You are on page 1of 27

GEOPHYSICS, VOL. 45, NO. 12 (DECEMBER 1980); P. 1753-1779, 20 FIGS.

Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Prestack partial migration

&doQan Yilmaz* and Jon F. ClaerboutJ:

Conventional seismic data processingcan be improved by modifying wide-offset data so that dipping events
stack coherently. A procedure to achieve this improvement is proposed here, which is basically a “partial”
migration of common offset sections prior to stack. It has an advantage over full migration before stack in
that, in the caseof the latter, the final product is a migrated section. However, the prestackpartial migration pro-
vides the interpreter with a high-quality common midpoint (CMP) stackedsection which can be subsequently
migrated.
The theory of prestackpartial migration is basedon the double square-rootequation, which describesseismic
imaging with many shots and receivers. The double square-root operator in midpoint-offset space can be
separatedapproximately into two terms, one involving only migration effects and the other involving only
moveout correction. This separationprovides an analysis of conventional processing. Estimation of errors in
the separationyields the equation for prestack partial migration.
Extension of the theory for separableapproximation to incorporate lateral velocity variation yields a signifi-
cant term proportional to the product of the first powers of offset, dip, and lateral velocity gradient. This term
was used to obtain a rough estimate of lateral velocity variation from a field data set.

INTRODUCTION
medium velocity (the exploding reflectors model).
Conventional processingis robustenoughto handle This makes it possible to migrate the stack section
seismic field data from many different parts of the with wave extrapolation techniques.
world. Even in areas with flat beds, however, offsets In recent years, considerable attention has been
can be quite large to accommodate for accurate given to tackling the problem with CMP stack so that
velocity estimation. Dipping events at large offset steeply dipping events could be delineated. Gardner
can have a degrading effect upon common midpoint et al (1974) described a procedure to migrate simul-
(CMP) stack. Moreover, in regions of complex struc- taneously and do velocity analysis of seismic data
tural setting, conventional processing has difficulty using an experimental model. Sattlegger (1975) and
delineating steeply dipping events. The underlying Dohr and Stiller (1975) demonstrated on some field
reasonfor all this is related to two crucial assumptions data the successof migration before stack to produce
upon which conventional processing is based. The a more enhanced section of the earth. Doherty and
first assumption is that the classic normal moveout Claerbout (1976) developed the early wave equation
(NMO) equation is derived for the stratified earth methods to downward continue moveout-corrected
(zero-dip) model. This allows one to apply moveout seismic wave fields for a more reliable velocity
correction to each CMP gather independently. The estimation in a nonlayered earth. Sherwood et al
other assumptionis that the CMP stack is equivalent (1976) implemented a migration before stack pro-
to a zero-offset section, which in turn is considered cedure by adjusting velocities for offset and success-
to be generatedby putting sourceson reflectors, tum- fully applied it on field data from the Gulf of Mexico.
ing them on at t = 0, and extrapolating the resulting However, it was realized that full migration of each
upcoming wave field to the surface using half the individual common offset section is computationally

Presented at the 49th InternationalSEG MeetingNovember7, 1979, in New Orleans. Manuscript received by the Editor
October 22, 1979; revised manuscript received March 20, 1980.
*Westem Geophysical Co., Box 2469, Houston, TX 77001; formerly Department of Geophysics, Stanford University
Stanford, CA 94305.
$Geophysics Department, Stanford University, Stanford, CA 94305.
0016-8033/80/1201-1753$03.00. 0 1980 Society of Exploration Geophysicists. All rights reserved.

1753
17!5i Yllmaz and~Cl&erbout

costly. Additionally, but also more importantly, sucha familiar term that is a function of offset, dip, and
procedure does not yield an unmigrated CMP stack lateral velocity gradient. A simple experiment on
sectionasan intermediateproduct. Moreover, velocity field data, in which a lateral velocity gradient of
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

estimation provides only limited~accuracy. An urn+ roughly (so0 ftisecjiioo0 ft -was measured; shows
gratedCMP stack section helps the interpreter a great that this term can be used for estimating lateral
deal in resolving spurious events on a migrated sec- velocity variation. In particular, it is potentially use-
tion due to inaccuratevelocities. Finally, Judsonet al ful in statics problem where the near-surface lateral
(1978) presented a procedure which produced an velocity gradient is often significant.
unmigrated CMP stack section in which steeply
dipping events are enhanced. Even though the theory
THE DOUBLE SQUARE-ROOT EQUATION
of this procedure, namely “Devilish,” is based on
AND RELATED OPERATORS
coarseray approximations (John Sherwood, personal
communication, 1979), it produced encouraging re- The DSR equation describes downward continua-
sults from a particular field data set. We propose a tion of both shots and receivers into the earth; thus,
systematic procedure, based on the wave equation, it is of fundamental importance to seismic imaging.
which will improve conventional processing by re- It is exact in the sensethat it handles all wave angles
moving the effect of offset on dipping events so that up to 90 degrees from vertical. Neglect of velocity
they stack coherently. The process is basically a gradients causesthe kind of amplitude errors which
partial migration of common offset sections prior arise from neglect of multiple reflections. Thus the
to stack. equation can be described as exact in the sense that
The theory of prestack partial migration is closely there are no timing errors which increase with angle.
related to the problem of seismic imaging. A real
The double square-root equation
seismic experiment involves many shots and re-
ceivers. The earth image from the wave field acquired Seismic imaging can be expressedin simple terms
by such an experiment is restored by downward con- as follows: Given the upcoming seismic wave field
tinuation of both shots and receivers. The reflectivity P(x, 0, t) recorded at the surface, we would like to
at any (x, z) is then seenin the limit of zero traveltime determine the reflectivity P(x, z, 0). This requires
and zero offset. The mathematical descriptionof such extrapolating the surface wave field down to depth z
an imaging process is given by the double square- and collecting it at t = 0. The extrapolationprocedure
root equation (DSR). It turns out that conventional in the Fourier domain is described by
processing can be understood as an approximation P(k,, z, w) = P(k,, O,u)erk~*, (la)
to DSR. Imposing the zero-dip and zero-offset as-
sumptions, we separate the DSR operator into two which is the solution to the one-way wave equation
terms (Sep), one involving the zero-dip moveout
correction applied in offset space and one involving $‘(k,. z, o) = ik,P(k,, z, w). (1’4
the migration of zero-offset section applied in mid-
point space. The physical implication of equations (1) is that,
We formally define the deviation operator (Dev) as given the wave field recordedat the surfaceP (x, 0, t),
the difference between DSR and Sep, which is effec- we may double Fourier transform over (x, t) and get
tively the error involved in the separationof the DSR P(k,, 0, w). Next, we simply multiply by the
operator. The recognition of a significant term in the all-pass filter exp(ik,z) to obtain the wave field
difference yields an equation for prestack partial P(k,, z, o) at depth z. Subsequent inverse Fourier
migration. A wide variety of model experiments and transformation over (k,, o) yields P(x, z, t) from
field data examples demonstrates the fact that con- which we obtain the earth image P(x, z, 0). In prac-
ventional CMP stack is considerably improved by tice (Stolt, 1978), however, mapping is done in the
incorporating this procedure into the mainstream of transform domain directly from (k,, o) to (k,, k,)
seismic data processing. In particular, the mute zone using
during stackis considerablyreduced, a better imaging
is achieved by migration after stack, and dipping k, = - ; (1 - X2)“2, (2)
events can be delineated much more successfully.
In the final section, we extend our theoretical where X = vk,/w. Mathematical details of wave
analysis on separation of the DSR equation to the field extrapolation are discussed in Appendix A.
case of lateral velocity variation. We discover an un- Notice that the normalized wavenumber X is a sine-
Prestack Partial Migration 1755

