2 Deconvolution

• Introduction • The Convolutional Model • The Convolutional Model in the Time Domain • The Convolu-
tional Model in the Frequency Domain • Inverse Filtering • The Inverse of the Source Wavelet • Least-Squares
Inverse Filtering • Minimum Phase • Optimum Wiener Filters • Spiking Deconvolution • Prewhitening • Wavelet
Processing by Shaping Filters • Predictive Deconvolution • Predictive Deconvolution in Practice • Operator
Length • Prediction Lag • Percent Prewhitening • Effect of Random Noise on Deconvolution • Multiple Atten-
uation • Field Data Examples • Prestack Deconvolution • Signature Deconvolution • Vibroseis Deconvolution
• Poststack Deconvolution • The Problem of Nonstationarity • Time-Variant Deconvolution • Time-Variant
Spectral Whitening • Frequency-Domain Deconvolution • Inverse Q Filtering • Deconvolution Strategies • Exer-
cises • Appendix B: Mathematical Foundation of Deconvolution • Synthetic Seismogram • The Inverse of
the Source Wavelet • The Inverse Filter • Frequency-Domain Deconvolution • Optimum Wiener Filters • Spiking
Deconvolution • Predictive Deconvolution • Surface-Consistent Deconvolution • Inverse Q Filtering • References

2.0 INTRODUCTION

Deconvolution compresses the basic wavelet in the
recorded seismogram, attenuates reverberations and
short-period multiples, thus increases temporal resolu-
tion and yields a representation of subsurface reflec-
tivity. The process normally is applied before stack;
however, it also is common to apply deconvolution
to stacked data. Figure 2.0-1 shows a stacked section
with and without deconvolution. Deconvolution has
produced a section with a much higher temporal res-
olution. The ringy character of the stack without de-
convolution limits resolution, considerably.
Figure 2.0-2 shows selected common-midpoint
(CMP) gathers from a marine line before and after de-
convolution. Note that the prominent reflections stand
out more distinctly on the deconvolved gathers. Decon-
volution has removed a considerable amount of ringy-
ness, while it has compressed the waveform at each of
the prominent reflections. The stacked sections associ-
ated with these CMP gathers are shown in Figure 2.0-
3. The improvement observed on the deconvolved CMP
gathers also are noted on the corresponding stacked sec-
tion.
Figure 2.0-4 shows some NMO-corrected CMP
gathers from a land line with and without deconvo-
lution. Corresponding stacked sections are shown in
Figure 2.0-5. Again, note that deconvolution has com-
pressed the wavelet and removed much of the reverber-
ating energy.
Deconvolution sometimes does more than just
wavelet compression; it can remove a significant part
of the multiple energy from the section. Note that the
stacked section in Figure 2.0-6 shows a marked improve-
160 Seismic Data Analysis

FIG. 2.0-1. Interpreters prefer the crisp, finely detailed appearance of the deconvolved section (right) as opposed to the
blurred, ringy appearance of the section without deconvolution (left). (Data courtesy Enterprise Oil.)

ment between 2 and 4 s after deconvolution.
To understand deconvolution, first we need to ex-
amine the constituent elements of a recorded seismic
trace (Section 2.1). The earth is composed of layers of
rocks with different lithology and physical properties.
Seismically, rock layers are defined by the densities and
velocities with which seismic waves propagate through
them. The product of density and velocity is called seis-
mic impedance. The impedance contrast between adja-
cent rock layers causes the reflections that are recorded
along a surface profile. The recorded seismogram can
be modeled as a convolution of the earth’s impulse re-
sponse with the seismic wavelet. This wavelet has many
components, including source signature, recording fil-
ter, surface reflections, and receiver-array response. The
earth’s impulse response is what would be recorded if
the wavelet were just a spike. The impulse response
comprises primary reflections (reflectivity series) and all
possible multiples.
Ideally, deconvolution should compress the wavelet
components and eliminate multiples, leaving only the
earth’s reflectivity in the seismic trace. Wavelet com-
pression can be done using an inverse filter as a de-
convolution operator. An inverse filter, when convolved
with the seismic wavelet, converts it to a spike (Section
2.2). When applied to a seismogram, the inverse filter
should yield the earth’s impulse response. An accurate
inverse filter design is achieved using the least-squares
method (Section 2.2).
The fundamental assumption underlying the de-
convolution process (with the usual case of unknown
source wavelet) is that of minimum phase. This issue is
dealt with also in Section 2.2.
The optimum Wiener filter, which has a wide
range of applications, is discussed in Section 2.3. The
Wiener filter converts the seismic wavelet into any de-
sired shape. For example, much like the inverse filter,
a Wiener filter can be designed to convert the seismic
 Deconvolution 161

FIG. 2.0-2. Note the prominent reflections on the deconvolved gathers (b). The reverberations would make it difficult to
distinguish prominent reflections on the undeconvolved gathers (a).
162 Seismic Data Analysis
FIG. 2.0-3. (a) The section obtained from the undeconvolved gathers of Figure 2.0-2a, and (b) the section obtained from
the deconvolved gathers of Figure 2.0-2b.
wavelet into a spike. However, the Wiener filter differs
from the inverse filter in that it is optimal in the least-
squares sense. Also, the resolution (spikiness) of the
output can be controlled by designing a Wiener pre-
diction error filter — the basis for predictive deconvo-
lution (Section 2.3). Converting the seismic wavelet into
a spike is like asking for a perfect resolution. In prac-
tice, because of noise in the seismogram and assump-
tions made about the seismic wavelet and the recorded
seismogram, spiking deconvolution is not always desir-
able. Finally, the prediction error filter can be used to
remove periodic components — multiples, from the seis-
mogram. Practical aspects of predictive deconvolution
are presented in Section 2.4, and field data examples are
provided in Section 2.5. Finally, time-varying aspects of
the source waveform — nonstationarity, are discussed in
Section 2.6.
The mathematical treatment of deconvolution is
found in Appendix B. However, several numerical ex-
amples, which provide the theoretical groundwork from
a heuristic viewpoint, are given in the text. Much of the
early theoretical work on deconvolution came from the
MIT Geophysical Analysis Group, which was formed in
the mid-1950s.
2.1 THE CONVOLUTIONAL MODEL
A sonic log segment is shown in Figure 2.1-1a. The
sonic log is a plot of interval velocity as a function of
depth based on downhole measurement using logging
tools. Here, velocities were measured between the 1000-
to 5400-ft depth interval at 2-ft intervals. The veloc-
ity function was extrapolated to the surface by a lin-
ear ramp. The sonic log exhibits a strong low-frequency
component with a distinct blocky character represent-
ing gross velocity variations. Actually, it is this low-
frequency component that normally is estimated by ve-
locity analysis of CMP gathers (Section 3.2).
In many sonic logs, the low-frequency component
is an expression of the general increase of velocity with
depth due to compaction. In some sonic logs, however,
the low-frequency component exhibits a blocky charac-
ter (Figure 2.1-1a), which is due to large-scale lithologic
variations. Based on this blocky character, we may de-
fine layers of constant interval velocity (Table 2-1), each
of which can be associated with a geologic formation
(Table 2-2).
 Deconvolution 163

FIG. 2.0-4. Some NMO-corrected gathers associated with the stacked sections in Figure 2.0-5, (a) before, (b) after deconvo-
lution. Deconvolution has removed the ringy character from the data.

Table 2-1. The interval velocity trend obtained from Table 2-2. Stratigraphic identification associated with
the sonic log in Figure 2.1-1a. layering described in Table 2-1.

Layer Layer Number Lithologic Unit

Number Interval Velocity,* ft/s Depth Range, ft
1 Limestone
1 21 000 1000 − 2000 2 Shaly limestone with gradual
2 19 000 2000 − 2250 increase in shale content
3 18 750 2250 − 2500 3 Shaly limestone
4 12 650 2500 − 3775 4 Sandstone
5 19 650 3775 − 5400 5 Dolomite

*The velocity in Layer 2 gradually decreases from the

top of the layer to the bottom.
164 Seismic Data Analysis

FIG. 2.0-5. Deconvolution helps distinguish prominent reflections with ease (b). However, on a section without deconvolution
(a), reflections are buried in reverberating energy. Selected CMP gathers for both sections are shown in Figure 2.0-4.

The sonic log also has a high-frequency component
superimposed on the low-frequency component. These
rapid fluctuations can be attributed to changes in rock
properties that are local in nature. For example, the
limestone layer can have interbeddings of shale and
sand. Porosity changes also can affect interval velocities
within a rock layer. Note that well-log measurements
have a limited accuracy; therefore, some of the high-
frequency variations, particularly those associated with
a first arrival that is strong enough to trigger one re-
ceiver but not the other in the log tool (cycle skips), are
not due to changes in lithology.
Well-log measurements of velocity and density pro-
vide a link between seismic data and the geology of the
substrata. We now explain the link between log mea-
surements and the recorded seismic trace provided by
seismic impedance — the product of density and veloc-
ity. The first set of assumptions that is used to build
the forward model for the seismic trace follows:
 Deconvolution
FIG. 2.0-6. CMP stacks with (a) no deconvolution before
stack, (b) spiking deconvolution before stack. Deconvolution
can remove a significant amount of multiple energy from
seismic data. (Data courtesy Elf Aquitane and partners.)
Assumption 1. The earth is made up of horizontal
layers of constant velocity.
Assumption 2. The source generates a compres-
sional plane wave that impinges on layer bound-
aries at normal incidence. Under such circum-
stances, no shear waves are generated.
Assumption 1 is violated in both structurally com-
plex areas and in areas with gross lateral facies changes.
Assumption 2 implies that our forward model for the
seismic trace is based on zero-offset recording — an un-
realizable experiment. Nevertheless, if the layer bound-
aries were deep in relation to cable length, we assume
that the angle of incidence at a given boundary is small
and ignore the angle dependence of reflection coeffi-
165
cients. Combination of the two assumptions thus imply
a normal-incidence one-dimensional (1-D) seismogram.
Based on assumptions 1 and 2, the reflection coeffi-
cient c (for pressure or stress), which is associated with
the boundary between, say, layers 1 and 2, is defined as
I2 − I1
c= , (2 − 1a)
I2 + I1
where I is the seismic impedance associated with each
layer given by the product of density ρ and compres-
sional velocity v.
From well-log measurements, we find that the ver-
tical density gradient often is much smaller than the
vertical velocity gradient. Therefore, we often assume
that the impedance contrast between rock layers is es-
sentially due to the velocity contrast, only. Equation
(2-1a) then takes the form:
v2 − v1
c= . (2 − 1b)
v2 + v1
If v2 is greater than v1 , the reflection coefficient would
be positive. If v2 is less than v1 , then the reflection
coefficient would be negative.
The assumption that density is invariant with
depth or that it does not vary as much as velocity is
not always valid. The reason we can get away with it
is that the density gradient usually has the same sign
as the velocity gradient. Hence, the impedance function
derived from the velocity function only should be cor-
rect within a scale factor.
For vertical incidence, the reflection coefficient is
the ratio of the reflected wave amplitude to the incident
wave amplitude. Moreover, from its definition (equation
2-1a), the reflection coefficient is seen as the ratio of
the change in acoustic impedance to twice the average
acoustic impedance. Therefore, seismic amplitudes as-
sociated with earth models with horizontal layers and
vertical incidence (assumptions 1 and 2) are related to
acoustic impedance variations.
The reflection coefficient series c(z), where z is the
depth variable, is derived from sonic log v(z) and is
shown in Figure 2.1-1b. We note the following:
(a) The position of each spike gives the depth of the
layer boundary, and
(b) the magnitude of each spike corresponds to the
fraction of a unit-amplitude downward-traveling
incident plane wave that would be reflected from
the layer boundary.
To convert the reflection coefficient series c(z) (Fig-
ure 2.1-1b) derived from the sonic log into a time series
c(t), select a sampling interval, say 2 ms. Then use the
velocity information in the log (Figure 2.1-1a) to con-
vert the depth axis to a two-way vertical time axis. The
166 Seismic Data Analysis
 Deconvolution
FIG. 2.1-2. A seismic source wavelet after onset takes the
form shown at top left. As the wavelet travels into the earth,
the amplitude level drops (geometric spreading) and a loss
of high frequencies occurs (frequency absorption).
result of this conversion is shown in Figure 2.1-1c, both
as a conventional wiggle trace and as a variable area
and wiggle trace (the same trace repeated six times to
highlight strong reflections). The reflection coefficient
series c(t) (Figure 2.1-1c) represents the reflectivity of a
series of fictitious layer boundaries that are separated by
an equal time interval — the sampling rate (Goupillaud,
1961). The major events in this reflectivity series are
from the boundary between layers 2 and 3 located at
about 0.3 s, and the boundary between layers 4 and 5
located at about 0.5 s.
The reflection coefficient series (Figure 2.1-1c) that
was constructed is composed only of primary reflec-
tions (energy that was reflected only once). To get a
complete 1-D response of the horizontally-layered earth
model (assumption 1), multiple reflections of all types
(surface, intrabed and interbed multiples) must be in-
cluded. If the source were unit-amplitude spike, then the
recorded zero-offset seismogram would be the impulse
response of the earth, which includes primary and mul-
tiple reflections. Here, the Kunetz method (Claerbout,
1976) is used to obtain such an impulse response. The
impulse response derived from the reflection coefficient
series in Figure 2.1-1c is shown in Figure 2.1-1d with
the variable area and wiggle display.
The characteristic pressure wave created by an im-
pulsive source, such as dynamite or air gun, is called the
signature of the source. All signatures can be described
as band-limited wavelets of finite duration — for exam-
ple, the measured signature of an Aquapulse source in
Figure 2.1-2. As this waveform travels into the earth,
its overall amplitude decays because of wavefront diver-
gence. Additionally, frequencies are attenuated because
of the absorption effects of rocks (see Section 1.4). The
progressive change of the source wavelet in time and
167
depth also is shown in Figure 2.1-2. At any given time,
the wavelet is not the same as it was at the onset of
source excitation. This time-dependent change in wave-
form is called nonstationarity.
Wavefront divergence is removed by applying a
spherical spreading function (Section 1.2). Frequency
attenuation is compensated for by the processing tech-
niques discussed in Section 2.6. Nevertheless, the simple
convolutional model discussed here does not incorporate
nonstationarity. This leads to the following assumption:
Assumption 3. The source waveform does not
change as it travels in the subsurface — it is sta-
tionary.
The Convolutional Model in
the Time Domain
A convolutional model for the recorded seismogram
now can be proposed. Suppose a vertically propagat-
ing downgoing plane wave with source signature (Figure
2.1-3a) travels in depth and encounters a layer bound-
ary at 0.2-s two-way time. The reflection coefficient as-
sociated with the boundary is represented by the spike
in Figure 2.1-3b. As a result of reflection, the source
wavelet replicates itself such that it is scaled by the re-
flection coefficient. If we have a number of layer bound-
aries represented by the individual spikes in Figures 2.1-
3b through 2.1-3f, then the wavelet replicates itself at
those boundaries in the same manner. If the reflection
coefficient is negative, then the wavelet replicates itself
with its polarity reversed, as in Figure 2.1-3c.
Now consider the ensemble of the reflection coeffi-
cients in Figure 2.1-3g. The response of this sparse spike
series to the basic wavelet is a superposition of the in-
dividual impulse responses. This linear process is called
the principle of superposition. It is achieved computa-
tionally by convolving the basic wavelet with the reflec-
tivity series (Figure 2.1-3g). The convolutional process
already was demonstrated by the numerical example in
Section 1.1.
The response of the sparse spike series to the basic
wavelet in Figure 2.1-3g has some important character-
istics. Note that for events at 0.2 and 0.35 s, we iden-
tify two layer boundaries. However, to identify the three
closely spaced reflecting boundaries from the composite
response (at around 0.6 s), the source waveform must be
removed to obtain the sparse spike series. This removal
process is just the opposite of the convolutional process
used to obtain the response of the reflectivity series to
the basic wavelet. The reverse process appropriately is
called deconvolution.
168 Seismic Data Analysis

FIG. 2.1-3. A wavelet (a) traveling in the earth repeats itself when it encounters a reflector along its path (b, c, d, e, f). The
left column represents the reflection coefficients, while the right column represents the response to the wavelet. Amplitudes
of the response are scaled by the reflection coefficient. The resulting seismogram (bottom right) represents the composite
response of the earth’s reflectivity (bottom left) to the wavelet (top right).

The principle of superposition now is applied to the
impulse response derived from the sonic log in Figure
2.1-1d. Convolution of a source signature with the im-
pulse response yields the synthetic seismogram shown
in Figure 2.1-4. The synthetic seismogram also is shown
in Figure 2.1-1e. This 1-D zero-offset seismogram is free
of random ambient noise. For a more realistic represen-
tation of a recorded seismogram, noise is added (Figure
2.1-4).
The convolutional model of the recorded seismo-
gram now is complete. Mathematically, the convolu-
tional model illustrated in Figure 2.1-4 is given by
x(t) = w(t) ∗ e(t) + n(t), (2 − 2a)
 Deconvolution
FIG. 2.1-4. The top frame is the same as in Figure 2.1-1d.
The asterisk denotes convolution. The recorded seismogram
(bottom frame) is the sum of the noise-free seismogram and
the noise trace. This figure is equivalent to equation (2-2a).
where x(t) is the recorded seismogram, w(t) is the ba-
sic seismic wavelet, e(t) is the earth’s impulse response,
n(t) is the random ambient noise, and ∗ denotes con-
volution. Deconvolution tries to recover the reflectivity
series (strictly speaking, the impulse response) from the
recorded seismogram.
An alternative to the convolutional model given by
equation (2-2a) is based on a surface-consistent spec-
tral decomposition (Taner and Coburn, 1981). In such
a formulation, the seismic trace is decomposed into the
convolutional effects of source, receiver, offset, and the
earth’s impulse response, thus explicitly accounting for
variations in wavelet shape caused by near-source and
near-receiver conditions and source-receiver separation.
The following equation describes the surface-consistent
convolutional model (Section B.8):
xij (t) = sj (t) ∗ hl (t) ∗ ek (t) ∗ gi (t) + n(t), (2 − 2b)
169
where xij (t) is a model of the recorded seismogram,
sj (t) is the waveform component associated with source
location j, gi (t) is the component associated with re-
ceiver location i, and hl (t) is the component associated
with offset dependency of the waveform defined for each
offset index l = |i−j|. As in equation (2-2a), ek (t) repre-
sents the earth’s impulse response at the source-receiver
midpoint location, k = (i + j)/2. By comparing equa-
tions (2-2a) and (2-2b), we infer that w(t) represents
the combined effects of s(t), h(t), and g(t).
The assumption of surface-consistency implies that
the basic wavelet shape depends only on the source
and receiver locations, not on the details of the ray-
path from source to reflector to receiver. In a transition
zone, surface conditions at the source and receiver lo-
cations may vary significantly from dry to wet surface
conditions. Hence, the most likely situation where the
surface-consistent convolutional model may be applica-
ble is with transition-zone data. Nevertheless, the for-
mulation described in this section is the most accepted
model for the 1-D seismogram.
The random noise present in the recorded seismo-
gram has several sources. External sources are wind mo-
tion, environmental noise, or a geophone loosely coupled
to the ground. Internal noise can arise from the record-
ing instruments. A pure-noise seismogram and its char-
acteristics are shown in Figure 2.1-5. A pure random-
noise series has a white spectrum — it contains all the
frequencies. This means that the autocorrelation func-
tion is a spike at zero lag and zero at all other lags. From
Figure 2.1-5, note that these characteristic requirements
are reasonably satisfied.
Now examine the equation for the convolutional
model. All that normally is known in equation (2-2a) is
x(t) — the recorded seismogram. The earth’s impulse
response e(t) must be estimated everywhere except at
the location of wells with good sonic logs. Also, the
source waveform w(t) normally is unknown. In certain
cases, however, the source waveform is partly known;
for example, the signature of an air-gun array can be
measured. However, what is measured is only the wave-
form at the very onset of excitation of the source array,
and not the wavelet that is recorded at the receiver.
Finally, there is no a priori knowledge of the ambient
noise n(t).
We now have three unknowns — w(t), e(t), and
n(t), one known — x(t), and one single equation (2-2a).
Can this problem be solved? Pessimists would say no.
However, in practice, deconvolution is applied to seismic
data as an integral part of conventional processing and
is an effective method to increase temporal resolution.
To solve for the unknown e(t) in equation (2-2a),
further assumptions must be made.
Assumption 4. The noise component n(t) is zero.
170 Seismic Data Analysis
FIG. 2.1-5. A random signal with infinite length has a flat
amplitude spectrum and an autocorrelogram that is zero
at all lags except the zero lag. The discrete random series
with finite length shown here seems to satisfy these require-
ments. What distinguishes a random signal from a spike
(1, 0, 0, . . .)?
Assumption 5. The source waveform is known.
Under these assumptions, we have one equation,
x(t) = w(t) ∗ e(t). (2 − 3a)
and one unknown, the reflectivity series e(t). In reality,
however, neither of the above two assumptions normally
is valid. Therefore, the convolutional model is examined
further in the next section, this time in the frequency
domain, to relax assumption 5.
If the source waveform were known (such as the
recorded source signature), then the solution to the de-
convolution problem is deterministic. In Section 2.2, one
such method of solving for e(t) is considered. If the
source waveform were unknown (the usual case), then
the solution to the deconvolution problem is statistical.
The Wiener prediction theory (Section 2.3) provides one
method of statistical deconvolution.
The Convolutional Model in
the Frequency Domain
The convolutional model for the noise-free seismogram
(assumption 4) is represented by equation (2-3a). Con-
volution in the time domain is equivalent to multiplica-
tion in the frequency domain (Section A.1). This means
that the the amplitude spectrum of the seismogram
equals the product of the amplitude spectra of the seis-
mic wavelet and the earth’s impulse response (Section
B.1):
Ax (ω) = Aw (ω)Ae (ω), (2 − 3b)
where Ax (ω), Aw (ω), and Ae (ω) are the amplitude spec-
tra of x(t), w(t), and e(t), respectively.
Figure 2.1-6 shows the amplitude spectra (top row)
of the impulse response e(t), the seismic wavelet w(t),
and the seismogram x(t). The impulse response is the
same as that shown in Figure 2.1-1d. The similarity in
the overall shape between the amplitude spectrum of
the wavelet and that of the seismogram is apparent. In
fact, a smoothed version of the amplitude spectrum of
the seismogram is nearly indistinguishable from the am-
plitude spectrum of the wavelet. It generally is thought
that the rapid fluctuations observed in the amplitude
spectrum of a seismogram are a manifestation of the
earth’s impulse response, while the basic shape is asso-
ciated primarily with the source wavelet.
Mathematically, the similarity between the ampli-
tude spectra of the seismogram and the wavelet suggests
that the amplitude spectrum of the earth’s impulse re-
sponse must be nearly flat (Section B.1). By examin-
ing the amplitude spectrum of the impulse response in
Figure 2.1-6, we see that it spans virtually the entire
spectral bandwidth. As seen in Figure 2.1-5, a time se-
ries that represents a random process has a flat (white)
spectrum over the entire spectral bandwidth. From close
examination of the amplitude spectrum of the impulse
response in Figure 2.1-6, we see that it is not entirely
flat — the high-frequency components have a tendency
to strengthen gradually. Thus, reflectivity is not entirely
a random process. In fact, this has been observed in the
spectral properties of reflectivity functions derived from
a worldwide selection of sonic logs (Walden and Hosken,
1984).
We now study the autocorrelation functions (mid-
dle row, Figure 2.1-6) of the impulse response, seismic
wavelet, and synthetic seismogram. Note that the au-
tocorrelation functions of the basic wavelet and seismo-
gram also are similar. This similarity is confined to lags
for which the autocorrelation of the wavelet is nonzero.
Mathematically, the similarity between the autocorrelo-
gram of the wavelet and that of the seismogram suggests
that the impulse response has an autocorrelation func-
tion that is small at all lags except the zero lag (Section
B.1). The autocorrelation function of the random series
in Figure 2.1-5 also has similar characteristics. How-
ever, there is one subtle difference. When compared,
Figures 2.1-5 and 2.1-6 show that autocorrelation of
 Deconvolution 171

FIG. 2.1-6. Convolution of the earth’s impulse response (a) with the wavelet (b) (equation 2-2a) yields the seismogram (c)
(bottom row). This process also is convolutional in terms of their autocorrelograms (middle row) and multiplicative in terms
of their amplitude spectra (top row). Assumption 6 (white reflectivity) is based on the similarity between autocorrelograms
and amplitude spectra of the impulse response and wavelet.

