Multiattribute seismic analysis
BRIAN RUSSELL and DAN HAMPSON, Hampson-Russell Software Ltd., Calgary, Alberta
JIM SCHUELKE and JOHN QUIREIN, Mobil E&P Technical Center, Dallas, Texas
M ultiattribute seismic analysis is a
Downloaded 07/26/13 to 18.7.29.240. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
broad term that encompasses all geo-
statistical methods that utilize more
than one attribute to predict some
physical property of the earth. Bivari-
ate geostatistics is obviously the sim-
plest subset of the multivariate tech-
niques and thus the standard
technique of cokriging could be called
multivariate geostatistics. However,
in this paper we will use the name to
refer to geostatistical methods that
use more than two variables. Even
then, there are many different meth-
ods that fall under this heading. There
are three major subcategories:
1) The extension of cokriging to
include more than one secondary
attribute to predict the primary
parameter.
2) Methods that use the covariance
matrix to predict the parameter
from a linearly weighted sum of
the input attributes.
3) Methods that use artificial neural
networks (ANNs) or nonlinear
optimization techniques to com-
bine attributes into an estimate
of the desired parameter.
This paper will consider only
techniques described under the sec-
ond point. Point (3) is discussed in a
series of papers in TLE Schultz et al,
(1994). For a discussion of methods
in point (1), refer to Multivariate Geo -
statistics by Wackernagel (Springer-
Verlag, 1995).
Using seismic attributes to estimate
log properties. The idea of using
multiple seismic attributes to predict
log properties was first proposed by
Schultz et al. in their series of TLE
articles. They point out that the tra-
ditional approach to deriving reser-
voir parameters from seismic data
has been to look for a physical rela-
tionship between the parameter to
be mapped and some attribute of the
seismic data, and then use that single
attribute over a 2-D line or 3-D
volume to predict the reservoir para-
meter. This could be termed deter-
ministic reservoir parameter predic-
tion. Examples include:
EDGENET
1) Analysis of amplitude changes in
the seismic trace itself, which
indicate changes in re f l e c t i o n
coefficient size. For example, if
the acoustic impedance shows a
dramatic change, as in a gas sand,
the seismic trace will display the
classic “bright spot”.
2) Extraction of acoustic impedance
from the seismic data by trace
inversion. If we simply integrate
the trace, the result will lack a
low frequency component due to
the effect of the seismic wavelet.
This is usually corrected by
adding the low frequency com-
ponent from a velocity model.
(Schultz et al. object to calling the
result a true seismic attribute,
since it has already been influ-
enced by the well log data).
3) The use of prestack data to extract
information about intercept and
gradient, and hence Poisson’s
ratio or shear wave reflectivity,
using the formula:
R(θ) = R P+G sin2θ,
where: R P = intercept; G = gradi-
ent; R S ≈(RP - G)/2; and ∆σ ≈ (RP
+ G)*4/9.
In the above formulation, sever-
al approximations have been
made. We have assumed that
Vp/Vs ≈ 1/2 and have dropped
the third term in the A k i -
Richards formula.
4) The use of instantaneous attrib-
utes derived from seismic data.
These attributes are based on the
definition of the complex trace,
which, in polar form, give us the
three classical seismic attributes:
the amplitude envelope, instan-
taneous phase, and instantaneous
frequency. Mathematically:
C(t) = s(t) + jh(t),
where C(t) = complex trace; s(t) =
A(t)cosφ(t) = the seismic trace;
h(t) = A(t)sinφ(t) = the Hilbert
transform trace; A(t) =(s 2 ( t ) +
h2(t))1/2 = the amplitude envelope;
φ(t) = tan-1(h(t)/s(t)) = the instan-
taneous phase; and ω(t) =
HTTP://WWW.EDGE-ONLINE. ORG
October 1997 THE LEADING EDGE
Well Seismic
Data Data
Log Attribute
Properties
Statistical
Relationships
Calibration
Residual
Correction
Interpretation
Figure 1. The data-driven statistical
interpretation (after Schultz et al.).
dφ(t)/dt = the instantaneous fre-
quency.
These are the three primary
attributes, but many more can be
derived from the basic three.
5) Attributes which are derived spa-
tially from multiple traces, such
as coherency.
Although relationships have been
inferred between these attributes and
reservoir parameters, the physical
basis is not always clear, and we may
want to derive statistical, rather than
deterministic, relationships. This
approach, which Schultz et al. call a
data-driven methodology, is summa-
rized in Figure 1.
In the most general case, we look
for a function that will convert m
different attributes into the desired
property. This may be written:
P(x,y,z) = F[A1(x,y,z,),....,
Am(x,y,z)] (1)
where P = the property as a function
of coordinates x, y, z; F = the func-
tional relationship; and Ai, i= 1, ... , m
= the m attributes.
The simplest possible case would
be a linearly weighted sum:
OCTOBER 1997
 P = wo + w 1 A1 + ....wmAm (2) (6) (N 3 4) matrix containing the
attribute values, and w is a (4 3 1)
where w i, i = 0, ...., m = the m + 1 φ1  1 I1 E1 F1  w0 
matrix with the unknown weights.
weights.  φ2  1 I2 E2 F2   w1  This can be solved by least-
Obviously, this raises two ques- = squares to give:
tions: Which attributes should we  M M M M M w2 
   IN E N FN  w3 
use, and how do we judge their qual- w = [M TM]-1MTP (8)
ity? The number and types of attrib- φ N  1
utes must be determined geophysi- where superscript T indicates trans-
cally, whereas the quality of the or: P = Mw (7) pose, and superscript -1 indicates
Downloaded 07/26/13 to 18.7.29.240. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