like quantity. We will define extrapolation operators Equation (8) describes downward continuation of
in the simple form involving only the cosine-like both shots and receivers into the earth. We will refer
to equation (7) as the DSR operator. For simplicity,
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

term Op(X) = (1 - X2)1’2, in which case k, =


- (o/v) OPCG. we leave out the scaling wavenumber w/v from the
Let usutilize these conceptsto downward continue definition of DSR.
a complete seismic experiment with many shots and The DSR operator given by equation (7) is separ-
receivers. We would like to downward continue both able in terms of shot-and-receiverwavenumbers. This
shots and receivers. First, consider a continuous means that one can first organize the wave field re-
stretch of receivers, each with a unique location g, corded at the surface into common shot gathers and
along the x-axis. Then, we can use equation (1) to- use the first part of the DSR operator to downward
gether with equation (2), by replacing x with g, to continue the receivers to depth z. Following this,
downward continue these receivers. We use the one can reorganize the already downward-continued
reciprocity principle and interchange shots and re- wave field into common receiver gathers and use the
ceivers. We then proceed similarly on a continuous second part of DSR to downward continue the shots
stretch of shots, each with a unique location s, along to depth z. Alternating between common receiver and
the x-axis. In shot-receiver space, equation (la) will common shot gathers, the entire seismic wave field
read as can be downward continued until imaging is ac-
complished. Although no approximation, apart from
PVC,,k,, z, WI = P(k,, k,, 0, w)e rkzz. (3) the stratifiedearth assumption,is made in this process,
When all shots and receivers are at depth J dz, the it is clear that computationally this approach can be
wave field has undergone a total phase shift of -o exhausting. In fact, today’s seismic data processing
./ (dt, + dt,). Following the mathematical develop- is essentially done in the space of midpoint y and
ment in Appendix A, the correspondingvertical wave- (half-)offset h rather than in shot-receiver (s, g)-
number then becomes space. We therefore would like to put DSR defined
by equation (7) into (y, h)-coordinates. This re-
k, = _ $1 _ (2)““” quires the following coordinate transformation:

(10)
+ [I - ($J2 JU2j (4)
The principle of invariance statesthat wave fields do
For convenience, we define the normalized shot-and- not change under a coordinate transformation; thus
receiver wavenumbers as
P(s, g, 2, t) = P’(Y, h, 2, t), (11)

El =:[::I~ (5a, b) where z and t are invariant under the coordinatetrans-


formation.
causing equation (4) to take the simple form Applying the chain rule to equation (1 l),

ap
-=
apt ay ap' ah
k, = - ; DSR(G, S), (6) --+xa,. (124
as ay as
where
and
DSR(G, S) = (1 - G2)“2 + (1 - S2)“2. (7)
ap
-=
ap' ay
--+__-_. ap' ah
Upon substitutingequation (6) into equation (3), we (12b)
have
ag ay ag ah ag
Using equation (lo), we first simplify these differ-
P(k,, k,, Z, w) entials, then Fourier transform both sidesof the result-
= P(k,, k,, 0, m)e-r?DSR(G,S)z, (8)
ing expressions, and finally canceling P = P’, we
which is the solution to the double square-root obtain the midpoint-offset wavenumbers in terms of
equation shot-receiver wavenumbers:

$ P(k,, k,, z, o) = - if DSR(G, S) [t] =i[::‘“,:]. (13a,b)

. PC&, k,, z, 0). (9) Multiplying by v/o, we normalize both sides and
1756 Yilmaz and Claerbout

substituteequation (5) to get assumed to simulate a wave field that would be re-
corded at zero offset. The equation used for down-
[z]=[:T:], (14a,b) ward extrapolation is the one-way wave equation
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(1b). In order to account for the one-way traveltime


where of the exploding reflectors model, however, the
velocity used in extrapolation is taken as half the
(15a, b) medium velocity. Thus, the vertical wavenumber
given by equation (2) is reexpressed as
Final substitutionof equation (14) into equation (7)
yields the DSR operator in midpoint-offset space: k, = - y [I _ (fk,“]‘;“, (18)

DSR(Y, H) = [1 - (Y + H)‘ll”
Since migration is done in midpoint space, we re-
+ [I - (Y - H)2]“2. (16) placed k, with k,. Some of the current techniques
Is equation (16) an improvement ,over equation of migration based on wave extrapolation utilize cer-
(7)? We have lost the property of separationin terms tain rational approximations to equation (18), and
of spatial frequencies. Notice the strong coupling be- some implement the exact form in the frequency
tween Y and H. For instance, Taylor expansions of domain.
the square roots in equation (16) will yield terms By now it is clear that conventional processing
containing cross products of the two wavenumbers. has an advantage over the exact theory represented
This in turn will require the recorded wave field, by the DSR equation in midpoint-offset space. Un-
which is now transformed into (y, h, r)-space, to be like the latter, the conventional approachis composed
handled in its entirety, a problem we would like to of two separable operators, namely the NM0 +
avoid. stack applied in offset space and migration applied
in midpoint space. However, we should be reminded
Conventional processing that such an advantage is based on zero-dip and
The conventional processing sequence comprises zero-offsetassumptions.Where do we go from here?
two major and crucial components. First, the data are On the one hand, we have an exact theory that can
organized into CMP gathers, and NM0 correction is handle all dips and offset angles, but it is difficult
applied on each gather. The equation used for move- to implement. On the other hand, we have a con-
out correction is ventional approach that has the convenient property
of separation, but it is based on assumptionswhich
At = , - t’ = t’ can be severe, particularly in a region with a complex
([1 + ($)2]li2 - 1). (17)
structuralsetting. Let us investigatewhether thesetwo
approachesare related in any way. We will go back
where t is the two-way traveltime for a given (half-) to the exact theory and make the same two as-
offset h, and t’ is the corresponding two-way zero- sumptionswhich underlie the conventional approach.
offset time Total moveout correction is the difference The zero-dip assumption implies that the earth
between t and t’. In this equation, v is taken to be model is stratified. The seismic energy recorded
the rms velocity at t’. Equation (17) is based upon over such an earth would be completely concentrated
the stratified earth (zero dip) assumption.The process at the zero spatial frequency, in this case k, = 0.
implied by equation (17) simulates mapping of the This in turn suggeststhat we set the normalized wave-
CMP gather onto zero offset. Following the NM0 number Y = 0 in DSR defined by equation (16).
correction, tracesof the CMP gather are stacked. This The resulting operator will be defined as the stacking
not only reduces the volume of the data but also en- (St) operator, where
hances the signal-to-noise (S/N) ratio.
St(H) = 2(1 - H2)“2 - 2. (19)
The CMP stack is regarded as an upcoming wave
field ,recorded zero offset. This allows us to obtain The factor -2 may at first appear to be a post facto
the migrated section by downward extrapolation. The addition to St(H). But, let us think what we want to
imaging condition is based on the so-called explod- do with this operator. The first part of St(H) takes
ing reflectors model (Loewenthal et al, 1974), in that each shot-receiver pair on a CMP gather down to a
at t = 0, the reflectors explode in unison sending an reflecting point following raypaths with cos(8) =
upcoming wave to the surface. Apart from being re- (1 - H2)“2. Next, we would like to come back
corded on a one-way time scale, this wave field is to the surface following a vertical path to the point
Prestack Partial Migration 1757
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 1. Real part of (y, t)-planes at z = 200,400,600, and 800 m superimposed.(a) DSR with h = 0 m, (b)
DSR with h = 400 m, and (c) Sep with h = 400 m. These frames represent impulse responsesof DSR and
Sep operators to point scatterersburied at the depths indicated.
1758 Yilmaz and Claerbout

halfway between the shot and receiver. This is simply Y and H, which would be present in the Taylor series
equivalent to a constantphaseshift (time retardation), expansion of DSR (Y, H). It is the absence of such
which accountsfor the factor -2.
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

terms that makes it possible to express Sep(Y, H)


