Beruflich Dokumente
Kultur Dokumente
INVERSE
R.
B.
CONVOLUTION
VOL.
SSiII,
NO
1 (FEBKUAKY,
1962).
p,,
4 1X, 14 Fl(;%
FILTERS*
RICEt
Tte difficult problem of trying to locate stratigraphic traps with the reflection seismograph would be simplified
(at least in good record areas) if it were possible to perform the inverse of the reflection process, i.e., to divide out
the reflection wavelet of which the record is composed, leaving only the impulses representing the reflection coefficients. This process has been discussed by Robinson under the title predictive decomposition, but his approach
requires that the basic composition wavelet be a one-sided, damped, minimum-phase time function. Most seismic
wavelets which we observe or are accustomed to working with (e.g. the symmetric Ricker wavelet) are not of this
class. The purpose of this paper is to discuss a digital computer approach to the problem. Finite, bounded inverse
filter functions are obtained which will reduce seismic wave forms to best approximations to the unit impulse in the
least squares sense.The degree of approximation obtained depends upon the time length of the inverse filter. Inverse
filter functions of moderate length produce approximate unit impulses whose breadths are 50 percent or less than
those of the original wavelets. Hence, these filters will increase resolution well beyond the practical limits of instrumental filters. Their effectiveness is more or less sensitive to variations in the peak frequency and shape of the
composition wavelet, and to interference, depending upon individual conditions. Although this sensitivity problem
can be solved to some extent through the proper design of the inverse filter, it is aggravated by the usual lack of
knowledge about the form of the composition wavelet.
Robinson (1957) has treated the inverse filtering problem under the title predictive decomposition. Starting with a seismic trace or portion
thereof, he gives theoretical and statistical methods for computing (1) one form of the seismic
wavelet composing the trace and (2) the prediction operator or inverse wavelet for effecting the
contraction to a sequence of spikes. His techniques are based on discrete or digitized time
functions which are amenable to treatment on
digital computers. This approach has the advantage that completely arbitrary wavelet shapes or
impulsive responses, which may be quite expensive if not impossible to simulate electroniically, are handled with ease.
Robinson restricts his attention to cases in
which the basic composition wavelet is a onesided, damped, minimum-phase time function, a
mathematical definition of which will be given
later. The reason for imposing this restriction is
that these are the only wavelets for which
bounded, one-sided inverses exist. However, most
seismic wavelets which we observe or are accustomed to working with (e.g. the theoretical
Ricker wavelet) are not of this restricted class.
INTRODUCTION
The
problem
of trying
to locate
stratigraphic
seismograph
is a difficult
of detail
Hence,
or resolution
the problem
is the
would be simplified
(at least
inverse
divide
record
out
applying
wavelet
the
point
reflection
the reflection
is composed,
representing
another
of
the
leaving
reflection
of view,
a filter
process,
wavelet
which
only
the
From
to
the
impulses
coefficients.
to a spike representing
and amplitude
i.e.,
of which
of
each reflection
the arrival
time
of that reflection.
In his paper dealing in part with wavelet contraction, Ricker (1953) discusses this problem
from an instrumental point of view. He was able
to design electronic filters which will reduce conventional seismic wavelets to 70 or 80 percent of
their original breadth. The application of such
filters to seismograms only achieves a partial
transformation back to the reflection coefficient
function, but a considerable improvement in resolution is effected.
* Presentedat the 30th Annual SEC Meeting, Galveston, Texas, November 9, 1960. Manuscript receivedby the
Editor June 9, 1961.
t The Ohio Oil Company, Littleton, Colorado.
4
inverse
f(t)
g(t)
Convolution
-80)
-+z-/-
h(t)= -6 (11=
laf(r,g(t-r)
-v-
dr
-a,
FIG. 1. The problem of determining an inverse filter
function, g(t), which will transform a given wavelet,
f(t), into a negative unit impulse, --6(t).
OF
THE
co
S
MATHEMATICAL
f(7)g(t - TW,
--m
(1)
BACKGROUND
PROBLEM
h(t) =
Filteti
h(t) =
t
S
_f(T)& -
TW.