the impulse response has a significantly large negative from the autocorrelation of the seismogram. For this
lag value following the zero lag. This is not the case type of deconvolution, Assumption 5, which is almost
for the autocorrelation of random noise. The positive never met in reality, is not required. But first, we need
peak (zero lag) followed by the smaller negative peak in to review the fundamentals of inverse filtering.
the autocorrelogram of the impulse response arises from
the spectral behavior discussed above. In particular, the
positive peak and the adjacent, smaller negative peak of 2.2 INVERSE FILTERING
the autocorrelogram together nearly act as a fractional
derivative operator (Section A.1), which has a ramp ef- If a filter operator f (t) were defined such that convolu-
fect on the amplitude spectrum of the impulse response tion of f (t) with the known seismogram x(t) yields an
as seen in Figure 2.1-6. estimate of the earth’s impulse response e(t), then
The above observations made on the amplitude
spectra and autocorrelation functions (Figure 2.1-6) im- e(t) = f (t) ∗ x(t). (2 − 4)
ply that reflectivity is not entirely a random process. By substituting equation (2-4) into equation (2-3a), we
Nonetheless, the following assumption almost always is get
made about reflectivity to replace the statement made
in assumption 5. x(t) = w(t) ∗ f (t) ∗ x(t). (2 − 5)
When x(t) is eliminated from both sides of the equation,
Assumption 6. Reflectivity is a random process. the following expression results:
This implies that the seismogram has the charac- δ(t) = w(t) ∗ f (t), (2 − 6)
teristics of the seismic wavelet in that their auto-
correlations and amplitude spectra are similar. where δ(t) represents the Kronecker delta function:
1, t = 0,
This assumption is the key to implementing the pre- δ(t) = (2 − 7)
0, otherwise.
dictive deconvolution. It allows the autocorrelation of
By solving equation (2-6) for the filter operator
the seismogram, which is known, to be substituted for
f (t), we obtain
the autocorrelation of the seismic wavelet, which is un-
known. In Section 2.3, we shall see that as a result of 1
f (t) = δ(t) ∗ . (2 − 8)
assumption 6, an inverse filter can be estimated directly w(t)
172 Seismic Data Analysis

the inverse filter, F (z), is obtained by polynomial di-

vision of the z-transform of the input wavelet, W (z),
given by equation (2-9) (Section A.2):
1 1 1
F (z) = = 1 + z + z2 + · · · · (2 − 11)
1 − 21 z 2 4
The coefficients of F (z) : (1, 12 , 41 , . . .) represent the
time series associated with the filter operator f (t). Note
that the series has an infinite number of coefficients, al-
though they decay rapidly. As in any filtering process,
in practice the operator is truncated.
FIG. 2.2-1. A flowchart for inverse filtering. First consider the first two terms in equation (2-
11) which yield a two-point filter operator (1, 12 ). The
Therefore, the filter operator f (t) needed to compute design and application of this operator is summarized in
the earth’s impulse response from the recorded seismo- Table 2-3. The actual output is (1, 0, − 14 ), whereas the
gram turns out to be the mathematical inverse of the ideal result is a zero-delay spike (1, 0, 0). Although not
seismic wavelet w(t). Equation (2-8) implies that the in- ideal, the actual result is spikier than the input wavelet,
verse filter converts the basic wavelet to a spike at t = 0. (1, − 12 ).
Likewise, the inverse filter converts the seismogram to a Can the result be improved by including one more
series of spikes that defines the earth’s impulse response. coefficient in the inverse filter? As shown in Table
Therefore, inverse filtering is a method of deconvolution, 2-4, the actual output from the three-point filter is
provided the source waveform is known (deterministic (1, 0, 0, − 18 ). This is a more accurate representation of
deconvolution). The procedure for inverse filtering is de- the desired output (1, 0, 0, 0) than that achieved with
scribed in Figure 2.2-1. the output from the two-point filter (Table 2-3). Note
that there is less energy leaking into the nonzero lags of
the output from the three-point filter. Therefore, it is
spikier. As more terms are included in the inverse filter,
The Inverse of the Source Wavelet
the output is closer to being a spike at zero lag. Since
the number of points allowed in the operator length is
Computation of the inverse of the source wavelet is ac- limited, in practice the result never is a perfect spike.
complished mathematically by using the z-transform
(Section A.2). For example, let the basic wavelet be
a two-point time series given by w(t) : (1, − 21 ). The
z-transform of this wavelet is defined by the following Table 2-3. Design and application of the truncated in-
polynomial: verse filter (1, 21 ) with the input wavelet (1, − 12 ).
1 Filter Design
W (z) = 1 − z. (2 − 9)
2 Input Wavelet w(t) : (1, − 12 )
The power of variable z is the number of unit time de- The z−Transform W (z) = 1 − 12 z
lays associated with each sample in the series. The first The Inverse F (z) = 1 + 21 z + 41 z 2 + · · ·
term has zero delay, so z is raised to zero power. The The Inverse Filter f (t) : (1, 12 , 41 , · · ·)
second term has unit delay, so z is raised to first power.
Hence, the z-transform of a time series is a polynomial Filter Application
in z, whose coefficients are the values of the time sam- Truncated Inverse Filter (1, 12 )
ples. Input Wavelet (1, − 21 )
A relationship exists between the z-transform and Actual Output (1, 0, − 14 )
the Fourier transform (Section A.2). The z-variable is Desired Output (1, 0, 0)
defined as Convolution Table:
z = exp −iω∆t , (2 − 10) 1 − 21 Output

where ω is angular frequency and ∆t is sampling inter- 1

1 1
2
val. 1
1 0
2
The convolutional relation in the time domain 1
1 − 14
2
given by equation (2-8) means that the z-transform of
 Deconvolution 173

Table 2-4. Design and application of the truncated in- Table 2-5. Design and application of the truncated in-
verse filter (1, 21 , 41 ) with the input wavelet (1, − 12 ). verse filter (−2, −4) with input wavelet (− 21 , 1).

Filter Design Filter Design

Input Wavelet w(t) : (1, − 21 ) Input Wavelet w(t) : (− 12 , 1)
The z−Transform W (z) = 1 − 12 z The z−Transform W (z) = − 12 + z
The Inverse F (z) = 1 + 21 z + 41 z 2 + ··· The Inverse F (z) = −2 − 4z − 8z 2 − · · ·
The Inverse Filter f (t) : (1, 21 , 14 , · · ·) The Inverse Filter f (t) : (−2, −4, −8, · · ·)

Filter Application Filter Application

Truncated Inverse Filter (1, 21 , 14 ) Truncated Inverse Filter (−2, −4)
Input Wavelet (1, − 21 ) Input Wavelet (− 12 , 1)
Actual Output (1, 0, 0, − 81 ) Actual Output (1, 0, −4)
Desired Output (1, 0, 0, 0) Desired Output (1, 0, 0)
Convolution Table: Convolution Table:
1 − 21 Output − 21 1 Output

1 1 −4 −2 1
1 1
4 2 −4 −2 0
1 1 −4 −2 −4
4 2 1 0

1 1
4 2 1 0
Table 2-6. Design and application of the truncated in-
1
4
1
2 1 − 81 verse filter (−2, −4, −8) with input wavelet (− 21 , 1).

Filter Design
Input Wavelet w(t) : (− 12 , 1)
The inverse of the input wavelet w(t) : (1, − 12 ) The z−Transform W (z) = − 12 + z
has coefficients that rapidly decay to zero (equation The Inverse F (z) = −2 − 4z − 8z 2 − · · ·
2-11). What about the inverse of the input wavelet The Inverse Filter f (t) : (−2, −4, −8, · · ·)
w(t) : (− 21 , 1)? Again, define the z−transform:
Filter Application
1 Truncated Inverse Filter (−2, −4, −8)
W (z) = − + z. (2 − 12)
2 Input Wavelet (− 12 , 1)
The z−transform of its inverse is given by the polyno- Actual Output (1, 0, 0, −8)
mial division: Desired Output (1, 0, 0, 0)
1 Convolution Table:
F (z) = 1 = −2 − 4z − 8z 2 − · · · (2 − 13) − 21 1 Output
−2 + z
As a result, the inverse filter coefficients are given by −8 −4 −2 1
the divergent series f (t) : (−2, −4, −8, · · ·). Truncate −8 −4 −2 0
this series and convolve the two-point operator with the −8 −4 −2 0
input wavelet (− 12 , 1) as shown in Table 2-5. The actual −8 −4 −2 −8
output is (1, 0, −4), while the desired output is (1, 0, 0).
Not only is the result far from the desired output, but
also it is less spiky than the input wavelet (− 21 , 1). The
reason for this poor result is that the inverse filter co- Least-Squares Inverse Filtering
efficients increase in time rather than decay (equation
2-13). When truncated, the larger coefficients actually A well-behaved input wavelet, such as (1, − 12 ) as op-
are excluded from the computation. posed to (− 12 , 1), has a z-transform whose inverse can
If we kept the third coefficient of the inverse filter be represented by a convergent series. Then the inverse
in the above example (equation 2-13), then the actual filtering described above yields a good approximation to
output (Table 2-6) would be (1, 0, 0, −8), which also is a zero-lag spike output (1, 0, 0). Can we do even better
a bad approximation to the desired output (1, 0, 0, 0). than that?
174 Seismic Data Analysis

Formulate the following problem: Given the input Table 2-7. Design and application of a two-term least-
wavelet (1, − 21 ), find a two-term filter (a, b) such that squares inverse filter (a, b).
the error between the actual output and the desired
output (1, 0, 0) is minimum in the least-squares sense. Filter Design
Compute the actual output by convolving the filter Convolution of the filter (a, b) with input wavelet
(a, b) with the input wavelet (1, − 12 ) (Table 2-7). The (1, − 21 ):
cumulative energy of the error L is defined as the sum 1 − 21 Actual Output Desired Output
of the squares of the differences between the coefficients
of the actual and desired outputs: b a a 1
2 a 2 b 2 b a b − a/2 0
L= a−1 + b− + − . (2 − 14)
2 2 b a −b/2 0
The task is to find coefficients (a, b) so that L takes Filter Application
its minimum value. This requires variation of L with Least-Squares Filter (0.95, 0.38)
respect to the coefficients (a, b) to vanish (Section Input Wavelet (1, −0.5)
B.5). By simplifying equation (2-14), taking the par- Actual Output (0.95, −0.09, −0.19)
tial derivatives of quantity L with respect to a and b, Desired Output (1, 0, 0)
and setting the results to zero, we get
5
a − b = 2, (2 − 15a)
2
and Table 2-8. Error in two-term inverse and least-squares
5 filtering.
−a + b = 0. (2 − 15b)
2 Input: (1, − 21 )
We have two equations and two unknowns; namely, the Desired Output: (1, 0, 0)
filter coefficients (a, b). The so-called normal set of equa-
tions (2-15a) and (2-15b) can be put into the following Actual Output Error
convenient matrix form Energy

5/2 −1 a 2
= . (2 − 16) Inverse Filter (1, 0, −0.25) 0.063
−1 5/2 b 0
By solving for the filter coefficients, we obtain (a, b) : Least-Squares Filter (0.95, −0.09, −0.19) 0.048
(0.95, 0.38). Design and application of this least-squares
inverse filter are summarized in Table 2-7.
To quantify the spikiness of this result and compare
it with the result from the inverse filter in Table 2-3, Table 2-9. Design and application of a two-term least-
compute the energy of the errors made in both (Table squares inverse filter (a, b).
2-8). Note that the least-squares filter yields less error
when trying to convert the input wavelet (1, − 12 ) to a Filter Design
spike at zero lag (1, 0, 0). Convolution of filter (a, b) with input wavelet (− 21 , 1):
We now examine the performance of the least- Actual Output Desired
squares filter with the input wavelet (− 12 , 1). Note that − 21 1 Output
the inverse filter produced unstable results for this
wavelet (Table 2-5). We want to find a two-term fil- b a −a/2 1
ter (a, b) that, when convolved with the input wavelet b a −b/2 + a 0
(− 21 , 1), yields an estimate of the desired spike out-
b a b 0
put (1, 0, 0) (Table 2-9). As before, the least-squares
Filter Application
error between the actual output and the desired output
Least-Squares Filter (−0.95, −0.19)
should be minimal.
Input Wavelet (−0.5, 1)
The cumulative energy of the error is given by
Actual Output (0.24, −0.38, −0.19)
a 2 b 2 Desired Output (1, 0, 0)
L= − −1 + − +a + b2 . (2 − 17)
2 2
 Deconvolution 175

Table 2-10. Error in two-term inverse and least- Table 2-11. Design and application of a two-term least-
squares filtering. squares inverse filter (a, b).

Input: (− 12 , 1) Filter Design

Desired Output: (1, 0, 0) Convolution of filter (a, b) with input wavelet (1, − 21 ):
Actual Output Error Actual Output Desired
Energy
1 − 12 Output
Inverse Filter (1, 0, −4) 16
b a a 0
Least-Squares Filter (0.24, −0.38, −0.19) 0.762 b a b − a/2 1
b a −b/2 0
Filter Application
By simplifying equation (2-17), taking the partial Least-Squares Filter (−0.09, 0.76)
derivatives of quantity L with respect to a and b, and Input Wavelet (1, − 21 )
setting the results to zero, we obtain Actual Output (−0.09, 0.81, −0.38)
5 Desired Output (0, 1, 0)
a − b = −1, (2 − 18a)
2
and
5 Table 2-12. Error in least-squares filtering.
−a + b = 0. (2 − 18b)
2
Input Wavelet: (1, − 21 )
Combine equations (2-18a,b) into a matrix form
Desired Output Actual Output Error
5/2 −1 a −1
= . (2 − 19) Energy
−1 5/2 b 0
By solving for the filter coefficients, we obtain (a, b) : (1, 0, 0) (0.95, −0.09, −0.19) 0.048
(−0.95, −0.19). The design and application of this filter
are summarized in Table 2-9.
(0, 1, 0) (−0.09, 0.81, −0.38) 0.190
Table 2-10 shows the results from the inverse filter
and least-squares filter quantified. The error made by
the least-squares filter is, again, much less than the error
made by the truncated inverse filter. However, both fil- By simplifying equation (2-20), taking the partial
ters yield larger errors for input wavelet (− 12 , 1) (Table
derivatives of quantity L with respect to a and b, and
2-10) as compared to errors for wavelet (1, − 12 ) (Table
2-8). The reason for this is discussed next. setting the results to zero, we obtain
5
a − b = −1, (2 − 21a)
2
Minimum Phase and
5
Two input wavelets, wavelet 1: (1, − 12 ) and wavelet 2: −a + b = 2. (2 − 21b)
(− 12 , 1), were used for numerical analyses of the inverse 2
filter and least-squares inverse filter in this section. The Combine equations (2-21a,b) into a matrix form
results indicate that the error in converting wavelet 1
to a zero-lag spike is less than the error in converting 5/2 −1 a −1
= . (2 − 22)
wavelet 2 (Tables 2-8 and 2-10). −1 5/2 b 2
Is this also true when the desired ouput is a de-
layed spike (0, 1, 0)? The cumulative energy of the error By solving for the filter coefficients, we obtain (a, b)
L associated with the application of a two-term least- : (−0.09, 0.76). The design and application of this filter
squares filter (a, b) (Table 2-11) to convert the input are summarized in Table 2-11.
wavelet (1, − 21 ) to a delayed spike (0, 1, 0) is Table 2-12 shows the results of the least-sqaures
a 2 b 2 filtering to convert the input wavelet (1, − 21 ) to zero-
L = a2 + b− −1 + − . (2 − 20) lag (Table 2-7) and delayed spikes (Table 2-11). Note
2 2
176 Seismic Data Analysis

Table 2-13. Design and application of a two-term least- Table 2-14. Error in least-squares filtering.
squares inverse filter (a, b).
Input Wavelet: (− 21 , 1)
Filter Design
Desired Output Actual Output Error
Convolution of filter (a, b) with input wavelet (− 12 , 1):
Energy
Actual Output Desired
− 21 1 Output (1, 0, 0) (0.24, −0.38, 0.19) 0.762

b a −a/2 0 (0, 1, 0) (−0.38, 0.81, −0.09) 0.190

b a −b/2 + a 1
b a b 0
Filter Application Now, evaluate the results of the least-squares in-
Least-Squares Filter (0.76, −0.09) verse filtering summarized in Tables 2-12 and 2-14.
Input Wavelet (−0.5, 1) Wavelet 1: (1, − 21 ) is closer to being a zero-delay spike
Actual Output (−0.38, 0.81, −0.09) (1, 0, 0) than wavelet 2: (− 21 , 1). On the other hand,
Desired Output (0, 1, 0) wavelet 2 is closer to being a delayed spike (0, 1, 0) than
wavelet 1. We conclude that the error is reduced if the
desired output closely resembles the energy distribution
that the input wavelet is converted to a zero-lag spike in the input series. Wavelet 1 has more energy at the
with less error, and the corresponding actual output onset, while wavelet 2 has more energy concentrated at
more closely resembles a zero-lag spike desired output. the end.
We now examine the performance of the least- Figure 2.2-2 shows three wavelets with the same
squares filter with the input wavelet(− 12 , 1). The cu- amplitude spectrum, but with different phase-lag spec-
mulative energy of the error L associated with the ap- tra. As a result, their shapes differ. (From Section 1.1,
plication of a two-term least-squares filter (a, b) (Table we know that the shape of a wavelet can be altered
2-13) to convert the input wavelet (− 12 , 1) to a delayed by changing the phase spectrum without modifying the
spike (0, 1, 0) is amplitude spectrum.) The wavelet on top has more en-
ergy concentrated at the onset, the wavelet in the mid-
a 2 b 2
dle has its energy concentrated at the center, and the
L= − + − + a − 1 + b2 . (2 − 23)
2 2 wavelet at the bottom has most of its energy concen-
By simplifying equation (2-23), taking the partial trated at the end.
derivatives of quantity L with respect to a and b, and We say that a wavelet is minimum phase if its en-
setting the results to zero, we obtain ergy is maximally concentrated at its onset. Similarly,
a wavelet is maximum phase if its energy is maximally
5
a − b = 2, (2 − 24a) concentrated at its end. Finally, in all in-between situa-
2
tions, the wavelet is mixed phase. Note that a wavelet is
and defined as a transient waveform with a finite duration —
5 it is realizable. A minimum-phase wavelet is one-sided
−a + b = −1. (2 − 24b)
2 — it is zero before t = 0. A wavelet that is zero for
Combine equations (2-24a,b) into a matrix form t < 0 is called causal. These definitions are consistent
with intuition — physical systems respond to an excita-
5/2 −1 a 2 tion only after that excitation. Their response also is of
= . (2 − 25)
−1 5/2 b −1 finite duration. In summary, a minimum-phase wavelet
By solving for the filter coefficients, we obtain (a, b) is realizable and causal.
: (0.76, −0.09). The design and application of this filter These observations are quantified by consider-
are summarized in Table 2-13. ing the following four, three-point wavelets (Robinson,
Table 2-14 shows the results of the least-squares 1966):
filtering to convert the input wavelet (− 21 , 1) to zero-
lag (Table 2-9) and delayed spikes (Table 2-13). Note Wavelet A : (4, 0, −1)
that the input wavelet is converted to a delayed spike Wavelet B : (2, 3, −2)
with less error, and the corresponding actual output Wavelet C : (−2, 3, 2)
more closely resembles a delayed spike desired output. Wavelet D : (−1, 0, 4)
 Deconvolution 177

FIG. 2.2-3. A quantitative analysis of the minimum- and

maximum-phase concept. The fastest rate of energy build-
up in time occurs when the wavelet is minimum-phase (A).
The slowest rate occurs when the wavelet is maximum-phase
(D).

up is significantly different for each wavelet. For exam-

ple, with wavelet A, the energy builds up rapidly close
to its total value at the very first time lag. The energy
for wavelets B and C builds up relatively slowly. Finally,
the energy accumulates at the slowest rate for wavelet
D. From Figure 2.2-3, note that the energy curves for
FIG. 2.2-2. A wavelet has a finite duration. If its energy
wavelets A and D form the upper and lower boundaries.
is maximally front-loaded, then it is minimum-phase (top).
If its energy is concentrated mostly in the middle, then it
Wavelet A has the least energy delay, while wavelet D
is mixed-phase (middle). Finally, if its energy is maximally has the largest energy delay.
end-loaded, then the wavelet is maximum-phase. A quanti-
tative analysis of this phase concept is provided in Figure
2.2-3.

Compute the cumulative energy of each wavelet

at any one time. Cumulative energy is computed by
adding squared amplitudes as shown in Table 2-15.
These values are plotted in Figure 2.2-3. Note that all
four wavelets have the same amount of total energy —
17 units. However, the rate at which the energy builds

Table 2-15. Cumulative energy of wavelets A, B, C,

and D at time samples 0, 1 and 2.

Wavelet 0 1 2

A 16 16 17
B 4 13 17
C 4 13 17 FIG. 2.2-4. All wavelets referred to in Figure 2.2-3 (A, B,
C, and D ) have the same amplitude spectrum as shown
D 1 1 17
above (Adapted from Robinson, 1966).
178 Seismic Data Analysis

Table 2-16. Autocorrelation lags of wavelets A, B, C,

and D.
Wavelet A
4 0 −1 Output
4 0 −1 17
4 0 −1 0
4 0 −1 −4
Wavelet B
2 3 −2 Output
2 3 −2 17
2 3 −2 0
2 3 −2 −4
Wavelet C
−2 3 2 Output
−2 3 2 17
−2 3 2 0
FIG. 2.2-5. Phase-lag spectra of the wavelets referred to in −2 3 2 −4
Figure 2.2-3. They have the common amplitude spectrum of Wavelet D
Figure 2.2-4 (Adapted from Robinson, 1966).
−1 0 4 Output
Given a fixed amplitude spectrum as in Figure 2.2- −1 0 4 17
4, the wavelet with the least energy delay is called min- −1 0 4 0
imum delay, while the wavelet with the most energy de- −1 0 4 −4
lay is called maximum delay. This is the basis for Robin-
son’s energy delay theorem: A minimum-phase wavelet
has the least energy delay. The spiking deconvolution operator is strictly the
Time delay is equivalent to a phase-lag. Figure 2.2- inverse of the wavelet. If the wavelet were minimum
5 shows the phase spectra of the four wavelets. Note phase, then we would get a stable inverse, which also is
that wavelet A has the least phase change across the minimum phase. The term stable means that the filter
frequency axis; we say it is minimum phase. Wavelet coefficients form a convergent series. Specifically, the
D has the largest phase change; we say it is maximum coefficients decrease in time (and vanish at t = ∞);
phase. Finally, wavelets B and C have phase changes therefore, the filter has finite energy. This is the case
between the two extremes; hence, they are mixed phase. for the wavelet (1, − 21 ) with an inverse (1, 21 , 41 , . . .). The
Since all four wavelets have the same amplitude inverse is a stable spiking deconvolution filter. On the
spectrum (Figure 2.2-4) and the same power spectrum, other hand, if the wavelet were maximum phase, then
they should have the same autocorrelation. This is ver- it does not have a stable inverse. This is the case for the
ified as shown in Table 2-16, where only one side of the wavelet (− 21 , 1), whose inverse is given by the divergent
autocorrelation is tabulated, since a real time series has series (−2, −4, −8, . . .). Finally, a mixed-phase wavelet
a symmetric autocorrelation (Section 1.1). does not have a stable inverse. This discussion leads us
Note that zero lag of the autocorrelation (Table 2- to assumption 7.
16) is equal to the total energy (Table 2-15) contained
in each wavelet — 17 units. This is true for any wavelet.
In fact, Parseval’s theorem states that the area under Assumption 7. The seismic wavelet is minimum
the power spectrum is equal to the zero-lag value of the phase. Therefore, it has a minimum-phase inverse.
autocorrelation function (Section A.1).
The process by which the seismic wavelet is com- Now, a summary of the implications of the under-
pressed to a zero-lag spike is called spiking deconvolu- lying assumptions for deconvolution stated in Sections
tion. In this section, filters that achieve this goal were 2.1 and 2.2 is appropriate.
studied — the inverse and the least-squares inverse
filters. Their performance depends not only on filter (a) Assumptions 1, 2, and 3 allow formulating the con-
length, but also on whether the input wavelet is mini- volutional model of the 1-D seismogram by equa-
mum phase. tion (2-2).
 Deconvolution 179

(b) Assumption 4 eliminates the unknown noise term Table 2-17. Autocorrelation lags of input wavelet
in equation (2-2a) and reduces it to equation (2- (1, − 21 ).
3a).
(c) Assumption 5 is the basis for deterministic decon- 1 − 12 Output
volution — it allows estimation of the earth’s re-
flectivity series directly from the 1-D seismogram 1 − 12 5
4
described by equation (2-3a).
(d) Assumption 6 is the basis for statistical deconvolu- 1 − 12 − 12
tion — it allows estimates for the autocorrelogram
and amplitude spectrum of the normally unknown
wavelet in equation (2-3a) from the known recorded
1-D seismogram.
(e) Finally, assumption 7 provides a minimum-phase Table 2-18. Crosscorrelation lags of desired output
estimate of the phase spectrum of the seismic (1, 0, 0) with input wavelet (1, − 21 ).
wavelet from its amplitude spectrum, which is es-
timated from the recorded seismogram by way of 1 0 0 Output
assumption 6.
1 − 12 1
Once the amplitude and phase spectra of the
seismic wavelet are statistically estimated from the 1 − 12 0
recorded seismogram, its least-squares inverse — spik-
ing deconvolution operator, is computed using opti-
mum Wiener filters (Section 2.3). When applied to the
wavelet, the filter converts it to a zero-delay spike.
In general, the elements of the matrix on the left
When applied to the seismogram, the filter yields the
earth’s impulse response (equation 2-4). In Section 2.3, side of equation (2-27) are the lags of the autocorre-
we show that a known wavelet can be converted into a lation of the input wavelet, while the elements of the
delayed spike even if it is not minimum phase. column matrix on the right side are the lags of the
crosscorrelation of the desired output with the input
wavelet.
Now perform similar operations for wavelet
2.3 OPTIMUM WIENER FILTERS
(− 12 , 1). By rewriting the matrix equation (2-19), we
obtain
Return to the desired output — the zero-delay spike
(1, 0, 0), that was considered when studying inverse and 5/4 −1/2 a −1
2 = . (2 − 28)
least-squares filters (Section 2.2). Rewrite equation (2- −1/2 5/4 b 0
16), which we solved to obtain the least-squares inverse Divide both sides by 2 to obtain
filter, as follows:
5/4 −1/2 a −1/2
5/4 −1/2 a 2 = . (2 − 29)
2 = . (2 − 26) −1/2 5/4 b 0
−1/2 5/4 b 0
Divide both sides by 2 to obtain The autocorrelation of wavelet (− 12 , 1) is given in
Table 2-19. The elements of the matrix on the left side of
5/4 −1/2 a 1 equation (2-29) are the autocorrelation lags of the input
= . (2 − 27)
−1/2 5/4 b 0 wavelet. Note that autocorrelation of wavelet (− 12 , 1) is
The autocorrelation of the input wavelet (1, − 21 ) is identical to that of wavelet (1, − 12 ) (Table 2-17). As
shown in Table 2-17. Note that the autocorrelation lags discussed in Section 2.2, an important property of a
are the same as the first column of the 2 × 2 matrix on group of wavelets with the same amplitude spectrum is
the left side of equation (2-27). that they also have the same autocorrelation.
Now compute the crosscorrelation of the desired The crosscorrelation of the desired output (1, 0, 0)
output (1, 0, 0) with the input wavelet (1, − 12 ) (Table with input wavelet (− 12 , 1) is given in Table 2-20. Note
2-18). The crosscorrelation lags are the same as the col- that the right side of equation (2-29) is the same as the
umn matrix on the right side of equation (2-27). crosscorrelation lags.
180 Seismic Data Analysis

Table 2-19. Autocorrelation lags of input wavelet The Wiener filter applies to a large class of prob-
(− 21 , 1). lems in which any desired output can be considered,
not just the zero-lag spike. Five choices for the desired
− 21 1 Output output are:

− 12 1 5
4 Type 1: Zero-lag spike,
Type 2: Spike at arbitrary lag,
− 21 1 − 12 Type 3: Time-advanced form of input series,
Type 4: Zero-phase wavelet,
Type 5: Any desired arbitrary shape.

