You are on page 1of 4

Signature after predictive deconvolution

John F. Parrish*, Periseis Company


Summary
Conventional predictive deconvolution is very good for
suppressing normal incidence water bottom reverberations.
However, the output signature can vary significantly with
the value selected for the prediction distance (lag).
Generalizing predictive deconvolution with relative entropy
deconvolution
concepts
can
provide
consistent
dereverberation filters for lags shorter than the length of the
wavelet kernel.

pairs of (magnitude = 0.99) z-domain zeros at 155, 185 and


250 Hz. The dashed line in Figure 1 shows the resulting,
minimum phase, time signature response of this synthetic
FIR instrument.

Kernel

Instrument

0.7
0.6
0.5

Introduction

However, concepts of seismic deconvolution processing


have evolved. Merely shortening the interpretation wavelet
is no longer enough. In order to interpret rock properties, it
is necessary to know the interpretation wavelet shape and
to maintain its amplitude and phase spectrum throughout an
entire seismic volume.

Amplitude

0.4

Classic papers describing the behavior of water


reverberations (Backus, 1959) and the calculation of
predictive deconvolution filters (Peacock and Treitel, 1969)
have provided rules of thumb for conventional seismic
deconvolution processing. These rules were invaluable in
shortening field wavelets enough to allow structural
interpretation of the subsurface.

0.3
0.2
0.1
0
-0.1
-0.2
-0.3
0

10

20

30

Time-ms

Figure 1: FIR instrument and kernel signatures.

Theory
Burg (1975 & 1967) clearly enunciated the maximum
entropy formulation for seismic deconvolution. This
spiking or whitening version of deconvolution is very
effective but does not always provide a known output
signature. In addition, the filter shape can be very sensitive
to the selected value of the noise level constraint. Burgs
student, John E. Shore (1979, 1980, and 1981) generalized
maximum entropy into a consistent relative entropy
formulation for spectral estimation and zero-lag
deconvolution (see Parrish, 1997 and 1999).
Predictive deconvolution can also be generalized by
relative entropy concepts. The result is a consistent output
signature for interpretation that is independent of the value
selected for the prediction distance or lag.
Example
In order to illustrate these concepts, a synthetic digital
seismogram has been constructed entirely from minimum
phase components, sampled at 2 milliseconds. A finite
impulse response (FIR) instrument is simulated by three

NOTE: actual instrument and hydrophone responses were


not used in this example because they have infinite impulse
responses (IIR). However, with appropriate spectral energy
constraints and with a sufficiently long (designature)
compensation filter, it would be possible to convert an
actual seismic instruments IIR into a close approximation
of a specified FIR digital response like the one chosen for
this example.
Source and receiver ghosts are synthesized with sea surface
reflection coefficients of 0.99 and two-way travel times of
8 ms. The solid line in Figure 1 shows the combined finite
impulse response of both ghosts convolved with the
instrument (as if recorded with an ideal SEG standard
polarity hydrophone). This FIR wavelet kernel for the
synthetic seismogram has exactly 15 non-zero samples
between 0 ms and 28 ms. With a 14 ms shift, it would be
suitable as a nearly zero phase interpretation wavelet.
The earth model has a sea bottom with a two way travel
time of 80 ms and a reflection coefficient of 0.3. This will

SEG Int'l Exposition and 72nd Annual Meeting * Salt Lake City, Utah * October 6-11, 2002

Signature after predictive deconvolution

generate a normal incidence water bottom multiple


sequence with the time and amplitude series shown in the
first and second columns respectively of Table 1. At normal
incidence, each subsurface reflection will be convolved
with the reverberation series shown in column 3 of Table 1.
Multiple
0.00000
0.30000
-0.09000
0.02700
-0.00810
0.00243
-0.00073
0.00022
-0.00007
0.00002
-0.00001
0.00000
-0.00000

Reverberation
1.00000
-0.60000
0.27000
-0.10800
0.04050
-0.01458
0.00510
-0.00175
0.00059
-0.00020
0.00006
-0.00002
0.00001

0.8

0.2

-0.2
0

50

100

150

200

Time-ms

Figure 3: Predictive deconvolution filter for lag = 30 ms.

Figure 4 shows the filtered output between 1000 ms and


1400 ms. Not only is the reverberation suppressed, but the
reflections signature (the interpretation wavelet) is an
undistorted copy of the models FIR wavelet kernel. The
prediction error filters in this example are nearly exact for
lags between 30 and 50 ms. In practice, the filtered output,
including the reflections signature, is very good for lags up
to and including 80 ms.

Amplitude

In order to avoid any significant interference from the


water bottom multiple in this example, the synthetic
seismogram contains only a single isolated reflection at a
two way travel time of 1 second with a positive reference
amplitude of 1.0. A 2 seconds long, synthetic seismogram
record (compensated perfectly for spherical spreading) was
constructed from these various components. The time
interval from 1000 ms to 1400 ms is displayed in Figure 2.

Amplitude

0.4

Table 1: Multiple and reverberation time and


amplitude series.

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
1000

0.6

Amplitude

Time
0
80
160
240
320
400
480
560
640
720
800
880
960

filter for a noise level constraint of 0.001 (0.1%). This


predictive deconvolution filter is manifestly a bandpassed
version of Backuss (1959) dereverberation operator for
this synthetic seismogram.

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
1000

1100

1200

1300

1400

Time-ms

1100

1200

1300

1400

Time-ms

Figure 2: Synthetic seismogram.

