1

Chapter 3
Deconvolution

The convolutional model

The convolutional model is used to explain how the seismic trace is formed.

The convolutional model approximates the earth by a linear system.

A linear system is one whose output o(t) is given by the convolution of its input i(t)
with its response r(t): o(t) = i(t) * r(t).

In the convolutional model, the systems (earths) response e(t) is a series of impulses
corresponding, in time and amplitude, to the reflection coefficients at layers
boundaries. e(t) is also known as the reflectivity (series).

The systems input w(t) is the source wavelet.

According to this model, the seismic trace s(t) is given by:

s(t) = w(t) * e(t).

A noise component n(t), if present, is additive; hence, the seismic trace becomes:
s(t) = w(t) * e(t) + n(t).

(3.1)

(3.2)

The forward convolutional model of the seismic trace is used to compute the
synthetic seismogram s(t) given the source wavelet and earths reflectivity.

The deconvolution (inverse convolutional model) is used to compute the earths

reflectivity e(t) given the seismic trace s(t).

The convolutional model of the seismic trace is widely accepted because it agrees
well with the observed seismic traces.

The basic convolutional model of the seismic trace assumes the following:
(1) The earth is made up of horizontal layers of constant velocity.
(2) The source generates a P-wave, which is reflected on layer boundaries at normal
incidence.
(3) The source waveform is stationary. That is, it does not change its shape as it
travels in the subsurface.
(4) The noise component n(t) is zero.
(5) The Earths reflectivity e(t) is random.
(6) The seismic wavelet is a minimum-phase wavelet.

Often, one or more of these assumptions might not be satisfied; in which case,
advanced techniques of deconvolution have to be implemented.

If the source wavelet in known, then the deconvolution is deterministic and we use
inverse filtering to find the earth response.

If the source wavelet in unknown, then the deconvolution is statistical and we usually
use predictive deconvolution to find the earth response.

Inverse filtering (spiking deconvolution)

The aim of this process is to compress the source wavelet w(t) into a zero-lag spike of
zero width (i.e., (t)). This means that we are eliminating w(t) and leaving only e(t)
in s(t).

That is, we want to find a filter f(t) such that:

w(t) * f(t) = (t).

f(t) is called the inverse filter for w(t).

(3.3)

Taking the FT of equation (3.3):

W(f) F(f) = 1,

(3.4)

where W(f), F(f), and 1 are the FTs of w(t), f(t), and (t), respectively.

From equation (3.4):

F(f) = 1/W(f) = [1/|W(f)|] exp[-w(f)],
|F(f)| = 1/|W(f)|, and
f(f) = -w(f),
where |W(f)| and w(f) are the amplitude and phase spectra of w(t) and |F(f)| and f(f)
are the amplitude and phase spectra of f(t).

Therefore, the amplitude spectrum of the inverse filter is the reciprocal of that of the
source wavelet whereas its phase spectrum is the negative of that of the wavelet.

The deconvolution is accomplished by convolving the inverse filter f(t) with the
seismic trace s(t):
f(t) * s(t) = [f(t) * w(t)] * e(t) = (t) * e(t) = e(t),
which is the earths response that we want to extract from the seismic trace.

The inverse filter of a minimum-phase wavelet has a minimum phase, too.

For small wavelets, the z-transform is used and equation (3.4) becomes:
W(z) F(z) = 1,
which means that: F(z) = 1 / W(z).

Example: Given a source wavelet w(t) = (1, -1/2), find its inverse filter f(t).
Step (1): Take the z-transform of w(t): W(z) = 1 (1/2)z.
Step (2): Use polynomial division to find F(z):
F(z) = 1 / W(z) = 1 / [1 (1/2)z] = 1 + (1/2)z + (1/4)z2 + (1/8)z3 + .

(3.5)

Step (3): Since polynomial division is infinite, the filter will be of infinite length
and truncation is required. We will truncate the division to n terms. For this
example, lets truncate at the second term (n = 2), giving F2(z) = 1 + (1/2)z.
Step (4): Find the truncated filter f2(t) = (1, ).