Table 2-20. Crosscorrelation lags of desired output

These desired output forms will be discussed in the fol-
(1, 0, 0) with input wavelet (− 21 , 1).
lowing sections.
1 0 0 Output The general form of the normal equations (2-30)
was arrived at through numerical examples for the spe-
cial case where the desired output was a zero-lag spike.
− 12 1 − 12
Section B.5 provides a concise mathematical treatment
of the optimum Wiener filters. Figure 2.3-1 outlines the
− 21 1 0
design and application of a Wiener filter.
Determination of the Wiener filter coefficients re-
quires solution of the so-called normal equations (2-
Matrix equations (2-27) and (2-29) were used to de- 30). From equation (2-30), note that the autocorrela-
rive the least-squares inverse filters (Section 2.2). These tion matrix is symmetric. This special matrix, called
filters then were applied to the input wavelets to com- the Toeplitz matrix, can be solved by Levinson recur-
sion, a computationally efficient scheme (Section B.6).
press them to zero-lag spike. The matrices on the left
To do this, compute a two-point filter, derive from it
in equations (2-27) and (2-29) are made up of the au-
a three-point filter, and so on, until the n-point filter
tocorrelation lags of the input wavelets. Additionally,
is derived (Claerbout, 1976). In practice, filtering algo-
the column matrices on the right are made up of lags
rithms based on the optimum Wiener filter theory are
of the crosscorrelation of the desired output — a zero-
known as Wiener-Levinson algorithms.
lag spike, with the input wavelets. These observations
were generalized by Wiener to derive filters that convert
the input to any desired output (Robinson and Treitel,
1980). Spiking Deconvolution
The general form of the matrix equation such as
equation (2-29) for a filter of length n is (Section B.5): The process with type 1 desired output (zero-lag
 r r1 r2 · · · rn−1    spike) is called spiking deconvolution. Crosscorrelation
0 a0 g0
of the desired spike (1, 0, 0, . . . , 0) with input wavelet
 r1 r0 r1 · · · rn−2  a1   g1 
    (x0 , x1 , x2 , . . . , xn−1 ) yields the series (x0 , 0, 0, . . . , 0).
 r2 r1 r0 · · · rn−3  a2 = g2 
 .  .   
 . .. .. ..
.
..  ..   .. 
. . . . .
rn−1 rn−2 rn−3 ··· r0 an−1 gn−1
(2 − 30)

Here ri , ai , and gi , i = 0, 1, 2, . . . , n − 1 are the auto-

correlation lags of the input wavelet, the Wiener filter
coefficients, and the crosscorrelation lags of the desired
output with the input wavelet, respectively.
The optimum Wiener filter (a0 , a1 , a2 , . . . , an−1 ) is
optimum in that the least-squares error between the ac-
tual and desired outputs is minimum. When the de-
sired output is the zero-lag spike (1, 0, 0, . . . , 0), then
the Wiener filter is identical to the least-squares inverse
filter. In other words, the least-squares inverse filter re- FIG. 2.3-1. A flowchart for Wiener filter design and appli-
ally is a special case of the Wiener filter. cation.
 Deconvolution 181

The generalized form of the normal equations (2-30) and (j). One way to extract the seismic wavelet, pro-
takes the special form: vided it is minimum phase, is to compute the spiking
 r r1 r2 ··· rn−1 
   deconvolution operator and find its inverse.
0 a0 1
 r1 r0 r1 ··· rn−2 
 a1   0  In conclusion, if the input wavelet is not minimum
 
   phase, then spiking deconvolution cannot convert it to a
 r2 r1 r0 ··· rn−3 
 a2 = 0 
 . 
  
 . .
.. .
.. ..
.  ...   ... 
..  perfect zero-lag spike as in frame (k). Although the am-
. . plitude spectrum is virtually flat as shown in frame (l),
rn−1 rn−2 rn−3 ··· r0 an−1 0
the phase spectrum of the output is not minimum phase
(2 − 31)
as shown in frame (m). Finally, note that the spiking
Equation (2-31) was scaled by (1/x0 ). The least- deconvolution operator is the inverse of the minimum-
squares inverse filter, which was discussed in Section 2.2, phase equivalent of the input wavelet. This wavelet may
has the same form as the matrix equation (2-31). There- or may not be minimum phase.
fore, spiking deconvolution is mathematically identical
to least-squares inverse filtering. A distinction, however,
is made in practice between the two types of filtering.
Prewhitening
The autocorrelation matrix on the left side of equation
(2-31) is computed from the input seismogram (assump-
tion 6) in the case of spiking deconvolution (statistical From the preceding section, we know that the ampli-
deconvolution), whereas it is computed directly from the tude spectrum of the spiking deconvolution operator is
known source wavelet in the case of least-squares inverse (approximately) the inverse of the amplitude spectrum
filtering (deterministic deconvolution). of the input wavelet. This is sketched in Figure 2.3-3.
Figure 2.3-2 is a summary of spiking deconvolution What if we had zeroes in the amplitude spectrum of the
based on the Wiener-Levinson algorithm. Frame (a) is input wavelet? To study this, apply a minimum-phase
the input mixed-phase wavelet. Its amplitude spectrum band-pass filter (Exercise 2-10) with a wide passband
shown in frame (b) indicates that the wavelet has most (3-108 Hz) to the minimum-phase wavelet of Figure
of its energy confined to a 10- to 50-Hz range. The au- 2.3-2, as shown in frame (h). Deconvolution of the fil-
tocorrelation function shown in frame (d) is used in tered wavelet does not produce a perfect spike; instead,
equation (2-31) to compute the spiking deconvolution a spike accompanied by a high-frequency pre-and post-
operator shown in frame (e). The amplitude spectrum cursor results (Figure 2.3-4). This poor result occurs be-
of the operator shown in frame (f) is approximately the cause the deconvolution operator tries to boost the ab-
inverse of the amplitude spectrum of the input wavelet sent frequencies, as seen from the amplitude spectrum
shown in frame (b). (The approximation improves as op- of the output. Can this problem occur in a recorded seis-
erator length increases.) This should be expected, since mogram? Situations in which the input amplitude spec-
the goal of spiking deconvolution is to flatten the out- trum has zeroes rarely occur. There is always noise in
put spectrum. Application of this operator to the input the seismogram and it is additive in both the time and
wavelet gives the result shown in frame (k). frequency domains. Moreover, numerical noise, which
Ideally, we would like to get a zero-lag spike, as also is additive in the frequency domain, is generated
shown in frame (n). What went wrong? Assumption 7 during processing. However, to ensure numerical sta-
was violated by the mixed-phase input wavelet shown in bility, an artificial level of white noise is added to the
frame (a). Frame (h) shows the inverse of the deconvo- amplitude spectrum of the input seismogram before de-
lution operator. This is the minimum-phase equivalent convolution. This is called prewhitening and is referred
of the input mixed-phase wavelet in frame (a). Both to in Figure 2.3-3.
wavelets have the same amplitude spectrum shown in If the percent prewhitening is given by a scalar, 0 ≤
frames (b) and (i), but their phase spectra are signif- ε < 1, then the normal equations (2-31) are modified as
icantly different as shown in frames (c) and (j). Since follows:
spiking deconvolution is equivalent to least-squares in-
verse filtering, the minimum-phase equivalent is merely  βr
the inverse of the deconvolution operator. Therefore, the 0 r1 r2 · · · rn−1 

a0
 
1
amplitude spectrum of the operator is the inverse of the  r1 βr0 r1 · · · rn−2 
 a1   0 
 
  
amplitude spectrum of the minimum-phase equivalent  r2 r1 βr0 · · · rn−3 
 a2 = 0  ,
 . 
  
as shown in frames (f) and (i), and the phase spectrum  . .. .. ..
.  ...   ... 
.. 
. . . .
of the operator is the negative of the phase spectrum rn−1 rn−2 rn−3 · · · βr0 an−1 0
of the minimum-phase wavelet as shown in frames (g) (2 − 32)
182 Seismic Data Analysis
 Deconvolution 183

FIG. 2.3-3. Prewhitening amounts to adding a bias to the amplitude spectrum of the seismogram to be deconvolved. This
prevents dividing by zero since the amplitude spectrum of the inverse filter (middle) is the inverse of that of the seismogram
(left). Convolution of the filter with the seismogram is equivalent to multiplying their respective amplitude spectra — this
yields nearly a white spectrum (right).

where β = 1 + ε. Adding a constant εr0 to the zero lag there is less error when converting wavelet (− 12 , 1) to the
of the autocorrelation function is the same as adding delayed spike (0, 1, 0) than to zero-lag spike (1, 0, 0).
white noise to the spectrum, with its total energy equal In general, for any given input wavelet, a series of
to that constant. The effect of the prewhitening level desired outputs can be defined as delayed spikes. The
on performance of deconvolution is discussed in Section least-squares errors then can be plotted as a function of
2.4. delay. The delay (lag) that corresponds to the least er-
ror is chosen to define the desired delayed spike output.
The actual output from the Wiener filter using this opti-
Wavelet Processing by Shaping Filters mum delayed spike should be the most compact possible
result.
Spiking deconvolution had trouble compressing wavelet
(− 21 , 1) to a zero-lag spike (1, 0, 0) (Table 2-14). In terms
of energy distribution, this input wavelet is more sim-
ilar to a delayed spike, such as (0, 1, 0), than it is to a Table 2-21. Crosscorrelation lags of desired output
zero-lag spike, (1, 0, 0). Therefore, a filter that converts (0, 1, 0) with input wavelet (− 12 , 1).
wavelet (− 12 , 1) to a delayed spike would yield less error
than the filter that shapes it to a zero-lag spike (Table 0 1 0 Output
2-14).
Recast the filter design and application outlined − 21 1 1
in Table 2-13 in terms of optimum Wiener filters by
following the flowchart in Figure 2.3-1. First, compute − 12 1 − 21
the crosscorrelation (Table 2-21). From Table 2-19, we
know the autocorrelation of the input wavelet. By sub-
stituting the results from Tables 2-19 and 2-21 into the
matrix equation (2-30), we get Table 2-22. Convolution of input wavelet (− 21 , 1) with
16 2
filter coefficients ( 21 , − 21 ).
5/4 −1/2 a 1
= . (2 − 33)
−1/2 5/4 b −1/2
− 12 1 Output
By solving for the filter coefficients, we obtain (a, b) :
16 2 2
( 21 , − 21 ). This filter is applied to the input wavelet as − 21 − 16
21 −0.38
shown in Table 2-22. As we would expect, the output is
the same as that of the least-squares filter (Table 2-13). 2
− 21 − 16 0.81
21
Note that, from Table 2-14, the energy of the least-
squares error between the actual and desired outputs 2
− 21 − 16 −0.09
21
was 0.190 and 0.762 for a delayed spike and a zero-
lag spike desired output, respectively. This shows that
184 Seismic Data Analysis

FIG. 2.3-4. (a) Minimum-phase wavelet, (b) after band-pass filtering, (c) followed by deconvolution. The amplitude spectrum
of the band-pass filtered wavelet is zero above 108 Hz (middle row); therefore, the inverse filter derived from it yields unstable
results (bottom row). The time delays on the wavelets in the left frames of the middle and bottom rows are for display
purposes only.

The process that has a type 5 desired output (any
desired arbitrary shape) is called wavelet shaping. The
filter that does this is called a Wiener shaping filter.
In fact, type 2 (delayed spike) and type 4 (zero-phase
wavelet) desired outputs are special cases of the more
general wavelet shaping.
Figure 2.3-5 shows a series of wavelet shapings that
use delayed spikes as desired outputs. The input is a
mixed-phase wavelet. Filter length was held constant
in all eight cases. Note that the zero-delay spike case
(spiking deconvolution) does not always yield the best
result (Figure 2.3-5a). A delay in the neighborhood of
60 ms (Figure 2.3-5e) seems to yield an output that is
closest to being a perfect spike. Typically, the process
is not very sensitive to the amount of delay once it is
close to the optimum delay. If the input wavelet were
minimum-phase, then the optimum delay of the desired
output spike generally is zero. On the other hand, if the
input wavelet were mixed-phase, as illustrated in Fig-
ure 2.3-5, then the optimum delay is nonzero. Finally, if
the input wavelet were maximum-phase, then the opti-
mum delay is the length of that wavelet (Robinson and
Treitel, 1980).
Can we not delay the desired spike output (Fig-
ure 2.3-5) and obtain a better result than we obtained
from spiking deconvolution? This goal is achieved by
applying a constant-time shift (60 ms in Figure 2.3-5)
to a delayed spike result. Better yet, the same result
can be obtained by shifting the shaping filter operator
as much as the delay in the spike and applying it to the
input wavelet. Such a filter operator is two-sided (non-
causal), since it has coefficients for negative and positive
time values. The one-sided filter defined along the posi-
tive time axis has an anticipation component, while the
filter defined along the negative time axis has a mem-
ory component (Robinson and Treitel, 1980). The two-
sided filter has an anticipation component and a mem-
ory component. Figure 2.3-6 shows a series of shaping
filterings with two-sided Wiener filters for various spike
delay values.
 Deconvolution
FIG. 2.3-5. Shaping filtering. (0) Input wavelet, (1) desired
output, (2) shaping filter operator, (3) actual output. Here,
the purpose is to convert the mixed-phased wavelet (0) to
a series of delayed spikes as shown in (a) through (h) by
using a one-sided operator (anticipation component only).
The best result is with a 60-ms delay (e).
Figure 2.3-7 shows examples of wavelet shaping.
The input wavelet represented by trace (b) is the same
mixed-phase wavelet as in Figure 2.3-6 (top left frame).
This wavelet is shaped into zero-phase wavelets with
three different bandwidths represented by traces (c),
(d) and (e). This process commonly is referred to as de-
phasing. Figure 2.3-7 shows another wavelet shaping in
which the input wavelet is converted to its minimum-
phase equivalent represented by trace (f). This conver-
sion is often applied to recorded air-gun signatures.
Figure 2.3-8 shows examples of a recorded air-
gun signature that was shaped into its minimum-phase
equivalent and into a spike. When the input is the
recorded signature, then the wavelet shapings in Fig-
ure 2.3-8 are called signature processing.
Wavelet shaping requires knowledge of the input
wavelet to compute the crosscorrelation column on the
right side of equation (2-30). If it is unknown, which is
the case in reality, then the minimum-phase equivalent
of the input wavelet can be estimated statistically from
the data. This minimum-phase estimate then is shaped
to a zero-phase wavelet.
185
FIG. 2.3-6. Shaping filtering. (0) Input wavelet, (1) desired
output, (2) shaping filter operator, (3) actual output. Here,
the purpose is to convert the mixed-phase wavelet (0) to a
series of delayed spikes as shown in (a) through (h) using a
two-sided operator (with memory and anticipation compo-
nents). The best result is obtained with a zero-delay spike
using a two-sided filter (a).
Wavelet processing is a term that is used with flex-
ibility. The most common meaning refers to estimat-
ing (somehow) the basic wavelet embedded in the seis-
mogram, designing a shaping filter to convert the esti-
mated wavelet to a desired form, usually a broad-band
zero-phase wavelet (Figure 2.3-8), and finally, applying
the shaping filter to the seismogram. Another type of
wavelet processing involves wavelet shaping in which
the desired output is the zero-phase wavelet with the
same amplitude spectrum as that of the input wavelet
(Figure 2.3-9). Note that this type of wavelet processing
does not try to flatten the spectrum, but only tries to
correct for the phase of the input wavelet, which some-
times is assumed to be minimum-phase.
Predictive Deconvolution
The type 3 desired output, a time-advanced form of the
input series, suggests a prediction process. Given the
input x(t), we want to predict its value at some future
time (t + α), where α is prediction lag. Wiener showed
186 Seismic Data Analysis

FIG. 2.3-7. Shaping filtering with various desired outputs. (a) Impulse response, (b) input seismogram. Here, (c), (d) and
(e) show three possible desired outputs that are band-limited zero-phase wavelets, while (f) shows a desired output that is
the minimum-phase equivalent of the input wavelet (b). Finally, (g) and (h) are desired outputs that are band-pass filtered
versions of (f).

that the filter used to estimate x(t+α) can be computed The prediction filter coefficients a(t) : (a0 , a1 ,
by using a special form of the matrix equation (2-30) a2 , a3 , a4 ) can be computed from equation (2-34) and
(Robinson and Treitel, 1980). Since the desired output applied to the input series x(t) : (x0 , x1 , x2 , x3 , x4 ) to
x(t + α) is the time-advanced version of the input x(t), compute the actual output y(t) : (y0 , y1 , y2 , y3 , y4 ) (Ta-
we need to specialize the right side of equation (2-30) ble 2-25). We want to predict the time-advanced form of
for the prediction problem. the input; hence, the actual output is an estimate of the
Consider a five-point input time series x(t) : series x(t + α) : (x2 , x3 , x4 ), where α = 2. The predic-
(x0 , x1 , x2 , x3 , x4 ), and set α = 2. The autocorrelation tion error series e(t) = x(t + α) − y(t) : (e2 , e3 , e4 , e5 , e6 )
of the input series is computed in Table 2-23, and the
is given in Table 2-26.
crosscorrelation between the desired output x(t+2) and
The results in Table 2-26 suggest that the error se-
the input x(t) is computed in Table 2-24. Compare the
ries can be obtained more directly by convolving the
results in Tables 2-23 and 2-24, and note that gi = ri+α
input series x(t) : (x0 , x1 , x2 , x3 , x4 ) with a filter with
for α = 2 and i = 0, 1, 2, 3, 4.
Equation (2-30), for this special case, is rewritten coefficients (1, 0, −a0 , −a1 , −a2 , −a3 , −a4 ) (Table 2-27).
as follows: The results for (e2 , e3 , e4 , e5 , e6 ) are identical (Tables
     2-26 and 2-27). Since the series (a0 , a1 , a2 , a3 , a4 ) is
r 0 r 1 r2 r3 r 4 a0 r2
called the prediction filter, it is natural to call the se-
 1 r r 0 r 1 r 2 r a
3  1  r 
    3  ries (1, 0, −a0 , −a1 , −a2 , −a3 , −a4 ) the prediction error
 r2 r1 r0 r1 r2   a2  =  r4  . (2 − 34)
     filter. When applied to the input series, this filter yields
r 3 r 2 r1 r0 r 1 a3 r5
r 4 r 3 r2 r1 r 0 a4 r6 the error series in the prediction process (Table 2-27).
 Deconvolution 187

Table 2-24. Crosscorrelation of desired output x(t +

α) : (x2 , x3 , x4 ), α = 2, with input x(t) : (x0 , x1 , x2 ,
x3 , x4 ).

g0 = x0 x2 + x1 x3 + x2 x4
g1 = x0 x3 + x1 x4
g2 = x0 x4
g3 =0
g4 =0

Table 2-25. Convolution of prediction filter a(t) : (a0 ,

a1 , a2 , a3 , a4 ) with input series x(t) : (x0 , x1 , x2 , x3 , x4 )
to compute actual output y(t) : (y0 , y1 , y2 , y3 , y4 ).

y0 = a0 x0
y1 = a1 x0 + a0 x1
y2 = a2 x0 + a1 x1 + a0 x2
y3 = a3 x0 + a2 x1 + a1 x2 + a0 x3
y4 = a4 x0 + a3 x1 + a2 x2 + a1 x3 + a0 x4

FIG. 2.3-8. Signature processing: (a) Recorded signature,

(b) desired output, (c) shaping operator, (d) shaped signa-
ture. The desired output is a zero-delay spike (top) and the
Table 2-26. The error series e(t) = x(t + α) − y(t) :
minimum-phase equivalent of the recorded signature (bot-
(e2 , e3 , e4 , e5 , e6 ), α = 2. For y(t), see Table 2-25.
tom).
e2 = x2 − a0 x0
e3 = x3 − a1 x0 − a0 x1
e4 = x4 − a2 x0 − a1 x1 − a0 x2
e5 = 0 − a3 x0 − a2 x1 − a1 x2 − a0 x3
e6 = 0 − a4 x0 − a3 x1 − a2 x2 − a1 x3 − a0 x4

Table 2-23. Autocorrelation lags of input series x(t) : Table 2-27. Convolution of prediction error filter coef-
(x0 , x1 , x2 , x3 , x4 ). ficients (1, 0, −a0 , −a1 , −a2 , −a3 , −a4 ) with input series
x(t) : (x0 , x1 , x2 , x3 , x4 ).

r0 = x20 + x21 + x22 + x23 + x24 e0 = x0

r1 = x0 x1 + x1 x2 + x2 x3 + x3 x4 e1 = x1
r2 = x0 x2 + x1 x3 + x2 x4 e2 = x2 − a0 x0
r3 = x0 x3 + x1 x4 e3 = x3 − a1 x0 − a0 x1
r4 = x0 x4 e4 = x4 − a2 x0 − a1 x1 − a0 x2
r5 =0 e5 = 0 − a3 x0 − a2 x1 − a1 x2 − a0 x3
r6 =0 e6 = 0 − a4 x0 − a3 x1 − a2 x2 − a1 x3 − a0 x4
188 Seismic Data Analysis

FIG. 2.3-10. A flowchart for predictive deconvolution using

a prediction filter.

of occurrence. According to assumption 6, anything else,

such as primary reflections, is unpredictable.
Some may claim that reflections are predictable as
well; this may be the case if deposition is cyclic. How-
ever, this type of deposition is not often encountered.
While the prediction filter yields the predictable com-
ponent (the multiples) of a seismic trace, the remaining
unpredictable part, the error series, is essentially the
reflection series.
Equation (2-34) can be generalized for the case of
an n-long prediction filter and an α-long prediction lag.
 r r1 r2 · · · rn−1 
  r 
0 a0 α
 r1 r0 r1 · · · rn−2 
 a1   rα+1 
 
  
 r2 r1 r0 · · · rn−3 
 a2 = rα+2 
 . .. .. . . 

