discrimination of signal and noise events on seismic
TRANSCRIPT
Scholars' Mine Scholars' Mine
Doctoral Dissertations Student Theses and Dissertations
1970
Discrimination of signal and noise events on seismic recordings Discrimination of signal and noise events on seismic recordings
by linear threshold estimation theory by linear threshold estimation theory
David Nuse Peacock
Follow this and additional works at: https://scholarsmine.mst.edu/doctoral_dissertations
Part of the Engineering Commons, and the Geophysics and Seismology Commons
Department: Geosciences and Geological and Petroleum Engineering Department: Geosciences and Geological and Petroleum Engineering
Recommended Citation Recommended Citation Peacock, David Nuse, "Discrimination of signal and noise events on seismic recordings by linear threshold estimation theory" (1970). Doctoral Dissertations. 2157. https://scholarsmine.mst.edu/doctoral_dissertations/2157
This thesis is brought to you by Scholars' Mine, a service of the Missouri S&T Library and Learning Resources. This work is protected by U. S. Copyright Law. Unauthorized use including reproduction for redistribution requires the permission of the copyright holder. For more information, please contact [email protected].
DISCRIMINATION OF SIGNAL AND NOISE EVENTS
ON SEISMIC RECORDINGS BY LINEAR
THRESHOLD ESTIMATION THEORY
by
DAVID NUSE PEACOCK, 1943-
A DISSERTATION
Presented t o the Faculty of the Graduate School of the
UNIVERSITY OF MISSOURI - ROLLA
In Partial Fulfillment of the Requirements for the Degree
DOCTOR OF PHILOSOPHY
in
GEOPHYSICAL ENGINEERING
1970
T2366 c.1 101 pages
193936
ABSTRACT
The object of this study is the investigation of a
linear threshold element technique for identifying surface
multiples on a single seismic trace. Traces of seismic
events were generated which contained primaries, surface
multiples, and various levels of Gaussian random noise.
Since it was necessary to separate the events as much as
possible, the traces were subjected to pulse-compression
deconvolution processing prior to LTE analysis. Mean
frequency, peak frequency, amplitude spectrum variance,
periodicity, and polarity were employed as pattern param
eters. A set of weights was found that would maximize
the moment of inertia of the S line distribution of the
patterns subject to the constraint that the sum of the
squared values of the weights was minimized. It is shown
that the problem of the maximization of the moment of
inertia reduces to the solution of a simple eigenvalue
problem. Furthermore, the optimum set of weights is the
eigenvector corresponding to the largest eigenvalue of a
matrix proportional to the autocovariance matrix of the
pattern vectors. The classes of patterns representing
primaries and multiples on traces with high signal-to-noise
ratios were clustered and separ~ted, making identification
by inspection a simple procedure. Clustering and separation
of classes on traces with low signal-to-noise ratios was
less than optimum.
ACKNOWLEDGEMENTS
The author wishes to express his appreciation first
of all to his wife and parents for their patience and
understanding during the course of this investigation.
He is indebted to Dr. Hughes M. Zenor, his dissertation
supervisor for his encouragement and support. Special
thanks are due to Dr. Gerald B. Rupert and to Dr. John
C. Robinson for their critical review of the manuscript.
The author's gratitude is also extended to Dr. Robinson
for providing subroutine TPLITZ used in this research
i i i
and to Dr. Frank J. Kern for his constructive suggestions
and helpful insight into some of the facets of this research.
ABSTRACT .
ACKNOWLEDGEMENTS
LIST OF FIGURES
LIST OF TABLES .
I. INTRODUCTION
TABLE OF CONTENTS
II. SEISMIC TRACE SYNTHESIZATION AND PROCESSING
TECHNIQUES .
A. Reflectivity Spikes
B. Wavelet Generation
C. Primary Generation
D. Multiple Generation .
E. Noise Generation
F. Processing
1. Autocorrelation Function
2. Power Spectral Estimate
3. Wavelet Reconstruction .
4. Inverse Filter Construction and
Deconvolution
III. PATTERN RECOGNITION ANALYSIS OF SYNTHETIC
SEISMOGRAMS .
A. Parameters Investigated
B. Determination of Weighting Vector
C. Analysis and Discussion
IV. SUMMARY AND CONCLUSIONS .
V. APPENDICES
lV
Page
ll
iii
vi
viii
1
5
5
8
9
10
12
17
17
18
25
25
36
37
42
47
58
61
A. Mathematics Relative to Development of
Deconvolution Operator
B. Subroutine to Compute Deconvolution
Operator
C. Linear Threshold Element Data
D. Linear Threshold Element Details
VI. BIBLIOGRAPHY
VII. VITA
\'
68
69
82
87
LIST OF FIGURES
Figure
1. Hypothetical reflectivity series
2. Symmetric Ricker wavelet
3. Reverberation model
4. Trace of primaries and multiples
5. Trace of primaries, multiples and random
noise. P = 10
6. Trace of primaries, multiples and random
noise. P = 5
7. Trace of primaries, multiples and random
noise. P = 2
8. Autocorrelation function of synthetic seismic
trace. P =oo.
9. Power spectral estimate of synthetic seismic
trace. P =oo.
10. Statistically derived wavelet,-y(t).P=oo.
11. Deconvolution operator for synthetic seismic
trace. P = oo .
12. Deconvolution operator for synthetic seismic
trace. P = 10
13. Deconvolution operator for synthetic se1sm1c
trace. P = 5
14. Deconvolution operator for synthetic seismic
trace. P 2
15. Deconvolved synthetic seismic trace. P =oo
16. Deconvolved synthetic seismic trace. P = 10
Vl
Page
5
9
10
11
14
15
16
19
24
24
27
28
29
30
32
33
17. Deconvolved synthetic seismic trace. P
18. Deconvolved synthetic seismic trace. P
5
2
19. Partial amplitude spectrum of event no. 9 from
\' l
:)4
35
figure 15. t 0 = 1118 ms . :;9
20. S line distribution produced with unit weights.
p = 00 •
21. S line distribution produced with unit weights.
p = 10
22. S line distribution produced with unit weights.
p = 5 .
23. S line distribution produced with unit weights.
p = 2 •
24. S line distribution produced with eigenvector
corresponding to A x P = 00 rna •
25. S line distribution produced with eigenvector
corresponding to A x P = 10 rna .
26. S line distribution produced with eigenvector
corresponding to A P = 5 max.
27. S line distribution produced with eigenvector
corresponding to Amax. P = 2
48
49
so
51
53
54
55
56
LIST OF TABLES
Table
I. Times, polarities and amplitudes of primary
spikes .
II. Pattern vector coding for periodicity
component
III. Pattern vector coding for polarity component
IV. Eigenvalues and eigenvectors of the matrix B
V. Times of the central extremum, t 0 , for events
1n the reference set .
VI. Times of the central extremum, t 0 , and pattern
vectors for events of interest .
VI I l
Page
6
41
41
44
70
74
I. INTRODUCTION
Exploration geophysicists and other interested scientists
are constantly searching for new and better methods of ob
taining and analyzing reflection seismic data. Furthermore,
the presence of a multitude of anomalous waveforms on seismic
records, such as surface waves, dispersed energy, multiples,
ghosts, diffractions etc. necessitates continual research
with the purpose of more fully identifying, understanding,
and controlling these forms of important seismic "noise".
The reader is referred to Olhovich (1964) for a comprehensive
outline of the major noise problems in reflection seismology.
The purpose of this research is to investigate the appli
cations of pattern recognition analysis techniques to the
problem of identifying multiply reflected events (multiples)
in reflection seismic prospecting, a problem which is con
sidered to be one of the most common and dominant of those
encountered by exploration geophysicists.
The recognition that multiple events were geophysically
possible, and, in fact, were quite common in occurrence, Has
first made by Leet (1937), who noticed their presence and
observed their basic characteristics on records from a well
which had been shot for velocity. The identification of
multiples as such was first made by Johnson (1943), who pro
posed a specific geologic model which was necessary and suf
ficient to explain their existence on seismograms from certain
regions of California.
problem were Ellsworth
Early investigators of the multiple
(19 4 8) ~ ,Johnson ( 19 4 8) , S 1 oat ( 19 4 8) ,
and Dix (1948), among others.
The concepts of statistical communication theory and
generalized harmonic analysis were first developed shortly
before, and during, World War II, and, subsequent to declass
ification, were published by Wiener (1949). Other precursors
in this area included Bartlett (1948, 1950) and !\icc (1944).
During the early 1950's, the application of statistical com
munication theory to the problems of reflection seismology
was led by Wadsworth et al (1953), and Jones and Morrison
(1954). The efforts of these and others, however, were
hindered by the unavailability of high speed data processing
equipment with large storage capacity. With the advent of such
equipment and supporting digital recording systems in the
early 1960's, a veritable technological revolution overcame
the industry. On an economical and practical basis, investi
gators now had the wherewithal to attack the nagging seismic
problems, multiples included, with the techniques of statisti
cal communication theory and allied concepts.
During the past few years a wealth of literature has
appeared concerning investigations which make usc of high
speed data processing techniques. Among the principal authors
are Robinson (1966, l967a; and Treitel 1964), Treitel (and
Robinson 1966; et al 1967), Schneider (et al 1964, et al
1965), Embree (et al 1963), Anstey (1960, 1964), and Anstey
and Newman (1966).
The field of pattern recognition developed from a late
1950's interest in "artificial intelligence" and "thinking
machines". In this respect~ Rosenblatt (1962) is recognized
as the father of the science of pattern recognition due to
his exhaustive study of a brain model~ the "perceptron".
Other early investigators included Sebestyen (1962), and
Marill and Green (1960). Almost from the outset, the science
was developed along two dissimilar lines. A geometrical or
"deterministic" approach was held to by Rosenblatt and others,
while Sebestyen and others conducted investigations based on
studies of the statistical properties of the images or pat
terns to be recognized. Recently a trend has developed
toward unification of these two approaches. Techniques have
been developed whereby the n features of all patterns may be
weighted with an optimum (in some sense) set of statistically-
derived feature weights. This brings about clustering of
similar patterns within the n-dimensional pattern space so
that simple thresholds may be used to isolate each cluster.
See~ e.g., Fischler et al (1962) and Minsky and Papert (1969).
A technique of this nature has been applied with some success
in the field of diagnostic medicine, e.g., by Stark et al
(1962)~ among others.
The application of pattern recognition analysis tech
niques to geophysical problems consists of one paper by
Mathieu and Rice (1969)~ who, with limited success, used the
4
statistical approach to differentiate between events on a
synthetic seismogram corresponding to shale and similar ones
corresponding to sandstone.
Chapter II of this study begins with a description of
the generation of a synthetic seismic trace consisting of
primaries and surface multiples. After addition of various
levels of random noise, this trace is subjected to a series
of mathematical processes leading to a pulse compression
deconvolved trace, which then provides the input for the
pattern recognition analysis. Chapter III consists of the
description of the linear threshold element technique ap
plied in this problem and the discriminatory results
obtained through its use. Chapter IV includes a discussion
of the results and recommendations for further research.
Appendix A contains the mathematical details of the devel
opment of the deQonvolution operator while Appendix B is a
listing of the Fortran subroutine used to obtain the operator.
Appendix C lists the input pattern vectors for the various
noise levels studied while Appendix D consists of the theoret
ical development of the multiple identification scheme.
II. SEISMIC TRACE SY0JTII1:SIZJ\TJON
AND PROCESSING TECII:'< I QULS
A. Reflectivity Spikes
The presence of energy reflecting horizons within the
crust of the earth can be represented mathematically for
seismic purposes by a time series of Dirac delta functions,
symbolized by 8 (t), which are commonly called "spikes".
