7.
3 Sources and deconvolution
The ideal source for reection seismology would produce a delta function or a very
short impulsive wavelet that would permit closely spaced reectors to be clearly
resolved. In practice, however, more extended sources must be used and the nite
source durations can cause complications in interpreting the data. For example, an
airgun is often used for marine seismic reection proling. This device is towed
behind a ship and res bursts of compressed air at regular intervals. This creates a
bubble that oscillates for several cycles before dissipating, producing a complicated
ringy source-time function (e.g., Fig. 7.6). The reection seismograms produced
by such a source will reproduce this source-time function for each reector. This is
not too confusing in the case where there are only a few, widely separated reectors.
However, if several closely spaced reectors are present then it becomes difcult
to separate the real structure from the source.
The combination of the Earth response with the source-time function is termed
convolution (see Appendix E) and may be written as
u(t) = s(t) G(t)
ts
0
s()G(t ) d,
(7.14)
where
u(t)
is
the
recorded
seismogram,
s(t)
is
the
effective
source-time
function
(i.e.,
what
is
actually
recorded
by
the
receiver;
we
assume
that
s(t)
includes
the
receiver
response
and
any
near-source
attenuation),
G(t)
is
the
Earth
response,
s()
.
7.3 S O U R C E S A N D D E C O N V O L U T I O N 189
(a)
(b)
(c)
Figure
7.6
(a)
A
schematic
example
of
a
typical
source-time
function
s(t)
produced
by
an
airgun
in a marine experiment. A series of bubble pulses are produced by pressure reverberations
within
the
water.
(b)
An
idealized
example
of
the
Earth
response
function
G(t)
showing
a
number
of reflected pulses. (c) The result of convolving (a) and (b).While single isolated reflectors can still
be identified, closely spaced reflectors produce a complex time series that cannot easily be
unraveled.
and
ts
is
the
duration
of
the
source.
Recovering
G(t)
from
u(t)
in
this
case
is
termed deconvolution and is often an important part of reection seismic processing.
However, it is not always clear how best to perform deconvolution and this has been
the subject of considerable research. The problem appears simpler in the frequency
domain (see Appendix E) where convolution is expressed as a product, that is,
u() = s()G(),
(7.15)
where
u()
,
s()
,
and
G()
are
the
Fourier
transforms
of
u(t)
,
s(t)
,
and
G(t)
.
Thus,
in principle, frequency-domain deconvolution is straightforward:
G() =
u()
(7.16)
The
desired
time
series
G(t)
can
then
be
obtained
from
the
inverse
Fourier
transform
of
G()
.
The
difculty
with
this
approach
is
that
(7.16)
is
exact
and
stable
only
for
noiseless
data
and
when
s()
does
not
go
to
zero.
In
practice,
some
noise
is
present
and
the
effective
source-time
function
is
usually
band-limited
so
that
s()
becomes
very small at the low- and high-frequency limits. These complications can cause
(7.16) to become unstable or produce artifacts in the deconvolved waveform. To
f(t) = a(t) b(t) =
190 7. R E F L E C T I O N S E I S M O L O G Y
address these difculties, various methods for stabilizing deconvolution have been
developed. Often a time-domain approach is more efcient for data processing, in
which case a lter is designed to perform the deconvolution directly on the data.
Although deconvolution is an important part of reection data processing, no
deconvolution method is perfect, and some information is invariably lost in the
process of convolution with the source-time function that cannot be recovered.
For this reason, it is desirable at the outset to obtain as impulsive a source-time
function as possible. Modern marine proling experiments use airgun arrays that
are designed to minimize the amplitudes of the later bubble pulses, resulting in
much cleaner and less ringy pulses than the example plotted in Figure 7.6a.
Another important source-time function is produced by a machine that vibrates
over a range of frequencies. This is the most common type of source for shallow
crustal proling on land and is termed vibroseis after the rst commercial applica-
tion of the method. The machine produces ground motion of the form of a modulated
sinusoid, termed a sweep,
v(t) = A(t)
sin[2
(f0 + bt)t
]
.
(7.17)
The
amplitude
A(t)
is
normally
constant
except
for
a
taper
to
zero
at
the
start
and
end
of the sweep. The sweep lasts from about 5 to 40 s with frequencies ranging from
about 10 to 60 Hz. The sweep duration is long enough compared with the interval
between seismic reections that raw vibroseis records are difcult to interpret. To
obtain
clearer
records,
the
seismograms,
u(t)
,
are
cross-correlated
with
the
vibroseis
sweep function.