.  . .   .. 
 . . ..  . 
. . . .
rn−1 rn−2 rn−3 · · · r0 an−1 rα+n−1
FIG. 2.3-9. Wavelet processing. An autocorrelogram (a), (2 − 35)
estimated from the seismic trace, is used after smoothing (b) Note that design of the prediction filters requires only
to compute the spiking deconvolution operator (d). Here (c) autocorrelation of the input series.
is just a one-sided version of (b). The inverse of the operator
There are two approaches to predictive deconvolu-
(d) is the minimum-phase wavelet (e), which is sometimes
assumed to be the basic wavelet contained in the original tion:
seismic trace. It is easy to compute its zero-phase equiv-
alent (f) and design a shaping filter (g) that converts the (1) The prediction filter (a0 , a1 , a2 , . . . , an−1 ) may be
minimum-phase wavelet (e) to the zero-phase wavelet (f). designed using equation (2-35) and applied on in-
The actual output is (h), which should be compared with put series as described in Figure 2.3-10.
(f). The zero-phase equivalent (f) has the same amplitude
(2) Alternatively, the prediction error filter (1, 0, 0, . . . ,
spectrum as the minimum-phase wavelet (e).
0, −a0 , −a1 , −a2 , . . . , −an−1 ) can be designed and
Why place so much emphasis on the error series? convolved with the input series as described in Fig-
Consider the prediction process as it relates to a seis- ure 2.3-11.
mic trace. From the past values of a time series up to
time t, a future value can be predicted at time t + α, Now consider the special case of unit prediction lag,
where α is the prediction lag. A seismic trace often has a α = 1. For n = 5, equation (2-35) takes the following
predictable component (multiples) with a periodic rate form:
 Deconvolution 189

(2-38) as follows:
    
r0 r1 r2 r3 r4 r5 b0 L
 r1 r 0 r 1 r2 r3 r4   b1   0 
    
 r2 r 1 r 0 r1 r2 r3   b2   0 
   =  .
 r3 r 2 r 1 r0 r1 r2   b3   0 
    
r4 r 3 r 2 r1 r0 r1 b4 0
r5 r 4 r 3 r2 r1 r0 b5 0
(2 − 39)
where b0 = 1, bi = −ai , and i = 1, 2, 3, 4, 5. This equa-
tion has a familiar structure. In fact, except for the scale
factor L, it has the same form as equation (2-31), which
yields the coefficients for the least-squares zero-delay in-
verse filter. This inverse filter is therefore the same as
the prediction error filter with unit prediction lag, ex-
cept for a scale factor. Hence, spiking deconvolution ac-
FIG. 2.3-11. A flowchart for predictive deconvolution using
tually is a special case of predictive deconvolution with
a prediction error filter.
unit prediction lag.
We now know that predictive deconvolution is a
general process that encompasses spiking deconvolu-
tion. In general, the following statement can be made:
Given an input wavelet of length (n + α), the prediction
error filter contracts it to an α-long wavelet, where α
is the prediction lag (Peacock and Treitel, 1969). When
α = 1, the procedure is called spiking deconvolution.
     Figure 2.3-12 interrelates the various filters dis-
r0 r1 r2 r3 r4 a0 r1
 r1 r0 r1 r2 r3   a1   r2  cussed in this chapter and indicates the kind of process
     they imply. From Figure 2.3-12, note that Wiener fil-
 r2 r1 r0 r1 r2   a2  =  r3  . (2 − 36)
     ters can be used to solve a wide range of problems. In
r3 r2 r1 r0 r1 a3 r4
r4 r3 r2 r1 r0 a4 r5 particular, predictive deconvolution is an integral part
of seismic data processing that is aimed at compressing
By augmenting the right side to the left side, we obtain: the seismic wavelet, thereby increasing temporal reso-
    lution. In the limit, it can be used to spike the seismic
  1 0
−r1 r0 r1 r2 r3 r4 wavelet and obtain an estimate for reflectivity.
 a0   0 
 −r2 r1 r0 r1 r2 r3     
 a  0
 −r3 r2 r1 r0 r1 r2   1  =   .
 a  0
−r4 r3 r2 r1 r0 r1  2   
a3 0
−r5 r4 r3 r2 r1 r0
a4 0
(2 − 37)
Add one row and move the negative sign to the column
matrix that represents the filter coefficients to get:
    
r0 r1 r2 r3 r4 r5 1 L
 r1 r0 r1 r2 r3 r4   −a0   0 
    
 r2 r1 r0 r1 r2 r3   −a1   0 
   =  .
 r3 r2 r1 r0 r1 r2   −a2   0 
    
r4 r 3 r 2 r1 r0 r 1 −a3 0
r5 r 4 r 3 r2 r1 r 0 −a4 0
(2 − 38)
where L = r0 − r1 a0 − r2 a1 − r3 a2 − r4 a3 − r5 a4 .
Note that there are six unknowns, (a0 , a1 , a2 , a4 , a5 , L),
and six equations. Solution of these equations yields
the unit-delay prediction error filter (1, −a0 , −a1 ,
−a2 , −a4 , −a5 ), and the quantity L — the error in the 2.3-12. A flowchart for interrelations between various de-
filtering process (Section B.5). We can rewrite equation convolution filters.
190 Seismic Data Analysis
2.4 PREDICTIVE DECONVOLUTION
IN PRACTICE
It now is appropriate to review the implications of the
assumptions stated in Sections 2.1 and 2.2 that under-
lie the process of deconvolution within the context of
predictive deconvolution.
(a) Assumptions 1, 2, and 3 are the basis for the convo-
lutional model of the recorded seismogram (Section
2.1). In practice, deconvolution often yields good
results in areas where these three assumptions are
not strictly valid.
(b) Assumption 3 can be relaxed in practice by consid-
ering a time-variant deconvolution (Section 2.6).
In this technique, a seismogram is divided into a
number of time gates, typically three or more. De-
convolution operators then are designed from each
gate and convolved with data within that gate. Al-
ternatively, time-variant spectral whitening can be
used to account for nonstationarity (Section 2.6).
(c) Not much can be done about assumption 4. How-
ever, noise can be minimized in the recording pro-
cess. Deconvolution operators can be designed us-
ing time gates and frequency bands with low noise
levels. Poststack deconvolution can be used in an
effort to take advantage of the noise reduction in-
herent in the stacking process.
(d) If the source wavelet were minimum-phase and
known (assumption 5), then a perfect result could
be obtained from deconvolution in the noise-free
case as in trace (c) of Figures 2.4-1 and 2.4-2.
(e) If assumption 6 were violated and if the source
waveform were not known, then you would have
problems as in trace (d) of Figures 2.4-1 and 2.4-2.
(f) The quality of the output from spiking deconvolu-
tion is degraded further when the source wavelet is
not minimum-phase as in Figures 2.4-3 and 2.4-4;
that is, when assumption 7 is violated.
(g) Finally, in addition to violating assumptions 5 and
7, if there were noise in the data, that is, when as-
sumption 4 is violated, then the result of the decon-
volution would be unacceptable as in Figure 2.4-5.
Figures 2.4-1 through 2.4-5 test our confidence in
the usefulness of predictive deconvolution. In reality,
deconvolution has been applied to billions of seismic
traces; most of the time it has yielded satisfactory re-
sults. Figures 2.4-1 through 2.4-5 emphasize the criti-
cal assumptions that underlie predictive deconvolution.
When deconvolution does not work on some data, the
most probable reason is that one or more of the above
assumptions has been violated. In the remaining part
of this section, a series of numerical experiments will be
performed to examine the validity of these assumptions.
The purpose of these experiments is to gain a basic un-
derstanding of deconvolution from a practical point of
view.
Operator Length
We start with a single, isolated minimum-phase wavelet
as in trace (b) of Figure 2.4-6. Assumptions 1 through 5
are satisfied for this wavelet. The ideal result of spiking
deconvolution is a zero-lag spike, as indicated by trace
(a). In this and the following numerical analyses, we
refer to the autocorrelogram and amplitude spectrum
(plotted with linear scale) of the output from each de-
convolution test to better evaluate the results. In Figure
2.4-6 and the following figures, n, α, and ε refer to oper-
ator length of the prediction filter, prediction lag, and
percent prewhitening, respectively. The length of the
prediction error filter then is n + α.
In Figure 2.4-6, prediction lag is unity and equal
to the 2-ms sampling rate, prewhitening is 0%, and op-
erator length varies as indicated in the figure. Short
operators yield spikes with small-amplitude and rela-
tively high-frequency tails. The 128-ms-long operator
gives an almost perfect spike output. Longer operators
whiten the spectrum further, bringing it closer to the
spectrum of the impulse response.
The action of spiking deconvolution on the seismo-
gram derived by convolving the minimum-phase wavelet
with a sparse-spike series is similar (Figure 2.4-7) to the
case of the single isolated wavelet (Figure 2.4-6). Recall
that spiking deconvolution basically is inverse filtering
where the operator is the least-squares inverse of the
seismic wavelet. Therefore, an increasingly better result
should be obtained when more and more coefficients are
included in the inverse filter.
Now consider the real situation of an unknown
source wavelet. Based on assumption 6, autocorrelation
of the input seismogram rather than that of the seismic
wavelet is used to design the deconvolution operator.
The result of using the trace rather than the wavelet
autocorrelation is shown in Figure 2.4-8. Deconvolution
recovers the gross aspects of the spike series, trace (a).
However, note that the deconvolved traces have spu-
rious small-amplitude spikes trailing each of the real
spikes. We see that increasing operator length does not
indefinitely improve the results; on the contrary, more
and more spurious spikes are introduced.
Very short operators produce the same type of
noise spikes as in Figures 2.4-7 and 2.4-8. Examine the
series of deconvolution tests in Figure 2.4-8 and note
that the 94-ms operator does the best job. Compare
the autocorrelogram of trace (b) in Figure 2.4-8 with
 Deconvolution 191

2.4-1. (a) Impulse response, (b) seismogram, (c) spiking deconvolution using known, minimum-phase wavelet, (d) deconvo-
lution assuming an unknown, minimum-phase source wavelet. Impulse response (a) is a sparse-spike series. For an unknown
source wavelet (in violation of assumption 4), spiking deconvolution yields a less than perfect result (compare (c) and (d)).

2.4-2. (a) Impulse response, (b) seismogram, (c) spiking deconvolution using known, minimum-phase source wavelet, (d)
deconvolution assuming an unknown, minimum-phase source wavelet. Impulse response (a) is based on a sonic log (Figure
2.1-1a). For the unknown source wavelet (in violation of assumption 4), spiking deconvolution yields a less than perfect result.
(Compare (c) and (d).
192 Seismic Data Analysis

2.4-3. (a) Impulse response, (b) seismogram, (c) deconvolution using a known, mixed-phase source wavelet, (d) deconvolution
assuming an unknown, mixed-phase source wavelet. Impulse response (a) is a sparse-spike series. For a mixed-phase source
wavelet (in violation of assumption 5), spiking deconvolution yields a degraded output (d), even when the wavelet is known
(c).

2.4-4. (a) Impulse response, (b) seismogram, (c) deconvolution using a known, mixed-phase source wavelet, (d) deconvolution
assuming an unknown, mixed-phase source. Impulse response (a) is based on the sonic log of Figure 2.1-1a. For the mixed-
phase source wavelet (in violation of assumption 5), spiking deconvolution yields a degraded output (d) even when the wavelet
is known (c).
 Deconvolution 193

2.4-5. (a) Impulse response, (b) seismogram with noise, (c) deconvolution assuming an unknown, mixed-phase source wavelet.
Impulse response (a) is based on the sonic log of Figure 2.1-1a. In the presence of random noise (in violation of assumption
3), spiking deconvolution can produce a result with nelation to the earth’s reflectivity (compare (a) to (c)).

that of trace (b) in Figure 2.4-6. Note that only the
first 100-ms portion represents the autocorrelation of
the source wavelet. This explains why the 94-ms oper-
ator worked best; that is, the autocorrelation lags of
trace (b) in Figure 2.4-8 beyond 94 ms do not represent
the seismic wavelet.
Consider the seismogram in Figure 2.4-9, where the
wavelet is assumed to be unknown. Deconvolution has
restored the spikes that correspond to major reflections
in the impulse response as in trace (b) with some suc-
cess. The 64-ms operator is a good choice.
The mixed-phase wavelet in Figure 2.4-10 shows
what can happen when assumption 7 is violated. The
wavelet in Figure 2.4-6 is the minimum-phase equiva-
lent of the mixed-phase wavelet in Figure 2.4-10. Both
wavelets have the same autocorrelograms and ampli-
tude spectra. Hence, the deconvolution operators for
both wavelets are identical. Because the minimum-
phase assumption was violated, deconvolution does not
convert the mixed-phase wavelet to a perfect spike. In-
stead, the deconvolved output is a complicated high-
frequency wavelet. Also note that the dominant peak in
the output is negative, while the impulse response has
a positive spike. This difference in the sign can happen
when a mixed-phase wavelet is deconvolved. Increasing
the operator length further whitens the spectrum; how-
ever, the 128-ms operator yields a result that cannot be
improved further by longer operators.
The seismogram obtained from the mixed-phase
wavelet and the sparse-spike series (used in the pre-
ceding figures) is shown in Figure 2.4-11. The 94-ms
operator gives the best result. This also is the case in
Figure 2.4-12, where both assumptions 6 and 7 are vi-
olated. The situation with the seismogram in Figure
2.4-13 is not very good. The spikes that correspond to
major reflections in the impulse response were restored;
however, there are some timing errors and polarity re-
versals. (Compare these results with those in Figure 2.4-
9 for the events between 0.2 and 0.3 s and 0.6 and 0.7
s.) The 64-ms operator gives an output that cannot be
improved by longer operators.
What kind of operator length should be used for
spiking deconvolution? To select an operator length,
ideally we want to use the autocorrelation of the un-
known seismic wavelet. Fortunately, the autocorrelation
of the input seismogram has the characteristics of the
wavelet autocorrelation (assumption 6). Therefore, it
seems appropriate that we should use part of the au-
tocorrelation obtained from the input seismogram that
most resembles the autocorrelation of the unknown seis-
mic wavelet. That part is the first transient zone in the
autocorrelation, as seen by comparing the autocorrela-
tions of trace (b) in Figure 2.4-6 and trace (c) in Figure
2.4-9. The autocorrelations of trace (b) in Figure 2.4-10
and trace (c) in Figure 2.4-13 suggest the same princi-
ple.
Prediction Lag
So far, we have learned that predictive deconvolution
has two uses: (a) spiking deconvolution — the case
of unit prediction lag, and (b) predicting the input
(text continues on p. 198)
194 Seismic Data Analysis

2.4-6. Test of operator length for a single, isolated input wavelet, where n = operator length, α = prediction lag, and ε =
percent prewhitening. (a) Impulse response, (b) seismogram with minimum-phase source wavelet.

2.4-7. Test of operator length where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with known, minimum-phase source wavelet.
 Deconvolution 195

2.4-8. Test of operator length where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with known, minimum-phase source wavelet.

2.4-9. Test of operator length where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with unknown, minimum-phase source wavelet.
196 Seismic Data Analysis

2.4-10. Test of operator length for a single, isolated input wavelet where n= operator length, α = prediction lag, and ε =
percent prewhitening. (a) Impulse response, (b) seismogram with mixed-phase source wavelet.

2.4-11. Test of operator length where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with known, mixed-phase source wavelet.
 Deconvolution 197

2.4-12. Test of operator length where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with unknown, mixed-phase source wavelet.

2.4-13. Test of operator length where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with unknown, mixed-phase source wavelet.
198 Seismic Data Analysis
seismogram at a future time defined by the prediction
lag. Case (b) is used to predict and attenuate multiples.
Now, the effect of the prediction lag parameter is
examined from an interpretive point of view. Consider
the single, isolated minimum-phase wavelet in Figure
2.4-14. Here, operator length and percent prewhiten-
ing are kept constant, while prediction lag is varied.
When prediction lag is equal to the sampling rate, then
the result is equivalent to spiking deconvolution. Predic-
tive deconvolution using a prediction lag greater than
unity yields a wavelet of finite duration instead of a
spike. Given an input wavelet of α + n samples, predic-
tive deconvolution using prediction filter with length n
and prediction lag α converts this wavelet into another
wavelet that is α samples long. The first α lags of the
autocorrelation are preserved, while the next n lags are
zeroed out. Additionally, the amplitude spectrum of the
output increasingly resembles that of the input wavelet
as prediction lag is increased (Figure 2.4-14). At a 94-
ms prediction lag, predictive deconvolution does noth-
ing to the input wavelet because almost all the lags of
its autocorrelation have been left untouched. This ex-
periment has an important practical implication: Under
the ideal, noise-free conditions, resolution on the output
from predictive deconvolution can be controlled by ad-
justing the prediction lag. Unit prediction lag implies
the highest resolution, while a larger prediction lag im-
plies less than full resolution. However, in reality, these
assessments are dictated by the signal-to-noise ratio.
2.4-14. Test of prediction lag for a single, isolated input wavelet where n= operator length, α = prediction lag, and ε =
percent prewhitening. (a) Impulse response, (b) seismogram with minimum-phase source wavelet.
The deconvolved output using a unit prediction lag con-
tains high frequencies; nevertheless, resolution may be
degraded if the high-frequency energy is mostly noise,
not signal.
In Figure 2.4-14, prediction lags of 8 and 22 ms
correspond to the first and second zero crossings on au-
tocorrelation of the input wavelet, respectively. The first
zero crossing produces a spike with some width, while
the second zero crossing lag produces a wavelet with a
positive and negative lobe.
The relationship between prediction lag and
whitening also holds for the sparse-spike series in Figure
2.4-15 and when the input wavelet is unknown (Figure
2.4-16).
The effect of prediction lag on the output from
predictive deconvolution of a synthetic seismogram,
which was obtained from the sonic log (Figure 2.1-1a),
is demonstrated in Figures 2.4-17 and 2.4-18. As the
prediction lag is increased, the output spectrum be-
comes increasingly less broadband. Predictive decon-
volution of seismograms constructed from the mixed-
phase wavelet again demonstrates that output resolu-
tion can be controlled by adjusting prediction lag (Fig-
ures 2.4-19 through 2.4-23).
If prediction lag is increased, then the output am-
plitude spectrum becomes increasingly band-limited.
The output also can be band-limited by applying a
band-pass filter on the spiking deconvolution output.
(text continues on p. 203)
 Deconvolution 199

2.4-15. Test of prediction lag where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with known, minimum-phase source wavelet.

2.4-16. Test of prediction lag where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with unknown, minimum-phase source wavelet.
200 Seismic Data Analysis

2.4-17. Test of prediction lag where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with known minimum-phase source wavelet.

2.4-18. Test of prediction lag where n= operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with unknown, minimum-phase source wavelet.
 Deconvolution 201

2.4-19. Test of prediction lag for a single, isolated input wavelet where n = operator length, α = prediction lag, and ε =
percent prewhitening. (a) Impulse response, (b) seismogram with mixed-phase source wavelet.

2.4-20. Test of prediction lag where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with known, mixed-phase source wavelet.
202 Seismic Data Analysis

2.4-21. Test of prediction lag where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Impulse
response, (b) seismogram with unknown, mixed-phase source wavelet.

2.4-22. Test of prediction lag where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with known, mixed-phase source wavelet.
 Deconvolution
2.4-23. Test of prediction lag where n = operator length, α = prediction lag, and ε = percent prewhitening. (a) Reflectivity,
(b) impulse response, (c) seismogram with unknown, mixed-phase source wavelet.
Are these two ways of band-limiting equivalent? Refer
to the results from both the minimum- and mixed-phase
wavelets in Figures 2.4-14 and 2.4-19, respectively. Note
that the output of the 22-ms prediction lag has an am-
plitude spectrum that is band-limited to approximately
0 to 100 Hz. However, the spectral shape within this
bandwidth is not a boxcar, but rather similar to that of
the input wavelet. The boxcar shape would be the case if
a band-pass filter (0 to 100 Hz) were applied to the out-
put of the spiking deconvolution (2-ms prediction lag).
Hence, spiking deconvolution followed by band-pass fil-
tering is not equivalent to predictive deconvolution with
a prediction lag greater than unity.
In conclusion, if prediction lag is increased, the out-
put from predictive deconvolution becomes less spiky.
This effect can be used to our advantage, since it allows
the bandwidth of deconvolved output to be controlled
by adjusting the prediction lag. The application of spik-
ing deconvolution to field data is not always desirable,
since it boosts high-frequency noise in the data. The
most prominent effect of the nonunity prediction lag is
suppression of the high-frequency end of the spectrum
and preservation of the overall spectral shape of the
input data. This effect is seen in Figures 2.4-18 and 2.4-
23, which correspond to the minimum-and mixed-phase
seismic wavelets. If prediction lag is increased further,
then the low-frequency end of the spectrum is affected
as well, making the output more band-limited.
203
Percent Prewhitening
The reasons for prewhitening were discussed in Sec-
tion 2.3. Consider the single, isolated minimum-phase
wavelet in Figure 2.4-24. Keep the operator length and
prediction lag constant and vary the percent prewhiten-
ing. Note that the effect of varying prewhitening is
similar to that of varying the prediction lag; that is,
the spectrum increasingly becomes less broadband as
the percent prewhitening is increased. Compare Figure
2.4-14 with Figure 2.4-24. Note that prewhitening nar-
rows the spectrum without changing much of the flat-
ness character, while larger prediction lag narrows the
spectrum and alters its shape, making it look more like
the spectrum of the input seismic wavelet. These char-
acteristics also can be inferred from the shapes of the
output wavelets. Prewhitening preserves the spiky char-
acter of the output, although it adds a low-amplitude,
high-frequency tail (Figure 2.4-24). On the other hand,
increasing prediction lag produces a wavelet with a du-
ration equal to the prediction lag (Figure 2.4-14).
The effect of prewhitening on the sparse-spike train
seismogram with a known and unknown minimum-
phase wavelet is shown in Figures 2.4-25 and 2.4-26,
respectively. The effect of prewhitening on deconvo-
lution of the synthetic seismogram obtained from the
sonic log (Figure 2.1-1a) is shown in Figures 2.4-27
and 2.4-28 for known and unknown minimum-phase
wavelets. Prewhitening tests using the mixed-phase
204 Seismic Data Analysis

2.4-24. Test of percent prewhitening for a single, isolated input wavelet where n = operator length, α = prediction lag, and
ε = percent prewhitening. (a) Impulse response, (b) seismogram with minimum-phase source wavelet.

2.4-25. Test of percent prewhitening where n = operator length, α = prediction lag, and ε = percent prewhitening. (a)
Impulse response, (b) seismogram with known, minimum-phase source wavelet.
 Deconvolution 205

2.4-26. Test of percent prewhitening where n = operator length, α = prediction lag, and ε = percent prewhitening. (a)
Impulse response, (b) seismogram with unknown, minimum-phase source wavelet.

2.4-27. Test of percent prewhitening where n = operator length, α = prediction lag, and ε = percent prewhitening. (a)
Reflectivity, (b) impulse response, (c) seismogram with known, minimum-phase source wavelet.
206 Seismic Data Analysis

2.4-28. Test of percent prewhitening where n = operator length, α = prediction lag, and ε = percent prewhitening. (a)
Reflectivity, (b) impulse response, (c) seismogram with unknown, minimum-phase source wavelet.

2.4-29. Test of percent prewhitening for a single, isolated input wavelet where n = operator length, α = prediction lag, and
ε = percent prewhitening. (a) Impulse response, (b) seismogram with mixed-phase source wavelet.
 Deconvolution
2.4-30. Test of percent prewhitening for a single, isolated input wavelet where n = operator length, α = prediction lag, and
ε = percent prewhitening. (a) Impulse response, (b) seismogram with minimum-phase source wavelet.
wavelet are shown in Figure 2.4-29. Finally, the com-
bined effects of a prediction lag that is greater than
unity and prewhitening for the single, isolated wavelet
are shown in Figure 2.4-30. These figures demonstrate
that prewhitening narrows the output spectrum, mak-
ing it band-limited. In particular, the tests in Figures
2.4-24 and 2.4-29 using the single, isolated minimum-
and mixed-phase wavelets suggest that spiking decon-
volution with some prewhitening is somewhat equiva-
lent to spiking deconvolution without prewhitening fol-
lowed by post-deconvolution broad band-pass filtering.
However, this is not exactly true, for prewhitening still
leaves some relatively suppressed energy at the high-
frequency end of the spectrum. From Figure 2.4-30, we
infer that predictive deconvolution with a prediction lag
greater than unity and with some prewhitening yields a
result somewhat equivalent to a spiking deconvolution
followed by band-pass filtering.
In conclusion, we can say that prewhitening yields
a band-limited output. However, the effect is less con-
trollable when compared to varying the prediction lag.
By varying prediction lag, we have some idea of the out-
put bandwidth, since it is related to prediction lag. The
smaller the prediction lag, the broader the output band-
width. Prewhitening is used only to ensure that numer-
ical instability in solving for the deconvolution operator
(equation 2-32) is avoided. In practice, typically 0.1 to
1% prewhitening is standard.
207
Effect of Random Noise on Deconvolution
We assume that the noise component in the recorded
seismogram is zero (assumption 4). The autocorrelation
of ideal random noise is zero at all lags except the zero
lag (Figure 2.1-5). Therefore, the effect of random noise
on deconvolution operators should be somewhat simi-
lar to the effect of prewhitening. Both effects modify the
diagonal of the autocorrelation matrix, making it more
dominant [equation (2-32)]. However, the noise compo-
nent also slightly modifies the nonzero lags of the au-
tocorrelation. Compare the autocorrelograms of traces
(b) in Figures 2.4-24 and 2.4-31. In Figure 2.4-24, an
isolated minimum-phase wavelet was considered, while
in Figure 2.4-31, random noise was added to the same
wavelet. The output wavelet shape from spiking decon-
volution of the noisy wavelet using a 128-ms operator is
similar to the output from spiking deconvolution of the
wavelet without noise, using the same operator length
but with, say, 20 percent prewhitening. This result has
practical importance — prewhitening is equivalent to
adding perfect random noise to the system. Since a
recorded seismogram always contains some amount of
random noise, only a minute amount, say 0.1 percent,
of the white noise needs to be added to the seismogram
for numerical stability.
The effect of random noise on the performance
of deconvolution is examined further in Figures 2.4-
32 and 2.4-33. These results should be compared with
208 Seismic Data Analysis

2.4-31. Effect of random noise on deconvolution performance. Input seismogram (b) associated with reflectivity (a) contains
a single, isolated wavelet (at around 0.2 s) buried in random noise. Here, n = operator length, α = prediction lag, and ε =
percent prewhitening.

