You are on page 1of 204

9 Earth Modeling in Depth

• Introduction • Inversion Methods for Data Modeling • Inversion Procedures for Earth Modeling • Velocity-Depth
Ambiguity • Model Representation and Visualization • Models with Horizontal Layers • Dix Conversion • Co-
herency Inversion • Near-Surface Layer with Laterally Varying Velocities • Model with Low-Relief Structure
• Stacking Velocity Inversion • Coherency Inversion • Velocity Resolution • Model with Complex Overbur-
den Structure • Image-Gathers • Constant Half-Space Velocity Analysis • Model Building • Time-to-Depth
Conversion • Time Structure Maps • Interval Velocity Maps • Depth Structure Maps • Calibration to Well Tops
• Layer-by-Layer Inversion • Structure-Independent Inversion • Model Updating • Residual Moveout Analysis •
Reflection Traveltime Tomography • Limitations in Resolving Velocity-Depth Ambiguity by Tomography • Turning-
Ray Tomography • Exercises • Appendix J: Data Modeling by Inversion • The Generalized Linear Inversion
• The GLI Formalism of Deconvolution • Applications of the GLI Technique • Dix Conversion • Map Processing •
Reflection Traveltime Tomography • Threshold for Velocity-Depth Ambiguity • References

9.0 INTRODUCTION

Seismic representation of an earth model in depth usu- gies to build an earth model (Section 9.4). Finally, we
ally is described by two sets of parameters — layer shall use residual moveout analysis and reflection to-
velocities and reflector geometries. Depth migration is mography to update an initial model (Section 9.5) and
the ultimate tool for delineation of reflector geometries thus obtain a final model that can be used to image
(Chapter 8). If layer velocities are determined accu- the subsurface in depth as discussed in Chapter 8. By
rately, reflector geometries can be recovered by iterative being equipped with methods and strategies for inter-
depth migration (Section 8.2). Difficulties in estimating val velocity estimation, model building, and updating
layer velocities with a required level of accuracy make (Chapter 9), and depth migration (Chapter 8), we shall
the earth model estimation a challenging task for the then study in Chapter 10 complete case studies for earth
geophysicist. modeling and imaging in depth.
In this chapter, first, we shall discuss practical in- Nearly all of the practical methods of layer velocity
version methods to estimate layer velocities (Sections estimation are based on ray theory, and more specifi-
9.1, 9.2, and 9.3). Next, by using appropriate combi- cally, on inversion of seismic traveltimes. Velocity esti-
nations of inversion methods, we shall develop strate- mation methods include Dix conversion of rms veloc-
1354 Seismic Data Analysis