It turns out that the NM0 shift given by equation in the separableform equation (22). The generalized
(17) is a stationary phase approximation to equation expression (21) for Sep(Y, H) allows us to define
(19). What the operator St(H) does is condensepri- other forms of separable approximations to DSR by
mary information on a CMP gather down to zero assigning nonzero values to Ho and/or to Ya.
offset. As far as this operator is concerned, stacking In conclusion, we made a rigorous analysis of con-
amountsto selecting off the zero offset and abandon- ventional processing and showed that it can be de-
ing all other offsets. Therefore, equation (19) is a veloped from the theory of the DSR equation. We
zero-dip NM0 + stack-type operator. discovered that conventional processing simply
Incorporating the zero-offset (h = 0) assumption utilizes a separableapproximation to DSR. The zero-
into the DSR operator is relatively more subtle. On a dip and zero-offset assumptionsare the basis for the
CMP gather, at and near h = 0, energy is essentially separability.
concentrated at zero spatial frequency, in this case How severe are these assumptions?We study the
kh = 0. In fact, the task that the NM0 correction following transfer functions
attemptsto accomplish is to push the primary energy
on a CMP gather toward kh = 0. So, setting the
normaiized offset wavenumber Hi = D in equation

exp i w DSR(k,,
[
kh, o)z
1 (244
(16), we define the exploding reflectors (ER) migra- and
tion operator as

ER(Y) = 2(1 - Y2)1’2. (20)


exp i f Sep(k,, kh, o)z
L 1 @W
Using the definition for Y (equation 15a), we obtain Inverse Fourier transforming expressions (24a) and
the zero-offset vertical wavenumber associatedwith (24b) over k,, kh and o yields the impulse response
equation (20) which turns out to be identical to equa- of each of the extrapolation filters (24a) and (24b).
tion (18). Thus, we conclude that the zero-offset Figure 1 shows the (y, t)-plane for constant h
operator ER (Y) derived from the double square-root and z. Each figure is actually a superpositionof four
equation is identical to the migration operator that is (y , t)-planes at different z-levels. These figuresrepre-
based on the ER model of conventional processing. sent the impulse responsesof DSR and Sep to point
Let us now define the following operator: scatterers buried at indicated depths z. The zero-
Sep(Y, H) = DSR(Y, Ha) + DSR(Ya, H) offset case (a) has hyperbolic, while the nonzero
- DSR(Yc,, Ho), (21) offset case (b) has the well-known table top trajec-
tories, as can be verified by the stationary phase ap-
where He and Ya are scalars. In order to grasp the
proximation to equation (16). The difference between
nature of the operator Sep (Y, H), let us consider the
the zero offset and the nonzero offset is quite evident,
simplest case: If He = Ye = 0, using the definition
particularly for the two shallow scatterers.The move-
for DSR given by equation (16), we have
out is also obvious as one comparesthe arrival times.
Sep(Y, H) = 2(1 - Y’)“’ Although the Sep operator handles the NM0 to a
+ 2(1 - H2)“2 - 2. (22) certain extent (perfect at the top of each trajectory),
it cannot account for the table-top trajectories in-
Adding equations (19) and (20) and comparing with
dicated by the nonzero offset section (b). Impulse
equation (22), we have
responsecharacteristicsclearly demonstrate that the
Sep(Y, H) = ER(Y) + St(H). (23) Sep operator does not handle wide offsets properly,
especially in the shallow portion of common offset
Hence, the operator Sep(Y, H) defined by equation
sections.
(23) describes the total downward continuation in-
THE DEVIATION OPERATOR
volved in conventional processing. Notice that if
Y = 0 we have, from equations (16) and (22), We really do not want to abandon the separable
Sep(Y = 0, H) = DSR(Y = 0, H) and if H = 0, form [equation (22)] unless there is a strong reason
we have Sep(Y, H = 0) = DSR(Y, H = 0). How- to do so, becauseit gives us the advantageof working
ever, if Y and H are both nonzero, the Sep (Y, H) either in midpoint or offset space independently. So
operatorfails to include terms with cross productsof we ask ourselves the following question: Can we
Prestack Partial Migration 1759

build a theory for the Sep(Y, H) operator such that it number of z-steps
will relax the zero-dip and zero-offset assumptions,
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

at least to a certain degree of accuracy?

Fundamental concepts
Common sense suggeststhat the answer must be
present in the discrepancy between the operators
DSR (Y, H) and Sep (Y, H). Therefore, we are
motivated to make the following formal definition:

Dev(Y, H) = DSR(Y, H) - Sep(Y, H). (25)

The operator Dev(Y, H) describes the deviation of


conventional processing from the exact theory of ;;b I
wave extrapolation. Substituting equations (16) and :
(22) into equation (25), we obtain the full expression - o
for this operator 03

Dev(Y, H) = [I - (Y + H)211’2
+ [l - (Y - H)2]“2
-2(1 - YZ)l’2
- 2(1 - H2)“2 + 2. (26) FIG. 2. Extrapolation and interpolation of wave fields
Equation (26) looks rather formidable. We would with the prestack partial migration equation. We
start at the bottom of the (z, t)-plane and work our
like to obtain a simpler expression for it. To second
way up in f. At each t-level, H is a fixed scalar. As
order in Y, i.e., 1 > H * Y, we expand the DSR and we move upward in section, we interpolate between
Sep operators using the square-root expansions pro- z-stepsof level t to extrapolate the wave field at level
vided in Appendix B. With equation (B- l), equa- t - At. Notice that the number of z-steps increases
tion (22) is expanded as follows: as we go up the section as shown in (a) for the post-
NM0 equation (42) and in (b) for the pre-NM0
Sep(Y, H) = - Y2 + 2(1 - H2)1’2. (27) equation (49). Poles are indicated by the arrows.

Similarly, using equation (B-2), equation (16) is


expanded as follows:

DSR(Y, H) z 2(1 - H2)“2


_ (1 _ ~2j-312~2.
(28) that fi does not depend upon the spatial frequency
We substitute equations (27) and (28) into equa- kh which couples adjoining offsets. We therefore
tion (25), and obtain the approximate form for the redefine the deviation operator by substituting fi
deviation operator for H in equations (25), (29), and (30). An approxi-
mate form of the operator Dev(Y, fi) can now be
Dev(Y, H) = [l - (1 - H2)-3’2]Y2. (29)
implemented in midpoint space.
Expanding equation (29) to second order in H yields How do we estimate A? To answer this question,
a simpler but coarser approximation in H: we recall the Fourier solution exp( -iwt + ikhh),
upon which His defined. Staying at a constantphase
Dev(Y, H) i= -;H2Y2. (30) implies

- o dt + k,,dh = 0
A closer look at these approximations will reveal
the fact that in both, H and Y are coupled. This de- or
feats. the whole point of improving the separable
form Sep(Y, H) of conventional processing. Specifi- (31)
cally, we would like to come up with a procedure
which, together with moveout correction, maps For a constant velocity medium, the moveout equa-
common offset sections onto zero offset more tion reads
accurately. The only possibility of utilizing the devia-
tion term is to replace H by an estimate of fi such (h2 + z2)1’2 = ;, (32)
1760 Yllmaz and Claerbout
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 3. Synthetic common offset sections over 8 point scatterersburied at depths between 150-1200 m at
intervals of 150 m in a constantvelocity medium (v = 3000 mlsec). (a) The zero-offset section (h 7 0), and
(b) common offset section with h = 400 m. Note the table-top appearanceof the responseof the shallow event
on the far offset section. Sampling interval is 4 msec and midpoint spacing is 12.5 m. The bandwidth is 6,
12-36, 48 Hz. Amplitudes decay as t-1’2.