Because of truncation, convolution of the truncated filter fn(t) with the wavelet will
not give the desired output d(t), which is in this case = (t).

The actual output y(t) is given by:

y(t) = fn(t) * w(t)

Apparently, y(t)

d(t) and the truncation error E is defined as:

E = i(di yi)2,

where di and yi are the ith samples of the desired and actual outputs, respectively.

Continuing the steps of the above example:

Step (5): Calculate y(t) = (1, ) * (1, -1/2) = (1, 0, ).
Step (6): Knowing that d(t) = (t) = (1, 0, 0), find E as:
E = (1-1)2 + (0-0)2 + (0-1/4)2 = 1/16 = 6.25 %.

The seismogram-wavelet relation

Since the earths response e(t) is a random series of impulses, its FT amplitude
spectrum is a constant E0:
|E(f)| E0.

(3.6)

Taking the FT of equation (3.1):

|S(f)| = |W(f)| |E(f)| = E0 |W(f)|.

(3.7)

Equation (3.7) means that the amplitude spectrum of the seismic trace is a scaled
version of the amplitude spectrum of the source wavelet.

Because of the random nature of e(t), its autocorrelation is zero anywhere except at
t = 0, where it is equal to the energy in e(t) given by:

re 0 = ei2 =e02 + e12 + L .

(3.8)

Therefore, re(t) can be represented as re(t) = re0 (t).

Using this fact about re(t), we can derive the relation between rs(t) and rw(t) as
follows:
rs(t) = s(t) s(t) = s(t) * s(-t) = [w(t) * e(t)] * [w(-t) * e(-t)]
rs(t) = [w(t) * w(-t)] * [e(t) * e(-t)] = rw(t) * re(t) = re0 (t) * rw(t).

The last relation can be written as:

rs (t) = re0 rw(t).

(3.9)

Equation (3.9) means that the autocorrelation of the seismic trace is a scaled version
of the autocorrelation of the source wavelet.

Optimum Wiener filters

Wiener optimum filtering involves designing a filter f(t) so that the error E between
the actual output y(t) and the desired output d(t) is minimum:

E = (d i y i ) 2 ,

(3.10)

where di and yi are the ith samples of the desired and actual outputs, respectively.

The actual output y(t) is:

y (t ) = f (t ) x(t ) ,

(3.11)

where x(t) is the input.

Substituting equation (3.11) into (3.10):

E = (d i { f (t ) * x(t )}i ) 2 .

(3.12)

where {f(t)*x(t)}i = yi is the ith sample of the y(t).

The goal is to compute the filter coefficients (f0, f1, ..., fn-1) so that the error E is
minimum; where the length of the filter, n, must be predefined. This is a typical
least-squares problem.

The minimum error is attained by setting the variation of E with respect to fj to zero:
E/fj = 0,

r
i =0

f = gj ,

j = 0, 1, 2, ..., (n-1).

j i i

(3.14a)

In matrix format, equation (13.4a) can be written as:

r
r

.
r

r
.

r
.

n 2

.
r

.
r

n 1

(3.13)

Applying equation (3.13) on (3.12) and simplifying, the result can be expressed as:
n1

j = 0, 1, 2, ..., (n-1).

n 3

n 2

n4

n 3

. . r f
. . r f

. . . .

. . . .
. . r f

. . r f
n 1

n2

n 2

n1

g
g

.
=
.
g

g

n2

n 1

(3.14b)

In equations (3.14a) and (3.14b), gj is the jth term of the crosscorrelation between d(t)
and x(t):

g j = {d (t ) x (t )} j ,

j = 0, 1, 2, ..., (n-1). (3.15a)

(Note that g(t) = d(t) x(t) x(t) d(t) and that we only need lags 0, 1,

(n-1)).

In equations (3.14a) and (3.14b), rj-i is the jth term of the autocorrelation of x(t):

rj = {x(t ) x(t )} j ,

j = 0, 1, 2, ..., (n-1).