The reflection coefficients, A-, of the set of horizons l
can be used to provide the amplitude and polarity charac-
teristics of the delta functions in order that the series
may be written as
5
f(t) == 2: 1\.o(t-T-) l l
(l) i
where for practicality only a finite quantity of time shifts
T- need be considered. l
A pictorial representation of a por-
tion of a hypothetical reflectivity series is as illustrated
1n figure 1. A6
Al A" J\4
l ;\C) J\11
C' L
01 1 I I I 1~10 I J 1 I I 'l'
A3 A12
As J\7 As
Figure 1. Hypothetical reflectivity series.
S:ince velocity generally increases with depth, the negative
spikes represent density-velocity inversions of the form
P- lv- 1< P- v-l+ l+ 1 l for normally incident longitudinal waves.
Table I. Times, polarities and amplitudes of primary spikes.
Time (ms.) Pol. Amp. Time (ms.) Pol. Amp. Time (ms.) Pol. Amp.
401 - .03841 2905 - .68638 2224 - .47185
1344 + . 1.5025 3918 - .69232 859 + . 77559
946 + .37073 1985 + .93882 1311 - .22791
1088 + .29316 2581 + .84390 2848 - .32324
247 - .40823 1032 - .46248 668 - .27220
3571 + .55618 3209 + .06620 762 + .03901
600 + . 74273 593 - .12976 1588 - .81425
726 - .20237 700 + .51423 2830 + . 75661
1933 + .34547 453 - .49192 1315 - .19867
1135 + . 40 46 0 3710 - .16022 3003 + . 79427
3926 + .41330 1052 + .57671 119 + .05785
3148 - .36088 3443 - .07415 1455 + .69384
809 + .95472 1118 + .50896 3225 + .54827
1715 - .63386 3790 + . 58515 3238 + .44751
1165 + .88296 3835 - .44876 1569 + . 79462
1878 + . 27076 3247 - .12200 615 + .70387
1568 + .21093 1270 + .51522 1247 + . 77538
2351 + .24331 1265 + .94982 2589 - .07734
86 2 + .68093 1662 + .10023 2207 + .84673
2 756 - . 71122 1537 + .49920 1790 + .63957 0'
Table I continued.
Time (ms.) Pol. Amp. Time (ms.) Pol. Amp. Time (ms.) Pol. Amp.
1856 - .44326 390 + .20431 3103 + .30712
3021 + .81310 3275 + .09300 3300 - . 70184
2745 - . 78237 3841 - .87038 3494 + . 72692
1759 - .94451 2356 - .80140 1746 + .51115 1935 + .22898 2655 + . 82295 1628 + .49932
15 76 - .23409 3730 + .38142 2899 - .38840 1983 + .63072 2490 + .00243 213 - .48489
2066 + .94430 1201 - .67417 3590 - . 73720
1979 - .86556 2791 - .42892 3655 - .38234
1682 + .76428 1635 + . 71887 3270 + .61458
3296 - .08094 3009 - .38142 302 + .23245
2783 + .38984 334 + .00146 3924 + . 31151
2067 + .16511 3196 + .02303 2038 + . 21188
1839 + .22663 3556 + .68587 1307 + .63902
232 + .5027o 1064 + . 21127 3955 + .85962
174 + .16121 1248 .69112 887 + .65790
3911 + .14326 30 70 + .80048 1801 + .00862
564 + . 72876 1606 .17496 1 79 7 + .58295
2315 - .02289 568 + . 71029 3135 + .b4554
131 + .06584 2270 + .87932 22 71 + .21346
8
For this particular problem a set of 120 spikes with
random amplitudes was considered to represent a realizable
series of primary reflection coefficients under the re-
striction that the polarity distribution consisted of 80
positive and 40 negative. The total number of spikes and
their polarity distribution are arbitrary; however, fnr
velocity generally increasing with depth, there should be
more positive spikes than negative. The time of occurrence
in milliseconds of each spike was arbitrarily selected
from a tabulation of random numbers \vithin the total time
length (4.0 seconds) of the trace. The times of occurrence,
polarities, and amplitudes of the spikes are listed in
table I.
B. Wavelet Generation
A symmetric wavelet as developed by Ricker (1945) and
of the form
RW (t) ( 2)
was computed where t is time and b 1s the wavelet breadth,
measured between the two maxima. This wavelet has a minimum
of -1 at t = 0, and maxima of 2 exp(-3/2), (approx. 0.45),
at t = + b/2. Ricker's original wavelet was normalized hy
4/~ to yield equation (2). For computational purposes,
the realistic value of 10 ms. was chosen for b and the \vavelet
was truncated at 30 ms. on hoth sides of its central axis of
symmetry. An illustration of a \vavc1ct of this form appears
in figure 2.
Figure 2. Symmetric Ricker >Vavelet.
C. Primary Generation
For continuous functions the general form of the convo-
lution process is X
1./J ( t ) =! h ( T ) g ( t - T ) d T -00
or equivalently, 'XJ
1./J(t) =/ g(T)h(t-T)dT (3)
-X
where t is a time shift. Considering h(t) as a "filter",
the output \j_J(t) consists of a linear combination of the
values of the input g(t) at all instants of time.
Lmploying the notation of \\.ainstein and Zubakov, (1962),
10
the discrete form of the convolution process is represented
as
N l/J ( t) T~l r ( T ) R W ( t - T ) (4)
where variables t and T are integers. By convolving a dis-
crete version of the modified Ricker wavelet of equation
(2) (sampled at 1 ms. intervals) with the set of primary
reflection coefficients, a synthetic seismic trace of pri-
mary events was obtained.
D. Multiple Generation
A strong event occurring at 0.564 sec. was then arbi-
trarily selected as the generating horizon of a series of
surface multiples. A reverberation model of the form useu
by the geophysical industry and illustrated in figure 3
was employed, where R represents the amplitude of the event
whose peak occurs at 0.564 sec.
0 0.564 sec.
R2
3 R
Figure 3. Reverberation model.
The alternation of polarity signifies a phase reversal at
each reflection of the multiple from the surface. The con-
volution of this multiple reflectivity series with another
"VA -0 - ~ " A 1\. Af'\AI A AA ...... AI\ I I t v vv •v V I 9F I vV I \( vr .. r~ v
0.0 0 .4 0 .8
I It A A A I • • "' A 1\ 0 - • ... f1 0 A 1\ A I - AA A II. I\ I I J\ A ~ ... J\l\ A ... ~ V y w•vv lJ'f y vv vy 0- • vv vvv-o.8 1.2 1.6
'f Jo.A-A 1 ~.A •u •. A.- *' ~a, ,.bf> 1 AAI\, AAA 1 A 1 • yl/fi' 4jtV ff 9HV V '7 V~ 9F v r -v VV 1.6 2 . 0 2.4
I I " 1\ I " A I A A ... " 0 Y'/1. J I A A. f4 " " • • V" A 'V y y V\(v lf"' • .. -v 1('{ 1 o v vir 2.4 2 .8 3.2
~- ~~....._ A • I 1\.A I J\1"'\AA I A • .,. AA *" 411+.. .111\ I r lj \IV I 1( 1V14V 44P 1( rq • yr v 3 . 2 3 . 6 4.0
Time (sec . )
Figure 4 . Trace of primaries and multiples .
......
......
1 2
modified symmetric Ricker wavelet of the same form as that
in equation (2) yielded a synthetic trace of multiple events.
However, more energy in relatively lower frequency bands was
incorporated into the trace by allowing the breadtl1 of the
wavelet to be 15 ms. and subsequently increasing the total
length of the wavelet to 80 ms.
The composite trace of primaries and multiples as shown
1n figure 4 was constructed by superimposing the two original
traces.
E. Noise Generation
To better simulate a seismic trace, random noise traces
were generated and added to the composite.
If the samples of a stationary noise time ser1es are
uncorrelated, i.e., statistically independent, it is shown
by 11ancock and Wintz (1966) that the covariance matrix of
the noise may be written as
(S)
where o 2 ]s the variance of the no]se samples and l lS the
identity matrix. The average signal-to-noise power ratio 1s
defined as
p (6)
Accordingly, several traces of Caussian random no1se h'cre
constructed using the IBM subroutine CAUSS (1967h).
trace had a mean value of zero and the variance was preset
l:l
so as to yield values of P equal to 10, 5, and 2. For this
study, the signal s was the trace of both the primary and
multiple events synthesized above, so that,more precisely,
equation (6) becomes
primary power + coherent noise power
incoherent noise power
A value of P equal to 2 is considered to be a physically
realizable minimum (Kern, 1970), and for this reason, higher
levels of random noise power were not considered.
The individual noise traces were then subjected to a
five-point moving average routine described by
i+2
ni = j~- 2 O.Znj (7)
where n. is the value of the noise trace at time sample 1. l
This "smoothing" process effectively filtered the noise
traces so as to make the frequency spectrum of each one
more nearly coincident with the frequency spectrum of the
trace of primaries and multiples.
Subsequently, each noise trace was superimposed on
the trace s(t) to yield the output traces shown as g(t)
1n figures 5, 6, and 7.
The traces g(t) arc synthetic representations of
field traces of a reflection seismic record which have
t • • 1\,11. -- .. . I A • 1\1\._1\1'\A... .... AA&It A/\.'4 • •v .. 9P , vv lrVf v• r- lJ
0 . 0 0 .4 0 . 8
1 A 1\. A A 0 1 .... ~ 1\ A 1\ • * fl - 1\ 1\. f\ ~ 11./l A A 1\ ... 1 A II. ~ A ,A A A . ~ '{T v ... v, v V \f y w: ~y v- , . y y yv~
0 . 8 1.2 1.6
V' AA-A ""-6 •• A ~-11 -· •• ~~ .-1111 ' "~\ Mil '_A I yirvv1\v ff 1441PV 'ij "f v v' -r 1JV
. 4
. • A I\. • 1\ I\ t A " • 1\ V' A J.. ~ A ft. t. A 1\. .... • VA A 'V r 1 VOV' V"' • - • " vr y v 41 vV' 2 . 4 2 .8 3 . 2
1\- A• /1. I -.. 1 0 AI\ 1 Jll,....AA lc: 1\ .... y 0 J\'t _,A AA AI\ \F V vv - y yyVv 41 V¥ V 4Py Vy v I
3. 2 3 . 6 4. 0
Time (s ec . )
Figu re 5 . Trace of primarie~ , mul tiples and rand om noise . P = 10 .
f-' ~
' - .. - A ., A - - + - .. A • ' A "= A 1'\ Jl. .... -Jt. A ~- A A .. - v ., .. y ~ "" ll vyy v«e> r- -, 0.0 0 . 4 0 . 8
' A A A A. -.-- - ~.A A. .. • A - -A A. A .._AA A ~A ... - - • .A A ,.~~~.. ~A A .... .._ fT- y -~'<1 v VYV w a. rr ~ 0. f y r 0. 8 1.2 1.6
'f .u- A,.. A "'f" .,-A-~ r· A, .,--''- ., ""'!. """· .,. J.. ... tvw1'Y'--F '¥ 1 ~, r ' , ... 1 .6 2 .0 2.4
... • fA. • AA. • AI ~A .,A . ..A ~- ,A.,_ fA. ., • .,. -A 'Y · T T •vrr - .., f¥ T • • ~ .. 2 . 4 2 . 8 3 . 2
~- .... ., 1\ ' - -' - AA • -~ .. A - ~ A .... ,. - .. ., ..A - v AA. :4 v-y vV* = f yy'V'\F «JF 4N ''*' 't Vf • • 1 u
3.6 3.2 Time ( se c . )
4 . 0
Fi gure 6. Trace o f p r imaries , mult ip les & r a ndom noise . P = 5 .
1-' (J1
Ia n ~litO .... ~A .,A 0 7 _ ~ rt rt .... en A 0007 • A I\__ Aft .a ... - __ ..Alt.. •-•- A.A. 4 0 .. v• *'VI i.<$1 v a u.v " =9fi li~Vf o:cr ,.- =use y' 0.0 0.4 0.8
., AAA,.._ ............. rt ~....~>.A a • A , -A.A. _A ~AA A._ AA ,. e n =• ...... & 0 -"A ...AAA --,_, ovy-r ~ v· • V v y'V y«Uq?" ... ~T T 4 r· ,~,v,.