Such a long field signature could obscure even a structural


interpretation of an earth seismogram. However, as
discussed earlier, a prediction operator should be able to
predict the (IIR) reverberation sequence. Figure 3 shows
the 190 ms (160 ms plus a lag of 30 ms) prediction error

Figure 4: Dereverberated output for lag = 30 ms.

For long or unknown field wavelets, Peacock and Treitel


(1969) suggested choosing a lag that corresponds to the
second zero crossing of the autocorrelation function. In
this example, the second zero crossing occurs between 14
ms and 16 ms. Figure 5 shows the prediction error filter for
a lag of 14 ms. The filter is distorted significantly by
including a portion of the autocorrelation of the FIR
wavelet kernel. Backuss dereverberation operator is no
longer recognizable.

SEG Int'l Exposition and 72nd Annual Meeting * Salt Lake City, Utah * October 6-11, 2002

Signature after predictive deconvolution

3
2

0.8

Amplitude

Amplitude

1
-1
-2
-3

0.4
0.2
0

-4
-5

-0.2
0

50

100

150

200

Figure 5: Predictive deconvolution filter for lag = 14 ms.

Nevertheless, the filtered output in Figure 6 shows that the


water bottom reverberation has indeed been suppressed.
Unfortunately, the reflections signature is distorted and no
longer nearly zero-phase. In addition, the signature rings
and extends beyond the lag time. This phase-rotated
interpretation wavelet could be adequate for most structural
interpretations but it might distort the interpretation of rock
properties and well log synthetics.

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
1000

1100

1200

50

100

150

200

Time-ms

Time-ms

Amplitude

0.6

1300

1400

Time-ms

Figure 6: Dereverberated output for lag = 14 ms.

The distortion of the dereverberation operator for short lags


can be avoided by reformulating predictive deconvolution.
Figure 7 shows a relative entropy predictive deconvolution
filter for a lag of 14 ms. The resulting filter is very close to
the classic predictive deconvolution filter for a lag of 30 ms
shown in Figure 3. Backuss dereverberation operator is
recognizable. The filtered output (not displayed) is
indistinguishable from that shown in Figure 4.

Figure 7: Relative entropy predictive deconvolution filter for


lag = 14 ms.

Conclusions
Classic papers describing the behavior of water
reverberations (Backus, 1959) and the calculation of
predictive deconvolution filters (Peacock and Treitel, 1969)
can provide rules of thumb for conventional seismic
deconvolution processing. By utilizing an example
seismogram synthesized with a finite impulse response,
wavelet kernel, these rules can be refined:
1. Predictive deconvolution can suppress reverberations
as long as the lag is less than or equal to the minimum
time of the water bottom reverberation sequence.
2. An isolated reflections signature is not distorted by
predictive deconvolution, as long as the lag is larger
than the length of the wavelet kernel.
3. The dereverberation filter changes shape as soon as
the lagged interval includes a significant portion of the
wavelet kernels autocorrelation.
4. Placing the lag at the second zero crossing is a
reasonable compromise but the reflections signature
will be distorted and it will extend beyond the lag.
Generalizing predictive deconvolution with relative entropy
deconvolution (Parrish, 1997 & 1999) concepts can provide
consistent dereverberation filters for lags shorter than the
length of the wavelet kernel. Actual field signatures can be
compensated to any convenient wavelet shape, including
those with infinite impulse responses, before applying a
relative entropy predictive deconvolution.
References
Backus, M. M., 1959, Water reverberationstheir nature
and elimination: Geophysics, v. 24, p. 233-261.

SEG Int'l Exposition and 72nd Annual Meeting * Salt Lake City, Utah * October 6-11, 2002

Signature after predictive deconvolution

Burg, J. P., 1975, Maximum entropy spectral analysis,


Ph.D. dissertation, Stanford University, Stanford, CA.
(University Microfilms No. 75-25, 499)
Burg, J. P., 1967, Maximum entropy spectral analysis,
Society of Exploration Geophysicists International
Exposition and 37th Annual Meeting.
Parrish, John F., 1997, Relative entropy spectrum
deconvolution, Society of Exploration Geophysicists
International Exposition and 67th Annual Meeting, Dallas.
Parrish, John F., 1999, Applying minimum relative entropy
spectrum deconvolution, Society of Exploration
Geophysicists International Exposition and 69th Annual
Meeting, Houston.
Peacock, K. L., and Treitel, S., 1969, Predictive
deconvolution: theory and practice: Geophysics, v. 34, p.
155-169.
Shore, John E., 1979, Minimum cross-entropy spectral
analysis, NRL-MR 3921, Naval Research Laboratory,
Washington, D. C., Jan.
Shore, J. E. and Johnson, R. W., 1980, Axiomatic
derivation of the principle of maximum entropy and the
principle of minimum cross-entropy, IEEE Trans. Inform.
Theory, IT-26, 26-37, Jan.
Shore, J. E., 1981, Minimum cross-entropy spectral
analysis, IEEE Trans. Acous. Speech Signal Processing,
ASSP-29, 230-237, Apr.
Shore, John E. and Johnson, Rodney W., 1983, Properties
of cross-entropy minimization, IEEE Trans. Inform.
Theory, IT-27, 472-482, July 1981. See also: comments
and corrections, IEEE Trans. Inform. Theory, IT-29, Nov.
Acknowledgments
Thank you Wulf Massell and EPIC Geophysical for
allowing me to investigate and test some of these concepts
on actual seismic records.

SEG Int'l Exposition and 72nd Annual Meeting * Salt Lake City, Utah * October 6-11, 2002

You might also like