(3.16)

(Note that r(t) = x(t) x(t) = r(-t)).

If gi and ri- are known, equation (3.14) can be solved uniquely to find the filter f(t).

The autocorrelation matrix of equation (3.14b) is a Toeplitz matrix that can be

inverted efficiently using the Levinson recursion procedure.

Equations (3.14a) and (3.14b) are known as the normal equations.

Inverse filtering by normal equations

As an application of the normal equations, we will use them to design the inverse
filter of the source wavelet such that the desired output is a zero-lag spike.

x(t) = w(t) = (w0, w1, ..., wn-1).

f(t) = (f0, f1, ..., fn-1).

d(t) = (1, 0, ..., 0).

r0 = w02 + w12 + ... + wn-12

r1 = w0 w1 + w1 w2 + ... + wn-2 wn-1
... ...

...

...

rn-1 = w0 wn-1

g0 = w0
g1 = 0
... ...
gn-1 = 0.

...

...

length of each = n.

Therefore, the normal equations become:

r
r

.
r

r
.

r
.
.

n 2

.
r

n 1

n3

n 2

n 4

n3

. . r f
. . r f

. . . .

. . . .
. . r f

. . r f
n 1

n 2

n 2

n 1

w
0
.
.
=
.
0

0
0

(3.17)

Predictive deconvolution

In exploration seismology, all we usually know is the seismic trace s(t).

Although the source wavelet is usually unknown, we can use the seismogram-wavelet
relation to compute a scaled version of it and use it instead.

The desired output is the earth response e(t), which is unknown because of its random
nature.

For these reasons, we cannot use the normal equations directly; however, we can still
use them if we require the desired output to be a time-advanced version x(t+) of the
input x(t), where is called the prediction lag, which is known.

For example, if x(t) = (x0, x1, x2, x3, x4), = 2, then: d(t) = x(t+2) = (x2, x3, x4).

Since d(t) = x(t+), g(t) = d(t) x(t) = x(t+) x(t) = r(t+). Therefore, g0 = r,
g1 = r+1, , gn-1 = r+n-1.

Inserting these values for g(t) into the normal equations, we get:

r
r

.
r

r
.

r
.
.

n 2

.
r

n 1

n 3

n 2

n 4

n3

. . r f
. . r f

. . . .

. . . .
. . r f

. . r f
n 1

n 2

n 2

n 1

r
r

.
=
.
r

r

+1

+n 2

+ n 1

(3.18)

The filter f(t) is called the prediction filter.

Applying f(t) to the input x(t) will give us the actual output y(t).

Lets define the prediction error series pu(t) as the difference between the desired and
actual outputs:
pu(t) = d(t) y(t) = x(t+) f(t)*x(t).

(3.19)

y(t) can be thought of as the predictable part of the seismogram (e.g., multiples), pu(t)
as the unpredictable (random) part of the seismogram, and d(t) is their sum.

It can be shown that pu(t) = e(t+), which is a time-advanced version of the earths
response that we are seeking.

Knowing , we can compute the earths response e(t).

Equation (3.19) can be written in the following form:

e(t+) = a(t) * x(t),
where a(t) is called the prediction error filter given by:
(-1) zeros
a(t) = (1, 0, 0, ..., 0, -f0, -f1, ..., -fn-1).
n filter coefficients

In general, given a source wavelet of length (n+), the prediction error filter
compresses it to an -long wavelet, where is the prediction lag.

(3.20)

10

It can be shown that, for a unit prediction lag, = 1, which yields a(t) = (1, -f0, -f1, ...,
-fn-1), the normal equations have the same structure as those used to calculate the
inverse filter of x(t) except for a scaling factor (equation 3.17).

Hence, the prediction error filter a(t) with unit-prediction lag and length n+1 is
equivalent to an inverse filter of the same length except for a scale factor.

Prediction error filters for prediction lag > 1, are used for multiple suppression.

Because of the general nature of the predictive deconvolution, we normally use it for
most of our deconvolution needs.