0.8 1.2 1 . 6
'f __ .u.A""_A "tf"··A-· ·- JA. . .,....Aa.-- .. -- A..,. MA ..... A.- .. v fY ... ,.,. ,-v . t'l -y- -·T ""f lr-
1.6 2.0 2.4
.. - • A A- - AA • AA ~ /L_, A .~ .A -.. .A.,_. fA. ~ ..... .,. .. & ...._A ~ CO 41 0 r OW' I U W41Jyy yV'-" 40.... - b ¥1¥ 17 yn; ~
2.4 2.8 3.2
it( ..... A ., A. aft~ •• fA ... --·A.- A .... ,. --A- .A~ ..... AA.. ._. ., • ..,. u GV. =r 0 f ·-yyvv v..; 4<> .. --r 1?F. vyr r =
3 . 2 3 . 6
Time (sec.)
Figure 7. Trace of primaries, multiples and random noise. P = 2.
4.0 1-' 0\
been recorded with no frequency filtering (i.e., wideband
recorded), and comprise the input to the processing schemes
which follow. For processing and analysis purposes, only
the form of the input wavelet was assumed known.
F. Processing
In order to isolate each event as much as possible
each trace g(t) was deconvolved with an inverse filter
constructed from the statistical information contained
within the trace itself. A description follows of the
various processes necessarY to calculate and apply an
1nverse filter or deconvolution operator. The mathemat-
ical development of the operator appears in Appendix A.
In order to construct a deconvolution operator for a
17
seismic trace, one must know the expression for the original
wavelet. which, when convolved with a reflectivity function,
yields the seismic trace. The wavelet is determined, accord-
ing to Robinson (1967b), by taking the inverse complex
Fourier transform of the square root of an estimate of the
power spectrum of the trace. This estimate may be obtained
by complex Fourier transforming the smoothed autocorrelation
function of the trace.
1. Autocorrelation Function
The autocorrelation of a continuous function h 1s
defined as 00
¢hh(T) = ~ h(t)h(t+T)dt (8)
-00
] 8
Using the discrete form of equation (8), represented by
the sampled data autocorrelation functions (ACF) of the
input traces were obtained where again the notation lS
that of Wainstein and Zubakov (1962) with t and T being
integer valued and the overbar signifying average value.
Because the ACF is an even function only ¢ (T) for T > 0 gg -
was calculated. Since the input traces had a sampling
interval of 1 ms., the value of ¢gg(T) was computed at 1 ms.
intervals, while, for practicality, the ACF's were truncated
at T = 260. The function ¢gg(T) for 0 < T < 260 with P = oo
(i.e., no random noise) is shown as figure 8. The ACF's for
P = 10, 5, and 2 are not shown since they appear much the
same as that in figure 8.
2. Power Spectral Estimate
If h(t) is considered to be the limiting case of a
periodic waveform whose period has approached infinity,
and if h(t) possesses finite energy such that
00 12 _[ lh(t) dt < M (10)
then the complex Fourier transform of h(t) is defined as
00
H(f) = J h(t) exp (-2njft)dt (11) -00
19
150
120
90
60
30
-30 Lag (ms.)
-60
Figure 8. Autocorrelation function -90
of synthetic seismic trace. p = 00
where f is frequency and j is the unit complex quantity.
The original time function, h(t), may be obtained from
its frequency domain representation, H(f), through the
inverse complex Fourier transform given by
H (t) 1 -z:;y--
00
~ H(f) exp (Znjft)df -00
H(f) is a complex quantity and hence may be uniquely
specified by its amplitude spectrum, IH(f)l' and its
phase spectrum,
e (f) -1 tan
Im rH (f)] Re [H (f)]
20
(12)
(13)
where Re and Im signify real and imaginary parts, respec-
tively.
For this investigation, the discrete Fourier transform
of the empirical autocorrelation function is given by
n-1
rgg (f) = T=~+l cj> ( T ) e Xp ( - 2 TI j f T ) gg (14)
considering ¢ (T) to be one cycle of a periodic function. gg
The frequency function f is called the Schuster periodo-
tram and is an asymptotically unbiased estimate of the
power spectrum of the trace g(t). "Unbiased estimate" is
interpreted to mean that the estimate, considered to be
a random variable, will have an expected value "close" to
the true value of the power spectrum. However, according
to Robinson (1967b) and others, the periodogram is not a
consistent estimate of the amount of energy present in
any given band of frequencies, i.e., it is not a consist-
21
ent estimate of the power spectral density. The definition
of "consistency" is that as given by Meyer (1965).
states that p is a consistent estimate of p if
lim Prob n-+ 00
where p 1s an estimate based on a sample x 1 ,
This
(15)
the parameter p. Bartlett (1948) states that the variance
of the periodogram at any given frequency does not approach
zero as the sample size, n, approaches infinity, and hence,
I~g(f) is inconsistent. A consistency theorem and discussion
regarding the limit of the variance as n approaches infinity
is presented in the above cited work by Meyer.
There are at least two ways to obtain a consistent
power spectral estimate. One way is to weight or "smooth"
the ACF in the time domain with a lag window function or
smoothing function and then take the Fourier transform of
the result to obtain the required estimate. Another way is
to compute the periodogram, and then through convolution,
perform a smoothing operation in the frequency domain to
arrive at the estimate. Since the two methods are mathe-
matically equivalent, time domain smoothing was chosen for
computational purposes.
22
Smoothing functions are, in general, quite well-known
and a wealth of literature describes their characteristics
and mathematical properties. Among the leading investigators
are Bartlett (1948,1950), Zaremba (1967), and Parzen (1967).
An easily understood exposition of the basic theory is that
authored by Blackman and Tukey (1959).
For reasons which follow, the particular smoothing
function chosen was the Bartlett or Fejer window, first
proposed by Bartlett (1948). This is defined as
I t I I t I wCt) = 1 - < t
' m t m
I t I (16)
0 > t m
Equation (16) represents an even triangular function,
the slope of whose sides is determined by tm' the truncation
point, and for which w(O) = 1. It has been shown by Parzen
(1967) and others that the location of the truncation point
affects both the resolution and the bias of the spectral
estimate. The t for this problem was chosen to be 64, m
that is, at the 64th lag of the ACF. The Bartlett window
is a member of a class of smoothing functions which, upon
application, always yield non-negative spectral estimates.
Other smoothing functions such as the Hamming do not neces
sarily yield non-negative estimates. See Blackman and Tukey
(1959). It can be shown that the non-negative property
depends upon the non-negativity of the spectral window,
i.e., of the Fourier transform of the lag window. The
Bartlett lag window has as its Fourier transform
23
W(f) = (sin nftm)
2 (17)
Tift m
Thus, smoothing the empirical ACF with a Bartlett window
not only yields an unbiased, consistent, and therefore
much less erratic, spectral estimate upon Fourier trans
formation, but also, as a "lagnappe", prevents the occur-
renee of any fallacious negative power.
Following the procedure outlined above, the smoothed
empirical autocorrelation functions weresubjected to a
discrete complex Fourier transformation to yield estimates
of the power spectrum for each signal-to-noise ratio
investigated. The fast Fourier transform (FFT) algorithm
as developed by Cooley and Tukey (1965) was utilized. The
details regarding its implementation and operation are
presented by Cooley et al (1967), Cochran et al (1967),
Welch (1967), Robinson (1968), and others. The actual
FFT subroutine used was that as published by Robinson
(196 7b). It was decided to use a 256 point FFT since this
would give a resolution in the frequency domain of approx-
imately 3.91 cycles per sec. The estimate of the power
spectrum for P = oo is shown in figure 9. Minor differences
appeared in the estimates for P = 10, 5, and 2.
1200
900
1-< <l>600 :s: 0
0..
300
1
(!.) 'lj
;:::$ 0 +->
·rl r-i p.. E
<::r::
-1
406£ 806£ 1206£ 1606£ 2006£
Frequency 6£ = 3.91 Hz.
Figure 9. Power spectral estimate
of synthetic seismic trace. P =oo.
40 80 120 160
Time (ms.)
Figure 10. Statistically derived
wavelet, - y(t). P = oo •
200
24
2406£
240
3. Wavelet Reconstruction
Through the use of the inverse FFT, the inverse
Fourier transform of the negative square root of each
power spectral estimate yielded the time function wavelet
-y (t). The wavelet obtained for P = oo is shown as figure
25
10. No significant variations were noted for the other
values of P. This wavelet is a time function derived from
the statistics of the seismic trace, and actually is an
"average" in some respect of the two distinct wavelets
originally utilized to synthesize the trace. The negative
root was taken since Ricker (1945) has shown that a sym
metric wavelet of the form represented by equation (2)
contains only negative Fourier cosine components. This
implies that it has a constant phase spectrum of TI.
4. Inverse Filter Construction and Deconvolution
Know~ng the wavelet as a time function one may then
construct a deconvolution operator which theoretically will
yield spikes when convolved with the seismic trace. Rice
(1962) says heuristically that one wishes to "perform the
inverse of the reflection process". Mathematically one
seeks a solution to the matrix equation
cpA = y (18)
where y is the negative of the normalized wavelet obtained
from the preceeding analysis, cp is its normalized autocor-
relation matrix and A is unknown. The manipulation of y
26
implies that its phase is everywhere zero and follows the
theory of Rice (1962). In equation (18) the symmetric
matrix¢ is of the Toeplitz form, i.e., all elements along
any particular diagonal are the same. This property greatly
reduces both the required computer storage space and the
computations necessary to obtain a solution.
It is generally accepted (Robinson, 1969) that an
operator length of approximately twice the effective wave-
let length yields sufficiently reliable results upon de-
convolution. Rice (1962) has hypothesized that deconvolution
efficiency is directly proportional to the length of the
operator; however, the accumulation of computer round-off
error nullifies this hypothesis. A 49 point solution to
equation (18) was obtained through the utilization of
subroutine TPLITZ, a listing of which appears in Appendix B.
The deconvolution operators for the four signal-to-noise
ratios are shown in figures 11, 12, 13, and 14.
Efforts were made to reduce the amount of high frequency
side lobes present in the deconvolution operator a(t) by
prewhitening the amplitude spectrum of the derived wavelet.
¢ (0) was increased by 10, 20, and 50 per cent with the yy
results that all three experiments yielded more smoothly
varying filters whose side lobes were reduced in both number
and amplitude. However, upon convolving these operators with
the seismic traces according to
27
15
12
9
6
3 (j)
"0 ;:j .j....l .,...; ..-1 0 p... s ~ 50
Lag (ms.)
-3
-6
-9
Figure 11. Deconvolution operator
for synthetic seismic trace. p = 00 •
28
15
12
9
6
3 Q)
'"0 ;::J
+-1 .,...; .--1 ~ 0 s
-< 50
Lag (ms.)
-3
-6
-9
Figure 12. Deconvolution operator
for synthetic seismic trace. p = 10.
29
15
12
9
6
3
(J)
'"d ;:l +-' . ...,
0 rl 50 0... s Lag (ms.) <r:
-3
-6
-9
Figure 13. Deconvolution operator
for synthetic seismic trace. p = 5.
30
15
12
9
6
3
C)
'""0 ;:J -1-l
·..-1 0 r-f 0.. 50 s
<r: Lag (ms.)
-3
-6
-9
Figure 14. Deconvolution operator
for synthetic seismic trace. P = 2.
31
N
d(t) = L g(-r)a(t--r) (19) T=l
an unacceptable lack of resolution existed in d(t) and for
this reason the raw ACF of the wavelet was employed for the
solution of equation (18) and is the operator a(t) in
equation (19). This yielded the output traces shown in
figures 15, 16, 17, and 18 for values of P equal to oo ,
10, 5, and 2 respectively.