FIG. 4. Impulse responsesof the pre- and post-NM0 partial migration operators. Imagine we had a common
offset section with wint-like resoonses(which are due to semiellidical reflectors). With h = 200 m. the zero-
offset mapping us&g the pre-NkO opeiator is shown in (a), and-with h = 400 k in (b). The responseof the
post-NM0 operator to the case of h = 400 m is shown in (c). We observe that the mapping involves a larger
number of midpoints for large values of offset (h) and for shallower events (small r). Notice the anisotropic
behavior near the pole indicated by the arrow in (b). Note the similarity of behavior to the responseof the pre-
NM0 operator. The pole for the post-NM0 operator is at t’ = 0 indicated by the arrow in (c).
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
1762 Yilmaz and Claerbout
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 6. A critical test of the performance of the prestack partial migration procedure is to image a common
offset section with and without applying this procedure. (a) The migrated zero-offset section shown in Figure
3a; (b) the migrated version of the moveout-corrected common offset section shown in Figure 3b; and (c) the
migrated version of the partially migrated common offset section shown in Figure 5d. Even though the 15
degree finite-difference algorithm was used for migration, no migration scheme can collapse the dominant
zero spatialfrequency as seen in (b). Note the better imaging obtained by partially migrated section (c). Under-
migration of very high dips in all three cases is due to the 15degree accuracy.

FIG. 7. Sensitivity of the partial migration operator to variation in velocity. The true velocity of the medium is
3000 m/set. Figure 5c after partial migration with v = 2400 misec (a), and v = 3600 m/set (b). Comparison
with Figure 5d, partial migration with the correct velocity, clearly indicates the processis quite robust, i.e.,
insensitiveto significant variations in velocity. This is a desirable property for sucha procedure since velocity
estimation is usually not completely accurate.
Prestack Partial Migration 1763
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 8. Sensitivity of the partial migration operator to offset value. True offset value is h = 400 m. (a) The
partial migration of the moveout-correctedsection shown in Figure 5c with a smaller offset value (h = 300 m),
as is (b), but with a larger offset value (h = 500 m). As we would expect, with a small offset value the section
is underprocessed,and with a large offset value it is overprocessed.

from which we may compute the stepout dt/dh, V


v,, = - (35)
cos (e) ’
dt
_=_ 4h
(33) where t3is regional dip and v is true medium velocity.
dh v’t’
In more general cases, fi could be determined by
Substitutingequation (33) into equation (3 1) and re-
ray tracing from theoretical velocity models or per-
calling the definition of equation (15b) for H, we
haps estimated by smoothing data. As will be seen
have the estimate
from model experiments, the main point to keep in
mind is that great accuracy is not required.
vt Our estimate for fi makes the coefficients of the
approximate Dev(Y, A) operators (29) and (30)
Equation (34) defines A in terms of surface offset scalar-like quantities. Computational cost of these
and two-way traveltime. This supportsour intuition two approximations is the same, but their accuracies
that we expect to see different stepouts dt/dh at differ. Therefore, we are inclined to choose equation
different offsets and times. The moveout equation (29), which is a better approximation in H.
(32) is also valid for stratified earth if we take v to be Adding Dev (Y, A) to Sep (Y, H), we have
the rms velocity down to depth z, in which case the
NewSep(Y, H) = Sep(Y, H) + Dev(Y, fi). (36)
velocityinequation(34) becomesv,, Noticealsothat
equatioh (34) is in the form that we wish, i.e., the Recalling that Sep (Y, H) represents conventional
offset value enters as a parameter into the deviation processing, it is proper to say that NewSep(Y, H)
operator. Another estimate for Z? could be made to represents “improved” conventional processing.
improve equation (34) by including a regional dip Actually, equation (36) can be directly obtained
factor. It is well known that therms velocity estimated from the definition of Sep(Y, H) given by equation
from CMP gathers has a dip dependency given by (21). We derived equation (22) by setting Ho = Y. =
(Levin, 1971) 0 in equation (21). Now, letting Ho = A, Y. = 0,
1764 Yilmaz and Claerbout
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 9. (a) A synthetic zero-offset and (b) a far offset section with h = 1000 m section over point scatterers
buried at several depths, z = 150-1500 m at intervals of 150 m, in a constant velocity medium (v = 3000
misec). The sampling interval is 4 msec and the midpoint interval is 12.5 m. The bandwidth is 6, 12-36,48 Hz,
and the amplitudes decay as t-l”

we have Adding equations (37b) and (37~) and comparing


the result with equation (37a), we have
Sep(Y, H; Ho = f?, Y, = 0) = DSR(Y, A)
+ DSR(0, H) - DSR(0, A). (37a) Sep(Y, H;H, = A, Y, = 0)
= Sep(Y, H;Ho = 0, Y, = 0)
Rewriting equation (21),
+ Dev(Y, I?).
Sep(Y, H; Ho = 0, Y, = 0) = DSR(Y, 0)
+ DSR(0, H) - DSR(0, 0). (37b) Comparing this result with equation (36), we see that

Now recall the definition for Dev(Y, 6) from NewSep(Y, H) = Sep(Y, H;Ho = fi, Y, = 0),
equation (25): (38)
Dev(Y, a) = DSR(Y, fi) - Sep(Y, A). which implies that NewSep is the old Sep with a
nonzero Ho = fi, which conventional processing
Setting H = fi in equation (37b) and substitutingfor
takes as zero.
Sep(Y, A),
Dev(Y, fi) = DSR(Y, fi) Prestack partial migration equation
- [DSR(Y, 0) + DSR(O, fi) We now describe the implementation of the ap-
- DSR(0, O)]. (37c) proximate form [equation (29)] of the deviation
Prestack Partlal migration 1765
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

II l~llllllllllllllllllllllllllllllllllllllllllllllllllllllllllll~llllil~l
IY

FIG. 10. CMP gathersover the model of Figure 9. (a) Raw gathers,(b) moveout-corrected gathers,and
(c) moveout-corrected and partially migrated gathers.Note the overcorrectedsteeplydipping events on
moveout-correctdagathers. H definedby equation(34) was not adjustedfor dip angle (0 = 0). In practice,it
is betterto adjustH for a regionaldip an le. The NM0 performssatisfactorilyfor flat or nearly flat eventsat
very smalloffsets[right sideof frame(b)B. Partialmigrationcorrectsfor dippingeventsandremovestheeffect
of far offsets[frame (c)l.

operator.The associatedvertical wavenumberis equation(18). The differencebetweenthem is that


the extrapolationvelocity in equation(40) is not the
k, = - f Dev(Y, fi). (39) mediumvelocitybut theonedefinedby equation(41).
Since this velocity is r-dependent, equation (40)
Making all the relevant substitutionsfrom equations cannot be implementedin o-space. Assuming no
(Ea), (29), and (34) and simplifying, we have the lateralvariationin velocity, it can be implementedin
final expression (k,, t)-space.
Referringto equation(41), Gis actuallydefinedas
a negative quantity since fi 5 1. Keeping this in
mind, equation (40) yields the following partial
where differentialequation:

