14  1
 A13  1
 A12  2
 A24  2
 A23  2
 A22 
             
 L4   0   A14   A13   0   A24   A23 

(A-8)

Equation (A-8) shows that the effect of adding the three point operator is exactly the
same as increasing the number of attributes by a factor of three, the additional attributes
being calculated by shifting the original attributes by –1 and +1 sample. We can now use
exactly the same least-squares formulation derived in the previous section.

The explicit result for this case is:

37
 1
 A A11 0  L 
 w1 (1)  A12 0   12   A12 0  1 
A11  
A13 A14 A13 A14
 w (0)    A A14   13
A A12 A A14   2 
L
 1   11 A12 A13
A A13 A12  
  11 A12 A13
L 
 w1 (1)  0 A11 A12 A13   14   0 A11 A12 A13   3 
  0 A14 A13    L4 


(A-9)

Or:

1
 4 2 4 3
  4 
  A1i  A1i A1i1  A1i A1i 2   A1i Li 1 
 w1 (1)  3 i 2 i 2 i 2
  i 24 
 w (0)    A A
4 4

  A  A1i A1i 1   A1i Li  (A-10)


2
 1 i 2
1i 1i 1
i 1
1i
i 2
 i 2

 w1 (1)  2 3 3   4 
 A1i A1i 1 A A1i 1  A 2
  A1i Li 1 
 i 1   i  2 
1i 1i
i 1 i 1

38

View publication stats