The
cross-correlation
f(t)
between
two
real
functions
a(t)
and
b(t)
is
dened
as
a(
t)b()
d,
(7.18)
where, following Bracewell (1978), we use the ve-pointed star symbol to denote
cross-correlation;
this
should
not
be
confused
with
the
asterisk
that
indicates
con-
volution. The cross-correlation integral is very similar to the convolution integral
but without the time reversal of (7.14). Note that
a(t) b(t) = a(t) b(t)
(7.19)
and that, unlike convolution, cross-correlation is not commutative:
a(t) b(t) = b(t) a(t).
(7.20)
7.4 M I G R A T I O N 191
Cross-correlation
of
the
vibroseis
sweep
function
v(t)
with
the
original
seismogram
u(t)
yields
the
processed
time
series
u (t)
:
u (t) = v(t) u(t) =
ts
v( t)u() d,
(7.21)
0
where
ts
is
the
sweep
duration.
From
(7.14)
and
replacing
s(t)
with
v(t)
,
we obtain
u (t) = v(t)
[
v(t) G(t)
]
(7.22)
= v(t)
[
v(t) G(t)
]
(7.23)
=
[
v(t) v(t)
]
G(t)
(7.24)
=
[
v(t) v(t)
]
G(t)
(7.25)
= v (t) G(t),
(7.26)
where we have used (7.19) and the associative rule for convolution. The cross-
correlation
of
v(t)
with
itself,
v (t) = v(t) v(t)
,
is
termed
the
autocorrelation
of
v(t)
.
This
is
a
symmetric
function,
centered
at
t =
0,
that
is
much
more
sharply
peaked
than
v(t)
.
Thus,
by
cross-correlating
the
recorded
seismogram
with
the
vi-
broseis
sweep
function
v(t)
,
one
obtains
a
time
series
that
represents
the
Earth
response convolved with an effective source that is relatively compact. These re-
lationships are illustrated in Figure 7.7. Cross-correlation with the source function
is a simple form of deconvolution that is sometimes termed spiking deconvolu-
tion. Notice that the resulting time series is only an approximation to the desired
Earth
response
function
G(t)
.
More
sophisticated
methods
of
deconvolution
can
achieve
better
results,
but
G(t)
can
never
be
recovered
perfectly
since
v(t)
is
band-
limited
and
the
highest
and
lowest
frequency
components
of
G(t)
are
lost
in
the
convolution.
7.4 Migration
Up to this point, we have modeled reection seismograms as resulting from reec-
tions off horizontal interfaces. However, in many cases lateral variations in structure
are present; indeed, resolving these features is often a primary goal of reection
proling. Dipping, planar reectors can be accommodated by modifying the NMO
equations to adjust for differences between the updip and downdip directions. How-
ever, more complicated structures will produce scattered and diffracted arrivals that
cannot be modeled by simple plane-wave reections, and accurate interpretation
192 7. R E F L E C T I O N S E I S M O L O G Y
(a)
(b)
(c)
(d)
Figure
7.7
(a)
An
example
of
a
vibroseis
sweep
function
v(t)
.
(b)
A
hypothetical
Earth
response
function
G(t)
.
(c)
The
result
of
convolving
(a)
and
(b).
(d)
The
result
of
cross-correlating
(a)
with
(c).
of data from such features requires a theory that takes these arrivals into account.
Most of the analysis techniques developed for this purpose are based on the idea
that velocity perturbations in the medium can be thought of as generating secondary
seismic sources in response to the incident waveeld, and the reected waveeld
can be modeled as a sum of these secondary wavelets.
7.4.1 Huygens principle
Huygens principle, rst described by Christiaan Huygens (c. 1678), is most com-
monly mentioned in the context of light waves and optical ray theory, but it is
applicable to any wave propagation problem. If we consider a plane wavefront
traveling in a homogeneous medium, we can see how the wavefront can be thought
to propagate through the constructive interference of secondary wavelets (Fig. 7.8).
This simple idea provides, at least in a qualitative sense, an explanation for the be-
havior of waves when they pass through a narrow aperture.
v
,
7.4 M I G R A T I O N 193
(a)
t + t
t
(b)
Figure 7.8 Illustrations of
Huygens principle. (a) A plane
wave
at
time
t + t
can
be
modeled as the coherent sum
of the spherical wavefronts
emitted by point sources on the
wavefront
at
time
t
.
(b)
A
small
opening in a barrier to incident
waves will produce a diffracted
wavefront if the opening is
small compared to the
wavelength.