attributes can be determined by matrix inverse.

using crossplots between the attrib- where P is an (N 3 1) matrix con- As a detailed computation, note
utes and the parameter to be mea- taining the known porosities, m is an that:
sured, and the associated correlation
coefficient.
∑ I ∑ E ∑ F   ∑φ 
−1
As an example, let us consider wo   N 1 1 1 1
(9)
porosity to be a weighted sum of
three attributes: acoustic impedance, w1   ∑ I1 ∑ I ∑ I E ∑I F   ∑ I φ 
2

=
1 1 1 1 1 1 1
amplitude envelope, and instanta- w2   ∑ E1 ∑ E I ∑ E ∑ E F  ∑ E φ 
2
neous frequency. In symbols: 1 1 1 1 1 1 1
  
φ(x,y) = w 0 + w1I(x,y) +
w3  ∑ F1 ∑ F I ∑ F E ∑ F  ∑ Fφ 
1 1 1 1 1
2
1 1

w2E(x,y) + w3F(x,y) (3)

where φ(x,y) = porosity; I(x,y) =

acoustic impedance; E(x,y) = ampli-
tude envelope; F(x,y) = instantaneous
frequency.

In the above example, the reser-

voir zone is given by the vertical
average of a spatially varying time
zone over a picked seismic surface.
The problem is now determining the
weights. One approach would be to
first calculate an estimate from the
well logs using kriging, but this
requires a fairly large number of well
logs for estimates to be reliable. A
second approach would be to esti-
mate the coefficient over longer
depth or time zones calculated at the
well logs themselves:

φ(t) = w 0 + w1I(t) + w2E(t)

+ w3F(t) (4) Figure 2. The initial five sonic logs.

where φ(t) = porosity at the well inte-

grated to time and calibrated to the
seismic data.
Least squares will find solutions
for both equations (3) and (4). That is

(5)

φ1 = w0 + w1I1 + w2 E1 + w3 F1
φ2 = w0 + w1I2 + w2 E2 + w3 FN
M M M M M
φN =w0 + w1IN + w2 EN + w3 FN

where we have N observed values,

either spatially or vertically. Notice
that equations in (5) can be written Figure 3. The seismic trace and three attributes at well 75.

EDGENET HTTP://WWW.EDGE-ONLINE. ORG OCTOBER 1997

October 1997 THE LEADING EDGE

 Equation (9) shows that the solu-
tion for four unknown weights will
involve a (434) matrix inversion,
regardless of the number of observed
points. The number of points simply
enters into the summation terms.
Note that this analysis assumes that
N>4. If N<4, we have an underde-
termined case. If N = 4, we can solve
the problem as a standard square
Downloaded 07/26/13 to 18.7.29.240. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Figure 4. Crossplot of seismic trace against sonic log.
Figure 5. Crossplot of integrated trace against sonic log.
Table. 1. Single attribute correlation results
Attribute
integrated
instantaneous frequency
instantaneous amplitude
seismic trace
EDGENET
October 1997
matrix inversion, without least-
squares (although errors in observa-
tion will have a major impact on this
solution). It is also important to note
the matrix in the right side of equa-
tions (8) and (9); written as MTM, is
one of the most powerful analysis
tools in data analysis, and is called
the covariance matrix. We see it in
Wiener-Levinson deconvolution,
Correlation
0.442492
0.186565
0.177764
0.0245846
HTTP://WWW.EDGE-ONLINE. ORG
THE LEADING EDGE
kriging, and a host of other applica-
tions.
The attributes used in the linear
sum can then be measured for good-
ness of fit with the parameter to be
predicted by use of the correlation
coefficient between each variable,
given by:
(10)
σ AP
PAP =
σ Aσ P
where σAP = covariance between A
and P; σA = standard deviation of A;
σP = standard deviation of P; A is the
attribute; and P is the parameter.
For the N attributes, the correla-
tion coefficients can be ordered from
best to worst. However, it should
be noted that when we determine
the best combination of attributes
to use in the final sum, the order of
the attributes used may not be the
same as the order of their correla-
tion coefficients. This may seem
strange at first, but consider a sim-
ple example. One of the attributes
could just be a linearly scaled ver-
sion of another attribute, and it
would obviously have just as good
a correlation coefficient. But adding
this attribute to the sum would not
improve the fit.
The quality of the fit can be deter-
mined by finding the RMS error
between the known parameter and
the estimated parameter. By using
this criterion, we can find the opti-
mal combination of attributes to use.
A case study. As an example of the
technique described in the previous
section, we will look at a case study
involving the prediction of sonic logs
at various locations along a seismic
line from Alberta. The attributes
used were as described in the previ-
ous section:
• Bandlimited acoustic impedance,
found by integrating the trace.
The reflection coefficients were
the trace samples themselves nor-
malized by dividing by the max-
imum absolute amplitude on the
trace and multiplying by 0.25.
• Amplitude envelope of the seis-
mic trace.
• Instantaneous frequency of the
seismic trace.
• The seismic time itself, used as a
ramp function to introduce the
DC bias and low frequency con-
tent.
Note that the porosity was actu-
ally derived from a set of sonic logs
OCTOBER 1997
 using Wyllie’s equation:

