GEOPHYSICS, VOL. 72, NO. 3 共MAY-JUNE 2007兲; P. P33–P38, 7 FIGS., 2 TABLES.

10.1190/1.2716717
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Annual Meeting Selection

Redundant and useless seismic attributes

Arthur E. Barnes1

geophysical meaning; avoid attributes with purely mathematical

ABSTRACT meaning.
Seismic attributes represent subsets of the information in seismic
Many seismic attributes are redundant or useless and con- data. Quantities that are not subsets of the data are not attributes.
fuse seismic interpretation more than they help. They are eas- Attributes that differ only in resolution are the same attribute; treat
ily recognized given a few guidelines and tools. Discarding them that way.
them leaves an attribute list that is both more manageable and Seismic attributes should not vary greatly in response to small
more honest. data changes. Avoid overly sensitive attributes.
Not all seismic attributes are created equal. Details of implemen-
tation can be important. Avoid poorly designed attributes.
The utility of a seismic attribute is readily judged through visual
INTRODUCTION inspection aided by crossplots, histograms, correlation, rank corre-
lation, principal components analysis, and spectral analysis.
Hundreds of seismic attributes have been invented and more ap- Standard correlation identifies linear relationships between at-
pear each year 共e.g., Brown, 1999兲. Their great number and variety tributes. Rank correlation is more robust as it can also identify non-
are daunting and make it difficult to choose which ones to use. linear relationships. Rank correlation is computed like standard cor-
Many seismic attributes duplicate each other, or are obscure, un- relation, except that the attribute values are first sorted in order of
stable, or unreliable, or are purely mathematical quantities, or are not value and then their ranks in the sort are correlated 共Isaaks and
really attributes at all. These unnecessary seismic attributes can be Srivastava, 1989, p. 31兲.
identified through inspection aided by crossplots, histograms, and Like standard correlation, principal components analysis identi-
correlation. Discarding redundant and useless attributes leaves a fies linear relationships between attributes 共Duda et al., 2001兲. Long
much-reduced set of attributes that is easier to use. applied for coherency filtering, principal components analysis is
also used for reducing the number of attributes needed for analysis.
In essence, it transforms a set of linearly related attributes into a new
METHOD set of unrelated attributes. If some of the transformed attributes lack
To distinguish useful seismic attributes from those of doubtful useful information, then there is duplication of information within
utility, review your attributes in the light of the following common- the original attributes.
sense principles.
Seismic attributes should be unique. You only need one attribute APPLICATION
to measure a seismic property. Discard duplicate attributes. Where
multiple attributes measure the same property, choose the one that Duplicate attributes are legion. Many basic seismic properties,
works best. If you can’t tell which one works best, then it doesn’t particularly amplitude, frequency, and discontinuity, are quantified
matter which one you choose. through a variety of similar seismic attributes.
Seismic attributes should have clear and useful meanings. If you Consider the most important seismic property, amplitude. There
don’t know what an attribute means, don’t use it. If you know what it are more than a dozen common amplitude attributes. Figure 1 com-
means but it isn’t useful, discard it. Prefer attributes with geologic or pares four of them: average reflection strength, rms amplitude, aver-

Manuscript received by the Editor August 7, 2006; revised manuscript received October 10, 2006; published online April 19, 2007.
1
Paradigm Geophysical, Kuala Lumpur, Malaysia. E-mail: [email protected].
© 2007 Society of Exploration Geophysicists. All rights reserved.