The bending of the ray paths at the edges of the gap is termed diffraction. The
degree to which the waves diffract into the shadow of the obstacle depends upon
the wavelength of the waves in relation to the size of the opening. At relatively long
wavelengths (e.g., ocean waves striking a hole in a jetty), the transmitted waves
will spread out almost uniformly over 180
. However, at short wavelengths the
diffraction from the edges of the slot will produce a much smaller spreading in the
waveeld. For light waves, very narrow slits are required to produce noticeable
diffraction. These properties can be modeled using Huygens principle by comput-
ing the effects of constructive and destructive interference at different wavelengths.
7.4.2 Diffraction hyperbolas
We can apply Huygens principle to reection seismology by imagining that each
point on a reector generates a secondary source in response to the incident wave-
eld. This is sometimes called the exploding reector model. Consider a single
point scatterer in a zero-offset section (Fig. 7.9). The minimum travel time is
given by
t0 =
2
h
(7.27)
v
.
4
h
2
2
=
1
2 2
2
194 7. R E F L E C T I O N S E I S M O L O G Y
x
source & receiver
*
h
d
point scatterer
x
t
0
diffraction hyperbola
Figure
7.9
A
point
scatterer will produce a
t
curved reflector in a
zero-offset section.
where
h
is
the
depth
of
the
scatterer
and
v
is
the
velocity
(assumed
constant
in
this
case).
More
generally,
the
travel
time
as
a
function
of
horizontal
distance,
x
,
is
given by
t(x) =
2
x
2
+
h
2
(7.28)
Squaring and rearranging, this can be expressed as
or
v
2
t
2
x
2
h
(7.29)
t
2
t
0
4
x
2
v
2
t
0
=
1
(7.30)
after
substituting
4
h
2
=
v
2
t
0
from
(7.27).
The
travel
time
curve
for
the
scattered
arrival has the form of a hyperbola with the apex directly above the scattering point.
Note that this hyperbola is steeper and results from a different ray geometry than
the NMO hyperbola discussed in Section 7.2 (equation (7.5)). The NMO hyperbola
describes travel time for a reection off a horizontal layer as a function of source
receiver distance; in contrast (7.30) describes travel time as a function of distance
Model
7.4 M I G R A T I O N 195
Zero-offset section
Figure 7.10 The endpoint of a horizontal reflector will produce a diffracted arrival in a
zero-offset section.The reflector itself can be modeled as the coherent sum of the diffraction
hyperbola from individual point scatterers.The diffracted phase, shown as the curved heavy line,
occurs at the boundary of the region of scattered arrivals.
away from a point scatterer at depth for zero-offset data (the source and receiver
are coincident).
7.4.3 Migration methods
Consider a horizontal reector that is made up of a series of point scatterers, each
of which generates a diffraction hyperbola in a zero-offset prole (Fig. 7.10). Fol-
lowing Huygens principle, these hyperbolas sum coherently only at the time of
the main reection; the later contributions cancel out. However, if the reector
vanishes at some point, then there will be a diffracted arrival from the endpoint
that will show up in the zero-offset data. This creates an artifact in the section that
might be falsely interpreted as a dipping, curved reector.
Techniques for removing these artifacts from reection data are termed migration
and a number of different approaches have been developed. The simplest of these
methods is termed diffraction summation migration and involves assuming that
each point in a zero-offset section is the apex of a hypothetical diffraction hyperbola.
The value of the time series at that point is replaced by the average of the data from
adjacent traces taken at points along the hyperbola. In this way, diffraction artifacts
are collapsed into their true locations in the migrated section. In many cases
migration can produce a dramatic improvement in image quality (e.g., Fig. 7.11).
Aproper implementation of diffraction summation migration requires wave prop-
agation theory that goes beyond the simple ideas of Huygens principle. In partic-
ular, the scattered amplitudes vary as a function of range and ray angle, and the
196 7. R E F L E C T I O N S E I S M O L O G Y
Figure 7.11 Original (top) and migrated (bottom) reflection data from a survey line across the
Japan trench (figure modified from Claerbout, 1985; data from the Tokyo University
Oceanographic Research Institute).