ities, inversion of stacking velocities, coherency inver- eling of the nonzero-offset traveltimes associated with
sion, and analysis of image gathers from prestack depth the CMP reflection event that corresponds to the base
migration. of the nth layer and determining the moveout velocity
As described in Section 8.0, velocity variations by fitting a hyperbola to the modeled traveltime trajec-
within the earth may be characterized in two ways tory. This procedure is repeated for a range of constant
— structure-dependent and structure-independent. A trial velocities, and the velocity that yields the mini-
structure-dependent earth model comprises geological mum discrepancy between the actual stacking velocity
formations with interfaces that coincide with distinct and the modeled moveout velocity is assigned to the
velocity contrasts. We encounter structure-dependent nth layer below the CMP location where the stacking
earth models in areas with extensional and compres- velocity inversion is being performed.
sional tectonics, and especially in areas with salt Coherency inversion (Landa et al., 1988) also re-
and overthrust tectonics. A structure-dependent earth quires time horizons picked from unmigrated CMP-
model usually requires a layer-by-layer estimation of stacked data. However, in lieu of stacking velocities as
layer velocities and delineation of reflector geometries for stacking velocity inversion, coherency inversion re-
that coincide with the layer boundaries themselves. quires analyzing CMP gathers themselves. Again, as-
A structure-independent earth model comprises ge- sume that a velocity-depth model already has been es-
ologic formations with interfaces that do not necessar- timated for the first n − 1 layers, and that we want
ily coincide with distinct velocity contrast. Structure- to estimate the layer velocity for the nth layer below
independent earth models often are associated with low- a CMP location. For a trial constant velocity assigned
relief structures or stratigraphic plays involving a de- to the nth layer, coherency inversion involves normal-
positional sequence with facies changes. A structure- incidence time-to-depth conversion of the time horizon
independent earth model may be estimated initially by associated with the base of the nth layer, then mod-
Dix conversion of rms velocities without requiring a eling of the nonzero-offset traveltimes associated with
layer-by-layer analysis. the CMP reflection event that corresponds to the base
The simplest method for estimating layer veloci- of the nth layer, and computing the semblance (Section
ties is Dix conversion of rms velocities (Dix, 1955). The 3.2) within a CMP data window that follows the mod-
method requires the rms velocities associated with the eled traveltime trajectory. This procedure is repeated
layer boundaries that are included in the earth model to for a range of constant trial velocities and the velocity
be constructed. The rms velocities ideally are estimated that yields the highest semblance value is assigned to
by prestack time migration (Section 5.3). Alternatively, the nth layer below the CMP location where the co-
a smoothly varying form of stacking velocities estimated herency inversion is being performed.
from dip-moveout corrected data may be a reasonable Time horizons used in normal-incidence time-to-
substitute for rms velocities (Section 5.2). Less desir- depth conversion as part of the stacking velocity in-
ably, stacking velocities themselves with a fair degree of version and coherency inversion procedures are picked
smoothing applied may be used in lieu of rms velocities. from unmigrated CMP-stacked data. Alternatively,
The Dix conversion formula (Section J.4) is valid for time horizons interpreted from the time-migrated vol-
horizontally layered earth models with constant layer ume of data can be unmigrated to obtain the time hori-
velocities and small offsets. For an earth model with zons equivalent to the time horizons picked from the un-
dipping layer boundaries and layer velocities with ver- migrated data. We circumvent the picking of prestack
tical and lateral variations, more accurate methods are reflection traveltimes in coherency inversion by measur-
required such as stacking velocity inversion, coherency ing the discrepancy between the modeled and actual
inversion and image-gather analysis. traveltimes by way of semblance. Similarly, we avoid the
Stacking velocity inversion (Thorson et al., 1985) picking of prestack reflection traveltimes in stacking ve-
requires time horizons picked from unmigrated CMP- locity inversion by measuring the discrepancy between
stacked data and stacking velocities at analysis loca- the modeled and actual stacking velocities.
tions. Assume that a velocity-depth model already has Stacking velocity inversion and coherency inver-
been estimated for the first n − 1 layers, and that we sion both take into account vertical velocity gradients
want to estimate the layer velocity for the nth layer which may be available from sonic logs. The methods
below a CMP location. For a trial constant velocity as- also honor ray bending at layer boundaries. While Dix
signed to the nth layer, the method involves normal- conversion assumes a hyperbolic moveout for the reflec-
incidence time-to-depth conversion of the time horizon tion event that corresponds to the base of the layer un-
associated with the base of the nth layer, then mod- der consideration, both interval velocity estimate from
Earth Modeling in Depth 1355