(1)
F(s) G(s),
(2)
where B(S), F(s), and G(s) are the Laplace transforms of h(t),f(t), and g(t), respectively. Since our
interest is in discrete representations, the Laplace
transforms may be replaced by the Dirichlet
series representations:
H(s) =
(3)
B,e-YcAt),
=I
I?. B. Rice
F(s) =
UCS,
v=l
G(s) =
x,e-vs,
and
(4)
v=l
H(s) =
b,ecYs.
v=l
F(u) =
avu,
(5)
=I
= F(u) G(u).
(6)
G(u) =
gsu,
s=-cc
which converges on the unit circle. Such an expansion is called two-sided in contrast to a
one-sided power series expansion. It is possible
in this instance, if the convergence is rapid
enough, to use a reasonable number of terms of
this two-sided expansion as the inverse wavelet,
but we shall not be concerned with this approach.
If F(u) has all its roots within the unit circle,
then the Laurent expansion reduces to the singular part
G(u) =
gsu,
Inverse
Convolution
Gn Ul +
L&f-1X?
a,
rz+
bm
zz
. +
a1
a2 LL +
a1
-rnL+1
a, xm+
xm-l
.xmtl
Filters
X,
=b
. .
a1
a,x,_,,+l
in addition to roots within the unit circle, F(u)
also has one or more roots on the unit circle, then
the above expansion for C(u) will not converge on
the unit circle, and hence it cannot in any way be
used as an inverse operator.
In any of the above cases, F(u) is non-minimum-phase (at least one root inside the unit
circle), and an attempted one-sided polynomial
division of H(u) by F(zt) will produce an unbounded inverse.
If, on the other hand, F(zL) has no roots on or
within the unit circle, it is called minimumphase. Then G(u) does have a convergent power
series expansion for / UI _<1, which can be used
for the inverse filter. This is the case treated by
Robinson (1957), as mentioned in the introduction.
In summary then, there is usually no one-sided
bounded inverse which will transform observed or
theoretical seismic wavelets into the unit impulse.
This problem may be examined from another
point of-view which may be more lucid and which
will form the basis for later remarks about the
solution. Again, if convolution is represented as
1 Filters, whether electronic or digital, are usually
termed realizable only when they permit one to work
in real time. However, if the filtering is done with
respectto nominal time (e.g. the time scaleon a recordedseismogram),then nonrealizable filters can be
used.Suchnonrealizable filters are in fact filters with
large time-delays.Usually, electronicfilters are usedfor
real time applicationsand digital computersfor nominal
time applications,although there are numerousexceptions.
x,-m+1
(7)
m+l
b,_,+l
b,,
...=
SQUARES
COMPUTATION
OF
INVERSE
FILTERS
R. B. Rice
AX-B==,
(8)
EE
= (AX
- B)(AX
= (XA
- B)
- B)(AX
= XAA
X
- B)
- XAB
(9)
- BAX
+ BB,
XAA
lJk
+ UAA
X
-
BAUk
UkAB
= 0 for all k,
(10)
(XAA
BA)u~
= Uk(-AAX
+ AB).
(10)
XAA
(XAA
BA)
= -
AAX
EE
= BB
(11)
AtA =
,:
+ AB
010
(LYMa,,-,,-1
or
BAX.
uo
where
AAX
= AB
(11)
which gives
m--k
o!k = c aiai+k,
k = 0, 1, . . . , 12-
X = (ilA)-AB
(12)
X = BA(AA)-I.
(13)
and
with o&= 0 for k > m- 1. Now the (Ykare the amplitudes of the autocorrelation of the original wavelet, with
CYo =
Substituting
m,
i=l
2
s-1
Ui2
Inverse
Convolution
aA,
2=1
AB
a&i+1
(16)
i=l
-a,
-a,-1 .
I
i,
(16)
I - a2m--m
I
the elements of which are the negative amplitudes
of the original wavelet in reverse order down to
the aZ,+,th one.
The above observations show that the normal
equations are easily formed from the amplitudes
of the original wavelet and its autocorrelation.
Hence, with a polynomial multiplication routine
to calculate the autocorrelation, least squares inverse filters can be computed using a standard
program for solving systems of linear equations.