2.4-32. Effect of random noise on deconvolution performance. (a) Reflectivity, (b) impulse response. Input seismogram (c)
with unknown, minimum-phase source wavelet is noise contaminated. Here, n = operator length, α = prediction lag, and ε
= percent prewhitening.
 Deconvolution
2.4-33. Effect of random noise on deconvolution performance. (a) Reflectivity, (b) impulse response. Input seismogram (c)
with unknown, mixed-phase source wavelet is noise contaminated. Here, n = operator length, α = prediction lag, and ε =
percent prewhitening.
their noiseless counterparts in Figures 2.4-9 and 2.4-
13, respectively. Observe that the noise component has
a harmful effect on deconvolution. For example, when
comparing Figures 2.4-9 and 2.4-32, note that the de-
convolution result from the noisy seismogram has spu-
rious spikes (for instance, between 0.5 and 0.6 s), which
could be interpreted as genuine reflections. Noisy field
data, which yield better stack when not treated by de-
convolution, have been noted. Only by testing can we
determine whether deconvolution performs satisfacto-
rily on data with a severe noise problem.
Multiple Attenuation
We have learned that a prediction filter predicts peri-
odic events, like multiples, in the seismogram. The pre-
diction error filter yields the unpredictable component
of the seismogram — the reflectivity series. For exam-
ple, consider the simple case of water-bottom multiples.
If the reflection coefficient of the water bottom is cw and
if water depth is equivalent to a two-way time tw , then
the time series is
1, 0, . . . , 0, −cw , 0, . . . , 0, c2w , 0, . . . , 0, −c3w , 0, . . .)
as represented by trace (b) in Figure 2.4-34. The sepa-
ration between the spikes is tw in trace (b). Note that
209
the periodicity in the time series (trace (b) or (c)) man-
ifests itself in the amplitude spectrum as periodic peaks
(or notches). The greater the spike separation in time,
the closer the peaks (or notches) in the amplitude spec-
trum.
The noise-free convolutional model for the seismo-
gram that contains the water-bottom multiples can be
written as
x(t) = w(t) ∗ m(t) ∗ e(t), (2 − 40)
where m(t) represents the water-layer reverberation
spike series as in trace (b) of Figure 2.4-34, and e(t) now
represents the earth’s impulse response excluding mul-
tiples associated with the water bottom. Predictive de-
convolution can suppress the periodic component m(t)
in the seismogram as demonstrated by trace (d) in Fig-
ure 2.4-3.
Note the two distinct goals for predictive deconvo-
lution: (a) spiking the seismic wavelet w(t), and (b) pre-
dicting and attenuating multiples m(t). The first goal
is achieved using an operator with unit prediction lag,
while the second is achieved using an operator with a
prediction lag greater than unity.
The autocorrelation of the input trace can be used
to determine the appropriate prediction lag for multiple
suppression. Periodicity associated with multiples is ev-
ident in the autocorrelogram of trace (c) in Figure 2.4-
34, as an isolated series of energy lobes in the neighbor-
210 Seismic Data Analysis

2.4-34. (a) Reflectivity, (b) impulse response, (c) seismogram. Two-step deconvolution aimed at attenuating the multiples,
then spiking the remaining primary wavelet (d) to (e). The process can be performed in the reverse order (f) to (g). Here, n
= operator length and α = prediction lag.

2.4-35. Predictive deconvolution for multiple attenuation. (a) Reflectivity, (b) impulse response, (c) seismogram. Two-step
deconvolution: Predictive deconvolution (d) followed by spiking deconvolution (e). Traces (f), (g), and (h) result when single-
step deconvolution is applied to the input trace (c), using the operator lengths n and prediction lags α, as indicated.
 Deconvolution
hood of 0.2 and 0.4 s. Prediction lag should be chosen
to bypass the first part of the autocorrelogram that rep-
resents the seismic wavelet. Operator length should be
chosen to include the first isolated energy packet in the
autocorrelogram. After applying predictive deconvolu-
tion, we are left with only the water-bottom primary
reflection. Isolated bursts in the autocorrelogram have
been suppressed, while periodic peaks in the amplitude
spectrum have been eliminated as shown in trace (d). If
desired, the basic wavelet can be compressed into a spike
as shown in trace (e) by applying spiking deconvolution
to the output of predictive deconvolution as shown in
trace (d). The sequence can be interchanged by first
applying spiking deconvolution as shown in trace (f)
followed by predictive deconvolution as shown in trace
(g).
By using a sufficiently long spiking deconvolution
operator, two goals are achieved in one step as seen in
trace (h). However, this approach can be dangerous if
primary reflections are unintentionally suppressed. This
is the case in Figure 2.4-35. Here, the water-bottom
reflection is followed by a deeper event at about 0.28
s as seen in trace (a). The impulse response contains
water-bottom multiples and the peg-leg multiples that
are associated with the deeper reflector as seen in trace
(b). The amplitude spectrum has peaks that come in
pairs, indicating the presence of two different periodic
components in the seismogram. Careful choice of pre-
dictive deconvolution parameters yields an output with
only the wavelets associated with the water bottom and
the deeper reflector as seen in trace (d). This is fol-
lowed by a spiking deconvolution that yields two spikes
representative of the water bottom and the deep pri-
mary as seen in trace (e). Spiking deconvolution alone
produces the reflection coefficient series and the spikes
that represent the multiples as seen in trace (f). If a
longer spiking deconvolution operator is used, then the
primary reflection easily can be eliminated as in trace
(g). If a predictive deconvolution operator is used with
an improper parameter choice, then again, the primary
reflection can be eliminated easily as in trace (h).
How can we ensure that no primaries are de-
stroyed by deconvolution? Examine the autocorrelo-
gram of trace (c) in Figure 2.4-35. The first 50-ms por-
tion represents the seismic wavelet. This is followed by
a burst between 50 to 170 ms that represents the cor-
relation of the water bottom and primary. The isolated
burst between 170 to 340 ms represents the actual mul-
tiple series (both the peg-legs and water-bottom mul-
tiples). The prediction lag must be chosen to bypass
the first part of the autocorrelogram, which represents
the seismic wavelet and possible correlation between
211
the primaries. The operator length must be chosen to
include the first isolated burst, in this case between 170
to 340 ms.
It is only with vertical incidence and zero-offset
recording that periodicity of the multiples is preserved.
Therefore, predictive deconvolution aimed at multi-
ple suppression may not be entirely effective when
applied to nonzero-offset data, such as common-shot
or common-midpoint data. Figure 2.4-36a shows a
common-shot gather with its autocorrelogram and av-
erage amplitude spectrum. The field record is prepared
for deconvolution by first applying t2 −scaling (Figure
2.4-36b) and muting the first arrivals associated with
largely guided waves (Figure 2.4-36c). The autocorrelo-
gram in Figure 2.4-36c indicates the presence of multi-
ples. Note on the shot record the water-bottom reflec-
tion is at 0.4 s at near offset. Additionally, there are
two strong primary reflectors at 0.6 and 1.4 s at near
offset; these primaries give rise to a first-order and peg-
leg multiples. Despite the flatness of the spectrum and
the attenuation of nonzero lags of the autocorrelogram
after deconvolution, the first-order multiples associated
with the water-bottom reflection and the peg-leg mul-
tiples associated with the primary reflection at 0.6 s
still persist in the record (Figures 2.4-36d). This occurs
because these events have large moveout which causes
significant departure from periodicity at nonzero offsets.
On the other hand, note that the peg-leg multiples as-
sociated with the primary reflection at 1.4 s have been
attenuated significantly by deconvolution. This event
has a very small moveout; thus, its perodicity is much
preserved.
Predictive deconvolution sometimes is applied to
CMP stacked data in an effort to suppress multiples.
The performance of such an approach can be unsat-
isfactory, because the amplitude relationships between
multiples often are grossly altered by the stacking pro-
cess, primarily because of velocity differences between
primaries and multiples. Also, geometric spreading com-
pensation by using primary velocity function adversely
affects the amplitudes of multiples on nonzero-offset
data.
There is one domain in which the periodicity and
amplitudes of multiples are preserved — the slant-stack
domain. In Section 6.3, the application of predictive de-
convolution to data in the slant-stack domain for mul-
tiple suppression is discussed.
2.5 FIELD DATA EXAMPLES
The deconvolution parameters now are examined using
field data examples. We shall discuss application of sta-
212 Seismic Data Analysis

2.4-36. (a) A common-shot gather, (b) after t2 −scaling and (c) after muting first arrivals which are largely guided waves,
and (d) after spiking deconvolution. The amplitude spectra averaged over the shot record are shown at the top and the
autocorrelograms are shown at the bottom.
 Deconvolution
2.5-1. An autocorrelation window test used to design de-
convolution operators. The solid bars indicate the window
boundaries. The entire 6-s length was included in (a). The
autocorrelograms are displayed beneath the records.
tistical deconvolution to pre- and poststack data. Ad-
ditionally, we shall discuss application of deterministic
deconvolution to marine data to convert the recorded
source signature to its minimum-phase equivalent, and
to land data recorded using a vibroseis source to convert
the autocorrelogram of the sweep signal to its minimum-
phase equivalent.
Prestack Deconvolution
Figure 2.5-1 shows a CMP gather that contains five
prominent reflections at around 1.1, 1.35, 1.85, 2.15,
and 3.05 s. The gather also contains strong reverber-
ations associated with these reflections. The examina-
tion of the deconvolution parameters will begin with an
213
2.5-2. Test of operator length. The corresponding autocor-
relogram is beneath each record. The window used in auto-
correlation estimation is shown in Figure 2-5-1c. (a) Input
gather. Deconvolution using prediction lag = 4 ms (spik-
ing deconvolution), 0.1% prewhitening, and prediction filter
operator lengths (b) 40 ms, (c) 80 ms, (d) 160 ms, (e) 240
ms.
analysis of the time gate to estimate the autocorrelation
function. A first gate selected may be the entire length
(6 s) of the record as seen in panel (a). The solid lines
on the CMP gathers refer to the gate start and end
times. The autocorrelogram of the record is shown at
the bottom of each panel. A second choice might be to
exclude the deeper part of the record where ambient
noise dominates. The start of the gate is chosen as the
first arrival path as shown in panel (b). A third choice
may be to exclude not only the deeper portion, but
also the early part of the record that contains energy
corresponding to the guided waves as shown in panel
(c). These waves travel within the water layer and are
not part of the signal reflected from the substrata.
214 Seismic Data Analysis
2.5-3. Test of prediction lag. The corresponding autocor-
relogram is beneath each record. The window used in auto-
correlation estimation is shown in Figure 2.5-1c. (a) Input
gather. Deconvolution using prediction filter operator length
= 160 ms, 0.1 percent prewhitening, and prediction lags (b)
12 ms, (c) 32 ms, (d) 64 ms, (e) 128 ms.
By comparing the autocorrelograms from these dif-
ferent windows, note that the third choice best repre-
sents the reverberatory character of the data as shown
in panel (c) over most of the offsets. All of the traces
in the autocorrelogram within approximately the first
150 ms have a common appearance. This early portion
of the autocorrelogram characterizes the basic seismic
wavelet contained in the data.
In general, the autocorrelation window should in-
clude part of the record that contains useful reflection
signal, and should exclude coherent or incoherent noise.
An autocorrelation function contaminated by noise is
undesirable since the deconvolution process is most ef-
fective on noise-free data (assumption 4).
Another aspect of the autocorrelation window is
length. Panel (d) of Figure 2.5-1 shows the autocorrel-
2.5-4. Test of percent prewhitening. The corresponding au-
tocorrelogram is beneath each record. The window used in
autocorrelation estimation is shown in Figure 2.5-1c. (a) In-
put gather. Deconvolution using prediction filter operator
length = 160 ms, prediction lag = 4 ms (spiking deconvolu-
tion), and percent prewhitening (b) 1 percent, (c) 4 percent,
(d) 16 percent, (e) 32 percent.
ogram estimated from a narrow window. The autocor-
relogram estimated from the narrower part of the time
gate (the right side of the record) in some data cases
may lack the characteristics of the reverberations, and
even those of the basic seismic wavelet.
In general, any autocorrelation function is biased;
that is, the first lag value is computed from, say, n
nonzero samples, the second lag value is computed from
n − 1 nonzero samples, and so on. If n is not large
enough, then there can be an undesirable biasing effect.
How large should the data window be to avoid such bias-
ing? If the largest autocorrelation lag used in designing
the deconvolution operator were m, an accepted rule of
 Deconvolution 215

2.5-5. (a) A common-shot gather, (b) after muting guided waves, (c) after t2 -scaling, and (d) after spiking deconvolution
using an operator length of 320 ms. The amplitude spectra (top) averaged over the shot record, and the autocorrelograms
(bottom) are used to to choose deconvolution parameters and evaluate the data after the application of deconvolution.
216 Seismic Data Analysis

thumb is that the number of data samples should be no

less than 8m.
Now that the autocorrelation window is deter-
mined, we examine operator length. In Figure 2.5-2,
prediction lag (4 ms, the same as the sampling rate) and
percent prewhitening (0.1%) are fixed. The autocorrel-
ograms (at the bottom of each gather) are displayed
for diagnostic purposes. From the analyses of the sin-
gle spike, sparse spike, and reflectivity models (Section
2.4), recall that the short (40-ms) operator leaves some
residual energy that corresponds to the basic wavelet
and reverberating wavetrain in the record. For a spiking
deconvolution with a 160-ms-long operator, no remnant
of the energy is associated with the basic wavelet and
reverberations. Any operator longer than 160 ms does
not change the result, significantly. From the output of
the 160-ms operator, note that the prominent reflec-
tions (at 1.1, 1.35, 1.85, and 2.15 s at the near-offset)
have been uncovered, the seismic wavelet has been com-
pressed, and the reverberations have been significantly
suppressed.
The effect of prediction lag now is examined. In
Figure 2.5-3, the 160-ms operator length and 0.1%
prewhitening are fixed, while prediction lag is varied.
If prediction lag were increased, the deconvolution pro-
cess would be increasingly less effective in broadening
the spectrum, and the autocorrelograms would contain
increasingly more energy at nonzero lags. In the ex-
treme, the deconvolution process is ineffective with a
128-ms prediction lag. In practice, common values for
the prediction lag are unity (spiking deconvolution) or
the first or second zero crossing of the autocorrelation
function (predictive deconvolution).
Finally, the percent of prewhitening is varied, while
the 4-ms prediction lag and 160-ms operator length are
fixed. These tests are shown in Figure 2.5-4. By increas-
ing the percent prewhitening, the deconvolution process
becomes less effective. The high end of the spectrum is
not flattened as much as the rest of the spectrum (Fig-
ure 2.4-24). Note that the autocorrelograms contain in-
creasingly more energy at nonzero lags with increasing FIG. 2.5-6. The shot record in Figure 2.5-5c after predic-
percent prewhitening. In practice, it is not advisable tive deconvolution using an operator length of 320 ms and
to assign a large percent of prewhitening. Typically, a a prediction lag of: (a) unit-prediction, (b) 8 ms, and (c)
value between 0.1 and 1 percent is sufficient to ensure 24 ms. The amplitude spectra (top) averaged over the shot
stability in designing the deconvolution operator (equa- record, and the autocorrelograms (bottom) are used to to
tion 2-32). choose deconvolution parameters and evaluate the data af-
We now examine the effect of operator length and ter the application of deconvolution.
prediction lag on amplitude spectrum. Figure 2.5-5a
shows a common-shot gather with its autocorrelogram within the passband of the recorded data and attenu-
and average amplitude spectrum. The field record is ation of the energy at nonzero lags of the autocorrel-
prepared for deconvolution by first muting the guided ogram. With prediction lag greater than unity (Figure
waves (Figure 2.5-5b) and applying t2 -scaling (Figure 2.5-6), for the same operator length, we note insufficient
2.5-5c). Figure 2.5-5d shows the same record after spik- flattening at the high-frequency end of the spectrum. At
ing deconvolution. Note the flattening of the spectrum larger prediction lags, note the insufficient flattening at
 Deconvolution 217

the low-frequency end of the spectrum. A very large

prediction lag causes the amplitude spectrum of the de-
convolved data remain similar to that of the input data
(compare Figures 2.5-6c with 2.5-5c).
The data sometimes must be preconditioned for de-
convolution. If the data were too noisy, then a wide
band-pass filter could be necessary before deconvolu-
tion. If there is significant coherent noise in the data,
dip filtering (Sections 6.2 and 6.3) can be applied be-
fore deconvolution so that coherent noise is not in-
cluded in the autocorrelation estimate. Alternatively,
time-variant spectral whitening (Section 2.6) can be ap-
plied to balance the spectrum before deconvolution.

Signature Deconvolution

In marine seismic exploration, the far-field signature of

the source array can be recorded. The idea is to ap-
ply a deterministic deconvolution to remove the source
signature, then to apply predictive deconvolution. The
convolutional model is given by
x(t) = s(t) ∗ w(t) ∗ e(t), (2 − 41)
where s(t) is the source signature recorded in the far-
field just before it travels down into the earth, which
has an impulse response e(t). Since s(t) is recorded,
an inverse filter can be deterministically designed, as
discussed in Section 2.2, then applied to the recorded
seismogram to remove it from equation (2-41). The un-
known wavelet w(t) includes the propagating effects in
the earth and the response of the recording system. This FIG. 2.5-7. Signature processing. A shaping filter is de-
signed to convert the recorded signature [s(t) in equation
remaining wavelet then is removed by the statistical
(2-41)] to its minimum-phase equivalent and applied to the
method of spiking deconvolution as discussed in Sec- input record (a). The output (b) has the same bandwidth
tion 2.3. Compare equation (2-41) with equation (2-3a) as the input (a). The output (b) then has been processed by
and note that the old w(t) of equation (2-3a) is split predictive deconvolution using operator length of 160 ms,
into two parts — the source signature s(t), which is and prediction lags of (c) 4 ms (spiking deconvolution), (d)
the known component, and the new w(t), which is the 12 ms, (e) 32 ms. Examine the autocorrelograms (bottom)
unknown component. and note that those of (a) and (b) should be identical.
There are two ways to handle s(t). One way is to
convert it to its minimum-phase equivalent followed by
predictive deconvolution (Figure 2.5-7). Another way (c) Apply the shaping filter to each trace in each
is to convert s(t) into a spike followed by predictive recorded shot record.
deconvolution (Figure 2.5-8). The process involves the (d) Apply predictive deconvolution to output data
following steps: from step (c).

(a) Estimate the minimum-phase equivalent of the
recorded source signature by computing the spiking
deconvolution operator (equation 2-39) and taking
its inverse.
(b) Design a shaping filter to convert the source sig-
nature to its minimum-phase equivalent or a zero-
delay spike (equation 2-30).
The results shown in Figures 2.5-7 and 2.5-8 (pan-
els (c)) should be compared with single-step statistical
deconvolution (Figure 2.5-9). Since the source was not
minimum-phase in this case, Figure 2.5-9b should be
better than Figure 2.5-9d. Is it?
Actually, results of signature processing depend
on the accuracy of the recorded signature. One should
218 Seismic Data Analysis
FIG. 2.5-8. Signature processing. A shaping filter is de-
signed to convert the recorded signature (s(t) in equation
2-41) to a spike and applied to the input record (a). The
output (b) then has been processed by predictive deconvo-
lution using an operator length of 160 ms and prediction
lags of (c) 4 ms (spiking deconvolution), (d) 12 ms, and (e)
32 ms. Note from the autocorrelograms (bottom) that the
output from signature processing (b) still contains a wavelet
component — the w(t) component in equation (2-41), that
still needs to be removed.
avoid signature processing of old marine data unless
there exists a concurrently recorded source signature.
Contemporary marine data almost always include the
recorded source signature at each shot location. Fig-
ure 2.5-10a shows a recorded water-gun signature with
its amplitude and phase spectra. The minimum-phase
equivalent is shown in Figure 2.5-10b, and the result of
signature deconvolution to convert the recorded wave-
form to its minimum-phase equivalent is shown in Fig-
ure 2.5-10c. While the amplitude spectrum is unaltered,
the phase spectrum is minimum phase.
Application of signature processing described in
Figure 2.5-10 to a recorded shot gather is shown in
Figure 2.5-11. Again, note that signature processing
FIG. 2.5-9. Signature processing compared with statistical
deconvolution. (a) Input shot record. (b) A shaping filter is
designed to convert the recorded signature (s(t) in equation
2-41) to its minimum-phase equivalent and applied to the
input record followed by spiking deconvolution. (This panel
is the same as Figure 2.5-7c.) (c) A shaping filter is designed
to convert the recorded signature (s(t) in equation 2-41) to
a spike and applied to the input record followed by spiking
deconvolution. (This panel is the same as Figure 2.5-8c.) (d)
Spiking deconvolution of the input record (a). The autocor-
relograms (bottom) suggest that the wavelet compression is
achieved in all three cases, (b), (c) and (d).
aimed at converting the recorded source signature to its
minimum-phase equivalent leaves the amplitude spec-
trum and autocorrelogram of the shot record unaltered
(Figures 2.5-11a,b). We can also observe from the cor-
responding CMP-stacked sections that the process has
not made any impact on the degree of vertical resolu-
tion; only a change in phase has taken place (Figure
2.5-12). Following the deterministic step of Figure 2.5-
11b, statistical deconvolution is applied to flatten the
spectrum (Figure 2.5-11c).
 Deconvolution
FIG. 2.5-10. (a) A water-gun source signature and its am-
plitude and phase spectra (top and middle graphs), (b) its
minimum-phase equivalent, and (c) the ouput of signature
processing to convert (a) to (b).
Vibroseis Deconvolution
The vibroseis source is a long-duration sweep signal in
the form of a frequency-modulated sinusoid that is ta-
pered on both ends. Just as a convolutional model was
proposed for the marine seismogram given by equation
(2-41), a similar convolutional model can be proposed
for the vibroseis seismogram
x(t) = s(t) ∗ w(t) ∗ e(t), (2 − 42)
where x(t) is the recorded seismogram, s(t) is the sweep
signal, w(t) is the seismic wavelet with the same mean-
ing as in equation (2-41), and e(t) is the earth’s impulse
response. Convolutions in equation (2-42) become mul-
tiplications in the frequency domain:
X(ω) = S(ω) W (ω) E(ω). (2 − 43)
219
In terms of amplitude A(ω) and phase φ(ω) spec-
tra, equation (2-43) yields
Ax (ω) = As (ω) Aw (ω) Ae (ω) (2 − 44a)
and
φx (ω) = φs (ω) + φw (ω) + φe (ω). (2 − 44b)
Crosscorrelation of the recorded seismogram x(t)
with the sweep signal s(t) is equivalent to multiplying
equation (2-44a) by As (ω) and subtracting φs (ω) from
equation (2-44b). The correlated vibroseis seismogram
x (t) therefore would have the following amplitude and
phase spectra:
A (ω) = A2s (ω) Aw (ω) Ae (ω) (2 − 45a)
and
φ (ω) = φw (ω) + φe (ω). (2 − 45b)
The inverse Fourier transform of A2s (ω)
yields the
autocorrelation of the sweep signal, which is called the
Klauder wavelet k(t). Returning to the time domain,
equations (2-45a,b) yield
x (t) = k(t) ∗ w(t) ∗ e(t). (2 − 46)
Figure 2.5-13 outlines the process of vibroseis cor-
relation where the seismic wavelet w(t) has been omit-
ted for convenience. Note that, following vibroseis cor-
relation, the sweep s(t) contained in the recorded seis-
mogram x(t) is replaced with its autocorrelogram —
the Klauder wavelet k(t).
Since it is an autocorrelation, the Klauder wavelet
is zero-phase. Convolution of k(t) with the assum-
ingly minimum-phase wavelet w(t) yields a mixed-phase
wavelet. Because spiking deconvolution is based on
the minimum-phase assumption, it cannot recover e(t)
properly from vibroseis data.
One approach to deconvolution of vibroseis data
is to apply a zero-phase inverse filter to remove k(t),
followed by a minimum-phase deconvolution to remove
w(t). The amplitude spectrum of the inverse filter is
defined as 1/ A2s (ω). In practice, problems arise because
of zeroes in the spectrum that are caused by the band-
limited nature of the Klauder wavelet. Inversion of an
amplitude spectrum, which has zeroes, yields an un-
stable operator (Section 2.3). To circumvent this prob-
lem, a small percent of white noise, say 0.1%, usually is
added before inverting the Klauder wavelet spectrum.
Another approach is to design a filter that converts
the Klauder wavelet to its minimum-phase equivalent
(Ristow and Jurczyk, 1975). A technique to compute
the minimum-phase spectrum from a given amplitude
spectrum is described in Section B.4 and is included
in the discussion on frequency-domain deconvolution.
220 Seismic Data Analysis
 Deconvolution 221

FIG. 2.5-12. (a) A portion of a CMP-stacked section that corresponds to the data in Figure 2.5-11a, and (b) after signature
processing of shot records as in Figure 2.5-11b to convert the recorded water-gun signature to its minimum-phase equivalent
as described in Figure 2.5-10. Note that signature processing was not aimed at wavelet compression; instead, it was done to
convert the source signature to its minimum-phase equivalent.