I I • A yA ' s • j I _,a tz Yt . ' . 1 • .. MJt • 4/Ji lll 0.0 0.4 0 . 8
1 #A ~~~ -. I •• - ., I "' .... • tt .,. . ..~ I -- r'" v f l -1" t t • T t , 0.8 1.2 1.6
Ill.,.. ~A ,1 .. wA W I r J... -'"'- I yA #Wh I 1 I • f I • T J ww-1.6~0 2.4
I I .... ' ... ' u ... 1 A Al . _.A.~ 1 " ,.. J t 1 ..,.., IT " 2.4 2.8 3. 2
... ·-· y ' . ' .., .,_ J, .. ,. y . .,, 4 ' ·-t ' ff )f" aw 1 -.JU •1 3.2 3.6 4. 0
Time (sec.)
Figure 15 . Deconvolved synthet i c seismic trace . P = oo.
~ N
I -·-.. I ·~··"'...,.,...I .... 1 W. I ...... '¥'\sial .J. .... • .I "'i-1"('••+ 't.,. U II ,, .. 0 - 0 0 .4 0.8'
0 . 8 1.2 1.6
~ ~ -A.-"fr I .,.,...,,, • ,. ... ~~~~ ,.,,'r I. I I ~ S: 't"l"'t :fo ;p .......
1 - 6 2 .0 2.4
l •. \ * "* fl ~ I -r I*' ,.,.. I ...M,.~ ..,.... .. ...,.,.... .... tr II.,. "'' v~ I ..,J 2 . 4 2.8 3 .2
"""'"'f":o"yt"':A II ol lloa ""' ""IIIII ............. I """ >IQII •• ··~· ~ • .. ""f• I ,.,. ~ ~' .,. ... ' - VJO "( 1 f" -· \lA! .. ..
3.2 3. 6 4 . 0
Time (sec.)
Figure 16 .De convo l ve d synthetic seismic trace . P = 10.
~
VI
I ............... •v••fl'to ..... I·IJt~ ,.,., ... (t .. .Y."ff'""'"' .,~. f .... 0.0 0.4 0.8
..., .. ,...,.~.,.. ,.,.,~ ... ·""'~ .... ,., ........... Jot~ .... ,J. ... .. 1 . 6 2.0 2. 4
.......... , nr• ... ..,. 0 .J ...... 'V' • • ., ..... ..,.,. .. .,, +r ",.."' •'f' .. .,J 2.4 2.8 3.2
.................... ~·· .. ~ . ...,..,,..,. , ... ,.,. .. 3.6 4.0
Time (sec.)
Figure 17. Deconvolved synthetic seismic trace. P = S.
~ ~
I ............ ~~, .... ,.,._,._,..At4M• '*•'~1"~,...,.,,.., .. 0.0 0 . 4 0.8
0 . 8 1.2 1.6
~·~~.-~, •'t"-t-·~··--.... ~~· ... ··.Jr... 1.6 2 . 0 2.4
•••rt * l1le ..... ..-.~1 rtcJ.J .. ~~ .. ...,J..,a.low: .. ,t .. tY•W: ,..,J 2 . 4 2. 8 3. 2
3 . 2 3.6 4.0
Time (sec.)
Figure 18. Deconvolved synthetic seismic trace . P = 2.
VI V1
III. PATTERN RECOGNITION ANALYSIS
OF SYNTHETIC SEISMOGRAMS
36
The following analysis may be termed a linear thresh
old element (LTE) classificatory scheme. A linear thresh
old element can be described as a mathematical device ,
which, when subjected to a pattern of n inputs, forms a
linear combination of the inputs in a manner which results
in a single output. The output is then compared to a thresh
old with the result that the input pattern is classified
into one of two classes depending on whether the output is
less or greater than the threshold.
An early investigator of LTE techniques was Fisher
(1938) who used weighted sums of measurements of properties
to differentiate between species of Irises. More recently,
Mattson and Dammann (1965) used LTE's to determine and
code subclasses in an n-dimensional pattern space.
In the discussion that follows, the term "pattern" will
be considered to refer to one of the 38 seismic events of
interest selected from each deconvolved trace of figure 15,
16, 17, or 18. A set of properties is used to characterize
each pattern, while the numbers associated with these proper
ties make up the pattern vector. Thus, if n properties of
each seismic event are measured and represented with n real
number (x1
, x2
, ... ,xn), these numbers then constitute a
pattern vector in an n-dimensional space. The events of
interest for the four signal-to-noise ratios are listed in
table VI in Appendix C.
A. Parameters Investigated
37
The parameters used to describe a pattern in this study
are peak frequency, mean frequency, amplitude spectrum vari
ance, periodicity, and polarity. A discussion of each param
eter follows.
It is observed that weathered materials near the surface
of the earth attenuate high frequency energy. Because a
surface multiple travels two or more times through the
weathered zone while a primary traverses the weathered layer
only once, it is to be expected that surface multiples con
tain relatively more low frequency energy than primaries.
Epinat'eva and Ivanova (1959) utilized this fact as a basis
for constructing low-pass filters to eliminate certain types
of multiples and Mateker (1965) givPs an example to substan
tiate the property of anoma1ous attenuation of high frequency
energy by the weathered zone.
One of the pattern parameters used was peak frequency,
that is, the frequency at which the maximum value of the
amplitude spectrum occurs. It was expected that primary
events would have higher peak frequencies than multiple
events. This parameter was measured from the amplitude
spectrum obtained for each event, the details of which will
be subsequently discussed.
38
The mean or average frequency was employed because it
was expected that primary events would exhibit higher means
than multiple events. The mean frequency for each pattern
was also obtained from its amplitude spectrum. The final
frequency parameter used in this study was the sample vari
ance (with respect to frequency) of the amplitude spectrum
of each event. The relationship n
LAiCfi-:oz
'2 i=l q =--------- (20)
was used to calculate this parameter. In this expression
Ai and fi represent the amplitude and frequency respectively
of spectrum sample i, tis the mean frequency, and n is the
number of amplitude spectrum samples.
The amplitude spectrum for each pattern was obtained
through a 256 point FFT of the 11 points (11 ms.) considered
to constitute an individual event. The 11 points consisted
of the central extremum, the 5 points immediately preceding,
and the 5 points immediately following the extremum.
The pattern parameters of mean and peak frequency are
considered to be statistically independent since it seems
reasonable to assume that the value of one parameter in no
way influences the value of the other. Figure 19 shows the
amplitude spectrum of a typical pattern selected from
figure 15.
8
6
Q) '1j
;:::j .j.J
•M 4 r-1 .
~ -:r:
2
2011f 4011£ 6011f 8011£
Frequency 11£ = 3.91 Hz.
Figure 19. Partial amplitude spectrum of event no. 9 from figure 15. t 0 = 1118 ms.
Vl 1.0
40
Another parameter used in this study was the "periodic
ity" of the trace. A group of 49 events was selected from
the first 2.0 seconds of the trace to serve as a reference
set to which the 38 patterns of interest were compared for
time periodicity. This search and identify operation can
be thought of as follows: Given an event in the reference
set whose central extremum occurs at time t 0 , is there an
event of interest (from the original 38) occurring at time
t = K t0
± o (21)
where K = 2, 3, ,n, and o = 0,1,2,3,4, or 5 ms? In
this equation, inclusion of an error term, o, allows for
the possible shift of the central extremum due to construc
tive or destructive interference from neighboring events.
If a pattern of interest was found to satisfy these criteria,
this fact was coded with respect to the value of o. The
codes are listed in table II. A code of zero was employed
to denote nonsatisfaction of the above criteria. It was
arbitrarily assumed that the patterns corresponding to pri
maries would have lower codes than those corresponding to
multiples. The reference events for the four signal-to-noise
ratios are listed in table V in Appendix C.
The final parameter included in each pattern vector
was "polarity". The extremum of each event exhibiting a
non-zero periodicity code was examined and coded for phase
Table II. Patt.e.rn vector coding for
periodicity component.
I o I (ms.)
0
1
2
3
4
5
K
odd
odd
even
even
Table III. Pattern vector coding for
polarity component.
Polarities of Reference Event and
Event of Interest
opposite
alike
alike
opposite
code
10
6
4
3
2
1
code
10
1
10
1
41
inversion according to the criteria listed in table III· ,
patterns having zero periodicity codes necessarily were
assigned zero polarity codes. Thus primaries were gen
erally represented by lower polarity codes than multiples.
The pattern vectors for the four signal-to-noise
ratios are listed in table VI in Appendix C.
B. Determination of Weighting Vector
Consider the mathematical representation of an LTE
given by
42
(22)
where xik
i ;:::: 1,
i :::: 1, ,n is the kth pattern vector and w. l
,n is a weighting vector. Each pattern vector
1s thus reduced to a single number Sk. A set of weights is
to be found such that the values of S for one class of
patterns will be markedly different from the S values of
the other class. This difference can be observed by con-
structing a histogram of S values which will be referred
to as an S line distribution.
It is shown in Appendix D that an optimum set of weights
1s one which will maximize the moment of inertia of the S
line distribution given by
p
M:::: L ( 2 3)
k;::::l
43
subject to the constraint that
n 2 L: w.
l = min. (2 4)
i=l
In equation (23) S is the mean S value and p is the total
population of patterns. Equation (23) is expanded and
rearranged to yield
n n
M = L 2: i=l j=l
w.b .. w. l lJ J (25)
where the b .. are elements of a matrix B which is proportionlJ
al to the sample covariance matrix of the pattern vectors
, ... ,xnk), k = 1, ... 'p. It is further shown
that the set of weights is the eigenvector corresponding
to the largest eigenvalue of the matrix B.
Each signal-to-noise ratio was analyzed twice: once
using all the pattern parameters previously mentioned and
once using all except the parameter of amplitude spectrum
variance. Accordingly, the matrix B with elements
b ... = lJ
p
I: k=l
(x.k-x.) (x.k-x.) l l J J
(26)
was computed for each value of P. The IBM subroutine EIGEN
(1967a) was used to compute the eigenvalues and their cor
responding eigenvectors which are listed in table IV. After
calculating the eigenvalues of B, a normalized version of
the eigenvector corresponding to the largest eigenvalue was
44
Table IV. Eigenvalues and eigenvectorst of the matrix B.
Eigenvalues +- 53951. (f) 12501. 654.3 161.7
!-< 0 .378 .922 .078 -.001 +J u .923 (j) -.383 .053 .011 > s:: -.052 -.025 .557 .829 (j)
bO -.060 .,...; -.046 .825 -.559
~
p =00 Variance not included.
Eigenvalues
56814. 35796. 1196.1 604.9 161.1 +-
(f) .238 .637 .678 .278 -.019 !-< 0 .911 .105 -.394 -.060 .021 .;..l
u -.047 -.029 -.156 (j) . 542 .824
> s:: -.052 -.398 -.278 . 7 7 4 -.566 (j)
bO .,...; -.329 . 762 -.533 -.164 .015 ~
p = 00. Variance included.
Eigenvalues +- 64710. 18123. 966.0 144.7 (f)
!-< 0 -.063 .997 -.043 - . 0 2 9 .;..l
u .997 .065 .042 .008 (j)
> s:: -.032 .046 (j)
.534 .844 bO
.019 .843 -.536 .,...; -.033 ~
p = 10. Variance not included.
t Eigenvectors are listed in the order: mean frequency, peak frequency, periodicity, polarity, amplitude spectrum variance.
45
Table .IV. continued.
Eigenvalues
9496 7. 29630. 4030.2 865.9 141.9 +-
IJ) -.181 .663 .723 .070 -.012 l-< 0
. 75 8 ~ .565 -.327 -.017 -.001 u <l) -.035 .025 -.067 .513 .855 ~ !=: -.032 .007 -.105 <l) . 848 -.518 0.0
·r-l -.625 .490 -.596 -.113 -.019 ~
p = 10. Variance included.
Eigenvalues
+-IJ) 60099. 34206. 890.2 146.4 l-< 0 -.095 .994 -.047 -.009 ~
u <l) .995 .096 .032 -.008 ~ !=: -.014 .032 .554 .832 <l)
0.0 •r-l -.034 .031 .831 -.555 ~
p = 5 . Variance not included.