P33
 P34 Barnes

age absolute amplitude, and maximum peak amplitude 共Appendix Discontinuity attributes based on principal components derived
A兲. They all look about the same. Correlations, crossplots, and prin- from principal components analysis of seismic data in an analysis
cipal components analysis of these maps indicate that they contain window masquerade as independent measures, yet tend to be well
nearly the same information 共Table 1; Figure 2兲. Similar analysis of correlated. Attributes defined as ratios of principal components,
average peak amplitude, average trough amplitude, and other ampli- such as the Karhunen-Loeve signal complexity 共a confusing name兲,
tude attributes confirms that most amplitude attributes are strongly often show the same picture. The first principal component normal-
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

correlated. Rarely is anything gained by using more than one of these ized by the total energy is identical to covariance discontinuity. If
as a general amplitude measure. Average reflection strength nearly you don’t already have a discontinuity attribute, then use the first
always suffices.
You may prefer average energy because it exhibits more contrast
than reflection strength. Use it, but recognize that average energy has 150

Average absolute amplitude

150
exactly the same information as rms amplitude and almost the same
as reflection strength 共see Appendix A兲. Its greater contrast is due to

rms amplitude
100
100
how it presents the information. Perhaps you want to use the maxi-
mum peak amplitude because you really are interested only in the 50
50
strongest positive events in a window. Use it for this purpose but not
as a general amplitude measure. 00 100 200 0 0 100 200
Consider a more involved seismic property, discontinuity. Figure 50 100 150 200 250 50 100 150 200 250
Reflection strength Reflection strength
3 compares four common discontinuity attributes based on correla-
tion, semblance, covariance, and weighted correlation 共Appendix
B兲. Their correlation coefficients all exceed 0.92 and their rank cor-
relation coefficients all exceed 0.95 共Table 2兲. Despite significant 400 400

computational differences and enthusiastic claims, these four dis-

Maximum peak amplitude

Maximum peak amplitude

continuity attributes are so nearly the same that it doesn’t matter
which one you use. 200 200

5 mi 5 mi

0 0 50 100 150 0
50 100 150 50 100 150
Reflection strength rms amplitude

Figure 2. Crossplots between the amplitude attributes of Figure 1.

Reflection strength rms amplitude 5 mi 5 mi

High
5 mi 5 mi
Amplitude

Correlation Semblance
Average absolute amplitude Maximum peak amplitude Low

Figure 1. Maps of four amplitude attributes computed in the same

100 ms window 共25 samples兲. 5 mi 5 mi

Table 1. Standard correlation coefficients between four

amplitude attributes. ARS is average reflection strength, rms
is rms amplitude, AAA is average absolute amplitude, and
MPA is maximum peak amplitude.

ARS rms AAA MPA

Covariance Weighted correlation
ARS 1 0.977 0.991 0.792
rms 0.977 1 0.978 0.857 Figure 3. Four discontinuity attributes based on correlation, sem-
blance, covariance, and correlation weighted by trace magnitude,
AAA 0.991 0.978 1 0.782 computed as maps in a 60 ms window 共15 samples兲. These attributes
MPA 0.792 0.857 0.782 1 are nearly identical. Discontinuous data is shown as blue, continu-
ous data as white.
 Redundant and useless seismic attributes P35

principal component attribute but discard the others; otherwise, dis- domains. Not all spectral attributes are so flexible. Spectral skew and
card them all. kurtosis must be computed in the frequency domain, but they have
Some attributes are not only similar, they are essentially identical. little inherent geologic or geophysical meaning, so you probably
They contain exactly the same information and differ merely in how don’t need them.
they present it. As already noted, rms amplitude and average energy
are identical. Other identical attributes include instantaneous phase
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