If the Klauder wavelet were converted to its minimum-

FIG. 2.5-13. An outline of vibroseis correlation.
phase equivalent, then equation (2-45b) would take the
form:
φ (ω) = φk (ω) + φw (ω) + φe (ω). (2 − 47)
If we assume that w(t) is minimum-phase and if we
make k(t) minimum-phase, then the result of their con-
volution also is minimum-phase. Spiking deconvolution
now is applicable since the minimum-phase assumption
is satisfied.
There is a 90-degree phase difference in some vi-
brator systems between the control sweep signal and
the baseplate response. As an option, we may want to
subtract out this phase difference. Figure 2.5-14 shows
the recommended sequence of operations for vibroseis
processing.
222 Seismic Data Analysis
2.5-14. A flowchart for vibroseis deconvolution.
Figure 2.5-15 shows how the flowchart in Figure
2.5-14 is used with a synthetic reflectivity series. By
including the step to convert the Klauder wavelet into
its minimum-phase equivalent before spiking deconvo-
lution, a closer representation of the impulse response
is produced as seen by comparing steps (k) and (l) with
(m).
Although sound in theory, the above scheme may
have problems in practice. Fundamental issues, such as
whether the convolutional model given in equation (2-
42) really represents what goes on in the earth, are not
resolved.
Vibroseis data often are deconvolved as dynamite
data, without converting the Klauder wavelet to its
minimum-phase equivalent. An example of deconvolu-
tion of a correlated vibroseis record is shown in Figure
2.5-16.
Despite the fact that the basic minimum-phase as-
sumption is violated for vibroseis data as compared to
explosive data, spiking deconvolution without conver-
sion of the Klauder wavelet to its minimum-phase equiv-
alent seems to work for most field data. Figure 2.5-17
shows a set of correlated vibroseis records before and
after spiking deconvolution. Prominent reflections after
deconvolution are enhanced and reverberations are at-
tenuated significantly. Nonetheless, the problem of ty-
ing vibrator lines to lines recorded with other sources,
say dynamite, is more difficult if the vibrator data have
not been phase-corrected. Field systems now exist to do
minimum-phase vibroseis correlation in the field.
Poststack Deconvolution
Poststack deconvolution often is considered for sev-
eral reasons. First, a residual wavelet almost always is
present on the stacked section. This is because none of
the underlying assumptions for deconvolution is com-
pletely met in real data; therefore, deconvolution never
can completely compress the basic wavelet contained in
prestack data to a spike. Second, since a CMP stack
is an approximation to the zero-offset section, predic-
tive deconvolution aimed at removing multiples may be
a viable process after stack. Figure 2.5-18 is an exam-
ple of poststack deconvolution applied to marine data.
After deconvolution, the spectrum is flattened, albeit
incompletely, the wavelet is compressed further and the
marker horizons are better characterized. Again, as the
prediction lag is increased, the flatness character of the
spectrum and thus vertical resolution is increasingly
compromised (Figure 2.5-19).
Figure 2.5-20 shows poststack deconvolution ap-
plied to land data. Note the significant improvement
in vertical resolution as it can be verified by the au-
tocorrelogram and average amplitude spectrum of the
data.
2.6 THE PROBLEM OF
NONSTATIONARITY
Figure 2.6-1 shows a CMP gather and its filtered ver-
sions before deconvolution. The filter scans show that
there is signal between 10 to 40 Hz. Note that higher fre-
quencies are confined to the shallower part of the gather.
The same field record after spiking deconvolution is
shown in Figure 2.6-2. Filter scans of the deconvolved
record also are shown in this figure. A comparison of
the records before and after deconvolution (Figures 2.6-
1 and 2.6-2) demonstrates the effects of the process; in
particular, compression of the wavelet and broadening
of the spectrum. The input signal level above 40 Hz is
relatively weaker than that below 40 Hz. Deconvolution
has attempted to reduce the differences between the sig-
nal levels within different frequency bands by flattening
the spectrum. The flattening, however, was more effec-
tive in the shallow part of the record than in the deeper
part.
As discussed in Sections 1.4 and 2.1, the source
wavelet is not stationary — its shape and bandwidth
change with traveltime (Figure 2.1-2). Specifically, at-
tenuation of high frequencies in the wavelet increases as
waves travel deeper in the subsurface. Although multi-
window deconvolution was used in this case (Figure 2.6-
2), spectral flattening was not achieved over the entire
length of the record because of the large degree of non-
stationarity of the data. Nonstationarity is primarily
the result of the effects of wavefront divergence and fre-
quency attenuation.
Deconvolution 223
224 Seismic Data Analysis

2.5-16. Deconvolution of a vibroseis record. Three windows were used. (a) Correlated vibroseis record and its autocorrelograms
(b); (d) spiking deconvolution output and its autocorrelograms (c).

The phenomenon of nonstationarity on stacked
data is exemplified in Figure 2.6-3. Following a single-
gate poststack spiking deconvolution, note that the av-
erage amplitude spectrum of the data does not indicate
a flat spectrum (Figure 2.6-3b). Instead, the spectral be-
havior is similar to data with predictive deconvolution
with a large prediction lag (Figure 2.6-3c). The strong
reflector in the neighborhood of 2.5 s separates the zone
with two different bandwidths — a broad-band signal
zone above, and a relatively narrow-band zone below.
The usual approach to reduce nonstationarity is to
apply processes designed to compensate for the effects
wavefront divergence and frequency attenuation before
deconvolution. Wavefront divergence is corrected for by
applying a geometric spreading function (Section 1.4).
As yet, a method to compensate for attenuation has not
been discussed. Attenuation is measured by a quantity
 Deconvolution 225

2.5-17. (a) Common-shot gathers recorded with vibroseis source. Geometric spreading correction and trace balancing have
been applied. (b) After spiking deconvolution.

called the quality factor Q. An infinite Q means that
there is no attenuation. This factor can change in depth
and in the lateral direction. If we had an analytic form
for an attenuation function, then it would be easy to
compensate for its effect. Several models for Q have
been proposed. The constant-Q model is quite plausi-
ble and the easiest to deal with (Kjartannson, 1979).
However, the big problem of estimating Q from seismic
data still remains. If a reliable Q value is available, say
from borehole measurements, then, as will be discussed
later in this section, inverse Q filtering can be applied
to data to compensate for the frequency attenuation.
226 Seismic Data Analysis

2.5-18. (a) A portion of a CMP-stacked section, and after spiking deconvolution using an operator length of: (b) 120 ms,
(b) 160 ms, (c) 220 ms, and (d) 320 ms. Note from the autocorrelograms (bottom) that much of the reverberating energy
is attenuated using a 320-ms operator length. The amplitude spectra averaged over the CMP stack (top) indicate that,
irrespective of how long the operator length is, spiking deconvolution in this case has failed to flatten the spectrum completely
within the passband. This is because of nonstationarity of the signal.

2.5-19. A portion of a CMP-stacked section as in Figure 2.5-7a after predictive deconvolution using an operator length of 320
ms and a prediction lag of: (a) 8 ms, (b) 12 ms, (c) 24 ms, (d) 32 ms, and (e) 48 ms. The amplitude spectra (top) averaged
over the CMP stack, and the autocorrelograms (bottom) are used to to choose deconvolution parameters and evaluate the
data after the application of deconvolution.
 Deconvolution
2.5-20. (a) A portion of a CMP-stacked section, and (b)
after spiking deconvolution using an operator length of 240
ms. The amplitude spectra (top) averaged over the CMP
stack, and the autocorrelograms (bottom) are used to choose
deconvolution parameters and evaluate the data after the
application of deconvolution.
Time-Variant Deconvolution
Nonstationarity was discussed in Sections 1.4 and 2.1.
The time-variant character of the seismic wavelet (Fig-
ure 2.1-2) often requires a multiwindow deconvolution.
Figure 2.6-4 is a field record that was deconvolved
227
2.6-1. A field record (far left panel) with its band-pass fil-
tered versions.
2.6-2. Spiking deconvolution applied to the field record in
Figure 2.6-1 (far left panel), followed by application of a
series of band-pass filters.
by using three time gates. The autocorrelograms from
gates 1, 2, and 3 are shown in Figure 2.6-5. Note the
difference in character of the reverberatory energy from
one gate to another. The shallow gate (1) has more high-
frequency signal than the middle gate (2); while the
middle gate has more high-frequency signal than the
deeper gate (3). For best results, we must design differ-
ent deconvolution operators from different parts of the
record and apply them to the corresponding time gates.
Up to three windows usually are sufficient to handle the
nonstationary character of the seismic signal.
228 Seismic Data Analysis

2.6-3. (a) A portion of a CMP-stacked section with its amplitude spectrum averaged over the CMP range (top) and autocor-
relogram (bottom); (b) after time-invariant spiking deconvolution; and (c) after time-invariant predictive deconvolution with
a prediction lag of 24 ms.
 Deconvolution 229

2.6-4. Three-window deconvolution. The solid bars indicate window boundaries. With data from each window, a deconvolution
operator is designed and applied to the data in that window. The operators are blended across the window boundaries. (a)
Input gather. Deconvolution using operator length of 160 ms and prediction lags of (b) 4 ms, (c) 12 ms, (d) 32 ms.

Another example of single- and multiwindow de-
convolution is shown in Figures 2.6-6 and 2.6-7. Here,
autocorrelograms from different gates do not show sig-
nificant variations. Therefore, it probably does not make
any difference whether a single or multigate deconvo-
lution is used. In Figures 2.6-6 and 2.6-7, the record
is shown after deconvolution followed by a wide band-
pass filter application. Since the amplitude spectrum of
the input data is flattened as a result of spiking de-
convolution, both the high-frequency ambient noise as
much as the high-frequency components of the signal
are boosted. Therefore, the output of spiking deconvo-
lution often is filtered with a wide band-pass operator.
A practical problem with time-variant deconvolu-
tion is limiting design gates to small time windows. Con-
sider, for instance, a three-window deconvolution of a 5-
s data. This means that at best an average gate kength
of 1.5 s at near offset and less than 1 s at far offset can
be used to design a deconvolution operator. To attain
good statistics in an autocorrelation estimate, an oper-
ator length no more than one-eighth to one-tenth of the
design gate, say 150 ms, should be considered. Hence, if
230 Seismic Data Analysis
 Deconvolution 231

2.6-6. Spiking deconvolution (c) on a shot record (a) followed by band-pass filtering (d). (b) Autocorrelograms before and
after deconvolution.

you need to use a longer operator, time-variant decon-
volution may have limited effectiveness in attenuating
reverberations and short-period multiples. A way to
account for nonstationarity while avoiding the short-
operator effect of multiwindow deconvolution follows.
Time-Variant Spectral Whitening
Frequency attenuation and a way to compensate for it
are illustrated in Figure 2.6-8. Let us assume that we
have an input seismogram with amplitudes decaying in
232 Seismic Data Analysis

FIG. 2.6-7. Three-window deconvolution on the same shot record as in Figure 2.6-6. In this case, there is no significant dif-
ference between the characters of the autocorrelograms estimated from three windows. (a) Input gather; (b) autocorrelograms
before and after spiking deconvolution; (c) three-window spiking deconvolution on (a); (d) band-pass filtering on (c).

time, as depicted. Now apply a series of narrow band-
pass filters to this trace. Examine the field record in Fig-
ure 2.6-1 and associate the filter panels with the traces
sketched in Figure 2.6-8. Note that the low-frequency
component of trace FL has a lower decay rate than
the moderate-frequency component FM . Likewise, the
moderate-frequency component FM has a lower decay
rate than the high-frequency component of the signal
FH . A series of gain functions, such as G1 , G2 , G3 , can
be computed to describe the decay rates for each fre-
quency band. This is done by computing the envelope
of the band-pass filtered traces (Figure 2.6-8). The in-
verses of these gain functions then are applied to each
frequency band and the results are summed. The am-
plitude spectrum of the resulting trace has thus been
whitened in a time-variant manner. This time-variant
spectral whitening process is outlined in Figure 2.6-
9. The number of the filter bands, the width of each
band, and the overall bandwidth of application of time-
variant spectral whitening are parameters that can be
prescribed for a particular application.
 Deconvolution
2.6-8. A schematic illustration of the rate of decay of the
frequencies in a seismic trace (Gibson and Larner, 1982).
Figures 2.6-10 and 2.6-11 show some field records
before and after spiking deconvolution, respectively.
Note that this process not only has compressed the
wavelet, but also has tried to suppress any reverber-
ations in the data. Refer to Figure 2.6-12 and note that
time-variant spectral whitening mainly has compressed
the wavelet without changing much of the ringy char-
acter of the data. Also note that little was done explic-
itly to the phase. Therefore, the action of time-variant
spectral whitening may be close to a zero-phase decon-
volution, although there is no rigorous theoretical proof
of this.
In practice, one of the main differences between
time-variant spectral whitening and conventional de-
convolution is that the former seems to be able to do
a better job of flattening the amplitude spectrum. This
difference can be significant for broad-band data with
large dynamic range.
Time-variant spectral whitening sometimes helps
attenuate ground roll on land records by way of its
spectral balancing effect. Note that, in Figure 2.6-13,
spiking deconvolution with different operator lengths
has failed to flatten the spectrum, sufficiently. On the
other hand, following spiking deconvolution, applica-
tion of time-variant spectral whitening has balanced the
spectrum and thus attenuated the ground-roll energy.
The ability of time-variant spectral whitening in
flattening the spectrum within the passband of stacked
data is observed in Figure 2.6-14. Note that spiking
deconvolution is fairly effective, but not sufficient, for
attaining a flat spectrum (Figure 2.6-14b). Following
233
2.6-9. A flowchart for time-variant spectral whitening.
time-variant spectral whitening, the spectrum is flat-
tened within the passband of the data as seen in Figure
2.6-14c.
It is a requirement to prepare stacked data input
to amplitude inversion with the broadest possible band-
width and flattest possible spectrum. Hence, a process-
ing sequence tailored for amplitude inversion almost al-
ways includes poststack deconvolution and time-variant
spectral flattening steps.
Frequency-Domain Deconvolution
Spectral flattening can be achieved by an alternate ap-
proach in the frequency domain. As discussed in Section
B.4, minimum-phase spiking deconvolution can be for-
mulated in the frequency domain. Alternatively, we can
flatten the amplitude spectrum without modifying the
phase. This is called zero-phase frequency-domain de-
convolution. When performed over multiple time gates
down the trace, it is essentially equivalent to time-
variant spectral flattening. If we only want to flatten
the spectrum, then the approach shown in Figure 2.6-
15 can be taken.
Although the domains of operations may differ,
both minimum-phase frequency-domain and Wiener-
Levinson deconvolution techniques should yield equiv-
alent results. Differences between the results shown in
Figures 2.6-11 and 2.6-16 mainly are due to their com-
putational aspects.
The zero-phase frequency-domain deconvolution
aimed at achieving time-variant spectral whitening re-
quires partitioning the input seismogram into small
time gates, as well as designing and applying the process
described in Figure 2.6-15 to each gate, individually.
Figure 2.6-17 shows the field records after zero-phase
frequency-domain deconvolution. The output is com-
parable to the time-variant spectral flattening output
shown in Figure 2.6-12.
234 Seismic Data Analysis

FIG. 2.6-10. CMP gathers with no deconvolution.

Inverse Q Filtering Wave attenuation usually is described by a dimen-

Frequency attenuation caused by the intrinsic attenu-
ation in rocks was discussed in Sections 1.4 and 2.1.
Attenuation causes loss of high frequencies in the prop-
gating waveform with increasing traveltime. This gives
rise to a nonstationary behavior in the shape of the
wavelets associated with reflection events at different
times.
sionless factor Q, which is defined by the ratio of the
mean stored energy to the energy loss over a period
of time that is equivalent to one cycle of a frequency
component of the waveform (Kjartansson, 1979). Time-
variant deconvolution and time-variant spectral whiten-
ing discussed in this section are processes that can cor-
rect for the time-varying effects of attenuation by spec-
tral flattening. A deterministic alternative to compen-
 Deconvolution 235

FIG. 2.6-11. Spiking deconvolution applied to the CMP gathers in Figure 2.6-10.

sate for frequency-dependent attenuation is provided by where ω is the angular frequency component associated
inverse Q filtering. with the input trace and τ is the time variable asso-
The amplitude spectrum of the inverse Q filter is ciated with the output trace from inverse Q filtering.
The phase spectrum of the inverse Q filter usually is as-
given by (Section B.9)
sumed to be minimum-phase, which can be computed
ωτ by taking the Hilbert transform of the amplitude spec-
A(ω, τ ) = exp( ), (2 − 48)
2Q trum (Section B.4).
236 Seismic Data Analysis

FIG. 2.6-12. Time-variant spectral whitening (TVSW) applied to the CMP gathers in Figure 2.6-10. Compare this with
Figures 2.6-11, 2.6-16, and 2.6-17.
Deconvolution 237
238 Seismic Data Analysis

FIG. 2.6-14. (a) A portion of a CMP stack and its amplitude spectrum averaged over the CMP range (top) and auto-
correlogram (bottom); (b) after spiking deconvolution, (c) followed by time-variant spectral whitening, and (d) time-variant
filtering.

Application of the inverse Q filter requires knowl-

edge of the attenuation factor Q, which usually is as-
sumed to be constant. A compilation of laboratory mea-
surements of Q for some rock samples is given by Table
2-28.
Note from Table 2-28 that most measurements have
been made at extremely high frequencies compared to
the typical frequency band of seismic waves. Never-
theless, by assuming frequency-independent Q factor
(Kjartansson, 1979), these measurements can still be
considered useful. Also note that the Q factor can vary
significantly for limestone, sandstone, and shale rock
samples of different origin.
The inverse of the amplitude spectrum defined by
FIG. 2.6-15. A flowchart for a frequency-domain deconvo- equation (2-48) can be used to obtain a quantitative
lution. measure of attenuation. In terms of frequency f , wave
 Deconvolution 239

FIG. 2.6-16. Minimum-phase frequency-domain deconvolution applied to the CMP gathers in Figure 2.6-10. Compare this
with Figures 2.6-11, 2.6-12, and 2.6-17.
240 Seismic Data Analysis

FIG. 2.6-17. Zero-phase frequency-domain deconvolution applied to the CMP gathers in Figure 2.6-10. Compare this with
Figures 2.6-11, 2.6-12, and 2.6-16.
 Deconvolution 241

Table 2-28. Intrinsic attenuation measurements in Note that the smaller the Q factor the shallower
rocks (adapted from Waters, 1981). the depth at which the amplitude drops to the specified
value of one-tenth of the original value at the surface
Rock Type Attenuation Frequency z = 0. For very large Q values, that is, for small atten-
Constant, Q Range (Hz) uation, the amplitude reduction to the specified value
does not take place until the wave reaches very large
Basalt 550 3,000-4,000
depths beyond the exploration objectives.
Granite 300 20,000-200,000
Unfortunately, there is no reliable way to estimate
Quartzite 400 3,000-4,000
the attenuation factor Q directly from seismic data.
Limestone I 200 10,000-15,000
Limestone II 50 2-120 At best, inverse Q filtering can be applied to post- or
Limestone III 650 4-18,000 prestack data (Section B.9) using a range of constant Q
Chalk 2 150 factors to create a Q panel, much like a filter panel (Sec-
Sandstone I 25 550-4,000 tion 1.1). The factor that yields the flattest spectrum
Sandstone II 125 20,000 in combination with other signal processing applications
Sandstone III 75 2,500-5,000 — deconvolution and time-variant spectral whitening,
Sandstone IV 100 2-40 is chosen as the optimum Q value.
Shale I 15 75-550
Shale II 75 3,300-12,800 Deconvolution Strategies
Throughout the development of deconvolution theory,
several alternatives have been proposed to better solve
velocity v and depth z = vτ , the inverse is the deconvolution problem. Still, predictive deconvolu-
tion is used more than the other methods, although
πf z
A−1 (f, v, z, Q) = exp(− ). (2 − 49) the minimum-phase and white reflectivity sequence as-
Qv
sumptions have been key issues of concern.
To determine how far in depth the wave has to travel Follow the common sequence for deconvolution of
before its amplitude reduces to, say, one-tenth of its marine data in Figures 2.6-18 through 2.6-22. Note the
amplitude at the surface z = 0, rewrite equation (2-49) presence of reverberations and short-period multiples
as follows: in the CMP-stacked data with no deconvolution ap-
2.3Qv plied (Figure 2.6-18). Signature processing, in this case,
z= . (2 − 50)
πf was done to convert the recorded source signature to
Note that the smaller the Q factor, the lower the its minimum-phase equivalent (Figure 2.6-19). There-
velocity and the higher the frequency, the shallower the fore, aside from phase, there is no difference between
depth at which the wave amplitude decays to a fraction the sections in Figures 2.6-18 and 2.6-19. Deconvolu-
of the wave amplitude at z = 0. Table 2-29 lists the z tion before stack has helped attenuate reverberations
values for a frequency of 30 Hz and a velocity of 3000 and short-period multiples and, to some extent, com-
m/s for a range of Q values. pressed the basic wavelet (2.6-20). The additonal step of
poststack deconvolution has restored much of the high
frequencies attenuated during stacking (Figure 2.6-21).
Table 2-29. Depth at which wave amplitude drops to Finally, time-variant spectral whitening has flattened
one-tenth of its original at the surface for a range of Q the spectrum within the passband of the data (Figure
values (equation 2-50). 2.6-22) and yielded a crisp section with high resolution.
v = 3000 m/s The same sequence can be followed in Figures 2.6-
f = 30 Hz 23 through 2.6-27. The CMP-stacked section includes
reflections associated with a shallow, low-relief sedimen-
Q Factor Depth in m tary strata (Figure 2.6-23). Following the signature pro-
cessing (Figure 2.6-24), observe the gradual increase in
25 1,830
the vertical resolution by prestack deconvolution (Fig-
50 3,660
ure 2.6-25), poststack deconvolution (Figure 2.6-26) and
100 7,325
time-variant spectral whitening (Figure 2.6-27).
250 18,312
500 36,625 The following formal processing sequence for de-
convolution theoretically should yield optimum results:
(text continues on p. 247)
242 Seismic Data Analysis

2.6-18. A portion of a CMP-stacked section with no deconvolution.

2.6-19. The section in Figure 2.6-18 with signature processing to convert the recorded source signature to its minimum-phase
equivalent.
 Deconvolution 243

2.6-20. The section in Figure 2.6-19 with prestack spiking deconvolution.

2.6-21. The section in Figure 2.6-20 with poststack spiking deconvolution.

244 Seismic Data Analysis

2.6-22. The section in Figure 2.6-21 with time-variant spectral whitening.

2.6-23. A portion of a CMP-stacked section with no deconvolution.

 Deconvolution 245

2.6-24. The section in Figure 2.6-23 with signature processing to convert the recorded source signature to its minimum-phase
equivalent.

2.6-25. The section in Figure 2.6-24 with prestack spiking deconvolution.

246 Seismic Data Analysis

2.6-26. The section in Figure 2.6-25 with poststack spiking deconvolution.

2.6-27. The section in Figure 2.6-26 with time-variant spectral whitening.