G = Y [l - [l - (z)2]-3’2]. (41)

Noticethatequation(40) is in theform of theretarded This is the prestack partial migration equation


H-degreeapproximationto the zero-offsetmigration appliedon commonoffsetsection,whichis not move-
1766 Yilmaz and Claerbout

out corrected. 5 is interpreted as the velocity of


propagation adjusted for offset in the manner de- t’ = I [ 1 - (2)‘]“‘, (45a)
scribed by equation (41).
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Equation (40) can also be used in the case of a and


stratified medium. It is convenient, however, to
h’=h. (45b)
make the conversion
We also have the principle of invariance
= dz
r=2 -, (43) P(y, h, 7, r) = P’(y, h’, 7, 2’).
I 0 v(z) (46)

which yields The critical partial differentiation is

p’~~~(l_[l-(~)2]-3’z]pyy, (44) Z=$$+$$. (47)

aP’/ah’ is negligible for moveout-corrected data.


where we used the rms velocity in equation (34). Using equations (45),
Notice, however, that rrms is really a function of
the two-way zero-offset traveltime. We therefore $ = [ 1 + (~)2]? (48)
need to transform equation (44) to NM0 coordinates
defined as follows: Inverting equation (45a) and using equation (48),

FIG. 1I. (a) Conventional stackof the 6fold CMP gathersshown in Figure lob. (b) Stackof the 6-fold partially
migrated CMP gathers shown in Figure 10~. Note the severe interference on the shallow portion of the con-
ventional stack; this part is muted during stack. The partially migrated CMP gathersproduce a more coherent
stackthat is comparable to the zero-offset section shown in Figure 9a. We also managed to narrow the mute
zone.
Prestack Partial Migration 1767

we obtain the moveout-corrected form of equation them properly, and (2) nonzero data for t < 2h/v
(44: would not fit a wave propagation model.
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i The 15-degree finite-difference computational star


P Ttl = -v Pm, (49) described by Claerbout (1976) can be used to solve
4
equations (42) and (49) numerically. Writing equa-
where tion (42) in difference form,

f =; (I - [I + (~)z]3’2j. 6,6tP = -a f&.&P, (51)


where 6 is the difference operator, and

[I +($)“]-‘-.‘ (50) (52)

A final consideration is that the pre-NM0 equa- The t-outer algorithm (Claerbout, 1976) was used in
tion (42) has a pole at t = 2h/v [equation (41)], the present analysis for two reasons:(1) less memory
whereas the post-NM0 equation (49) has a pole at is required, and (2), more important, fi is constant
t’ = 0 [equation (50)]. At first, the presence of the along the extrapolation path. Notice that the extra-
poles seems worrisome. However, we may stop the polation coefficient “a” defined by equation (52) is
extrapolation at the poles for two reasons:(1) angles a function of offset and traveltime using (41). Nu-
become so large that our equations cannot handle merically, for stability reasons, we prefer a to be a

FIG. 12. Migration of (a) the conventional stack shown in Figure 1la, and (b) that of the stack improved by
the prestack partial migration procedure shown in Figure 1lb. It is apparent that we achieve better imaging
with the prestack partial migration.
1766 Vilmaz and Claerbout

global constant over the entire (z, t)-plane. This property for an algorithm that operates mostly on
suggeststhat the number of z-stepswill vary with time wide offset angles (large h or small t). although it
t. Solving equation (52) for AZ, we determine the does necessitate an interpolation procedure to re-
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

numberof z-stepsat a given time t. Figure 2 is a graph sample the (z, y)- or (z, k,)-plane. The interpola-
of the number of z-steps as a function of time for tion problem is described in detail in Appendix C.
pre- and post-NM0 equations. Notice that the num-
ber of z-steps increasesrapidly near the pole. Figure Model experiments
2 suggeststhat we startwith a small number of z-steps We will test the performance of the prestackpartial
at the bottom of the (z, t)-frame, i.e., we use a very migration procedure, both before and after moveout
coarse extrapolation interval, which is not how we correction. We also would like to determine some
would proceed using a finite-difference migration of its properties, such as sensitivity to velocity varia-
scheme. As we go up the section in time we extra- tion and offset value. An ideal testing model is point
polate at much finer intervals. This is a desirable scatterers buried at various depths. Such a model

FIG. 13. Conventional processingof the Gulf data: moveout correctionplus stackplus a band-passwith cut-off
frequencies 6, 12-48, 72 Hz. Note that the steeply dipping events are not stacked properly, as verified by
Figure 14.
Prestack Partial Migration 1769
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

5
a b
FIG. 14. Moveout-corrected CMP gathers from the Gulf data. Flat events (a) have been corrected properly,
but dipping events (b) are obviously overcorrected. Offset increases to the right.

would consistof a wide range of dips. Figure 3 shows can be used for better imaging. However, no migra-
common offset sections with h = 0 and h = 400 m tion algorithm can possibly collapse the energy at or
over such an earth model. near k,, = 0 as seen in Figure 6b.
Figure 4 shows the impulse responsesof the pre- How sensitiveis the operatorDev (Y, fi) to velocity
and post-NM0 partial migration equations, which variation and offset value? Figure 7 shows that a
clearly demonstratesthe role of the offset; the larger Z20 percent deviation from the true medium velocity
the offset and shallower the event, the wider the range made no significant difference. However, a ? 25
of midpoints involved in mapping to zero offset. We percent change in offset value did have an effect on
can also see the response near the pole that pro- the output, as shown in Figure 8.
duces evanescentenergy. Let us now simulate the entire sequenceof seismic
Figure 5 illustrates the comparative performance dataprocessingby consideringa 6-fold point scatterer
of the pre- and post-NM0 differential equations. model. Offset values are h = O-1000 m with an
These and subsequentmodel studies were done in increment of 200 m. The zero and the farthest offset
the (k,, r&domain. The post-NM0 operatoris clearly sections are shown in Figure 9 for comparison.
superior to the pre-NM0 operator. As we near the Figure 10a shows every twelfth CMP gather starting
pole t’ = 0, we observe the anisotropic dispersion. from the left and going to the center of the model.
A proper test of the performance of our equations is First let us follow the conventional route. Moveout
to migrate the outputsof Figures 5b and 5d, and com- corrected CMP gathers are shown in Figure lob. We
pare the results with the migrated zero-offset section can clearly see the effects of steep dip and large off-
of Figure 3a. Figure 6 clearly shows the effect of the sets on the NM0 correction. The CMP gather from
post-NM0 operatoron the performance of migration. the central trace where dip is zero is perfectly
Here we usedthe 15degree migration scheme for the flattened. Moving up to the left, however, we see
sake of convenience. Certainly other schemes, such the overcorrected gathers [Levin effect (Doherty,
as the 45-degree or the frequency domain algorithms, 1975)]. The resulting CMP stack is shown in Figure
Yllmaz and Claerbout
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a b
FIG. 15 Moveout-corrected CMP gathersafter partial migration. (a) The processdid not affect flat events, but
(b) corrected dipping events so that they stack coherently.