Eigenvalues
78074. 37840. 12354. 877.1 139.4 +-
IJ) -.222 .859 -.460 -.036 -.002 l-< 0 .807 .427 .407 .020 -.016 ~ u .543 <l) -.020 .028 .016 .839 ~ !=: -.034 .018 -.014 .838 -.544 <l) 0.0
·r-l -.546 .281 .788 -.028 -.020 ~
p = 5 . Variance included.
46
Table IV continued.
Eigenvalues
+- 70 58 7. 24872. 1019.8 160.4 Vl H 0 -.014 .999 -.040 -.029 +J u .999 .016 .022 .015 (])
> s:: -.024 (])
.045 .497 . 866 OfJ -.012 .020 . 86 7 -.498 •r-i
lJ-4
p = 2. Variance not included.
Eigenvalues
+- 96306. 37550. 6149.0 918.0 158.0 Vl
. 716 .676 .039 -.018 H -.170 0 -!-} .790 .510 -.340 -.025 .008 u (])
-.043 .479 > -.030 .029 .876 s:: (]) -.018 .019 OfJ
-.088 .871 -.482 •r-i
-.646 -.096 -.015 lJ.4 -.588 .476
p = 2. Variance included.
entered in equation (22) and an S value for each pattern
obtained. These were plotted as points on an s line and
presented in figures 24 through 27, forming the basis for
the pattern recognition analysis.
C. Analysis and Discussion
The measured values of the frequency parameters and
47
was
are
the assigned codes of the polarity and periodicity parameters
are such that the multiple events should have lower S values
than primaries. Thus, identification of separate clusters
of primaries and multiples should be possible upon visual
examination of the S line distributions provided.
In figures 20 through 27, circles are used to denote
primaries and triangles are used to denote multiples.
Figures 20 through 23 illustrate, for comparison, the
results for four signal-to-noise ratios using non-optimum
(unit) weighting vectors. The two classes of patterns,
primaries and multiples are not readily separable. This lS
true even under the ideal signal-to-noise ratio ofoofor the
case where all five parameters were employed.
Figure 24a shows very good clustering of the multiples
and sufficient separation of the two classes so as to permit
identification by inspection. For the same signal-to-noise
ratio, employment of all five parameters results in acceptable
clustering of the multiples but a scattering of anomalous
primaries between the two clusters of dissimilar events
(figure 24b). In figure 2Sa it is observed that two primaries
d - multiple 0
- primary
-1.0 -0.6
5
4 (!)
3 '-HU 0~
(!)
2 IJ) 1-< .6. (!)$-<
1 s ;:::l ~ ·rl u
E--<U
-0.2 0.0 0.2
Normalized S
(a) Variance not included.
-1.0 -0.6
6
5
-0.2 0.0 0.2
Normalized S
(b) Variance included.
48
00
000::0::00 oo
a:x:oo~ooo
0.6 1.0
0
a:o
a:o
0 o:ocoo 0
0.6 1.0
Figure 20. S line distribution produced
with unit weights. P = oo •
,6. - multiple
o - primary
-1.0 -0.6
5
4 (])
lH t)
3 0 s:: (]) 0
2 Vl !-< (]) !-< 0 s ;:::s
1 •r-i t)
E-< t)
0
-0. 2 0.0 0.2
Normalized S
(a) Variance not included.
6
5
4 (])
lH t)
3 0 s::
(]) Vl !-<
2 (]) !-< s ;:::s
1 •r-i t)
E-<U 0
0
-1.0 -0.6 -0.2 0.0 0.2
Normalized S
(b) Variance included.
49
00000
0
0.6 1.0
0
0
0
0 GIXX> 0
0 D,ax/so d:J.
<X> ax>~
0.6 1.0
Figure 21. S line distribution produced
with unit weights. P = 10.
~- multiple
o - primary
-1.0 -0.6
-1.0 -0.6
5
4
3
2
1
Q) '-HU 0 .::
Q) til!--< Q) !--< ~ ;::l
.,...; u E-<U
0
-0.2 0.0 0.2
Normalized S
0
co
0 0:00
o .C,.o~
oo.C.oco~
0.6
(a) Variance not included.
5
4
3
2
1
Q) '-HU 0 .::
Q) til !--< Q) !--< ~ ;::l
.,...; u E-< u
0
-0.2 0.0 0.2
Normalized S
(b) Variance included.
ooo
0 a:tn> (])
0 cn:tXlXtX>
0 O~(J)
0.6
Figure 22. S line distribution produced
with unit weights. P = 5.
50
1.0
1.0
a- mu1 tip1e
o - primary
-1. 0 -0.6
5
4
3
2
1
Q) 4-IU 0 s::
Q) til 1-< Q) 1-< s ::I
·rl u E--< u
-0.2 0.0 0.2
Normalized S
(a) Variance not included.
5
4 Q)
3 4-IU OS::
Q)
2 tllH Q)H
1 S::l
•rl u t--<U
-1.0 -0.6 -0.2 0.0 0.2
Normalized S
(b) Variance included.
51
0 co 000
0 f). 0/:S;D 0 ClD
0.6 1.0
0 0
0.6 1.0
Figure 23. S line distribution produced
with unit weights. P = 2.
52
exhibit S values in the same range as four of the multiples
and that the S line distribution appears to show three rather
distinct clusters of patterns. Figure 25b shows more intra
cluster dispersion than does figure 25a. The S line distri
bution for the four parameter case with P = 5 mistakenly
clusters two primaries as multiples, while in the five param
eter case intracluster dispersion again appears unacceptable.
Figure 27a and 27b indicate the deleterious effects of the
high level of random noise. Figure 27a shows an equal
number of primaries and multiples in the same cluster as
well as two multiples surrounded by a scattering of primaries.
Figure 27b shows practically no separation of the two classes
as well as poor clustering of the primary events.
In general, less intracluster dispersion and more
intercluster dispersion were obtained by not utilizing the
variance of the amplitude spectrum as a pattern descriptor.
Both of these characteristics would seem to be desirable
for good clustering and separation.
Several attempts were made to find combinations of
two or three of the pattern parameters which would result
in good clustering of the pattern classes when plotted on
an S line. Limited success was attained by using peak
frequency and mean frequency, but only for the trace which
had a signal-to-noise ratio of 00 •
In this case the S line
6.- multiple
o- primary
lH 0
U)
(J)
s •r-i
(J)
u r::::: (J) !-; !-; ;::j u
5 0
4 00
3 £::. 00:0
2 £::. 0 ooxa:oo E--< u 1 ~ 0 0::0 0 0 ()XO)OCO)J ro
-1. 0 -0.6 -0.2 0.0 0.2 0.6 1.0
Normalized S
(a) Variance not included.
~i 0
(J) CX)
lH u 0 r:::::
(J) 0 Oa:x:>
U) !-; (J) !-; A 0 00 OCXX) s ;::j I
•.-1 u ltiO. A 0 00 0 E--<U
0 0 CXXlO CXXXXX> 0 0
0
-1.0 -0.6 -0. 2 0.0 0. 2 0.6 1.0
Normalized S
(b) Variance included.
Figure 24. S line distribution produced with
eigenvector corresponding to A P = oo • max.
53
~- multiple
o - primary
-1.0 -0.6
5
4 (j) 4-lu
03 0 ~
(j) rn H
.&2 (!) H 00 s ;:j
L::.a> 1 ·r-1 u oo.c:..oo E-<U
0
-0.2 0.0 0.2
Normalized S
(a) Variance not included.
5 (!)
4-1 u 0 ~
0 4 (!) (f) H
3 (j) H
0 s ;:j •r-1 u
0 0 <XDcO E-< u 0
-1.0 -0.6 -0.2 0.0 0.2
Normalized S
(b) Variance included.
0
0
CX)
000:0 0
0 «XXX> cr:xxDCX> CX)
0.6
0.6
Figure 25. S line distribution produced with
eigenvector corresponding to A P = 10. max.
54
0
1.0
1.0
L1- multiple 0
- primary 5
4
3
<!) 4-iU 0 I=!
0
0
0 2
/':::LX:s) 1
<!) (f) !-< <!) !-< s ::s
•rl u E-U 000~ oco 000(11)])000 0
0
-1.0 -0.6 -0.2 0.0 0.2 0.6 1.0
Normalized S
(a) Variance not included.
54-i (])
u 0 I=!
0 4 <!) U) !-<
0 0 0 3 <!) !-< s ::s ·rl u
0 0 0 0 0 00 00 E-U 0
AO L» 0 0 0 0 OX> cxx:x:m:o 0 0
-1.0 -0.6 -0.2 0.0 0.2 0.6 1.0
Normalized S
(b) Variance included.
Figure 26. S line distribution produced with
eigenvector corresponding to A P = 5. max.
55
~- multiple
o- primary Q) 4-1U 0 s::
Q) Ul 1-<
3 a.> 1-< E ~
2 •r-1 u E-<U
0
56
0
0 o oa:o
oo o o o cr:. oLP o o o:x:o o cxx:o oo o
-1.0
-1. 0
-0.6 -0.2 0.0 0.2
Normalized S
(a) Variance not included.
0
00 000~ 00 co
-0.6
0
0 0
0 o a:oo
Q)
5 4-1 u 0 s::
Q) 4 Ul 1-<
a.> 1-<
3 E ~ •r-1 u
2 E-< g o o o a:x:ooooo o o
-0.2 0.0 0.2
Normalized S
(b) Variance included.
0.6
0.6
Figure 27. S line distribution produced with
eigenvector corresponding to A P = 2. max.
1.0
1.0
showed not two, but three clusters of patterns, viz., one
cluster of multiples and two clusters of primaries. For
57
the traces with values of P equal to 10, 5, and 2, no success
with a reduced number of pattern parameters could be attained.
Because there are over 120 combinations of the five parameters
for the four signal-to-noise ratios it was impractical to
investigate all such combinations. However, it is felt that,
in general, acceptable results are not obtainable with three
or less parameters.
In figures 24 through 27 a threshold has not been in
dicated since no a priori information regarding threshold
location is known. If a priori S line distribution infor-
mation for a given signal-to-noise ratio is provided from
previous analyses, it would seem feasible to specify the
probabilities of the occurrence of Type I and Type II errors
and thereby select a threshold before analyzing a trace of
seismic data. If such a priori information has not been
provided one would not be able to select a threshold in
accordance with any preset statistical confidence level.
Instead, one could only assume that S line clustering and
intercluster separation is indicative of the existence of
two classes of patterns.
ss
IV. SUMMARY AND CONCLUSIONS
The object of this study was the investigation of a
linear threshold element technique for identifying surface
multiples on a single seismic trace. Traces of seismic
events were generated which contained primaries, surface
multiples, and various levels of Gaussian random noise.
The traces were subjected to pulse-compression deconvolution
processing in preparation for LTE analysis. The pattern
parameters employed were mean frequency, peak frequency,
amplitude spectrum variance, periodicity and polarity. A
set of weights was found that would maximize the moment of
inertia of the S line distribution of the patterns subject
to the constraint that the sum of the squared values of
the weights was minimized. The classes of patterns repre-
senting primaries and multiples on traces with high signal
to-noise ratios were clustered and separated, making identi-
fication by inspection a simple procedure. Clustering and
separation of classes on traces with low signal-to-noise
ratios was less than optimum.
The results of figures 24 through 27 generally in
dicate that an LTE classificatory scheme provides a usable
tool for identification of surface multiples on single
seismic traces. The figures indicate that the "performance"
of the technique deteriorates with increased random noise.
The rather poor results obtained with the LTE technique for low
59
signal-to-noise ratios is attributed to the relatively poor
performance of the deconvolution operator. It is felt that
a matched filter such as that described by Wainstein and
Zubakov (1962) would provide better resolution of individ
ual events on traces with low signal-to-noise ratios.