 Deconvolution
(a) Apply a geometric spreading compensation func-
tion to remove the amplitude loss due to wavefront
divergence.
(b) Apply an exponential gain or minimum-phase in-
verse Q filter (Hale, 1982; Hargreaves and Calvert,
1991). This compensates for frequency attenuation.
(c) Optionally apply signature processing to marine
data. For vibroseis data, apply the filter that con-
verts the Klauder wavelet to its minimum-phase
equivalent.
(d) Apply predictive deconvolution to compress the
basic wavelet and attenuate reverberations and
short-period multiples. If required, apply surface-
consistent deconvolution (Section B.8). This ac-
counts for the near-surface variations effect on the
wavelet because of inhomogeneities in the vicinity
of sources and receivers. In step (b), the vertical
variations effect on the wavelet is handled.
(e) Apply predictive deconvolution after stack to
broaden the spectrum and attenuate short-period
multiples.
(f) Apply time-variant spectral whitening. This pro-
vides further flattening of the spectrum within the
signal bandwidth without affecting phase.
The idea is to do as much deterministic deconvolu-
tion as possible. Inverse Q filtering, signature deconvo-
lution, and the filter that converts the Klauder wavelet
to its minimum-phase equivalent are deterministic op-
erators. Any remaining issues then are handled by sta-
tistical means. However, for most data cases, just doing
the geometric spreading correction followed by predic-
tive prestack and poststack deconvolution is adequate.
EXERCISES
Exercise 2-1. Write the z-transform of wavelet
(1, 0, − 14 ). Design a three-term inverse filter and apply
it to the original. Hint: The z-transform of the wavelet
can be written as a product of two doublets, (1, − 12 )
and (1, 12 ).
Exercise 2-2. Consider the following set of
wavelets:
Wavelet A: (3,-2,1)
Wavelet B: (1,-2,3)
(a) Plot the percent of cumulative energy as a func-
tion of time for each wavelet. Use Robinson’s en-
ergy delay theorem to determine the minimum- and
maximum-phase wavelet.
(b) Set up matrix equation (2-31) for each wavelet,
compute the spiking deconvolution operators, then
apply them.
247
(c) Let the desired output be (0, 0, 1, 0). Set up ma-
trix equation (2-30) for each wavelet, compute the
shaping filters, and apply them. Find that the error
for wavelet B with the delayed spike is smaller.
Exercise 2-3. Consider wavelet A in Exercise 2-
2. Set up matrix equation (2-32), where ε = 0.01, 0.1.
Note that ε = 0 already is assigned in Exercise 2-2. As
the percent prewhitening increases, the spikiness of the
deconvolution output decreases.
Exercise 2-4. Consider a multiple series associ-
ated with a water bottom with a reflection coefficient
cw and two-way time tw . Design an inverse filter to
suppress the multiples. [This is called the Backus fil-
ter (Backus, 1959)].
Exercise 2-5. Consider an earth model that com-
prises a water-bottom reflector and a deep reflector at
two-way times of 500 and 750 ms, respectively. What
prediction lag and operator length should you choose
to suppress (a) water-bottom multiples, and (b) peg-
leg multiples?
Exercise 2-6. Refer to Figure 2-6.9. Consider the
following three bandwidths — low (FL ), medium (FM )
and high (FH ), for TVSW application:
FL : 10 to 30 Hz
FM : 30 to 50 Hz
FH : 50 to 70 Hz
What kind of slopes should you assign to each band-
width so that the output trace has an amplitude spec-
trum that is unity over the 10-to-70-Hz bandwidth?
Exercise 2-7. If the signal character down the
trace changes rapidly (strong nonstationarity), should
you consider narrow or broad bandwidths for the filters
used in TVSW?
Exercise 2-8. Consider a minimum-phase wavelet
and the following two processes applied to it:
(a) Spiking deconvolution followed by 10-to-50-Hz
zero-phase band-pass filtering.
(b) Shaping filter to convert the minimum-phase
wavelet to a 10-to-50-Hz zero-phase wavelet.
What is the difference between the two outputs?
Exercise 2-9. How would you design a minimum-
phase band-pass filter operator?
Exercise 2-10. Consider (a) convolving a
minimum-phase wavelet with a zero-phase wavelet, (b)
convolving a minimum-phase wavelet with a minimum-
phase wavelet, and (c) adding two minimum-phase
wavelets. Are the resulting wavelets minimum-phase?
Exercise 2-11. Consider the sinusoid shown in
Figure 1-1 (frame 1) as input to spiking deconvolution.
What is the output?
Exercise 2-12. Order the panels in Figure 2.E-1
with increasing prediction lag.
248 Seismic Data Analysis

2.E-1. The shot record shown in Figure 2.4-36c after predictive deconvolution using an operator length of 480 ms, and four
different prediction lags (Exercise 2-12). The amplitude spectra averaged over the shot record are shown at the top and the
autocorrelograms are shown at the bottom.
 Appendix B
MATHEMATICAL FOUNDATION OF DECONVOLUTION

B.1 Synthetic Seismogram

Consider an earth model that consists of homogeneous horizontal layers with thicknesses cor-
responding to the sampling interval. The seismic impedance associated with a layer is defined
as I = ρv, where ρ is density and v is the compressional-wave velocity within the layer. The
instantaneous value of seismic impedance for the kth layer is given by
Ik = ρk vk . (B − 1)
For a vertically incident plane wave, the pressure amplitude reflection coefficient associated
with an interface is given by
Ik+1 − Ik
ck = . (B − 2)
Ik+1 + Ik
Assume that the change of density with depth is negligible compared to the change of velocity
with depth. Equation (B-2) then takes the form
vk+1 − vk
ck = . (B − 3)
vk+1 + vk
If the amplitude of the incident wave is unity, then the magnitude of the reflection coefficient
corresponds to the fraction of amplitude reflected from the interface.
With knowledge of the reflection coefficients, we can compute the impulse response of a
horizontally layered earth model using the Kunetz method (Claerbout, 1976). The impulse
response contains not only the primary reflections, but also all the possible multiples. Finally,
convolution of the impulse response with a source wavelet yields the synthetic seismogram.
Although random noise can be added to the seismogram, the convolutional model used here
to establish the deconvolution filters does not include random noise. We also assume that the
source waveform does not change as it propagates down into the earth; hence, the convolutional
model does not include intrinsic attenuation. The convolution of a seismic wavelet w(t) with the
impulse response e(t) yields the seismogram x(t)
x(t) = w(t) ∗ e(t). (B − 4)
By Fourier transforming both sides, we get
X(ω) = W (ω) E(ω), (B − 5)
where X(ω), W (ω), and E(ω) represent the complex Fourier transforms of the seismogram,
the source waveform, and the impulse response, respectively. In terms of amplitude and phase
spectra, the Fourier transforms are expressed as
X(ω) = Ax (ω) exp[iφx (ω)], (B − 6a)

W (ω) = Aw (ω) exp[iφw (ω)], (B − 6b)

and
E(ω) = Ae (ω) exp[iφe (ω)]. (B − 6c)
250 Seismic Data Analysis

By substituting equations (B-6) into equation (B-5), we have

Ax (ω) = Aw (ω) Ae (ω) (B − 7a)
and
φx (ω) = φw (ω) + φe (ω). (B − 7b)
Hence, in convolving the seismic wavelet with the impulse response, their phase spectra are
added, while the amplitude spectra are multiplied.
We assume that the earth’s impulse response can be represented by a white reflectivity
series; hence, its amplitude spectrum is flat
Ae (ω) = A0 = constant. (B − 8)
By substituting into equation (B-7a), we obtain
Ax (ω) = A0 Aw (ω). (B − 9)
This implies that the amplitude spectrum of the seismogram is a scaled version of the amplitude
spectrum of the source wavelet.
We now examine the autocorrelation functions r(τ ) of x(t), w(t), and e(t). Autocorrelation
is a measure of similarity between the events on a time series at different time positions. It is a
running sum given by the expression
N −1
1
re (τ ) = et et+τ , τ = 0, ∓1, ∓2, . . . , ∓(N − 1), (B − 10)
N t=0

where τ is time lag. A random time series is an uncorrelated series. (Strictly, it is uncorrelated
when it is continuous and infinitely long.) Therefore,
re (τ ) = 0, τ =0 (B − 11a)
and
re (0) = r0 = constant. (B − 11b)
Equation (B-11) states that the autocorrelation of a perfect random series is zero at all lags
except at zero lag. The zero-lag value actually is the cumulative energy contained in the time
series:
r0 = e20 + e21 + · · · + e2N −1 . (B − 12)
Consider the z-transform of the convolutional model in equation (B-4):
X(z) = W (z) E(z). (B − 13)
By putting 1/z in place of z and taking the complex conjugate, we get
X(1/z) = W (1/z) E(1/z), (B − 14)
where the bar denotes the complex conjugate. By multiplying both sides of equations (B-13)
and (B-14), we get
X(z) X(1/z) = W (z) E(z) W (1/z) E(1/z) . (B − 15)
By rearranging the right side,
X(z) X(1/z) = W (z) W (1/z) E(z) E(1/z) . (B − 16)
Finally, by definition, equation (B-16) yields
r x = rw ∗ r e , (B − 17)
 Deconvolution 251

where rx , rw , and re are the autocorrelations of the seismogram, seismic wavelet, and impulse
response, respectively. Based on the white reflectivity series assumption (equation B-11), we
have
r x = r0 rw . (B − 18)
Equation (B-18) states that the autocorrelation of the seismogram is a scaled version of that
of the seismic wavelet. We will see that conversion of the seismic wavelet into a zero-lag spike
requires knowledge of the wavelet’s autocorrelation. Equation (B-18) suggests that the autocor-
relation of the seismogram can be used in lieu of that of the seismic wavelet, since the latter
often is not known.

B.2 The Inverse of the Source Wavelet

A basic purpose of deconvolution is to compress the source waveform into a zero-lag spike so
that closely spaced reflections can be resolved. Assume that a filter operator f (t) exists such
that
w(t) ∗ f (t) = δ(t), (B − 19)
where δ(t) is the Kronecker delta function. The filter f (t) is called the inverse filter for w(t).
Symbolically, f (t) is expressed in terms of the seismic wavelet w(t) as
1
f (t) = δ(t) ∗ . (B − 20)
w(t)
The z-transform of the seismic wavelet with a finite length m + 1 is (Appendix A)
W (z) = w0 + w1 z + w2 z 2 + · · · + wm z m . (B − 21)
The z-transform of the inverse filter can be obtained by polynomial division:
1
F (z) = . (B − 22)
W (z)
The result is another polynomial whose coefficients are the terms of the inverse filter
F (z) = f0 + f1 z + f2 z 2 + · · · + fn z n + · · · . (B − 23)
Note that the polynomial F (z) in equation (B-23) has only positive powers of z; this means
f (t) is causal. If the coefficients of F (z) asymptotically approach zero as time goes to infinity,
so that the filter has finite energy, we say that the filter f (t) is realizable. If the coefficients
increase without bound, we say that the filter is not realizable. In practice, we prefer to work
with a causal and realizable filter. Such a filter, by definition, also is minimum-phase. If f (t) is
minimum-phase, then the seismic wavelet w(t) also must be minimum-phase. Finally, to apply
the filter with a finite length n + 1, the polynomial F (z) must be truncated. Truncation of the
filter operator induces some error in spiking the seismic wavelet.
The inverse of the seismic wavelet also can be computed in the frequency domain. By Fourier
transforming equation (B-19), we get
W (ω) F (ω) = 1. (B − 24)
By substituting equation (B-6b), we obtain
1
F (ω) = . (B − 25)
Aw (ω) exp [iφw (ω)]
We express the Fourier transform of the inverse filter F (ω) as
F (ω) = Af (ω) exp[iφf (ω)], (B − 26)
252 Seismic Data Analysis

and compare it with equation (B-25) to get

1
Af (ω) = (B − 27a)
Aw (ω)
and
φf (ω) = −φw (ω). (B − 27b)
Equations (B-27) show that the amplitude spectrum of the inverse filter is the inverse of that
of the seismic wavelet, and the phase spectrum of the inverse filter is negative of that of the
seismic wavelet.

B.3 The Inverse Filter

Instead of the polynomial division procedure described by equation (B-22), consider a different
approach to derive the inverse filter. Start with the z-transform of the autocorrelation of the
seismic wavelet:
Rw (z) = W (z) W (1/z), (B − 28)
and the z-transform of equation (B-19):
W (z) F (z) = 1, (B − 29)
from which we get
1
W (z) = . (B − 30)
F (z)
By substituting into equation (B-28), we obtain
Rw (z) F (z) = W (1/z). (B − 31)
Since the wavelet is a real-time function,
rw (τ ) = rw (−τ ). (B − 32)
Consider the special case of a three-point inverse filter (f0 , f1 , f2 ). We will assume that the
seismic wavelet w(t) is minimum-phase (causal and realizable); hence, its inverse f (t) also is
minimum-phase. The z-transform of f (t) is
F (z) = f0 + f1 z + f2 z 2 . (B − 33a)
The z-transform of its autocorrelogram rw (τ ) is, by way of equation (B-32),
Rw (z) = · · · + r2 z −2 + r1 z −1 + r0 + r1 z + r2 z 2 + · · ·
(B − 33b)

The z-transform of w(t) is

W (z) = w0 + w1 z + w2 z 2 + · · · + wm z m ,
therefore,
W (1/z) = w0 + w1 z + w2 z 2 + · · · + wm z m .
(B − 33c)

By substituting equations (B-33a), (B-33b), and (B-33c) into equation (B-31), we obtain
r2 z −2 + r1 z −1 r0 + r1 z + r2 z 2 f0 + f1 z + f2 z 2 = w0 + w1 z −1 + w2 z −2 . (B − 34)
0
To solve for (f0 , f1 , f2 ), we identify the coefficients of powers of z. The coefficient of z yields
r0 f0 + r1 f1 + r2 f2 = w0 ,
 Deconvolution 253

the coefficient of z 1 yields

r1 f0 + r0 f1 + r1 f2 = 0,

while the coefficient of z 2 yields

r2 f0 + r1 f1 + r0 f2 = 0.

When put into matrix form, these equations for the coefficients of powers of z yield

     
r0 r1 r2 f0 w0
 r1 r0 r1   f1  =  0  . (B − 35)
r2 r1 r0 f2 0

Note that w0 , which equals w0 for the usual case of a real source wavelet, is the amplitude of
the wavelet at t = 0. There are four unknowns and three equations. By normalizing with respect
to f0 , we get
     
r 0 r1 r 2 1 L
 r1 r0 r1   a1  =  0  , (B − 36a)
r 2 r1 r 0 a2 0

where a1 = f1 / f0 , a2 = f2 / f0 , and L = w0 / f0 . We now have three unknowns, a1 , a2 , and L,

and three equations. The square matrix elements on the left side of the equation represent the
autocorrelation lags of the seismic wavelet, which we do not know. However, the autocorrelation
lags from equation (B-18) of the seismogram that we do know can be substituted.
For the general case of an n-point inverse filter (f0 , f1 , f2 , . . . , fn ), equation (B-36a) takes
the form
r r1 r2 ··· rn   1   L 
0
 r1 r0 r1 · · · rn−1   a1   0 
    
 r2 r1 r0 · · · rn−2   a2  =  0  . (B − 36b)
 . . . .. .   .  .
 . .. .. . ..   ..   .. 
.
rn rn−1 rn−2 · · · r0 an−1 0

The autocorrelation matrix in equation (B-36b) is of a special type. First, it is a symmetric

matrix; second, its diagonal elements are identical. This type of matrix is called a Toeplitz
matrix. For n number of normal equations, the standard algorithms require a memory space
that is proportional to n2 and CPU time that is proportional to n3 . Because of the special
properties of the Toeplitz matrix, Levinson devised a recursive scheme (Claerbout, 1976) that
requires a memory space and CPU time proportional to n and n2 , respectively (Section B.6).

B.4 Frequency-Domain Deconvolution

We want to estimate a minimum-phase wavelet from the recorded seismogram. The inverse of the
wavelet is the spiking deconvolution operator. We start with the autocorrelation of the seismic
wavelet w(t) in the frequency domain:

Rw (ω) = W (ω) W (ω), (B − 37)

254 Seismic Data Analysis

where W (ω) is the complex conjugate of the Fourier transform of w(t). Since w(t) normally is
not known, Rw (ω) is not known. However, based on an assumption of white reflectivity series
(equation B-18), we can substitute the autocorrelation function of the seismogram in equation
(B-37).
Define a new function U (ω):

U (ω) = ln Rw (ω) . (B − 38)

When both sides of equation (B-38) are exponentiated:

Rw (ω) = exp U (ω) . (B − 39)

Suppose that another function, φ(ω), is defined and that equation (B-39) is rewritten as
(Claerbout, 1976)
1 1
Rw (ω) = exp U (ω) + iφ(ω) exp U (ω) − iφ(ω) (B − 40)
2 2
When equation (B-40) is compared with equation (B-37), we see that
1
W (ω) = exp U (ω) + iφ(ω) . (B − 41)
2
We know Rw (ω) from equation (B-18) and therefore we know U (ω) from equation (B-38).
To estimate W (ω) from equation (B-41), we also need to know φ(ω). If we make the minimum-
phase assumption, φ(ω) turns out to be the Hilbert transform of U (ω) (Claerbout, 1976). To
perform the Hilbert transform, first inverse Fourier transform U (ω) back to the time domain.
Then, double the positive time values, leave the zero-lag alone, and set the negative time values
to zero. This operation yields a time function u+ (t), which vanishes before t = 0. Then, return
to the transform domain to get
1
U + (ω) = U (ω) + iφ(ω) , (B − 42)
2
where U + (ω) is the Fourier transform of u+ (t). Finally, exponentiating U + (ω) yields the Fourier
transform W (ω) of the minimum-phase wavelet w(t) (equation B-41).
Once the minimum-phase wavelet w(t) is estimated, its Fourier transform W (ω) is rewritten
in terms of its amplitude and phase spectra,

W (ω) = A(ω) exp iφ(ω) . (B − 43)

The deconvolution filter in the Fourier transform domain is

1
F (ω) = . (B − 44)
W (ω)
By substituting equation (B-43), we obtain the amplitude and phase spectra of this filter:
1
Af (ω) = (B − 45a)
A(ω)
and

φf (ω) = −φ(ω). (B − 45b)

Since the estimated wavelet w(t) is minimum phase, it follows that the deconvolution filter
[whose Fourier transform is given by equation (B-44)] also is minimum phase. By inverse Fourier
transforming, equation (B-44) yields the deconvolution operator. The frequency-domain method
of deconvolution is outlined in Figure B-1.
 Deconvolution 255

FIG. B-1. Flowchart for frequency-domain deconvolution.

Figure B-2 illustrates the frequency-domain deconvolution that is described in Figure B-1.
The panels in Figure B-2 should be compared with the corresponding results of the Wiener-
Levinson (time-domain) method in Figure 2-20. As expected, there is virtually no difference
between the two results.
From the discussion so far, we see that the spiking deconvolution operator is the inverse of
the minimum-phase equivalent of the seismic wavelet. The process of estimating the minimum-
phase equivalent of a seismic wavelet is called spectral decomposition. The minimum-phase
wavelet, once computed, can be inverted easily to get the deconvolution operator. To avoid
dividing by zero in equation (B-45a) and to ensure that the filter is stable, a small number
often is added to the amplitude spectrum before division. This is called prewhitening. Equation
(B-45a) then takes the form
1
Af (ω) = . (B − 46)
A(ω) + ε
The amplitude spectrum also can be smoothed to get a more stable operator. Smoothing the
spectrum is analogous to shortening the equivalent time-domain deconvolution operator.
Zero-phase deconvolution can be implemented conveniently in the frequency domain. To do
this, the phase spectrum given by equation (B-45b) is set to zero and thus yields a deconvolution
operator that flattens the amplitude spectrum of the input seismogram, but does not alter the
phase.

B.5 Optimum Wiener Filters

The following concise discussion of optimum Wiener filters is based on Robinson and Treitel
(1980). Consider the general filter model in Figure B-3. Wiener filtering involves designing the
256 Seismic Data Analysis
 Deconvolution 257

FIG. B-3. Wiener filter model.

filter f (t) so that the least-squares error between the actual and desired outputs is minimum.
Error L is defined as
L= (dt − yt )2 . (B − 47)
t

The actual output is the convolution of the filter with the input
yt = ft ∗ xt . (B − 48)
When equation (B-48) is substituted into equation (B-47), we get
2
L= dt − fτ xt−τ . (B − 49)
t τ

The goal is to compute the filter coefficients (f0 , f1 , . . . , fn−1 ) so that the error is minimum.
Filter length n must be predefined. The minimum error is attained by setting the variation of
L with respect to fi to zero:
∂L
= 0, i = 0, 1, 2, . . . , (n − 1). (B − 50)
∂fi
By expanding the square term in equation (B-49), we have
2
L= d2t − 2 dt fτ xt−τ + fτ xt−τ . (B − 51)
t t τ t τ

By taking the partial derivatives and setting them to zero, we get

∂L
= −2 dt xt−i + 2 fτ xt−τ xt−i = 0, (B − 52)
∂fi t t τ
or
fτ xt−τ xt−i = dt xt−i ,
τ t t (B − 53)
i = 0, 1, 2, . . . , (n − 1).
By using

xt−τ xt−i = ri−τ (B − 54a)

t

and
dt xt−i = gi (B − 54b)
t
258 Seismic Data Analysis

for each ith term, we get

fτ ri−τ = gi , i = 0, 1, 2, . . . , (n − 1). (B − 55)

τ

When put into matrix form, equation (B-55) becomes

 r r1 r2 · · · rn−1   f0   g0 
0
 r1 r0 r1 · · · rn−2   f1   g1 
 . ..     
 . .. .. ..
.   ..  =  ..  . (B − 56)
. . . . . .
rn−1 rn−2 rn−3 ··· r0 fn−1 gn−1

Here, ri are the autocorrelation lags of the input and gi are the lags of the crosscorrelation
between the desired output and input. Since the autocorrelation matrix is Toeplitz, the optimum
Wiener filter coefficients fi can be computed by using Levinson recursion (Claerbout, 1976).
The least-squares error involved in this process now is computed. Expressed again, equation
(B-51) becomes
2
Lmin = d2t − 2 dt fτ xt−τ + fτ xt−τ . (B − 57)
t t τ t τ

By substituting the relationships

dt xt−τ = gτ (B − 58a)
t

and

xt xt−τ = rτ (B − 58b)
t

into equation (B-57), we get

Lmin = d2t − 2 fτ gτ + fτ fi xt−τ xt−i (B − 59)

t τ τ i t

or

Lmin = d2t − 2 fτ gτ + fτ fi rτ −i . (B − 60)

t τ τ i

Finally, by using equation (B-55), we get

Lmin = d2t − fτ gτ . (B − 61)

t τ

B.6 Spiking Deconvolution

Consider the desired output to be the zero-delay spike dt : (1, 0, 0, . . .). Given the input series
xt : (x0 , x1 , x2 , . . .), equation (B-56) takes the form
 r r1 r2 · · · rn−1   f0   x0 
0
 r1 r0 r1 · · · rn−2   f1   0 
 . ..     
 . .. .. ..
.   ..  =  ...  . (B − 62)
. . . . .
rn−1 rn−2 rn−3 ··· r0 fn−1 0
 Deconvolution 259

Divide both sides by f0 to obtain

 r r1 r2 · · · rn−1   1   v 
0
 1r r 0 r1 · · · rn−2   a1   0 
 . ..     
 . .. .. ..
.   ...  =  ...  , (B − 63)
. . . .
rn−1 rn−2 rn−3 ··· r0 an−1 0
where ai = fi /f0 , i = 1, 2, . . . , n − 1, and v = x0 /f0 . Equation (B-63) is solved for the unknown
quantity v and the filter coefficients (a1 , a2 , . . . , an−1 ). Since the desired output is a zero-delay
spike, the filter (1, a1 , a2 , . . . , an−1 ) describes a spiking deconvolution process.
The solution to equation (B-63) can be obtained efficiently using the Levinson recursion
(Claerbout, 1976). Start with the solution of equation (B-63) for the two-term filter (1, a1 ), then
solve for the filter (1, a1 , a2 ) and so on. Equation (B-63) for n = 2 takes the form
r0 r1 1 v
= . (B − 64)
r1 r0 a1 0
Write out the simultaneous equations:
r0 + r1 a1 = v
and
r1 + r0 a1 = 0,
which yield the filter coefficient a1 and the unknown variable v from the first iteration:
a1 = −r1 /r0 (B − 65a)
and
v = r0 + r1 a1 . (B − 65b)
Now write equation (B-63) for the three-term filter (1, a1 , a2 ):
    
r0 r 1 r 2 1 v
 r1 r0 r1   a1  =  0  . (B − 66)
r2 r 1 r 0 a2 0
We want to solve for the filter coefficients (a1 , a2 ) by using the results of the previous step
(equations B-65a,b). First, rewrite equation (B-64) by adding a row in the following manner:
    
r 0 r 1 r2 1 v
 r1 r0 r1   a1  =  0  , (B − 67)
r 2 r 1 r0 0 e
where
e = r2 + r1 a1 . (B − 68)
Rewrite equation (B-67) by changing the order of the rows:
    
r0 r1 r2 0 e
 r1 r0 r1   a1  =  0  , (B − 69)
r2 r1 r0 1 v
Multiply both sides of equation (B-69) with variable c, yet to be determined, and subtract the
result from equation (B-67):
    
r0 r1 r2 1 v − ce
 r1 r0 r1   a1 − ca1  =  0  , (B − 70)
r2 r1 r0 −c e − cv
260 Seismic Data Analysis

Finally, compare equation (B-70) with equation (B-66), and note that
   
1 1
 a1  =  a1 − ca1  (B − 71a)
a2 −c
and
   
v v − ce
 0  =  0 . (B − 71b)
0 e − cv
Solve for c and v :
e
c= , (B − 72a)
v
and
2
e
v =v 1− . (B − 72b)
v
Hence, the new filter coefficients (a1 , a2 ) are (equations B-71a)
e
a1 = a1 − a1 (B − 73a)
v
and
e
a2 = − . (B − 73b)
v
This recursive scheme is repeated to determine the next set of filter coefficients:

(a) Compute v and e using the autocorrelation lags of the input series and the present filter
coefficients (equations B-67).
(b) Compute the next set of filter coefficients (equations B-71a).
(c) Compute a new value for v (equation B-72b).