φ = (∆t - ∆tma)/(∆tf -∆tma), (11)

where ∆t = sonic transit time; ∆tma =

matrix transit time; and ∆tf = fluid
transit-time.
The sonic logs are shown in Figure
2. Only one of them, well 75, is actu-
ally a measured sonic log. The others
Downloaded 07/26/13 to 18.7.29.240. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

were derived from well 75 stretching

and squeezing the log so that their
derived synthetics matched the seis-
mic trace from the same location. The
bars on the plot show the zone of
interest, which covers a series of cre-
taceous events, but does not extend
into the higher velocity Devonian
events. Figure 3 shows the attributes
calculated for well 75, where we have
included a display of the seismic trace.
The reason that it was felt that the
integrated trace should be used rather
than the original trace is found when
we crossplot the attributes against the
Figure 6. Crossplot of amplitude trace against sonic log. P-wave sonic.
Figures 4 through 7 show the
crossplots of the seismic trace, inte-
grated trace, amplitude envelope,
and well 75 sonic. Notice that the
seismic trace displays the poorest
correlation when correlated against
the sonic log values. When all 5 wells
are taken into account, the situation
is magnified even more. Table 1
shows the averaged correlations of
all 5 wells, with the seismic trace by
far the worst.
Next, the multiattribute analysis
was done, using a 30 point, symmet-
rical operator. Table 2 displays the
numerical results, showing the resid-
ual error after the addition of each
attribute. Figure 8 shows the final
result, with the predicted trace
superimposed on the original trace.

Conclusions. We have shown a fair-

ly straightforward example of the
use of multivariate geostatistics. We
have concentrated on the prediction
of a well log parameter using a
weighted sum of seismic attributes.
Figure 7. Crossplot of instantaneous frequency against sonic log. The result, shown in Figure 8, is sur-
prisingly good when you consider
that no geological information was
used to predict the sonic logs; i.e.,
Table 2. Multiattribute analysis showing errors they were recreated simply by a lin-
ear convolutional sum of three seis-
Attribute Error mically derived attributes. The linear
operators were created by “training”
1 integrated 0.0118387177228541 the program using the techniques of
2 time 0.00915190953145179 geostatistics. This simple example
3 instantaneous amplitude 0.00833451304106493 was just to demonstrate the concept.
4 instantaneous frequency 0.00768988746816777 By using more attributes (with the
caution that there is an optimum

EDGENET HTTP://WWW.EDGE-ONLINE. ORG OCTOBER 1997

October 1997 THE LEADING EDGE

Downloaded 07/26/13 to 18.7.29.240. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 8. The final result. Modeled well logs in red.

number that should not be exceeded)
and other techniques (such as ANNs,
the prediction can get even closer.
Obviously, there are many other
applications of the technique that we
have just shown. We are preparing a
sequel to this paper in which we will
show a 3-D case study. It is our feel-
ing that the introduction of geosta-
tistical techniques to our industry
gives us a powerful tool which will
challenge the deterministic model of
seismic processing and lithology pre-
diction that has been in use for so
many years.
References. P. S. Schultz, S. Ronen,
M. Hattori, and C. Corbett, 1994,
“Seismic guided estimation of log
properties, Parts 1, 2, and 3”: (TLE,
1994). Hans Wackernagel, 1995, Mul -
tivariate Geostatistics: Springer-Ver-
lag, (1995). LE
Corresponding author: Brian Russell, 403-
266-3225; brian@hampson-russell.com

EDGENET HTTP://WWW.EDGE-ONLINE. ORG OCTOBER 1997

October 1997 THE LEADING EDGE