A special case of interest is the one for which
the original wavelet is symmetric; m, n, an8
N( =n-m+
1) are odd; and the negative unit
impulse is placed at the $(n+l)st
point. Then,
for n>2m-1,
the right-hand vector (16) will be
symmetric around the center value -a;(,+i),
with zeros at the top and bottom, and the inverse
filter function will be symmetric. If the inverse
filter is taken to be the same length as the original
b,=-1),
there are
wavelet (N=m, n=2m-1,
no zeros at the upper right- and lower left-hand
corners of AA (15). Also the right-hand members
(16) consist precisely of the negative amplitudes
of the original wavelet with no additional zeros.
The case of symmetric wavelets and symmetric
inverses is the one dealt with most frequently in
the illustrations to follow. For this case, the m
normal equations (11) can be reduced to $(m+ 1)
independent equations, thereby greatly reducing
the computation time and the amount of computer storage required.
ILLUSTRATIVE
Figure
resulting
tive unit
wavelets
EXAMPLES
LEAST SQUARES
INVERSE
RICKER
WAVELET
Filters
APPROXIMATE
UNIT IMPULSE
-+lk.OI
75
cps
set
37.5 cps
FE. 3. Least squaresinversesand approximateunit impulsesfor 75 and 37.5 cps Ricker wavelets.
R. B. Rice
10
FREQUENCY
I%.
cps
of Figure 3.
The tails of the Ricker wavelets have been cut
off at a level of 0.002 percent of the maximum
amplitude. In each case, the number of points in
the inverse is the same as the number defining the
Ricker wavelet.
The resulting approximation to the negative
unit impulse for the 75 cps case is narrower and
has less ripple on either side of the center trough
than the one for the 37.5 cps case. However, it
should be noted that in each instance the center
trough of the Ricker wavelet has been reduced to
RICKER
WAVELET
LEAST SQUARES
INVERSE
APPROXIMATE
UNIT IMPULSE
-+lk50 cps
.OI
set
25 cps
lr
FIG. 5. Least squaresinversesand approximate unit impulsesfor 50 and 25 cps Ricker wavelets.
Inverse
Convolution
an
:
I
50 cps &
INVERSE
,
I
)I=
0
50
II
150
200
250
FREQUENCY, cps
FIG. 6. Amplitude spectraof 50 and 25 cps inverses
of Figure 5.
75
RICKER
WAVELET
cps
INVERSE
Filters
11
when the number of points in the inverse is increased from 17 to 25, then to 41. The improvement in the shape of the unit impulse approximation is quite striking although the breadth, which
is limited by the cut-off frequency, remains essentially constant. The spectra for these inverses
are presented in Figure 8. Sate that, as the inverse increases in length, the spectrum exhibits
an increasingly steeper slope starting at higher
and higher frequencies.
These results indicate that the least squares inverse filters can be made as nearly perfect as one
desires, except for the limitations imposed on the
size of systems of normal equations that can be
solved on todays digital computers. Experience
to date indicates that double-precision arithmetic
(18 or 20 decimal digits) is required for about
20th order or larger systems. Round-off error is
more troublesome than in most cases because
there are few or no zeros present in the coefficient
matrix.
APPROXIMATE
UNIT
IMPULSE
R. 8. Rice
12
I.Or
SPECTRA OF 75cps
INVERSES
.9c
/I
II
1,
;
17pt
25pt
41 pt
.7
?
7 6
:I
II
II
2
a .5
:
.
:;
!I
.._f,
200
150
100
50
FREQUENCY
FIG. 8.
II
1;
2;o
cps
mograms back to the reflection coefficient functions from which they were derived. The results
will be indicative of possibilities under ideal condictions; i.e., assuming that the basic composition
wavelet is known, is invariant with respect to
time and that there is no interference present.
Trace (c) of Figure 9 represents a reflection coefficient function obtained from an actual continuous velocity log on a Nebraska well, assuming
constant density. Trace (a) is the synthetic seismogram obtained by convolving the 75 cps symmetric Ricker wavelet of Figure 3 with Trace (c),
and Trace (b) is the result of applying the 75 cps
inverse of Figure 3 to Trace (a). The detailed
agreement between Traces (b) and (c) is excellent.