Treitel and Robinson (1969) discuss the advantages of using
matched filters for deconvolving low signal-to-noise ratio
seismic traces.
It appears from this study that the "proper" selection
of pattern parameters effectively governs the power of the
technique. This limitation is of the same form as that for
pattern recognition techniques applied to other problems.
The question arises: "What properties does one need to
measure in order to specify a pattern?" This question must
remain largely unanswerable since the answers are most prob-
ably data dependent. The fact that virtually no success
could be attained with a reduced number of pattern parameters
implies that two dissimilar clusters of the input patterns
do not exist in the reduced pattern space. This does not
rule out the possibility that two dissimilar clusters of
another set of input patterns might exist in some two or
three dimensional pattern space.
If the process of frequency dependent attenuation is
accepted, it would seem more correct to employ a time-varying
attenuation process rather than the time-invariant process
60
used in this study. With time-invariant frequency attenua-
tion, the variance of the amplitude spectrum is felt to be,
at best, a pattern descriptor of minor importance. There
does not appear to be any physical reason to suggest that
the variance of the amplitude spectrum of a primary should
be different from that of a multiple. If however, a time
varying attenuation process were assumed for the multiples,
one would generally expect their amplitude spectra to be
different in shape from the spectra of the primaries. This
difference in shape might be evident not only in the variance,
but also in higher-ordered statistics of the amplitude spec-
trum. The simplest time-varying attenuation process might
well be an exponential of the form exp(-aft) where f and t
are frequency and time, respectively, and a is a constant.
More physical realism could be obtained by considering
a seismic trace to contain non-surface-generated as well as
surface-generated multiples which have geometrically simpler
propagation paths.
On more general terms, the concepts of pattern recog
nition analysis should be applied to other problems of re
flection seismology, measuring space variations in addition
to time variations. Some examples of such problems might be
the identification of diffractions generated from faults and
the general problem of automatic reflection picking. Initial
efforts on the latter subject have been made independently by
Hileman et al (1968) and other research workers.
V. APPENDICES
APPENDIX A
MATHEMATICS RELATIVE TO DEVELOPMENT
OF DECONVOLUTION OPERATOR
61
For continuous functions, the autocorrelation of an
input function, h, is defined by
00
¢hh(T) ~~ h(t)h(t+T)dt. -00
(A-1)
Changing the variable of integration with the substitutions
dx == dt,
~ _:1-_ 00 ' +oo -
one obtains 00
¢hh(T) :::: J( h(X-T)h(x)dX. -~
This may be written as 00 J h(:x)h[x+(-T)jdx == ¢hh(-T)
-00
which implies that
(A-2)
(A- 3)
(A- 4)
(A- 5)
so that the autocorrelation of a continuous function is an
even function.
Using the notation of Wainstein and Zubakov (1962), the
discrete autocorrelation is
¢gg(T) ~ g(t)g(t+T),
where the variables t and T are integers.
(A-6)
62
Since the process g is assumed to be time stationary, the
expected value required in equation (A-6) is independent
of time t so that
g(t)g(t+T) = g(O)g(T) (A- 7)
and
g(t)g(t+T) = g(-T)g(O) (A- 8)
which implies that
cjJ ( T) = </J (-T) (A- 9) gg gg
for discrete functions.
In order to obtain equation C 1~, first consider a
rectangular pluse defined by
u(t) = ~ c:m 2t2 :m) = 0 elsewhere
The autocorrelation of u(t) is given by
!¥-IT! <Puu(T) =J( u(t)u(t+T)dt
m 2
which integrates to
(-t <T<t ) m-- m
= 0 elsewhere.
(A -10)
(A-ll)
(A-12)
63
The autocorrelation of u(t) equals the autoconvolution of
u(t) since u(t) is symmetric. The Fourier transform of u(t)
lS
2
=Ff-m [
-j'ITft j'ITft J e m -e m
-j2'ITf
_rr sin 'ITftm =l~ 'IT£ Fm [si:f::tm J
(A- 13)
(A- 14)
Since convolution in the time domain corresponds to multi-
plication in the frequency domain, the Bartlett lag window
of equation (A-12) has as its Fourier transform the square
of the function in equation (A-14). That is,
W(f) = tm[sin 'ITftm ]2 'ITftm (A-15)
Equation (A-15) is identical to equation (17) in Chapter II.
An interesting property of the Fourier transform pair
given by equations (16) and (17) is that smoothing an
empirical autocorrelation function with equation (16) results
in non-negative spectral estimates. Consider the autocor-
relation function of a seismic trace g(t) represented by
00
¢(1:) =I g(t)g(t+T)dt (A-16) -00
¢ may be represented in a Fourier series as
64
7T
¢ ( T) = 2; f e j n T J ¢ ( x) e - j n x dx -do -7T
(A-17)
Since it is physically impossible to perform the summation
for doubly infinite values of n, the partial sum of this
series may be written as
N 7T
<P ( T ) "' 2; -~ e j n T I ¢ ( x) e - j nx dx
which, according to Wiener (1949), may be expressed as
7T
1 j sin [x (N+~)] -- ¢(x+T) 21T X
-7T sin Cz) dx .
This represents a weighted average of ¢(x) such that
7T
_1_/ sin [ x (N+~)] = 1 21T • (X)
-7T Sln z
(A-18)
(A-19)
(A- 2 0)
which implies that the total weight of all points is unity.
However, the weights can be negative as well as positive or
zero. Therefore, the Fourier coefficients, which are
actually lines in the periodogram of ¢, may become negative
by considering only a finite portion of the autocorrelation.
Since power is inherently positive the periodogram is not
considered to be a good estimate of the power spectrum.
Now consider the Cesaro partial sum of equation (A-17):
(A- 21)
The Cesaro method of partial summation is discussed in
Olmsted (1961) and other texts on advanced calculus.
65
Equation (A-21) can be written as
rr Nx 1 J sin2 2 2TIN ¢(x+T)
2 X
-rr sin 2 dx (A-22)
again according to Wiener (1949).
weighted average of <P(x) such that
This represents another
1 2TIN
7T Nx
f sin2 2 . 2 X
Sln Z -7T
dx = 1 , (A- 2 3)
which is of the same form as equation (A-20) except that the
weighting is non-negative. This implies that the estimate of
the Fourier coefficients of <P will be non-negative.
Therefore, assuming that the above arguments will with-
stand the transition from continuous to discrete autocorrela-
tion functions, one may "smooth" the discrete autocorrelation
function of equation (A-6) with the Bartlett lag window of
equation (A-12) to obtain the Cesaro approximation to <P given
by
The
$ gg ( T) " ~ gg ( T) ( 1-1 ~~ ) 0
IT I< t -m
discrete Fourier tm
......
~(f) = 2: T=-t m
transform of equation (A-24)
( 1 _tjmTj) e-j2TifT <Pgg(T)
which will necessarily be non-negative.
(A-24)
is then
(A-25)
66
In the design of a deconvolution operator one seeks a
set of filter weights a(t) which, when convolved with the
seismic trace, will yield a set of spikes as approximations
of reflection coefficients. This problem may be thought of
in more detail as follows: Find a filter such that its
convolution with a given wavelet will compress the wavelet
as nearly as possible into a spike. In symbols, consider a
spike m(t+s) and a wavelet y(t). Find the filter a(T) such
that N
X (t) = L a(T)y(t-T) (A-26)
T=l
will approximate m(t+s) with the smallest error. Thus, the
error N
s = x(t)-m(t+s) = L a(T)y(t-T)-m(t+s) (A- 2 7)
T=l
is to be made as small as possible (in the mean square sense).
Utilizing the overbar to denote expected value,
-- 2 s2 = x 2 (t)-2x(t)m(t+s)+m (t+s)
N N
= m2 (t+s)+ L L: a(T)a(n)y(t-T)y(t-n) T=l n=l
N -ZT~ a(T)m(t+s)y(t-T) .
Since the auto- and crosscorrelation functions
m2 (t+s) = ¢rnrn(O),
y(t-T)y(t-n) = ¢YY(T-n),
(A- 28)
(A- 29)
(A-30)
(A-31)
and
m(t+s)y(t-T) = ¢ (T+s) ym
67
(A-32)
are evident, one may simplify equation (A-29) to obtain
N = ¢mm (0) + L
T=l N
N L a(T)a(n)¢YY(T-n) n=l
-2 L ·a(T)¢ (T+s). T=l ym
(A-33)
The requirement of minimization of the mean squared error is
now accomplished by imposing the restriction
= 0 T=l, ... ,N
on equation (A-33) to obtain
N L a(n)¢YY(T-n) = <~>ym(T+s)
n=l
(A-34)
(A-35)
forT= 1, ... ,N. Realizing that the crosscorrelation of a
spike with the input wavelet yields a reversed wavelet, one
may express equation (A-35) in matrix form as
<PA = y (A-36)
where in this case, the reversed wavelet is the same as the
original because the wavelet is symmetric. Equation (A-36)
is identical to equation 08) of Chapter II.
APPENDIX B
SUBROUTINE TO COMPUTE DECONVOLUTION OPERATOR
SUBROUTINE TPLITZ(JM,R,GAMMA,A) DIMENSION R(SO),C(90),GAMMA(SO),CC(90),A(SO),AA(90) DOUBLE PRECISION A,R,GAMMA,C,CC,AA A(1)=GAMMA(1)/R(1) CC (1)=R(2)/R(1) JK=JM-1 DO 20 M=1,JK I=M+1 IF (M- 1) 12 , 12 , 7
7 SUMA=O.O SUMB=O.O DO 9 K=2,M SUMA=SUMA+C(K-1)*R(K) KK=K-1 MM=M-KK
9 SUMB=SUMB+C(KK)*R(MM+1) CC(1)=(R(I)-SUMA)/(R(1)-SUMB) DO 10 K=2 ,M NN=M-K CC(K)=C(K-1)-CC(1)*C(NN+1)
10 CONTINUE 12 SUMD=O.O
SUMC=O.O DO 13 K=1,M NN=M-K SUMC=SUMC+A(K)*R(NN+2) SUMD=SUMD+CC(K)*R(NN+2)
13 CONTINUE AA(I)=(GAMMA(I)-SUMC)/(R(1)-SUMD) DO 16 K=1,M AA(K)=A(K)-CC(K)*AA(I) A (K) =AA (K) C (K) =CC (K)
16 CONTINUE A(I}=AA(I)
20 CONTINUE RETURN END
69
APPENDIX C
LINEAR THRESHOLD ELEMENT DATA
70
Table V. Times of the central extremum, to, for events in the reference set. p = oo.
Event to Event to Number (ms.) Number (ms.)
1 213 26 1249
2 232 27 1265
3 247 28 1270
4 302 29 1307
5 390 30 1315
6 453 31 1455
7 564 32 1537
8 568 33 1569
9 600 34 1588
10 615 35 1606
11 668 36 1628
12 700 37 1635
13 809 38 1662
14 860 39 1682
15 887 40 1693
16 946 41 1715
17 1032 42 1746
18 1052 43 1759
19 1088 44 1790
20 1118 45 1797
21 1128 46 1856
22 1135 47 1934
23 1165 48 19 79
24 1201 49 1984
25 1246
71
Table V continued. p = 10.
Event to Event to Number (ms.) Number (ms.)
1 214 26 1249
2 231 27 1265
3 246 28 12 71
4 301 29 1307
5 390 30 1315
6 453 31 1455
7 564 32 1537
8 567 33 1569
9 601 34 1588
10 615 35 1607
11 669 36 1628
12 700 37 1634
13 809 38 1662
14 860 39 1682
15 887 40 1691
16 945 41 1715
17 1032 42 1747
18 1052 43 1759
19 1088 44 1791
20 1118 45 1797
21 1128 46 1856
22 1136 47 1933
23 1165 48 19 79
24 1200 49 1984
25 1246
72
Table V continued. p ::: 5.
Event t Event t Number (m~.) Number cm2.)