lla. Ideally, we would like to have a zero-offset offsets varying between 1037-12,740 ft. The pro-
section like the one shown in Figure 9a. Let us now cessing sample rate is 4 msec, and total record length
include the prestack partial migration process in the is 5 sec. The number of midpoints is 282, with an
conventional sequence. Applying the post-NM0 interval of 83 ft.
operator on moveout-corrected common offset sec- The processing sequence is as follows.
tions, we have the CMP gathersshown in 1 (gure 10~.
Notice the considerable improvement in the wide (1) At every sixth midpoint, velocity estimation
offset and steep dip regions. In the upper part of the was done by the semblance technique. The velocity
CMP gathers, we see events that are not corrected function was chosen which stacked flat or near-flat
properly. This is becausethese events have dips that events best, free of any dip effects.
are beyond the range of dips our operator can handle. (2) CMP gathers were moveout corrected, simul-
The improved CMP stack is shown in Figure 1lb. taneously stacking every four adjacent offsets. Thus,
The usual practice is to mute the top portion of the we reduced the size of the data by creating substacks
stack. By improving the CMP stack, we manage to while enhancing the SIN ratio. Considerable subse-
narrow the mute zone considerably. By doing so, quent computational effort was also saved in this
we are able to recover primary energy under the inter- way. Certainly, substacksare allowable for only a
ference zone shown in Figure 1la. Migrations of the narrow range of offset values.
CMP stacks with and without the post-NM0 partial (3) Figure 13 shows the conventional stack of
migration are shown in Figure 12. Better imaging is moveout-correctedCMP gathers. Referring to Figure
achieved with prestack partial migration. 14, we seethat dipping events are over-corrected, thus
resulting in poor coherency in stack.
Field data example (4) The moveout-corrected common offset sec-
A field data set provided by Digicon, Inc. from the tions were partially migrated using equation (49). in
Gulf of Mexico was processed using the prestack which fi was adjustedfor a regional dip of 45 degrees.
partial migration equation. The data are 4%fold with The effect of this processis seen on the CMP gathers
Prestack Partlal Migration 1771

shown in Figure 15. Note that flat events are left as lateral velocity variation. In particular, we will study
they are, but on dipping events the offset effect was the Sep(Y, H) operator and investigate terms that
removed. may be of significance.
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(5) Subsequentstacking (Figure 16) is superior to


the conventional stack shown in Figure 13. Dip- The double square-root equation
ping events present in the far-offset data are well We would like to define explicitly the DSR equa-
preserved. tion for laterally variable media. This in effect implies
that we distinguish between the velocity along the
LATERAL VELOCITY VARIATION
shot axis and the velocity along the receiver axis.
Lateral variation in velocity is of great concern in This allows us to use equation (1) to extrapolate
areaswith complex structural settings or rapid facies wave fields in media with lateral velocity variations.
changes. With this motivation, we would like to It is convenient to express the associated vertical
extend the theory developed earlier to incorporate wavenumber k, of equation (2) as

FIG. 16. Stack of the partially migrated CMP gathers. Comparison with the conventional stackshown in Figure
13 indicatesconsiderableimprovement of dippingevents. Processingsequenceis identical to that of Figure 13,
with addition of partial migration before stack. H was adjusted for a regional dip of 8 = 45 degrees.
1772 Yllmaz and Claerbout

Making all the relevant substitutions,


k, = - 0 (f - $,l’, (53)
NewSep = [M(s) - (Y - ti)2]1’2
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Writing equation (53) once for receivers and once + [M(g) - (Y + I)2]l’2
for shots, then summing, yields + (M(s) - H2)“2 + (M(g) - H2)l”
- (M(s) - liy - (M(g) - A2)“2.
(60)
The third and fourth squareroots are NMO-like terms
similar to equation (19), i.e., the stacking operator.
(54)
Abandoning them, we are left with the retarded
migration part:
Using equation (13), we put equation (54) into
midpoint-offset space Mig = [M(s) - (Y - fi)2]1’2
+ [M(g) - (Y + A)2]1’2
k, = - o I[-& - (?)‘I”
’ - (M(s) - As)“2 - (M(g) - A2)“2.
(61)
Referring to the second order square-root expansion
+ [-& - (%,‘]‘-“,. (55)
provided in Appendix B [equation (B-3)], we expand
For brevity, we leave the variables s and g in the the first two terms and obtain
velocity function. For convenience, we redefine the
normalized wavenumbers

[i] =+-[:I. Wa,b) (62)


We make the following further definitions: where

Mb) = -& (574 C(s) = (M(s) - A2)1’2, (63a)


and
and
C(g) = (M(g) - A2)1’2. (63b)
M(g) = J- W’b) The second term on the right-hand side of equation
v(d2’ (62) is a typical 15-degree-like migration equation
where M is the square of the slowness function. which is adjusted for offset using equation (63). But
the first term is entirely a stranger to us
Upon substitutionof equations(56) and (57), equa-
tion (55) takes the form
A Y. (64)
I
k, = - o {[M(s) - (Y - H)2]“2
+ M(g) - (Y + w211. (58) Substituting equations (56a) and (63), we have

Omitting the scaling frequency w, we define the DSR k, = - o NewMig


operator in laterally varying media as
1
DSR(Y, H) = [M(s) - (Y - H)2]“2
+ [M(g) - (Y + H)2]1’2. (59)
kz = - 1
(M(s) _ fi2)1'2

The separable approximation :65)


We defined the improved conventional processing
by the NewSep(Y, H) operator. Let us derive this Notice that the new definitions of Y and H [equa-
operator using the newly defined DSR(Y, H) given tion (56)] are the old Y and H [equation (15)] divided
by equation (59). We make use of the short-cut by v. We may also define the new A as the old fi
derivation equation (38): [equation (34)] divided by v; thus

NewSep(Y, H; Ho = f?, Y, = 0) = DSR(Y, I!?) &?!I,


+ DSR(0, H) - DSR(0, A). v2t
Prestack Partial Migration 1773

Using the new definition for fi and assuming it is


small, equation (65) becomes
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

k = _ h v(s) - v(g) k
z v2 Y’
t
which transforms to

p = _ h v(s)- v(g) p
z
t
v2 Y’ (66)
2.5
where v is considered to be the average of v(s) and
v(g). This equationimplies a pure lateral shift. Notice
the coefficient is strongly dependent upon the offset
value and the lateral velocity gradient. For zero
offset, this equation meansno operation. However, in
a laterally varying medium, the thin lens term of the
migration process still needs to be applied even to
zero offset. If we imagine a point scatterer in a
medium with lateral velocity, equation (66) implies
the following: If the top of the diffraction hyperbola
on a common offset section (hr) is located under
midpoint (yr), then on a common offset section (hs),
the top will be located under a different midpoint 2.5
(y2). The lateral shift (y2 - yr) will depend upon
how strongthe lateral velocity gradient is and how far
apart the two offsets are (h2 - hr).
Let us investigatethe significance of equation (66).
During the processing of the field data set described FIG. 17. Portions of the two moveout-corrected and
in the previous section, it was observed that some of partially migrated common offset sections from the
the dipping events showed evidence that some lateral Gulf data. Notice the phase shift of nearly 112 wave-
phase shift from one common offset to another was length across a lateral extent of roughly 4000 ft be-
tween midpoints 100-150. The event marked is an
still presenteven after moveout correction and partial
indicator of this phase shift possibly due to lateral
migration. An example is shown in Figure 17. One variation in velocity.
possibleexplanation is that there may be lateral varia-
tion in velocity due to overpressurized shale or
prominent structural complexity.
First let us transform equation (66) to NM0 co-
ordinates. The transformation defined by equation P(k,, z, W) = P(k,, 0, CO)~-~~~Y’. (70)
(59) will simply change the coefficient of equation
Given the phase shift @ = C k,z, we can estimate
(66):
the lateral velocity gradient. Solving for v(s) - v(g)
P, = -c P,, (67) in equation (68) and substituting k, = (o/v) sin (0)
where, with the small-offset approximation, and z = vt’/2, we obtain