Huygens secondary sources are given, for a three-dimensional geometry, by the time
derivative of the source-time function (in the frequency domain this is described
by
the
factor
i
,
a
/
2
(90-degree)
phase
shift
with
amplitude
proportional
to
frequency). In the case of a two-dimensional geometry, the secondary sources are
the half-derivative of the source function (a 45-degree phase shift with amplitude
scaled by the square root of frequency). These details are provided by Kirchhoff
theory, which is discussed later in this chapter. The diffraction hyperbola equation
assumes a uniform velocity structure, but migration concepts can be generalized
to more complicated velocity models. However, it is important to have an accu-
rate velocity model, as use of the wrong model can result in undermigrated or
overmigrated sections.
t
2
(x)
=
t
0
2
+
x
2
2
2
v
n
i=
1
v
i
ti
i=
1
ti
,
7.5 V E L O C I T Y A N A L Y S I S 197
In common practice, data from seismic reection experiments are rst processed
into zero-offset sections through CMP stacking. The zero-offset section is then
migrated to produce the nal result. This is termed poststack migration. Because
CMP stacking assumes horizontal layering and may blur some of the details of the
original data, better results can be obtained if the migration is performed prior to
stacking. This is called prestack migration. Although prestack migration is known
to produce superior results, it is not implemented routinely owing to its much greater
computational cost.
7.5 Velocity analysis
Knowledge of the large-scale background seismic velocity structure is essential for
reection seismic processing (for both stacking and migration) and for translating
observed events from time to depth. Often this information is best obtained from
results derived independently of the reection experiment, such as from borehole
logs or from a collocated refraction experiment. However, if such constraints are not
available a velocity prole must be estimated from the reection data themselves.
This can be done in several different ways.
One approach is to examine the travel time behavior of reectors in CMP gathers.
From
(7.7),
we
have
for
a
reector
overlain
by
material
of
uniform
velocity
v
:
v
2
(7.31)
=
t
0
+
u
2
x
2
,
(7.32)
where
u =
1
/v
is
the
slowness
of
the
layer.
From
observations
of
the
NMO
offsets
in
a
CMP
gather,
one
can
plot
values
of
t
2
versus
x
2
.
Fitting
a
straight
line
to
these
points
then
gives
the
intercept
t
0
and
the
slope
u
2
=
1
/v
2
.
Velocity
often
is
not
constant with depth, but this equation will still yield a velocity, which can be shown
to be approximately the root-mean-square (rms) velocity of the overlying medium,
that
is,
for
n
layers
2
n
2
(7.33)
where
ti
is
the
travel
time
through
the
i
th
layer.
Another method is to plot NMO corrected data as a function of offset for different
velocity models to see which model best removes the range dependence in the data
or produces the most coherent image following CMP stacking. As in the case of the
198 7. R E F L E C T I O N S E I S M O L O G Y
t
2
(x
2
)
plotting
method,
this
will
only
resolve
the
velocities
accurately
if
a
reasonable
spread in sourcereceiver offsets are available. Zero-offset data have no direct
velocity resolution; the constraints on velocity come from the NMO offsets in the
travel times with range. Thus, wider sourcereceiver proling generally produces
better velocity resolution, with the best results obtained in the case where receiver
ranges are extended far enough to capture the direct refracted phases. However,
even zero-offset data can yield velocity information if diffraction hyperbolas are
present in the zero-offset proles, as the curvature of these diffracted phases can
be used to constrain the velocities. One approach is to migrate the section with
different migration velocities in order to identify the model that best removes the
artifacts in the prole.
7.5.1 Statics corrections
Often strong near-surface velocity heterogeneity produces time shifts in the records
that can vary unpredictably between sources and stations. This could be caused
by topography/bathymetry or a sediment layer of variable thickness. The result-
ing jitter in the observed reected pulses (Fig. 7.12) can hinder application
of stacking and migration techniques and complicate interpretation of the results.
Thus, it is desirable to remove these time shifts prior to most processing of the
results. This is done by applying timing corrections, termed statics corrections, to
the data. In the case of the receivers, these are analogous to the station terms (the
Figure 7.12 Small time shifts on individual records produce offsets in reflectors in CMP gathers
(left plot) that prevent coherent stacking of these phases in data processing.These shifts can be
removed by applying static corrections (right plot).
7.6 R E C E I V E R F U N C T I O N S 199
average travel time residual at a particular station for many different events) used
in travel time inversions for Earth structure. Statics may be computed by tracking
the arrival time of a reference phase, such as a refracted arrival. Often automatic
methods are applied to nd the time shifts that best smooth the observed reectors.
The goal is to shift the timing of the individual records such that reectors will stack
coherently. This problem is tractable since the time shifts are generally fairly small,
and solutions for the time shifts are overdetermined in typical reection experiments
(multiple receivers for each source, multiple sources for each receiver).