1 214 26 1249
2 231 27 1265
3 246 28 12 71
4 301 29 130 7
5 390 30 1315
6 453 31 1455
7 564 32 1537
8 567 33 1569
9 601 34 1588
10 615 35 1604
11 669 36 1628
12 700 37 1634
13 809 38 1662
14 860 39 1682
15 887 40 1691
16 945 41 1715
17 1032 42 1747
18 1052 43 1759
19 1088 44 1791
20 1118 45 1797
21 1128 46 1856
22 1136 47 1933
23 1165 48 1979
24 1200 49 1984
25 1246
73
Table V continued. p = 2.
Event to Event to Number (ms.) Number (ms.)
1 214 26 1249
2 231 27 1265
3 246 28 1269
4 301 29 1307
5 399 30 1315
6 453 31 1455
7 564 32 1537
8 569 33 1569
9 610 34 1588
10 615 35 1607
11 669 36 1628
12 700 37 1634
13 809 38 1662
14 860 39 1682
15 886 40 1691
16 945 41 1715
17 1031 42 174 7
18 1052 43 1759
19 1088 44 1791
20 1118 45 179 7
21 1128 46 1856
22 1134 47 1933
23 1165 48 1979
24 1200 49 1984
25 1246
Table VI-A. Times of the central extremum, to, and pattern vectors for events of interest. p = 00
Event to Mean Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance 1 213 106.6 105.8 0 0 34.1 2 232 103.9 98.0 0 0 32.4 3 247 106.2 105.4 0 0 34.7 4 453 109.6 108.9 0 0 36.0 5 564 121. 8 43.3 0 0 94.5 6 568 123.6 43.3 0 0 9 7. 2 7 809 109.6 108.6 0 0 36.0 8 860 108.5 66.8 0 0 80.0 9 1118 106.2 93.7 0 0 54.0
*10 1128 89.0 8. 5 10 10 88. 7 11 1135 153.1 99.6 6 1 104.4 12 1201 109.3 108.6 6 10 35.4 13 1246 130. 5 128.5 0 0 31. 7 14 1249 145.0 125.0 0 0 59.6 15 1265 142.3 51.9 0 0 89.6 16 1682 122.9 114.0 0 0 41.4
*17 1693 76.9 10.0 6 10 58.3 18 1715 103.6 101.9 1 10 32. 5 19 1934 107.6 94.1 0 0 54.4
'-l * Denotes a multiple event. .j::,.
Table VI-A continued.
Event to Mean Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance 20 19 79 164.3 101.9 3 1 106.9 21 1984 144.0 98.0 0 0 97.1 22 2207 107.3 107.8 0 0 34.4 23 2224 105.8 98.8 6 10 35.3
*24 2256 78.2 9.8 10 10 77.4 25 2270 105.3 101.9 10 1 3 7. 9 26 2581 108.4 106.2 6 10 35.4 27 2655 109.6 107.4 0 0 36.0 28 2745 117.9 110.1 0 0 41. 7 29 2756 119.1 109.7 0 0 43.1 30 2783 103.8 45.7 6 1 41.9 31 2791 103.4 81.2 0 0 38.1
*32 2819 93.9 5.6 6 10 79.4 33 2830 113.8 112.5 0 0 36.3 34 2848 101.9 94.1 0 0 36.0 35 3003 137.9 105.0 3 10 88.5 36 3008 164.4 81.2 0 0 116.7 37 3021 110.9 106.2 6 1 39.3
*38 3384 57. 7 15.8 10 10 26.2 * Denotes a multiple event
'-1 (Jl
Table VI-B. Times of the central extremum, to, and pattern vectors for events of interest. p = 10.
Event to Mean Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance
1 214 152.5 82.9 0 0 130.9 2 231 179.5 102.7 0 0 117.1 3 246 203.3 90.2 0 0 159.7 4 453 165.0 119.9 0 0 105.6 5 564 155.6 48.4 0 0 161.0 6 567 158.0 46.1 0 0 129.9 7 809 135.3 106.6 0 0 87.0 8 860 101.3 39.4 2 10 68.0 9 1118 151.6 0.0 0 0 138.9
*10 1128 144.6 0.0 10 10 175.7 11 1136 185.9 106.6 4 1 147.1 12 1200 163.5 120.7 4 10 104.9 13 1246 16 7. 3 135.5 0 0 90.9 14 1249 166.2 111. 7 0 0 86.2 15 1265 142.7 59. 7 0 0 89.2 16 1682 189.4 113.6 0 0 114.3
*17 1691 179.6 53.5 6 10 180.3 18 1715 145.4 110.5 1 10 86.9 19 1933 142.1 51.2 0 0 121.6
'-J * Denotes a multiple event. 0\
Table VI-B continued.
Event to Mean Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance 20 19 79 165.7 102.3 0 0 111. 8 21 1984 150.3 98.0 0 0 104.3 22 2207 155.7 88.6 0 0 113.1 23 2224 158.6 101.5 0 0 129.7
*24 2255 169.5 0.0 6 10 195.5 25 2270 137.8 109.3 4 1 82.4 26 2581 145.8 110.1 6 10 102.5 27 2655 160.1 112.5 0 0 122.4 28 2746 163.6 101.5 0 0 123.0 29 2756 163.4 116.0 0 0 100.0 30 2782 16 7. 2 86. 7 10 1 124.1 31 2791 154.4 0.0 0 0 139.8
*32 2820 195.3 0.0 10 10 212.6 33 2831 172.6 108.9 2 10 138.0 34 2848 171.3 58.6 0 0 160.7 35 3003 161.1 107.8 10 10 139.2 36 3009 184.7 80.8 6 10 137.5 37 3021 147.7 108.9 0 0 98. 2
*38 3384 235.0 0.0 10 10 207.8
* Denotes a multiple event. '-.1 '-.1
Table VI-C. Times of the central extremum, to, and pattern vectors for events of interest. p = 5.
Event to Mean. Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance
1 214 161.4 84.7 0 0 142.0 2 231 193.2 109.3 0 0 120.3 3 246 219.2 97.6 0 0 163.9 4 453 176.7 129.3 0 0 112.9 5 564 169.7 53.9 0 0 177.6 6 567. 171. 7 54.3 0 0 136.5 7 809 143.7 106.6 0 0 97.2 8 860 104.6 54.7 2 10 70.6 9 1118 162.7 0.0 0 0 151.0
*10 1128 163.2 0.0 10 10 193.0 11 1136 196.8 105.4 4 1 159.0 12 1200 176.9 125.0 4 10 113.7 13 1246 174.6 137.1 0 0 97.0 14 1249 173.8 120.3 0 0 91. 7 15 1265 150.9 65.2 0 0 99.8 16 1682 201.8 113.2 0 0 117.0
*17 1691 191.3 58.2 6 10 184.6 18 1715 154.0 115.2 3 1 90.9 19 1933 152.1 53.9 0 0 131.9 '-J
00 * Denotes a multiple event.
Table VI-C continued.
Event t Mean Peak Perio~ Pol- Var-Number (mg.) Freq. Freq. dicity arity iance
20 1979 166.8 108.2 0 0 112.3 21 1984 152.9 100.0 0 0 105.4 22 2208 166.3 80.8 0 0 125.2 23 2225 177.6 104.3 0 0 150.5
*24 2255 186.6 0.0 6 10 199.5 25 2270 157.7 115.2 4 1 101.5 26 2581 155.1 113.2 6 10 112.2 27 2655 172.4 116.8 0 0 134.6 28 2744 169.0 116.8 0 0 125.5 29 2756 176.0 119.9 0 0 110.0 30 2782 183.4 96.8 10 1 134.4 31 2791 166.6 0.0 0 0 153.3
*32 2 816 244.7 0.0 2 10 146.9 33 2831 178.0 105.0 2 10 149.1 34 2848 187.4 70.7 0 0 170.8 35 3003 168.7 105.8 10 10 149.5 36 3009 194.3 84. 7 6 10 144.0 37 3021 157.4 110.1 0 0 108.9
*38 3382 298.6 84.7 10 10 145.3
* Denotes a multiple event. '-.1 \.0
Table VI-D. Times of the central extremum, to, and pattern vectors for events of interest. P = 2.
Event to Mean Peak Pe rio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance
1 214 165.8 69.9 0 0 149.6 2 231 206.0 113.6 0 0 117.9 3 246 239.8 95.3 0 0 168.3 4 453 194.4 130.4 0 0 124.9 5 564 192.9 54.3 0 0 200.1 6 569 213.9 62.5 0 0 158.8 7 809 155.1 105.8 0 0 114.4 8 860 106.4 42.2 2 10 74.6 9 1118 174.6 0.0 0 0 168.3
*10 1128 187.3 0.0 10 10 211.1 11 1134 193.3 76.9 2 1 174.4 12 1200 193.3 132.0 2 10 121.5 13 1246 181.9 138.2 0 0 104.1 14 1249 180.8 117. 2 0 0 97.8 15 1265 162.0 62.9 0 0 117.9 16 1682 216.3 113.2 0 0 121.9
*17 1691 200.9 58.6 6 10 193.2 18 1715 160.0 116.4 1 10 87.5 19 1933 163.8 31. 2 0 0 147.2
00
* Denotes a multiple event. 0
Table VI-D continued.
Event to Mean Peak Perio- Pol- Var-Number (ms.) Freq. Freq. dicity arity iance 20 1979 168.6 105.4 0 0 117.6 21 1984 157.8 98.4 0 0 112.6 22 2208 178.5 0.0 0 0 138.5 23 2223 161.9 9 7. 6 0 0 119.3
*24 2255 201. 8 0.0 6 10 202.8 25 2270 163.8 113.2 4 1 105.3 26 2581 162.5 114.0 6 10 121.4 27 2654 191.7 119.5 2 10 153.0 28 2746 182.5 98.4 0 0 143.9 29 2756 190.3 117.9 0 0 129.3 30 2 782 206.4 94.1 10 1 145.3 31 2 791 178.5 0.0 4 1 171.4
*32 2820 214.9 0.0 10 10 199.8 33 2831 203.8 105.8 2 10 162.3 34 2848 202.7 28.1 3 1 184.1 35 3003 182.1 106.6 10 10 170.5 36 3009 20 7. 6 82.0 6 10 152.5 37 3021 169.6 109.3 0 0 120.9
*38 3382 249.4 70.3 10 10 182.4
* Denotes a multiple event. 00 1-1
APPENDIX D
LINEAR THRESHOLD ELEMENT DETAILS
A linear threshold element may be represented by
n
where x. 1 ' i = 1,
s = 2: i=l
,n is
x.w. 1 1
a pattern vector and w. 1
82
(D-1)
i = 1, ... ,n is a weighting vector such that the S value of
one class of patterns will be markedly different from the S
values of another class of patterns if, indeed, two classes
of patterns exist. Consider a total population of p patterns
and let the ith component of the kth pattern by represented
Then the S value for the kth pattern is given by
'
and the mean S value over the population is
p n
S = _1_ L L: x.kw. p k=l i=l 1 1 '
(D-2)
(D- 3)
which, upon interchanging the order of summation, may be
writ ten as
s =
or more simply,
n 2: i=l
p w.-
1- E 1 p k=l
n s = 2:
i=l x.w.
1 1
(D-4)
(D- 5)
where x. l
is the mean of the ith element of (x1
, 'X ) • n
83
If two separated classes exist in the n dimensional space,
the moment of inertia of the population will be a maximum
with respect to some axis within the space. Therefore, the
moment of inertia of the S line distribution, given by p
M L: k=l
(D-6)
will also become large if a weighting vector wi, i = 1, ... ,n
has forced the S values to be different for the two classes.
The value of M, however, can be made as large as desired
by allowing lw-1, i = 1, l
,n to increase without bound.