Predictive deconvolution in practice

Autocorrelation window (w)

The choice of deconvolution parameters depends largely on the characteristics of the

autocorrelation of the seismic trace.

Therefore, it is important to choose a suitable autocorrelation window (gate) that will

be used to calculate the deconvolution parameters.

The autocorrelation window should include the part of the record that contains useful
reflection signal and should exclude coherent (e. g., ground roll) or incoherent noise
(e. g., later parts of the record).

The length of the autocorrelation window should be eight times the largest operator
length that will be used for that dataset.

11

Operator length (n)

In order to decide upon the operator length, we need the autocorrelation function of
the source wavelet.

If the source wavelet is unknown, then we can use the autocorrelation of the seismic
trace instead.

The first transient zone of the autocorrelation function is the part that mostly
represents the autocorrelation of the source wavelet

The operator length should be selected so that it is approximately equal to the length
of the first transient zone of the autocorrelation.

The optimum operator length should not leave considerable amount of energy in the
autocorrelogram.

Prediction lag ()

Given a source wavelet of length (n+), the predictive deconvolution, using

prediction filter with length n and prediction lag , converts this wavelet into another
wavelet that is samples long.

For spiking deconvolution, = 1 should be selected.

Predictive deconvolution with prediction lags > 1 is used to predict and suppress
multiples.

If prediction lag is increased, the output from predictive deconvolution becomes less
spiky because the amplitude spectrum becomes more band-limited.

Applying a band-pass filter after deconvolution is not equivalent to predictive

deconvolution with > 1 prediction lag.

12

The application of spiking deconvolution to field data is not always desirable, since it
boosts high-frequency noise in the data. In this case, we can use > 1 prediction
lags.

For suppressing multiples, the selection of depends on the type of multiple:

Long-path multiples: should be selected equal to the beginning of first multiple
on the autocorrelogram.
Short-path multiples: should be selected equal to the first or second zero
crossing on the autocorrelogram.

In case of multiples, we deconvolve the trace in the following sequence:

1. Multiple suppression deconvolution.
2. Spiking deconvolution.

Prewhitening ()

The amplitude spectrum of the spiking deconvolution operator (inverse filter) is

approximately the reciprocal of that of the source wavelet.

If the amplitude spectrum of the source wavelet had zeros, then the amplitude
spectrum of inverse filter will be unstable (i. e., it will be infinite at these points).

To ensure numerical stability, we introduce an artificial level of random (white) noise

before deconvolution. This process is called prewhitening.

Prewhitening is achieved by adding white noise, with a very small variance, to the
amplitude spectrum of the source wavelet.

The magnitude of prewhitening is measured as a percentage of the zero-lag

autocorrelation function r0.

13

In practice, typically 0.1 0.3 percent (i.e., 0.001r0 - 0.003r0) prewhitening is

standard in processing.

Figure.

Effect of random noise on deconvolution

Field data practically contain a noise component. If this component is small, then we
can ignore its effect.

Alternatively, we can delay the deconvolution until we stack the data, so that the S/N
ratio is improved.

However, if the noise component is considerable even after stacking, then we must
test to see if deconvolution can enhance the resolution of the data without destroying
its interpretive content.

Multiple suppression by predictive deconvolution

Prediction lags of > 1 are used for multiple suppression.

An example of multiple situation is the case of water-bottom multiples.

If the reflection coefficient of the water bottom is R and if the water depth is
equivalent to a two-way time = T0, then the time series is:
(1, 0, ..., 0, -R, 0, ..., 0, R2, 0, ..., 0, -R3, ...).
where the separation between the nonzero elements is = T0/t-1, where t is the
sampling interval of the data (Figure).

If the length of the first transient zone of the autocorrelogram is T1, then our optimum
choices are: n = T1 and = T0.

14

Nonstationarity and deconvolution

Because the earth filters high frequencies, the shape of the source wavelet changes as
it travels in the earth.

If the wavelet changes its shape considerably with depth, a time-variant

deconvolution should be used, where the record is segmented into 3 - 4 parts and each
part is used to construct its own deconvolution parameters.