This recursive process yields the Wiener filter (1, a1 , a2 , . . . , an−1 ) of the desired length n.

B.7 Predictive Deconvolution

Suppose that the desired output in the filter model of Figure B-3 is a time-advanced version of
the input, d(t) = x(t + α). We want to design a Wiener filter f (t) that predicts x(t + α) from
the past values of the input x(t). In this special case, the crosscorrelation function g becomes

gτ = dt xt−τ = xt+α xt−τ = xt xt−(α+τ ) . (B − 74)

t t t

By definition, we have

rτ = xt xt−τ . (B − 75)
t

For the α + τ lag, equation (B-75) becomes

rα+τ = xt xt−(α+τ ) . (B − 76)

t

Combine equations (B-74) and (B-76) and note that rα+τ = gτ . By substituting this result into
equation (B-56), we get the set of normal equations that must be solved to find the prediction
 Deconvolution 261

filter (f0 , f1 , . . . , fn−1 ):

 r r r2 · · · rn−1   f0   rα 
0 1
 r1 r0 r1 · · · rn−2   f1   rα+1 
 . ..     .
 . .. .. ..
.   ..  =  ..  (B − 77)
. . . . . .
rn−1 rn−2 rn−3 ··· r0 fn−1 rα+n−1
For a unit prediction lag, α = 1, equation (B-77) takes the form:
 r r1 r2 · · · rn−1   f0   r1 
0
 r1 r0 r1 · · · rn−2   f1   r2 
 . ..     
 . .
.. .
.. ..
.   ..  =  ...  . (B − 78)
. . .
rn−1 rn−2 rn−3 · · · r0 fn−1 rn
By augmenting the right side to the square matrix on the left side, we have
   
 −r 1
1 r0 r1 r2 · · · rn−1  0
 f0   0 
 −r2 r1 r0 r1 · · · rn−2     
 . .. .. .. .. ..   f1  =  0  (B − 79)
 .     
. . . . . .  ..   ... 
.
−rn rn−1 rn−2 rn−3 · · · r0 0
fn−1
By adding one row, then putting the negative sign to the filter column, we obtain
r r1 r2 ··· rn   1   L 
0
 r1 r0 r1 · · · rn−1   −f0   0 
    
r
 2 r 1 r 0 · · · rn−2   −f1  =  0  . (B − 80)
 . ..   .. 
   
 . .. .. ..
.   ... 
. . . . .
rn rn−1 rn−2 · · · r0 −fn−1 0

We now have n + 1 equations and n + 1 unknowns — (f0 , f1 , . . . , fn−1 , L). From equation (B-80)

L = r0 − r1 f0 − r2 f1 − · · · − rn fn−1 . (B − 81)

By using equation (B-61), the minimum error associated with the unit prediction-lag filter
can be computed as follows. Start with

dt = xt+1 (B − 82)

so that

d2t = r0 (B − 83a)
t

and

gt = rt+1 . (B − 83b)

Substituting into equation (B-61), we get

Lmin = r0 − (r1 f0 + r2 f1 + · · · + rn fn−1 ), (B − 84)

which is identical to quantity L given by equation (B-81). Therefore, when we solve equation
(B-80), we compute both the minimum error and the filter coefficients.
Equation (B-77) gives us the prediction filter with prediction lag α. The desired output is
xt+α . The actual output is xt+α , which is an estimate of the desired output. This desired output
262 Seismic Data Analysis

is the predictable component of the input series; i.e., periodic events such as multiples. The error
series contains the unpredictable component of the series and is defined as

et+α = xt+α − xt+α = xt+α − fτ xt−τ . (B − 85)

τ

We assume that the unpredictable component et is the uncorrelated reflectivity we want to

extract from the seismogram. By taking the z-transform of equation (B-85), we have
z −α E(z) = z −α X(z) − F (z) X(z) (B − 86)
or
E(z) = 1 − z α F (z) X(z). (B − 87)
Now define a new filter a(t) whose z-transform is
A(z) = 1 − z α F (z) (B − 88)
so that, when substituted into equation (B-87), we have

E(z) = A(z) X(z). (B − 89)

The corresponding time-domain relationship is
e(t) = a(t) ∗ x(t). (B − 90)
After defining e(t) as reflectivity, equation (B-90) states that by applying filter a(t) to the
input seismogram x(t), we obtain the reflectivity series. Since computing the reflectivity series
is a goal of deconvolution, prediction filtering can be used for deconvolution. The time-domain
form a(t) of the filter defined by equation (B-88) is
α−1

at = (1, 0, 0, · · · , 0, −f0 , −f1 , · · · , −fn−1 ). (B − 91)

The filter a(t) is obtained from the prediction filter f (t), which is the solution to equation
(B-77). We call a(t) the prediction error filter. For unit-prediction lag, the filter coefficients given
by equation (B-91) are of the form
at = (1, −f0 , −f1 , · · · , −fn−1 ). (B − 92)
This is the same as the solution of equation (B-80). Moreover, equation (B-80) is equivalent
to equation (B-36) for n = 2. Thus, we can conclude that a prediction error filter with unit-
prediction lag and with n + 1 length is equivalent to an inverse filter of the same length except
for a scale factor.

B.8 Surface-Consistent Deconvolution

Deconvolution can be formulated as a surface-consistent spectral decomposition (Taner and

Coburn, 1981). In such a formulation, the seismic trace is decomposed into the convolutional
effects of source, receiver, offset, and the earth’s impulse response, thus explicitly accounting
for variations in wavelet shape affected by near-source and near-receiver conditions and source-
receiver separation. Decomposition is followed by inverse filtering to recover the earth’s impulse
response. The assumption of surface-consistency implies that the basic wavelet shape depends
only on the source and receiver locations, and not on the details of the raypath from source to
reflector to receiver.
 Deconvolution 263

The convolutional model discussed in Section 2.1 is described by

x(t) = w(t) ∗ e(t) + n(t), (B − 93)

where x(t) is the recorded seismogram, w(t) is the source waveform, e(t) is the earth’s impulse
response that we want to estimate, and n(t) is the noise component.
A postulated surface-consistent convolutional model is given by

xij (t) = sj (t) ∗ hl (t) ∗ ek (t) ∗ gi (t) + n(t), (B − 94)

where xij (t) is a model of the recorded seismogram, sj (t) is the waveform component associated
with source location j, gi (t) is the component associated with receiver location i, and hl (t) is
the component associated with offset dependency of the waveform defined for each offset index
l = |i − j|. As in equation (B-93), ek (t) represents the earth’s impulse response at the source-
receiver midpoint location, k = (i + j)/2. By comparing equations (B-93) and (B-94), we infer
that w(t) represents the combined effects of s(t), h(t), and g(t). Cambois and Stoffa (1992) offer
an alternative to equation (B-94) in which the offset term is ignored.
To illustrate a method of computing sj (t), hl (t), ek (t), and gi (t), assume n(t) = 0 and
Fourier transform equation (B-94):

Xij (ω) = Sj (ω) Hl (ω) Ek (ω) Gi (ω). (B − 95)

This equation can be separated into the following amplitude spectral components:

X ij (ω) = S j (ω) H l (ω) E k (ω) Gi (ω), (B − 96a)

and phase spectral components (Section A.1):

φij (ω) = φsj (ω) + φhl (ω) + φek (ω) + φri (ω). (B − 96b)

If the minimum-phase assumption is made, only the amplitude spectra (equation B-96a) need
to be considered.
Equation (B-96a) now can be linearized by taking the logarithm of both sides:

X̃ij (ω) = S̃j (ω) + H̃l (ω) + Ẽk (ω) + G̃i (ω). (B − 97)

The left side is the logarithm of the amplitude spectrum X ij (ω) of the modeled input trace as
in the left-hand side of equation (B-96a), and the right-hand terms are the logarithms of the
amplitude spectra of the individual components as in the right-hand side of equation (B-96a).
An alternative model equation is given by Cary and Lorentz (1993) based on the work by
Morley and Claerbout (1983):

X̃ij (ω) = S̃j (ω) + H̃l (ω) + Ẽk (ω) + G̃i (ω), (B − 98a)

where

X̃ij (ω) = X̃ij (ω) − X̃avg (ω). (B − 98b)

The spectral component X̃avg (ω) is associated with the average amplitude spectrum for the
entire data set. The terms on the right-hand side of equation (B-98) now correspond to residual
spectral components for each source, offset, midpoint, and receiver location.
The spectral components on the right-hand side of equation (B-98a) can be computed by
least-squares error minimization. For each frequency ω, equation (B-98a) is written for each trace
of each CMP gather in the data set. Consider a data set with ns shot locations and nc channels,
so that the total number of traces is ns ×nc . For ns ×nc values of the actual spectral components
X̃ij at frequency ω, and ns shot locations, nr receiver locations, ne midpoint locations, and nh
264 Seismic Data Analysis

offsets, we have the following set of model equations:

 . 
..
 
 S̃j 
   . 
  .. 

   .. 
    
    . 

 .     H̃l  
 ..   
     .. 
 X̃
 ij
  
 = ···1··· 1··· 1··· 1··· . . (B − 99a)
 .   
  .. 
 .    . 
 .    
    k 
Ẽ
    .  
   .. 
 . 
 . 
 . 
 
 G̃i 
..
.
Write equation (B-99a) in matrix notation:
X̃ = Lp, (B − 99b)
where X̃ is the column vector of m-length on the left-hand side in equation (B-99), L is the
sparse matrix with dimensions (ns × nc ) × (ns + nh + ne + nr ) and p is the column vector of
(ns + nh + ne + nr )-length on the right-hand side of the same equation. Except the four elements
in each row, the L matrix contains zeros.
We want to estimate for each frequency ω the model parameters p such that the difference
between the actual spectral component X̃ and the modeled spectral component X̃ is minimum
in the least-squares sense.
The error vector v is defined as the difference between the modeled and the actual spectral
component for each frequency ω
v = X̃ − X̃ . (B − 100a)
Substitute equation (B-99b) into equation (B-100a) to obtain
v = X̃ − Lp. (B − 100b)
Following Lines and Treitel (1984), the least-squares solution for equation (B-100b) can be
determined. First, the cumulative squared error C is expressed as
C = vT∗ v. (B − 101a)
where T is for transpose and ∗ is for complex conjugate. By substituting for v from equation
(B-100b), we get
C = (X̃ − Lp)T∗ (X̃ − Lp). (B − 101b)
Minimization of C with respect to p requires that
∂C ∂C ∂C ∂C
= = = = 0.
∂ S̃j ∂ H̃l ∂ Ẽk ∂ G̃i
This requirement yields the desired least-squares solution:
p = (LT∗ L)−1 LT∗ X̃. (B − 102)
Application of the least-squares minimization to surface-consistent prediction-error filtering
is given by Levin (1989). A practical scheme for solving equation (B-98a) is based on the Gauss-
Seidel method. In this scheme, each term on the right-hand side of equation (B-98a) is computed
 Deconvolution 265

by the following set of recursive equations:

nr
1
S̃jm = {X̃ij − H̃lm−1 − Ẽkm−1 − G̃m−1
i }, (B − 103a)
nr i

ns
1
G̃m
i = {X̃ij − H̃lm−1 − Ẽkm−1 − S̃jm−1 }, (B − 103b)
ns j

ne
1
H̃lm = {X̃ij − S̃jm−1 − Ẽkm−1 − G̃m−1
i }, (B − 103c)
ne
k

and
nh
1
Ẽkm = {X̃ij − S̃jm−1 − H̃lm−1 − G̃m−1
i }, (B − 103d)
nh
l

where m is the iteration index. The solutions in equations (B-103) are based on the orthogonality
of the shot and receiver axes, and the orthogonality of the midpoint and offset axes. Equations
(B-103) can be modified as follows:
nr nr
1 1
S̃jm = {X̃ij } − {H̃lm−1 − Ẽkm−1 − G̃m−1
i }, (B − 104a)
nr i
nr i

ns ns
1 1
G̃m
i = {X̃ij } − {H̃lm−1 − Ẽkm−1 − S̃jm−1 }, (B − 104b)
ns j
ns j

ne ne
1 1
H̃lm = {X̃ij } − {S̃jm−1 − Ẽkm−1 − G̃m−1
i }, (B − 104c)
ne ne
k k

and
nh nh
1 1
Ẽkm = {X̃ij } − {S̃jm−1 − H̃lm−1 − G̃m−1
i }. (B − 104d)
nh nh
l l

This modification enables us to compute and store the sum of the spectral components of input
data Xij , thus circumventing the need for storing the individual spectral components Xij
(Cary and Lorentz, 1993). The process is iterated until an index m that attains the least-squares
minimization.
The parameter vector p that contains the spectral components S̃j , G̃i , H̃l , and Ẽk , which
are associated with the source and receiver locations, offset dependency, and earth’s impulse
response, is solved for each frequency component ω using equations (B-104). Results from all
frequency components are then combined to obtain the terms in equation (B-98a). The surface-
consistent spiking deconvolution operator to be applied to each trace in the data set is then
the minimum-phase inverse of sj (t) ∗ gi (t) ∗ hl (t). In the case of predictive deconvolution with a
desired prediction lag, for each source, receiver and, midpoint location, a deconvolution operator
is computed by using the autocorrelograms of the terms sj (t), gi (t), and hl (t). To each trace in
the data set, these operators are then applied in a cascaded manner.
As a by-product of the derivation of the surface-consistent spectral decomposition equation
(B-97), trace amplitudes themselves can be corrected for in a surface-consistent manner (Taner
and Koehler, 1981). Sum the individual terms in equation (B-97) over the frequencies:

X̃ij (ω) = S̃j (ω) + H̃l (ω) + Ẽk (ω) + G̃i (ω). (B − 105a)
ω ω ω ω ω
266 Seismic Data Analysis

Each term yields a scalar that is related to the source, receiver, offset and midpoint locations:
x̃ij = s̃j + h̃l + ẽk + g̃i . (B − 105b)
Once computed in the same manner as the terms in equation (B-98a), these scalars are then
applied to individual traces in the data set for surface-consistent amplitude corrections.
In practice, application of surface-consistent deconvolution to field data usually involves
two terms, only — the source term sj (t) and the receiver term gi (t). In a transition zone,
surface conditions at source and receiver locations may vary significantly from dry to wet surface
conditions. Hence, the most likely situation where surface-consistent amplitude corrections and
deconvolution may be required is transition-zone data. Figure B-4 shows a field data example of
surface-consistent deconvolution. Note the variations in the autocorrelograms from one source
location to the next and from one receiver location to the next. Differences in reflection continuity
are observed within the first 1 s on the stacked sections created by application of conventional
trace-by-trace deconvolution and surface-consistent deconvolution.

B.9 Inverse Q Filtering

Consider a 1-D seismogram that represents a compressional plane wave that propagates vertically
downward in a homogeneous medium with intrinsic attenuation. This plane wave is expressed
as the solution to the scalar wave equation:
1 ∂2P ∂2P
= , (B − 106)
v 2 ∂t2 ∂z 2
where P (t, z) is the plane wave represented by the 1-D seismogram — a CMP-stacked trace, t
is the traveltime, z is the depth variable and v is the wave velocity. We shall assume that the
wave velocity is constant.
To solve equation (B-106), first, Fourier transform in the time direction:
ω2 ∂2P
P = , (B − 107)
v2 ∂z 2
where P (ω, z) is the Fourier tramsform of the wavefield P (t, z), and ω is the angular frequency.
The upcoming wave solution is then given by
ω
P (ω, z) = P (ω, z = 0) exp(−i z). (B − 108)
v
To include amplitude decay in wave propagation in a medium with intrinsic attenuation,
the wave velocity is defined as a complex variable:
v = α + iβ. (B − 109)
Substitute equation (B-109) into equation (B-108) to get
ω
P (ω, z) = P (ω, z = 0) exp(−i z). (B − 110)
α + iβ
By simple algebra, rewrite equation (B-110) as follows:
ωα ωβ
P (ω, z)=P (ω, z = 0) exp(−i z) exp(− 2 z). (B − 111)
α2 + β 2 α + β2
For most rocks, the assumption that β is much smaller than α can be made. As a result, equation
(B-111) can be simplified as follows:
ω ωβ
P (ω, z) = P (ω, z = 0) exp(−i z) exp(− 2 z). (B − 112)
α α
 Deconvolution 267

FIG. B-4. Surface-consistent deconvolution applied to field data: autocorrelograms of (a) the source
term, and (b) the receiver term as in equation (B-94), (c) conventional trace-by-trace prestack decon-
volution, and (d) surface-consistent deconvolution using the autocorrelation estimates as in (a) and (b)
(Analysis by Duane Dopkin).
268 Seismic Data Analysis

Now, define a vertical time variable τ equivalent to the depth variable z via z = ατ , and rewrite
equation (B-112):
β
P (ω, τ ) = P (ω, τ = 0) exp(−iωτ ) exp(−ω τ ). (B − 113)
α
Assume an attenuation constant Q that is independent of frequency ω (Kjartansson, 1979):
1 β
= , (B − 114)
2Q α
and substitute into equation (B-113) to obtain
ωτ
P (ω, τ ) = P (ω, τ = 0) exp(−iωτ ) exp(− ). (B − 115)
2Q
Note from equation (B-115) that the higher the frequency, the greater the attenuation.
For a nondissipative medium, β = 0; hence, equation (B-114) states that Q is infinite. As a
result, equation (B-115) takes the special form
P (ω, τ ) = P (ω, τ = 0) exp(−iωτ ). (B − 116)
The amplitude spectrum of the inverse Q filter is thus given by the exponential scaling function
ωτ
A(ω, τ ) = exp( ). (B − 117a)
2Q
The phase spectrum can either be set to zero, or more appropriately, assumed to be minimum-
phase. In the latter case, it can be computed by taking the Hilbert transform of the amplitude
spectrum given by equation (B-117a) (Section B.4):
φ(ω, τ ) = H{A(ω, τ )}, (B − 117b)
where H represents the Hilbert transform.
By combining the amplitude and phase spectra given by equations (B-117a,b), we define
the minimum-phase inverse Q filter as
W (ω, τ ) = A(ω, τ ) exp{−iφ(ω, τ )}. (B − 118)
The inverse Q filtering equation (B-115) now takes the form
P (ω, τ ) = P (ω, τ = 0) exp(−iωτ )W (ω, τ ). (B − 119)
Note that the time variable t is associated with the input trace P (t, τ = 0) and the time variable
τ is associated with the output trace P (t = 0, τ ) after the application of the inverse Q filter.
To apply the filter W (ω, τ ) to a trace P (ω, τ = 0), define a time step ∆τ and write equation
(B-119) in its recursive form:
P (ω, k∆τ ) = P [ω, (k − 1)∆τ ] exp(−iω∆τ ) W (ω, ∆τ ), (B − 120)
where k = 1, 2, . . . , n, with n number of time steps (number of samples in the input trace).
Equation (B-120) can now be used to describe a procedure for inverse Q filtering:

(a) Fourier transform the input trace P (t, τ = 0) to obtain the complex transform function
P (ω, τ = 0).
(b) Define a time step ∆τ and apply the linear phase shift to P (ω, τ = 0) by multiplying with
the exponential exp(−iω∆τ ).
(c) Specify a constant Q and apply the inverse Q filter given by equation (B-117a,b) that
represents the inverse Q filter.
(d) Repeat steps (a), (b), and (c) for all frequencies.
(e) Sum over all frequencies to obtain the inverse Q-filtered wavefield at time step ∆τ given by
P (t = 0, ∆τ ).
(f) Repeat step (d) for all time steps k∆τ , k = 1, 2, . . . , n, to obtain the inverse Q-filtered
wavefield P (t = 0, τ ) at all times τ .
 Deconvolution 269

FIG. B-5. Inverse Q filtering applied to field data (Saatcilar, 1996): Portion of a stacked section with
(a) no deconvolution, (b) inverse Q filtering, (c) inverse Q filtering followed by deconvolution, and (d)
deconvolution, only; (e) average amplitude spectrum of the data shown in (a); (f) average amplitude
spectrum of the data shown in (b); (g) average amplitude spectrum of the data shown in (c).

Inverse Q filtering often is applied to data using a constant Q factor. An efficient scheme
for a vertically varying Q factor is described by Hargreaves and Calvert (1991).
Figure B-5 shows an example of inverse Q filtering applied to field data. Compare the
stacked sections and the average amplitude spectra with no deconvolution, inverse Q filtering,
inverse Q filtering followed by deconvolution and deconvolution, only. An inverse Q filter restores
the high-frequency components of signal subjected to intrinsic attenuation by the propagation
medium. The deconvolution that follows the inverse Q filtering then easily flattens the spectrum
within the passband.
270 Seismic Data Analysis

REFERENCES

Backus, M. M., 1959, Water reverberations: Their nature and elimination: Geophysics, 24, 233-261.
Cambois, G. and Stoffa, P. L., 1992, Surface-consistent deconvolution in the log-Fourier domain:
Geophysics, 57, 823-840.
Cary, P. W. and Lorentz, G. A., 1993, Four-component surface-consistent deconvolution: Geo-
physics, 58, 383-392.
Claerbout, J. F., 1976, Fundamentals of geophysical data processing: McGraw-Hill Book Co.
Gibson, B. and Larner, K. L., 1982, Comparison of spectral flattening techniques: unpublished
technical document, Western Geophysical Company.
Goupillaud, P., 1961, An approach to inverse filtering of near-surface layer effects from seismic
records: Geophysics, 26, 754-760.
Hale, I. D., 1982, Q-adaptive deconvolution: Stanford Expl. Proj., Rep. No. 30, 133-158.
Hargreaves, N. D. and Calvert, A. J., 1991, Inverse Q filtering by Fourier transform: Geophysics,
56, 519-527.
Kjartansson, E., 1979, Constant Q-wave propagation and attenuation: J. Geophys. Res., 84, 4737-
4748.
Levin, S. A., 1989, Surface-consistent deconvolution: Geophysics, 54, 1123-1133.
Lines, L. R. and Treitel, S., 1984, Tutorial: A review of least-squares inversion and its application
to geophysical problems: Geophys. Prosp., 32, 159-186.
Morley, L. and Claerbout, J. F., 1983, Predictive deconvolution in shot-receiver space: Geophysics,
48, 515-531.
Peacock, K. L. and Treitel, S., 1969, Predictive deconvolution – theory and practice: Geophysics,
34, 155-169.
Ristow, D. and Jurczyk, D., 1975, Vibroseis deconvolution: Geophys. Prosp., 23, 363-379.
Robinson, E. A. and Treitel, S., 1980, Geophysical signal analysis: Prentice-Hall Book Co.
Saatcilar, R., 1996, An algorithm for Q-filtering: J. Seis. Expl., 5, 157-168.
Taner, M. T. and Coburn, K., 1981, Surface-consistent deconvolution: Presented at the 51st Ann.
Internat. Mtg., Soc. Expl. Geophys.
Taner, M. T. and Koehler, F., Surface-consistent corrections: Geophysics, 46, 17-22.
Treitel, S. and Robinson, E. A., 1966, The design of high-resolution filters: Inst. Electr. Electron.
Eng., GE-4, 1.
Walden, A. T. and Hosken, J. W. J., 1984, An investigation of the spectral properties of primary
reflections coefficients: Presented at the 46th Ann. Mtg. Eur. Assoc. Expl. Geophys.
Waters, K. H., 1981, Reflection seismology: Second edition, John Wiley & Sons.
Yilmaz, O., 1974, The problems of resolution and reverberations in reflection seismology: J. Geo-
phys. Soc. Turkey, 5, 2.