coherency inversion and stacking velocity inversion are Inversion Methods for Data Modeling
based on nonhyperbolic CMP traveltime modeling.
Both stacking velocity inversion and coherency in- Historically, the term seismic inversion often has been
version can be considered accurate for velocity-depth used within the context of acoustic impedance estima-
models with smoothly varying reflector geometries and tion from a broad-band time-migrated CMP-stacked
lateral velocity variations greater than the effective ca- data. This narrow meaning commonly is referred to
ble length associated with the layer boundary under as trace inversion. In practice, however, seismic inver-
consideration. As for conventional stacking velocity es- sion has a broader scope of applications which can be
timation (Section 3.2), the accuracy in interval velocity grouped in two categories — data modeling and earth
estimation from Dix conversion, stacking velocity in- modeling.
version, and coherency inversion are all influenced by What we do in seismic data processing described
the reflector depth, magnitude of the velocity, and the in Chapters 1 through 7 is based largely on data mod-
cable length. Specifically, the deeper the reflector, the eling. An observed seismic wavefield can be described
larger the layer velocity above, and the shorter the ca- in two parts — traveltimes and amplitudes. Seismic
ble length, the less accurate is the interval velocity es- amplitudes are more prone to the detrimental effects
timate. of noise as compared to traveltimes. Hence, in seismic
To estimate, update and verify velocity-depth mod- inversion, we almost always treat traveltimes and am-
els for targets beneath complex overburden structures, plitudes separately. When modeling the observed data,
such as those associated with overthrust and salt tec- we either model the traveltimes or amplitudes. When
tonics, ultimately, we have to do image-gather analysis modeling the earth, again, we use the traveltimes, as in
(Faye and Jeannaut, 1986; Reshef and Kessler, 1989; structural inversion (Chapter 10), or amplitudes, as in
Reshef, 1997, 2001). An image gather is the output from stratigraphic inversion (Chapter 11).
prestack depth migration and is a true (common depth- It is most appropriate to provide a summary list of
point) CDP gather at a surface location. Stacking of applications of seismic inversion for data modeling that
image gathers yields an earth image in depth. If the are spread throughout Chapters 1 to 7.
velocity-depth model is correct, then events on an im-
age gather are flat. In this respect, an image gather can (a) Deconvolution is based on modeling a one-
be considered like a moveout-corrected CMP gather, ex- dimensional (1-D) seismogram by optimum Wiener
cept the vertical axis on an image gather is in depth. filtering for a minimum-phase estimate of the
An event on an image gather with a moveout in- source wavelet, to predict multiples, and obtain an
dicates an erroneously too low or too high velocity. By estimate of white reflectivity series (Section 2.1).
examining a panel of image gathers from the same loca- (b) We model traveltime deviations on moveout-
tion but with different constant trial velocities for the corrected CMP gathers to estimate surface-
layer under consideration, one can pick the velocity that consistent shot and receiver residual statics (Sec-
yields a flat event and assign it as the velocity of the tion 3.3).
layer above. Image gathers also can be used to make (c) We model refracted arrival times to estimate,
residual corrections to velocity estimates at analysis lo- again, surface-consistent shot and receiver inter-
cations. This normally is done by first converting the cept time anomalies, and thus obtain shot and re-
gather to the time domain, performing residual move- ceiver refraction statics (Section 3.4).
out velocity analysis, and converting back to the depth (d) One type of formulation of the discrete Radon
domain. The resulting residual correction should favor- transform is by generalized linear inversion. The
ably improve the power of the stack obtained from im- discrete Radon transform is used to model a CMP
age gathers and yield an updated velocity-depth model. gather so as to attenuate multiples and random
The model updating based on the image-gather analy- noise, while compensating for missing data and fi-
sis usually is repeated until residual moveouts on im- nite cable length in recording (Section 6.4).
age gathers are reduced to a minimum (Al-Yahya, 1989; (e) We model the seismic signal represented by re-
Deregowski, 1990). flection events assumed to be linear from trace
In the following sections, we shall review the ve- to trace and attenuate random noise uncorrelated
locity estimation methods described above by using from trace to trace by using spatial prediction fil-
three synthetic model data sets with varying complex- ters (Section 6.5).
ity. These are models with horizontal layers, low-relief (f) Based on the same data modeling concept, we de-
structure and complex structure. Emphasis will be on sign spatial prediction filters to perform trace in-
practical aspects of the velocity estimation techniques. terpolation (Section 7.2).
1356 Seismic Data Analysis

(g) Data modeling also can be used in the design of Table 9-1. A set of inversion procedures for earth mod-
a three-dimensional (3-D) dip-moveout correction eling in depth to estimate layer velocities and delineate
operator which accounts for irregular spatial sam- reflector geometries.
pling and undersampling of recorded data (Section
7.2). Layer Velocities Reflector Geometries