If one were able to do this well on an actual siesmogram, certainly there would be much less difficulty in making detailed, accurate stratigraphic
and lithologic interpretations from seismic records
However, in practice there are many complicating
factors which will be discussed below.
.Ol SEC.
RICKER
(c)-REFLECTION
FUNCTION
COEFFICIEN
(d)-Trace
(e) x INVERSE OF
37.5 cps RICKER WAVELET
(e)-
Inverse
Convolution
Trace (e) of Figure 9 is the synthetic seismogram obtained from the same reflection coefficient
function (c) using the 37.5 cps symmetric Ricker
wavelet of Figure 3 , and Trace (d) is the result
of filtering this trace with the 37.5 cps inverse.
In this case, some of the detail has been lost, but
the agreement with Trace (c) is still fairly good.
In Figure 10, similar results, based on the same
reflection coefficient (Trace (c)), are shown for the
cases of the 50 and 25 cps Ricker wavelets and
their inverses of Figure 5. The 50 cps inverse produces quite good agreement (Trace (b)) with the
original reflection coefficient function, but the 25
cps inverse does not restore much of the detail.
Better results can be obtained in any case by using
a longer inverse.
In the application of the inverse filtering technique to actual seismograms, there are a number
of factors which may significantly affect the results. These include variations in the frequency
and character of the composition wavelet, and
the presence of interference. Some of these effects
have been investigated synthetically to obtain
preliminary estimates of their significance and to
evaluate partially the possibilities of overcoming
the problems which they generate.
For the purpose of illustrations, an arbitrary
distinction will be drawn between variations in
frequency which preserve wavelet form and
changes in wavelet shape which leave the peak
frequency unaltered. This distinction may be difficult to find in practice where variable earth fdtering corresponding to different reflection times,
differential effects due to weathering changes, and
variations in shooting conditions will usually affect both peak frequency and wavelet shape.
However, the assumption of linearity permits
these additive effects to be studied individually.
I?ffect of Variations in Peak Frequency
Filters
13
ties (Figure 6). For example, the 10 cps component for the 50 cps inverse is five times the 50 cps
component. Hence, as the peak frequency of the
Ricker wavelet is decreased and the amount of
high-frequency content above 100 cps becomes
insignificant, the low-frequency portion of the inverse spectrum becomes dominant.
On the other hand, if the peak frequency of the
wavelet is greater than that on which the inverse
is based, as in the two cases on the right side of
Figure 11, the additional high frequencies are
amplified too much.
The over-all results indicate that the effect of
variations in frequency of the order of + 10 percent is not serious. If the inverse is based on a peak
frequency which is too high, the resolution will
suffer proportionately. If the peak frequency is
too low, the amount of ripple will increase.
E$ect of Variations
in Wavelet Shape
Next consider variations in wavelet shape corresponding to changes in the form of the amplitude and/or phase spectra of the composition
wavelet which leave the peak frequency unaltered
Obviously, there is no end to the number of different kinds of variations that could be considered
However, space permits the illustration of two
types.
Figure 12 shows the approximate unit impulses
obtained by applying the 17-point, 75 cps inverse
of Figure 3 to symmetric 75 cps Ricker wavelets
with 17, 13, and 11 nonzero values, respectively.
It will be noted that this increase in the sharpness of cut-off of the tails of the wavelet has no
adverse effect whatever. This is a desirable property of the least squares computation technique
which would not hold for some other inverse filter
calculation methods.
In Figure 13, the traces on the right are the
result of convolving the slightly asymmetric
wavelets on the left with the inverses (Figures
3 and 5) for the corresponding symmetric Ricker
wavelets. The asymmetric wavelets have been
obtained by adding a saw-tooth function to the
symmetric Ricker wavelets with the amplitude of
positive and negative peaks of the saw-tooth
function being about 4 percent of the maximum
trough amplitude of the Ricker wavelet. Note
that the effect of the asymmetry is very slight in
the 75 cps case, but that it increases with decreasing frequency, becoming completely overriding in the 25 cps case. This is explained by the
R. B. Rice
14
.Ol SEC.