2 @ vz
C = ; [1 - ; c3j2] ‘(‘) izv(‘). (68) v(s) - v(g) = Av =
sin (0)~
!
For the sake of a crude estimation, let us consider
(71)
a constant lateral velocity gradient. This will allow
us to write an analytical solution to equation (67). where 0 can be regarded as the average dip over the
Fourier transforming equation (67) in t and y, P = distance s - g.
P(k,, z, w), we obtain Let us plug some numbers into the expression (71).
Referring to Figure 17, between midpoints 100-150,
P, = -iCk,P. (6%
we have @ = rr radians, h = 5200 ft, t’ = 2.5 set,
The solution to equation (69) is v(t’) = 7500 ftisec, 0 = 15 degrees, and w =
1774 Yilmaz and Claerbout

2a X (25 Hz). The resulting value for A v is approxi- ACKNOWLEDGMENTS


mately 500 ft/sec per 1000 ft of lateral distance. Even
We are indebted to John Sherwood who developed
this crude estimate gives an idea as to the significance
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the concept of prestack partial migration, to Robert


of equation(64). A closer study of this term may lead
W. Clayton who played a large part in the develop-
to a procedure by which one can accurately estimate
ment of the double square-rootequation, and to David
lateral variation in velocity. In particular, the statics
Brown who clarified our understanding of the
problem can now be reconsideredto take into account separability of operators.
offset dependency.
This work was supportedin part by the sponsoring
members of the Stanford Exploration Project (SEP)
SUMMARY AND CONCLUSION
and in part by the Scientific and Technical Research
We reviewed the fundamentals of extrapolation of Council of Turkey (TUBITAK) and The Turkish
wave fields which are acquired via seismic experi- Petroleum Corp. (TPAO). We are thankful to all
ment with many shotsand receivers. Mathematically, of them. The field data set was provided by Digicon,
the wave field extrapolation is described by the DSR Inc., who first brought to our attention a devilishly
equation. We determined that conventional process- clever procedure which is a shortcut to migration
ing is equivalent to a separable approximation to the before stack.
DSR operator. Underlying this separation are two
assumptions-zero-dip NM0 and zero-offset
migration. Through the separable approximation,
downward extrapolation is achieved in two in- REFERENCES
dependent steps: seismic data are moveout corrected Claerbout, J. F., 1976, Fundamentals of geophysicaldata
and stacked in offset space, and then migrated in processing: New York, McGraw-Hill Book Co.. Inc.
midpoint space. 1978, Class notes: Stanford University, ‘Depart-
ment of Geophysics.
To the conventional processing scheme, we added Conte, S. D., 1965, Elementary numerical analysis: New
partial migration before stack. The theory for this York, McGraw-Hill Book Co., Inc.
Doherty, S. M., 1975, Structure independent velocity
process is also based on the DSR equation. The estimation: Ph.D. thesis, Stanford University.
algorithm emergesfrom the deviation operator, which Doherty, S. M., and Claerbout, J. F., 1976, Structure in-
is formally defined as the difference between the DSR dependent velocity estimation: Geophysics, v. 41, p.
850-881.
operator and its separable form. Prestack partial Dohr, G. P., and Stiller, P. K., 1975, Migration velocity
migration is applied on common offset sections in- determination: part II: Applications: Geophysics, v. 40,
p. 6-16.
dividually. It removes the effect of offset on dipping
Gardner, G. F., French, W. S., and Matzuk, T., 1974,
events, thus producing a more coherent stack. The Elements of migration and velocity analysis: Geo-
processis robust, i.e., insensitive to any significant physics, v. 39, p. 81 l-825.
Judson, D. R., Schultz, P. S., and Sherwood, J. W. C.,
velocity variation. 1978, Equalizing the stacking velocities of dipping events
In the case of lateral velocity variation, a new via Devilish: Presented at the 48th AM& International
term emerged from our analysis of the separable SEG Meeting, October 30, in San Francisco.
Levin, F. K., i971, Apparent velocity from dipping inter-
approximation. This term involves pure lateral shift face reflections: geophysics v. 36. o. 510-516.
due to nonzero offset and the lateral velocity gradient. Loewenthal, D., Lu, i.: Robe&, R., and Sherwood,
J. W. C., 1976, The wave equation applied to migration:
A rough estimate (500 ft/sec difference over a dis-
Geophys. Prosp., v. 24, p. 380-399.
tance of loo0 ft) on a field data set indicates that a Sattlegger, J. W., 1975, Migration velocity determination:
lateral velocity estimation scheme may be formulated part I: Philosophy: Geophysics, v. 40, p. l-5.
Sherwood, J. W. C., Schultz, P. S., and Judson, D. R.,
based on the theory developed in the previous sec- 1976, Some recent developments in migration before
tion. In particular, the offset angle and lateral stack: Released by Digicon, Inc.
velocity gradient are both significant near surface. Stolt, R. H., 1978, Migration by Fourier transform: Geo-
physics, v. 43, p. 23-48.
Therefore, the theory has important implications for Yilmaz, O., 1979, Pre-stack partial migration: Ph.D. thesis,
the statics problem. Stanford University.
Prestack Partial Migration 1775

APPENDIX A
WAVE FIELD EXTRAPOLATION
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

We start with the scalar wave equation, which as may be readily verified by substituting equation
describesthe propagation of the compressionalwave (A-4) into equation (A-5). Let us define the vertical
field P(x, z, I) in a medium with velocity v(x, z) and wavenumber as
constant material density
k, = _ !!![l _ (%)“l”“. (A-6)
a2
;;;i+-&”
a2
v2
2

at2 1 P=O. (A-1) VL \w/ J

With this, equation (A-4) takes the simple form

P(k,, z, o) = P(k,, 0, co)eLkzz. (A-7)


We would like to determine the reflectivity P (x, z, 0),
given the upcoming wave field P (x, 0, t) recorded at
the surface. This requires extrapolating the surface Our main objective here is to interpret equation
wave field down to depth z and collecting it at (A-7) as a tool for downward extrapolating the wave
t = 0. field given at the surface. The Fourier transform
It is to our advantageto decompose the wave field allowed us to decompose this wave field into mono-
into monochromatic plane waves with different chromatic plane waves, each traveling at a different
angles of propagation from the vertical. Therefore, angle from the vertical. We identify these plane
we would like to work in the Fourier domain when- waves by attaching them to a unique (w, k,) pair of
ever possible. We may Fourier transform the wave numbers. Let us now consider one of these plane
field over time j. If we assume no lateral velocity waves as depicted in Figure A- 1. Imagine that this
variation, we may also Fourier transform over the plane wave passed point P at t = 0, traveled up-
horizontal axis x. Thus, we have ward, and was recorded by a receiver at the surface
point A at time 1. We would like to take the waveform
rk,x-id& dj,
P(k,, z, 0) = P(x, Z, r)e at point A sitting on the wavefront at time t back
II
(A-2a) to where it actually was, i.e., the reflection point P.
It seems sensible to travel back using the same ray-
and inversely,

P(x, z, t) = P(k,, z, co)e-(ikxx-Lwt) dk,do.


II
(A-2b)

Applying the differential operator of equation (A- 1)


onto (A-2b), we get

-$PL z, o) + c$ - k:) P(k,, z, w) = 0.

(A-3)
For simplicity, let us further assume the constant
velocity case. We will look into the case of stratified
earth later in the section. The upcoming wave solu-
tion to equation(A-3) can be immediately recognized
as

P(k,, z, w) = P(k,, 0, 0) .