and cosine of the phase, and dip-azimuth 共reflection dip combined Table 2. Rank correlation between the four discontinuity
with reflection azimuth兲 and shaded relief. Choose one and discard maps of Figure 3.
the other.
Cosine of the phase removes all amplitude information and re- Weighted
sembles a strong amplitude gain 共Figure 4兲. Treat it more as a process Correlation Semblance Covariance correlation
than as an attribute.
Watch out for attributes that have multiple names. Reflection Correlation 1 0.988 0.975 0.981
strength, trace envelope, and instantaneous amplitude are the same Semblance 0.988 1 0.959 0.967
attribute. Slope and dip are used interchangeably, as are continuity, Covariance 0.975 0.959 1 0.998
coherence, discontinuity, and similarity. Eigen-structure discontinu- Weighted 0.981 0.967 0.998 1
ity is the same as covariance discontinuity. Arc length is sometimes Correlation
called reflection heterogeneity. Response attributes are called wave-
let attributes. The quadrature trace and the imaginary trace are the
same — but the quadrature trace is a 90° phase rotation and is not re-
ally an attribute since it does not subset the information. a)
Attributes such as arc length and Karhunen-Loeve signal com- 0.0
plexity lack clear and useful meaning and are useless. Average in-
stantaneous phase is also useless because the more instantaneous
phase is averaged, the more it tends towards zero. Average un-
Time (s)
wrapped instantaneous phase is scarcely better. The slope of the in-
stantaneous frequency may have clear mathematical meaning but it
lacks useful geologic or geophysical meaning. Avoid it. The same is
true of dominant frequencies derived from maximum entropy spec-
tral decomposition. Response phase and response frequency are use-
less if you insist that they describe the seismic source wavelet as ad- 1.5
vertised because they fail on real data 共White, 1991兲. Use them em- 2 km
pirically, recognizing what they really record. Response phase
b)
records the apparent phase of reflections at envelope peaks. Re-
0.0
sponse frequency applies a nonlinear filter to smooth the instanta-
neous frequency. Response frequency has utility, but weighted aver-
age frequency is smoother and has simpler meaning.
Time (s)

Resolution is an important parameter of most attributes. It is deter-

mined by the size of the window in time and space in which the at-
tribute is computed. Changing an attribute’s resolution does not cre-
ate a new kind of attribute. This is obvious for attributes like rms am-
plitude and energy half-time. Less obviously, it is also true of instan-
taneous frequency and weighted average frequency. You might rea-
sonably use both to investigate targets of different size, but they 1.5
remain the same attribute. 2 km
Average frequency, rms frequency, bandwidth, and a few other Figure 4. A seismic line processed with 共a兲 cosine of the phase, and
spectral attributes can be computed equally well in the time domain 共b兲 a strong amplitude gain 共28 ms or 7 sample operator length兲.
or in the frequency domain. It is pointless to compute them in both They are almost indistinguishable.

Synthetic data Apparent polarity Response phase Figure 5. Illustration of instability in apparent po-
0 0 0 larity. The synthetic data has three reflections and a
small amount of random noise. The top reflection
has positive polarity, the bottom reflection has neg-
ative polarity, and the middle reflection is a com-
Time (ms)

Time (ms)

Time (ms)

posite of two reflections 4 ms apart. The composite

reflection looks like a single reflection with 90° of
phase for which the apparent polarity flips random-
ly. The same problem occurs on real data. Every
twentieth trace is overlain in wiggle format. Red is
positive polarity, blue is negative.
400 400 400
 P36 Barnes

Attributes that are sensitive to small perturbations in the data are phase instead. Attributes that count the integral events in an interval
unstable. Apparent polarity is an example. It is defined as the sign of are also unstable as they are sensitive to small changes in interval
the seismic data at envelope peaks scaled by the envelope peak and definition. Avoid them.
held constant in each interval around a peak. It works fine for clean Unsupervised waveform classification is more sensitive to details
zero-phase data that is free of reflection interference, but it is ambig- of the analysis window than most other attribute methods. Differ-
uous for thin-bed reflections, which have an apparent phase of ences due to small changes in the window are often significant. In
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