Hence, combining a constraint of the form
n 2 L: w.
i=l l min.
with equation (D-6) to yield n - 2 L (Sk-S)
k=l n A = L: i=l
w. l
2
(D- 7)
(D- 8)
results in a ratio, the maximization of which will tend to
maximize the moment of inertia while tending to minimize the
sum of the squares of w .. l
Before considering the maximization of A, an important
observation and some simplification of equation (D-8) are in
order. Utilizing equations (D-2) and (D-5), equation (D-6),
which is the numerator of equation (D-8), may be written as
p
M = L k=l [ ~ (x.k-x. )w. J . 1 l l l l=
84
(D-9)
Upon expansion, rearrangement of the summand, and inter
changing the order of summation, one obtains
If
then
n M= L
i=l
n L: j=l
p
p w.w. L (x.k-x.)(x.k-x.)
l J k=l l l J J
L (x.k-x.)(x.k-x.) =b .. k= 1 l l J J l J
n n
M = L i=l
L: j=l
w. b .. w. l lJ J
(D-10)
(D-11)
(D-12)
Here the b .. are elements of a matrix which is proportional to lJ
the sample covariance matrix of the pattern vectors xk
(xlk' ... ,xnk), k = 1, ... ,p.
In the light of equation (D-12), ratio (D-8) may be
written as n L i=l
Differentiating equation
n
L: j=l n 2: i=l
(D-13)
b .. w.w. lJ l J
2 w· l
(D-13)
with respect to each w., and 1
equating the result to ·zero will yield the conditions required
to maximize A. Accordingly,
ClA
aw. 1
= 0 (D-14)
85
implies that
n n n 2: 2: b .. w. -A 2: w. = o i=l j=l lJ 1 i=l l
(D-15)
Equation (D-15) corresponds in matrix form to the relation
T (B-AI)W = 0 (D-16)
where I is the identity matrix. Equation (D-16) can have a
non-trivial solution only if
IB-AII = 0 (D-17)
Thus, the problem of maximizing equation (D-8) for A is
reduced to the problem of finding the largest eigenvalue
of the matrix B, methods for which, being well-known, will
not be discussed here. The set of weights w., i = 1, ]_
which will produce this maximum is the eigenvector corres-
ponding to A where the eigenvectors are related to the max.
eigenvalues by
'n'
(D-18)
The matrix B is real and symmetric and therefore the
eigenvalues of B are necessarily real. A proof of this can
be found in DeRusso et al (1966).
The eigenvector W corresponding to the eigenvalue Amax.
is the set of weights which maximizes the moment of inertia
of the S line distribution while minimizing the sum of the
squares of w .. l
These weights, when used in equation (D-1),
86
result in a variable S such that two classes of patterns in
an n dimensional space are displayed as two clusters on the
S line.
87
VI. BIBLIOGRAPHY
ANSTEY, N.A. (1960): Attacking the problems of the synthetic
seismogram. Geophy. Prosp., 8, p.242-259.
(1964): Correlation techniques: a review. Geophy.
Prosp., 12, p.355-382.
and P. NEWMAN (1966): The sectional auto-correlogram
and the sectional retro-correlogram. Geophy. Prosp.,
14, p.389-426.
BARTLETT, M.S. (1948): Smoothing periodograms from time
series with continuous spectra. Nature, 161, p.686-687.
(1950): Periodogram analysis and continuous spectra.
Biometrika, 37, p.l-16.
BLACKMAN, R.B. and J.W. TUKEY (1959): The Measurement of
Power Spectra. New York, Dover Pub. 190 p.
COCHRAN, W.T. et al (1967): What is the fast Fourier trans
form? Inst. Elect. Electron. Engr. Trans., AU-15,
p.45-55.
COOLEY, J.W. and J.W. Tukey (1965): An algorithm for the
machine calculation of complex Fourier series. Math.
of Comp., 19, p.297-301.
COOLEY, J.W. et al (1967): Application of the fast Fourier
transform to computation of Fourier integrals, Fourier
series, and convolution integrals. Inst. Elect. Electron.
Engr. Trans., AU-15, p.79-84.
DERUSSO, P.M. et al (1966): State Variables for Engineers.
New York, John Wiley. p.239.
88
DIX, C.H. (1948): The existence of multiple reflections.
Geophy., 13, p.49-50.
ELLSWORTH, T.P. (1948): Multiple reflections. Geophy., 13,
p. 1-18.
EMBREE, P. et al (1963): Wide-band velocity filtering - the
pie-slice process. Geophy., 28, p. 948-974.
EPINAT'EVA, A.M. and L.A. IVANOVA (1959): High-frequency
filtering as a means of eliminating multiple reflec
tions. Izv. Acad. Sci. U.S.S.R., Geophy. Ser., p. 361-
371. (English translation by Amer. Geophy. Union, p.
244-252).
FISCHLER, M. et al (1962): An approach to general pattern
recognition. Trans. Inst. Rad. Engr., IT-8, p. 564-573.
FISHER, R.A. (1938): The use of multiple measurements in
taxonomic problems. Annals of Eugenics, 7, p. 179-188.
HANCOCK, J.C. and P.A. WINTZ (1966): Signal Detection Theory.
New York, McGraw-Hill. p.56.
HILEMAN, J.A. et al (1968): Automated static corrections.
30th meeting of the European Association of Explora
tion Geophysicists, Salzburg, Austria. 24 p.
IBM CORP. (1967a): Subroutine EIGEN. Srstem/360 Scientific
Subroutine Package (360A-CM-03X) Version II, Third
Edition. White Plains, N.Y., p. 56-57.
(1967b): Subroutine GAUSS. Srstem/360 Scientific
Subroutine ?ackage (360A-CM-03X) Version II, Third
Edition. White Plains, N.Y., p. 54.
89
JOHNSON, C.H. (1943): Discussion of "An analysis of abnormal
reflexions" by L.E. Deacon. Geophy., 8, p. 10-13.
(1948): Remarks regarding multiple reflections.
Geophy., 13, p. 19-26.
JONES, H.J. and J.A. MORRISON (1954): Cross-correlation
filtering. Geophy., 19, p. 660-683.
KERN, F.J. (1970): Personal Communication.
LEET, L.D. (1937): A plutonic phase in seismic prospecting.
Bul. Seis. Soc. Amer., 27, p. 97-98.
MARILL, T. and D.M. GREEN (1960): Statistical recognition
functions and the design of pattern recognizers. Trans.
Inst. Rad. Engr., EC-9, p. 472-477.
MATEKER, E.J. (1965): A Treatise on Some Aspects of Modern
Exploration Seismology. Okla. City, Pan American
Petroleum Corp., p. 32-33.
MATHIEU, R.L. and G.W. RICE (1969): Multivariate analysis
used in the detection of stratigraphic anomalies from
seismic data. Geophy., 34, p. 507-515.
MATTSON, R.L. and J.E. DAMMANN (1965): A technique for
determing and coding subclasses in pattern recog
nition problems. IBM Journ., 9, p. 294-302.
MEYER, P.L. (1965): Introductory Probability and Statistical
Applications. Reading, Mass., Addison-Wesley. p. 263-
264.
MINSKY, M. and S. PAPERT (1969): Perceptrons. Cambridge,
Mass., M.I.T. Press. 258 p.
90
OLHOVICH, V.A. (1964): The causes for noise in seismic re
flection and refraction work. Geophy., 29, p. 1015-1030.
OLMSTED, J.M.H. (1961): Advanced Calculus. New York,
Appleton-Century-Crofts. p. 540.
PARZEN, E. (1967): Time Series Analysis Papers. San Francisco,
Holden-Day. 565 p.
RICE, R.B. (1962): Inverse convolution filters. Geophy., 28,
p. 4-18.
RICE. S.O. (1944): Mathematical analysis of random noise.
Bell System Tech. Journ., 23, p. 282-332.
RICKER, N. (1945): The computation of output disturbances
from amplifiers for true wavelet inputs. Geophy., 10,
p. 207-220.
ROBINSON, E.A. and S. TREITEL (1964): Principles of digital
filtering. Geophy., 29, p. 395-404.
ROBINSON, E.A. (1966): Multichannel z-transforms and minimum
delay. Geophy., 31, p. 482-500.
(1967a): Predictive decomposition of time series
with application to seismic exploration. Geophy., 32,
p. 418-484.
(1967b): Multichannel Time Series Analysis with
Digital Computer Programs. San Francisco, Holden-Day.
298 p.
ROBINSON, J. C. (1968): HRVA - A velocity analysis technique
for seismic data. Ph.D. thesis, University of Missouri,
Rolla, Mo. 104 p.
91
ROBINSON, J.C. (1969): Personal communication.
ROSENBLATT, F. (1962): Principles of Neurodynamics. Washing
ton, D.C., Spartan Books. 616 p.
SCHNEIDER, W.A. et al (1964): A new data-processing technique
for the elimination of ghost arrivals on reflection
seismograms. Geophy., 29, p. 783-805.
SCHNEIDER, W.A. et al (1965): A new data-processing technique
for multiple attenuation exploiting differential normal
moveout. Geophy., 30, p. 348-362.
SEBESTYEN, G.S. (1962): Decision-Making Processes in Pattern
Recognition. New York, Macmillan Co. 159 p.
SLOAT, J. (1948): Identification of echo reflections. Geophy.,
13, p. 27-35.
STARK, L. et al (1962): Computer pattern recognition tech
niques: electrocardiographic diagnosis. Comm. Assoc.
Comp. Mach., 5, p. 527-532.
TREITEL, S. and E.A. ROBINSON (1966): Seismic wave propaga
tion in layered media in terms of communication theory.
Geophy., 31, p. 17-32.
(1969): Optimum digital filters for signal to noise
ratio enhancement. Geophys. Prosp., 17, p. 248-293.
TREITEL, S. et al (1967): Some aspects of fan filtering.
Geophy., 32, p. 789-800.
WADSWORTH, G.P. et al (1953): Detection of reflections on
seismic records by linear operators. Geophy., 18, p.
539-586.
92
WAINSTEIN, L.A. and V.D. ZUBAKOV (1962): Extraction of
Signals from Noise. Englewood Cliffs, N.J., Prentice
Hall. 382 p.
WELCH, P.D. (1967): The use of fast Fourier transform for
the estimation of power spectra: a method based on
time averaging over short, modified periodograms.
Inst. Elect. Electron. Engr. Trans., AU-15, p. 70-73.
WIENER, N. (1949): Extrapolation, Interpolation and
Smoothing of Stationary Time Series. Cambridge, Mass.,
M.I.T. Press. 128 p.
ZAREMBA, S.K. (1967): Quartic statistics in spectral analy
sis; in Spectral Analysis of Time Series, ed. by B.
Harris. New York, John Wiley. p. 47-79.
93
VI I. VITA
David Nuse Peacock is the son of Walter H. Peacock and
Helen Nuse Peacock. Born March 16, 1943 in Washington, D.C.,
he received his primary education at St. Peter's School and
his secondary education at Point Pleasant Beach High School,
both in Point Pleasant Beach, New Jersey. He obtained his
B.S. in Geology from the University of Missouri School of
Mines and Metallurgy, Rolla, in 1964 and his M.S. in Geophys
ical Engineering from the University of Missouri at Rolla in
1966. In September, 1966 he enrolled in the Mining and Petro
leum Engineering Department of the University of Missouri at
Rolla as a Ph.D. candidate in Geophysical Engineering. During
his graduate studies he was employed for two summers by Geo
physical Service, Inc. in New Orleans, La. and Dallas,Tex.
His studies were partly financed through a National Science
Foundation Traineeship and a Graduate Assistantship in the
Mining and Petroleum Engineering Department, where he was a
laboratory assistant in Geophysical Engineering. lle held
the position of Assistant Instructor in Mathematics at U.M.R.
for the academic year 1968-1969. He is a member of the Euro
pean Association of Exploration Geophysicists, American Asso
ciation for the Advancement of Science, New Jersey Academy of
Science, and is an associate member of the Society of Explor
ation Geophysicists. In June, 1969 he married the former
Donna Ruth Koch of Rolla, Missouri.
1.93936