(A-4)
which is also the solution to the one-way wave
equation

f Pk, z, o) = -i $ - k:)“” P(k,, z, co),

(A-5) FIG. A-l. On extrapolating wave fields.


1776 Yilmaz and Claerbout

path. This means that downward continuation does which is exactly what equation (A-7) implies, except
not change the horizontal wavenumber k,. Let us 1that equation (A-7) was derived for constant v.
move the wavefront so that the waveform at A is How do we know that we reached our destination
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

now at A’ at depth AZ. If we actually had a buried point P and did not passbeyond it? Seismic imaging
receiver at A ’ , it would have recorded our plane wave is not completed unless we impose a stopping con-
at t - At, where At is the traveltime between dition onto downward continuation. Here, we simply
A - A’. In other words, going down AZ, we changed terminate the process when our clock, which mea-
the traveltime by -At. From the geometry of Figure sures
A- I, we have

At = t cos(B) AZ, (A-8)


t-
I At,

reads zero traveltime. This completes our discussion


where v/cos(t3) is known as the vertical phase of the fundamentals of seismic imaging.
velocity. What we know about the plane wave is its Let us now turn our attention to the case of a strati-
k, and o. Suppose the distance between A - A ’ is fied earth. Since we have not Fourier transformed P
X wavelength. At time (t - At), the wavefront over z, the one-way wave equation (A-5) is also
intersects the x-axis at a distance of A, from A. valid for v(z)
Referring to the geometric relation in Figure A- 1,

A 2 P(k,, z, o) = -i (-$ - k:)1’2P(k,, z, w),


- = sin(E)). (A-9a)
A, (A-l 1)

Thus, using the definitions A = 2n/(o/v) and in which case equation (A-6) becomes
h, = 2v/k,, we obtain
k,(z) = - --$ [I - (+j21Li2. (A-12)
sin(e) = $, (A-9b)
Equation (A-l 1) has the following solution
and
P(k,, z, 0) = P(k,, 0, 4 .
cos(0) = [I - (?r]“‘, (A-9c)
. exp [i [ k,(z)dz], (A-13)

where w/v is the wavenumber along the raypath and


which immediately verifies equation (A-l 1) when
k, is the wavenumber along the x-axis. Substituting
substituted.However, in order for this to be a proper
equation (A-9c) into equation (A-8), we have
wave solution, we also would like it to satisfy the
two-way scalar wave equation (A-3). From equa-
At = i [1 - /J$)‘]“‘A~. (A-1Oa)
tion (A-l 1),

As we move down, we do not want to change the d2 dk, (z) dv (z)


-- P + i k,(z)%,
wave amplitude. Given the change in traveltime -A t d22p=i dv (z) dz
by equation (A-IOa), the corresponding phase shift
where P = P(k,, z, w). Substituting equation
will simply ‘be --WAC Atea& Ar-step of descent,
(A- 11) for dP/dz,
we may assign a different velocity v(z) to the wave-
form. This suggests that we change the angle (0) d2 dk, (z) dv (z)
- - P - k;(z)P.
from the vertical as we continue our journey along d22p=i dv(z) dz
the raypath down to the destinationpoint P. The total
If we ignore the velocity gradient dv(z)/dz, the final
phase shift to which the waveform has been subject
expression is
when we arrive at P is -J odt. In Fourier domain,
all we have to do is multiply the transformed surface
wave field P(k,, 0, o) by
$ P + k;(z) P = 0.

(A-lob)
Prestack Partial Migration 1777

Substitutingequation (A- 12) into this expression, we stratified earth model if we neglect the vertical gra-
obtainthe scalarwave equation(A-3). Consequently, dient of material velocity.
the DSR equation can be extended to the case of the
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

APPENDIX B
SECOND ORDER SQUARE-ROOT EXPANSIONS

B-l [l - (a + A)2]1/2
Consider the simplest case
2 = (1 - a2P2 [I - 5 - 2t1 A2,2,2]-

(1 - 4 l/Z=pA_A (B-l)
2 8’ (B-2)
where 1 + A. B-3

B-2 To second order in A, i.e., b > a + A,

To second order in A, i.e., 1 > a S=-A, [b2 - (a + A)2]1'2


= [(b2 - a2) - (2uA + A2)]1'2.
[l - (a + A)2]'/2 = [I - (a2 + 2aA + A2)]1'2.
Similarly to the development of equation (B-2),
a could be close to 1; therefore, factoring (1 - a2)
from the square root prior to expansion, we get ~ 03 - a2)1/2 [I - 2by’a;2]1’2
[l - (a + A)2]'/2
5 (b2 - a2)lf2.
5 (1 - a2)1/2 [1 - 2y_+a;2]
1’2
2aA+A2 (2uA + A2)2

Applying equation (B-l) to the square root on the


’ - 2(b2 - a2)- 8(b2 a2)2 1
right, Finally,
[I - (a + A)2]1'2 [b2 - (a + A)2]"2

= (1 - a2)1’2 [I - i;“;‘JatI = (b2 - u2)1/2 [ 1 - e2

- (2aA + A2)' b2A2


8(1 - a2)2 I .
-
2(b2 - uy2 1 (B-3)
Expanding the right-hand side and ignoring terms of Setting b = 1, equation (B-3) reduces to equation
higher than secondorder in A, the final expression is (B-2), as should be the case.

APPENDIX C
WAVE FIELD INTERPOLATION
Figure C-l illustrates different numerical tech- of Aitken’s algorithm for interpolation (Conte, 196.5),
niques for the interpolation required during extra- andthe techniqueof cubic spline (L. Morley, personal
polation in z. Although the nearestneighbor technique communication) perform equally well; the latter,
is the cheapest, it is obviously an undesirablechoice, however, is computationally more attractive. Model
becausecoarse interpolation creates artificial diffrac- tests and field data processing were done using the
tors, thereby causing a noisy appearance. Linear cubic spline technique for interpolation.
interpolation could be considereda possible means of Figure C-l illustrates a problem of concern to us.
interpolation in production-oriented implementa- Notice that near the pole there is actually a discon-
tion of prestack partial migration. The four-point tinuity. This is due to the limitation on the number
Lagrangian technique, which is a specialized form of z-steps that can be handled by the computer.
1778 Yilmaz and Claerbout
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

n
-Y

FIG. C-l. Illustrating possible means of interpolation required during resampling of the (z, y)-plane with a
different AZ(~) at each time sample [equation (52)]. (a) Nearest neighbor results in undesired numerical arti-
facts; (b) linear interpolation may be considered for production-oriented algorithms even though it has some
undesirablefeatures, as can be seen on the margins; (c) the four-point Lagrangian, and (d) cubic spline are the
most desirable types of interpolation. Arrows indicate the time beyond which the number of z-steps is fixed
due to limitation in data storage. This caused underprocessingof the data above the arrows.
Prestack Partial Migration 1779
Downloaded 05/17/15 to 130.132.123.28. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. C-2. Insufficient number of z-stepsin (a) and undersamplingin A z in (b) clearly show incomplete mapping
of the top event to zero offset. Dispersion near the top event is partly anisotropic, which is due to the limited
accuracy of equation (42) very near the pole at t = 2h/v.

Theoretically, one needs to take an immense number the pole, t approaches2h/v. This implies extremely
of z-steps near the pole. If we make an attempt to large offset angles, which are not accurately handled
allow more z-steps, we introduce dispersion, as by our equation, and it explains the anisotropic
shown in Figure C-2a. If we take coarser z-steps, the dispersion near the pole in Figure C-2a.
resulting output is poorer (Figure C-2b). As we near

You might also like