around 90° 共Figure 5兲. Discard apparent polarity and use response Figure 6, a one-sample shift in the analysis window causes a large
change in the attribute maps produced by the Kohonen self-organiz-
a) b) ing feature map 共KSOFM兲. Somewhat surprisingly, for this same ex-
5 mi 5 mi
ample a competing algorithm, K-means clustering, finds the same
patterns as the KSOFM, but their maps appear markedly different
because they follow different rules in assigning class numbers to
waveforms. Assigning the same colors to the same waveforms
shows that the competing maps found nearly the same classes. For
basic pattern recognition it matters little which method is used.
However, the KSOFM naturally and logically orders its waveform
classes, unlike K-means clustering, and is therefore preferable as its
maps are easier to interpret.
c) d)
Beware of differences between programs for generating seismic
5 mi 5 mi
attributes. Aside from incorrect algorithms, which plague instanta-
neous frequency, the same attribute produced by competing pro-
grams can differ substantially due to implementation details. One
such detail regards windowing, which is the way an algorithm se-
lects seismic data from an interval. Attributes are filters, and like any
filter they should employ tapered windows to reduce Gibbs effects.
Nontapered or “boxcar” windows are nonetheless widely used. Box-
car windows give rise to banding in the time domain and ringing in
the frequency domain, which are the Gibbs effects. In contrast, ta-
Figure 6. Waveform maps produced in a 60 ms window 共15 sam-
ples兲. A specific color indicates similar waveforms on a map. Be- pered windows produce sharper images and smoother power spec-
tween maps, the same color usually corresponds to different wave- tra. Figure 7 demonstrates this with energy half-time. Attributes with
forms. 共a兲 KSOFM centered at time A, 共b兲 KSOFM centered at time ringy spectra are poorly designed. Avoid them.
B = A + 4 ms 共1 sample兲, 共c兲 K-means at time B, and 共d兲 K-means Incidentally, energy half-time measures amplitude change. Opti-
at time B with color scale chosen to match the map in 共b兲. The differ-
ences caused by the small change in window exceed those caused by mistic claims that it indicates lithology are wrong. Prefer standard
the competing algorithm. measures of amplitude change to energy half-time 共see Appendix
A兲.
a) 5 km 0
Frequency (Hz)
100
0.0
–5
–10
Ringing spectrum
CONCLUSIONS
dB power
Time (ms)

–15
1.0 There are too many duplicate attributes, too many attributes with
–20
–25 obscure meaning, and too many unstable and unreliable attributes.
–30 This surfeit breeds confusion and makes it hard to apply seismic at-
2.0 tributes effectively. You do not need them all. Review your seismic
attributes and reduce them to a much smaller subset. Discard dupli-
b) Frequency (Hz) cate and dubious attributes, prefer attributes with intuitive geologic
5 km 0 100
or geophysical meaning, understand resolution, distinguish process-
0.0
–5 es from attributes, and avoid poorly designed attributes. The subset
–10 remaining is both more manageable and more honest.
Smooth spectrum
–15
dB power
Time (ms)

–20
1.0 –25
–30
–35 ACKNOWLEDGMENTS
–40
–45
2.0 I thank Seitel Data Ltd. for permission to publish the seismic data
shown in Figures 1, 3, and 6; I thank Landmark for permission to
Figure 7. Energy half-time computed with 共a兲 a boxcar window, and publish the paper; and I thank Paradigm for supporting the revision. I
共b兲 a Hamming window. Both windows are 60 ms long 共15 sam-
ples兲. The Hamming window prevents spectral ringing and produces also thank an unknown reviewer and Dengliang Gao for exception-
a sharper image. ally helpful and detailed reviews.
 Redundant and useless seismic attributes P37

APPENDIX A APPENDIX B

AMPLITUDE ATTRIBUTES CONTINUITY ATTRIBUTES

The amplitude attributes of Figure 1 are computed in an interval of Bahorich and Farmer 共1995兲 introduced a measure of seismic con-
N samples taken about a horizon or constant time. Average reflection tinuity based on crosscorrelations between three traces in an L pat-
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

strength ā is defined as tern. Their measure is readily extended to compute continuity in an

N
analysis window encompassing an arbitrary number of traces. The
1
兺 an ,
resulting correlation continuity attribute is similar in form to sem-
ā = 共A-1兲 blance continuity and comparable to covariance continuity 共Marfurt
N n=1
et al., 1998; Gersztenkorn and Marfurt, 1999兲. It is further general-
where an is the instantaneous reflection strength at time index n. The ized by weighting the constituent correlations to produce a weighted
rms amplitude xrms is defined as the square root of the average of the correlation continuity attribute. I provide the mathematics for these
squared trace values xn: two continuity measures without corrections for reflection dip. I im-