-ww(a) -Trace
(c)
WAVELET
x 50 cps
RICKER
(b)-Trace
(a) x INVERSE
OF
50 cps RICKER
WAVELET
(c)-REFLECTION
FUNCTION
coEFFiciENT
id)-Trace
(e) x INVERSE
OF
25 cps RICKER WAVELET
(e)-Trace
(c) x 25
WAVELET
cps
RICKER
FIG. 10. Results of applying 50 and 25 cps inverses of Figure 5 to synthetic seismograms
computed from a CVL on a Nebraska well.
x 44 cps R W 4
f-
x56.!5cpsRdb
V
n
FIG. 11. Convolutions of 50 cps inverse of Figure 5 with Ricker wavelets of different peak frequencies.
Inverse
75 cps RICKER
Convolution
Filters
15
APPROXIMATE
INVERSE
WAVELET
.Oi
17-Point
SEC.
UNIT
IMPULSE
I--
--T--
I3-Point
I:
I I-Point
v
I
1;
FIG. 12. Effect of approximateunit impulse for 75 cps caseof cutting off tails of Ricker wavelet
fact that the saw-tooth function added in each
case contains a larger proportion of high-frequency
components than the Ricker wavelets. These components are not significant in the extreme highfrequency range where the peak of the 7.5 cps inverse spectrum occurs, but are increasingly differentially amplified by the lower frequency inverses.
Most types of variations in wavelet shape would
not add such high-frequency components and
hence would not be so troublesome. If the varied
forms of the wavelet are known, the inverse
filter can be designed to circumvent the difficulties
as much as possible, although the amount of resolution effected may have to be comprised. The
biggest problem, however, is that of determining
the form of the composition wavelet. The standard technique based on the power spectrum of the
autocorrelation of the assumed stationary time
of Random Noise
Ft. B. Rice
16
MODIFIED
RICKER
WAVELETS
75
cps
50
cps
CONVOLUTIONS
OF WAVELETS
ON LEFT WITH
INVERSES OF SYMMETRIC
RICKER
WAVELETS
OF SAME FREQUENCY
Inverse
Convolution
SYNTHETIC
SEISMOGRAMS
TRACE
Filters
17
LEFT-HAND TRACES
X50 cps INVERSE FILTER
TRACE I
+25 cps NOISE
TRACE I
4-50 cps NOISE
TRACE I
3-75 cps NOISE
FIG. 14. Effect of random noiseof different frequencieson resultsof applying 50 cpsinverse of
Figure 5 to 50 cps synthetic seismogram.
high enough and the inverse filter long enough,
this requirement can usually be fulfilled. If this is
not accomplished, the results, as in the lower
right-hand trace, will be self-explanatory.
It should be noted that significant reading errors introduced in digitizing seismograms will
have the same effect as high-frequency interference. Hence, it is necessary to develop some
procedure for digitizing traces with sufficient accuracy. Commercial oscillographic readers are
usually employed.
SUMMARY
AND
CONCLUSIONS
seismograms well beyond the limits that are practical with instrumental filters. It has been shown
that the effectiveness of these filters is more or less
sensitive to variations in the peak frequency and
shape of the wavelet composing the seismogram,
and to interference. This sensitivity can be overcome to some extent by designing the filter so that
it will not amplify unwanted frequencies. In many
instances, the amount of resolution effected will
have to be compromised with the sensitivity problem, which is aggravated by the usual lack of
knowledge about the form of the composition
wavelet. Nevertheless, the method appears to
hold enough promise as a new seismic interpretation tool in the search for stratigraphic traps to
warrant further study of its application to field
records.
R. B. Rice
18
ACKNOWLEDGMENTS
The author
Mr.
is greatlv
R. L. Massey,
the work
thank
indebted
to his assistant,
on this paper.
Dr. E. A. Robinson
He would
also like
to
reading
of the manuscript.
REFERENCES
Ba Hli, F., 1954, .4 general method for time domain network synthesis: Transactions of the IRE Professional Group on Circuit Theory, v. CT-l, n. 3, p.
21-28.