冑
plicitly use a boxcar window operator, though windows could be ta-
N pered to avoid Gibbs effects. For both measures, continuity C ranges
1
xrms = 兺 x2 .
N n=1 n
共A-2兲 from 0 for perfectly discontinuous data to 1 for perfectly continuous
data. A corresponding measure of discontinuity D is formed as D
= 1 − C.
Average absolute amplitude xa is given by
I simplify the mathematics by expressing seismic traces as vec-
N tors. The zero-lag crosscorrelation of two traces xi and x j, each with
1
xa = 兺 兩xn兩. 共A-3兲 N samples, is given by
N n=1
N
Maximum peak amplitude is the magnitude of the largest trace value
in the interval. Average energy Ē is the total energy of the trace divid-
xi · x j = 兺
k=1
xikx jk , 共B-1兲
ed by the number of samples, and is defined by
N where k is the trace sample index. The energy Ei of a trace xi is the ze-
1
Ē = 兺 x2 .
N n=1 n
共A-4兲 ro-lag autocorrelation of the trace:

兺
Average energy is the square of the rms amplitude.
Energy half-time measures where in a time interval the seismic Ei = xi · xi = x2ik . 共B-2兲
k=1
energy is concentrated. As used here, it is defined as the percentage
of the interval length at which the center of gravity of the data occurs.
This center of gravity is the average time tc defined by A circumflex denotes a unit vector. The unit vector x̂i is the trace vec-
tor xi normalized by the square root of its energy:
N

兺 tnx2n xi
n=1 x̂i = 共B-3兲
tc = N . 共A-5兲 冑E i .
兺
n=1
x2n
The average trace of a set of M traces is the vector average xa given
by
Energy half-time Eht is this average time, referenced from the start
time of the interval, t1, expressed as a percentage of the total interval
length: M
1
xa = 兺 xi . 共B-4兲
tc − t1 M i=1
Eht = 100% · , 共A-6兲
tN − t1 A correlation continuity measure C can be defined as the average
of the squared crosscorrelations between every trace with the aver-
where tN is the time at the end of the interval.
age trace:
Relative amplitude change, ␴共t兲, is the time rate of change of the
reflection strength a共t兲 normalized by the reflection strength:
M
1
␴共t兲 =
1 da共t兲
. 共A-7兲
C= 兺 共x̂i · x̂a兲2 .
M i=1
共B-5兲
a共t兲 dt
Energy half-time closely resembles an averaged relative amplitude Weighted correlation continuity is the average of the squared cross-
change. correlations between every trace with the average trace, with each
 P38 Barnes

crosscorrelation weighted by trace energy Ei. Thus, a weighted cor- REFERENCES

relation continuity attribute C can be expressed as
Bahorich, M., and S. Farmer, 1995, 3-D seismic discontinuity for faults and
stratigraphic features: The coherence cube: 65th Annual International
M Meeting, SEG, Expanded Abstracts, 93–96.
兺 共xi · x̂a兲2 Brown, A. R., 1999, Interpretation of three-dimensional seismic data, 5th
ed.: AAPG Memoir, 42, American Association of Petroleum Geologists.
Downloaded 02/19/15 to 35.8.191.249. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i=1 Duda, R. O., P. E. Hart, and D. G. Stork, 2001, Pattern classification, 2nd ed.:
C= M . 共B-6兲 Wiley Interscience.

兺
Gersztenkorn, A., and K. J. Marfurt, 1999, Eigenstructure-based coherence
Ei computations as an aid to 3-D structural and stratigraphic mapping: Geo-
i=1 physics, 64, 1468–1479.
Isaaks, E. H., and R. M. Srivastava, 1989, An introduction to applied geosta-
tistics: Oxford University Press.
This continuity measure closely approximates covariance continu- Marfurt, K. J., R. L. Kirlin, S. L. Farmer, and M. S. Bahorich, 1998, 3-D seis-
mic attributes using a semblance-based coherency algorithm: Geophysics,
ity, but is computationally faster and does not assume zero-mean 63, 1150–1176.
traces. White, R. E., 1991, Properties of instantaneous seismic attributes: The Lead-
ing Edge, 10, 26–32.