Most data modeling applications are based on the Dix conversion of vertical-ray time-to-depth
theory of generalized linear inversion. Although each rms velocities conversion (vertical stretch)
of the applications listed above is treated in the ap-
pendixes of the respective chapters, for completeness, a stacking velocity image-ray time-to-depth
mathematical summary based on the generalized linear inversion conversion (map migration)
inversion theory is provided in Sections J.1, J.2 and J.3.
coherency inversion poststack depth migration

image-gather analysis prestack depth migration


Inversion Procedures for Earth Modeling

Practical methods for estimating layer velocities and de-


lineating reflector geometries can be appropriately com-
of reflector geometries. A mild-to-moderate lateral ve-
bined to form inversion procedures to construct earth
locity variation is associated with a zero-offset diffrac-
models in depth from seismic data. Listed in Table 9-1
tion response that is represented by a skewed, but al-
are four such combinations.
most hyperbolic traveltime trajectory (Figure 8.0-12).
These combinations are ordered from top to bot-
A strong lateral velocity variation is associated with a
tom with an increasing level of accuracy. Also, for a
zero-offset diffraction response that is represented by
given combination, the methods for layer velocity esti-
a distorted, nonhyperbolic traveltime trajectory (Fig-
mation and reflector geometry delineation are compat-
ure 8.0-11). A severe lateral velocity variation is associ-
ible. This means that, for instance, if you think you ated with a zero-offset diffraction response that is repre-
can afford the accuracy of stacking velocity inversion sented by a complex, multivalued traveltime trajectory
to estimate the layer velocities, it should suffice to per- (Figure 8.0-13).
form image-ray depth conversion to define the reflector
geometries. Nevertheless, in practice, you may wish to
(a) Vertical Stretch: A combination of Dix conversion
choose other combinations of the methods from the left-
of stacking velocities to estimate layer velocities
and right-hand columns. For instance, the combination
and vertical-ray time-to-depth conversion of time
of Dix conversion with image-ray depth conversion is
horizons picked from a time-migrated volume of
another inversion procedure that is used widely. Also, data to delineate reflector geometries: This is a
you may be compelled to apply an inversion procedure procedure appropriate for cases with negligible ray
that involves multiple combinations. For instance, in bending at layer boundaries, gentle dips, and lat-
areas where salt tectonics has caused formation of di- eral velocity variations judged to be within the
apiric structures, the earth model may be estimated in bounds of time migration.
three parts — the overburden above the salt diapir, the (b) Map Migration: A combination of stacking velocity
salt diapir itself, and the substratum. You may then use inversion to estimate layer velocities and image-ray
coherency inversion combined with poststack depth mi- time-to-depth conversion of time horizons picked
gration to estimate the overburden model, and image- from a time-migrated volume of data to delineate
gather analysis combined with prestack depth migra- reflector geometries: This is a procedure appropri-
tion to define the base-salt geometry and estimate the ate for cases with moderate ray bending at layer
substratum model. boundaries, moderate vertical velocity gradients,
The following are the general guidelines for imple- and moderate lateral velocity variations.
menting the procedures listed in Table 9-1. For conve- (c) Poststack Depth Migration: A combination of co-
nience, we shall refer to these procedures with keywords herency inversion to estimate layer velocities and
— vertical stretch, map migration, poststack depth mi- poststack depth migration to delineate reflector
gration, and prestack depth migration. The primary con- geometries: This is a procedure appropriate for
sideration in the choice for an inversion procedure is the cases with significant ray bending at layer bound-
degree of lateral velocity variations and the complexity aries and significant vertical velocity gradients,
Earth Modeling in Depth 1357

and strong lateral velocity variations with sharp them in the form of zero-offset sections as shown in
changes in reflector curvatures. Figure 9.0-2. While there are differences between the
(d) Prestack Depth Migration: A combination of velocity-depth models in Figure 9.0-1, differences in the
image-gather analysis to estimate and update layer zero-offset traveltime sections are less than marginal.
velocities, and stacking of image gathers to delin- Any differences that we observe in the traveltime sec-
eate reflector geometries: This is a procedure ap- tions may be attributed largely to the digitizing errors
propriate for cases with significant ray bending at that have occurred in building the velocity-depth mod-
layer boundaries, and severe lateral velocity varia- els. We may, within the bounds of ray theory associated
tions associated with salt and overthrust tectonics. with reflectors with reasonable curvature and shape, re-
tain one of the four zero-offset traveltime sections and
The inversion methods listed in Table 9-1 are used discard the other three sections shown in Figure 9.0-2.
We may also conclude that the zero-offset traveltime
to estimate an initial earth model in depth. Seismic in-
section we have kept is consistent with all four of the
version also is used to update the estimated model (Sec-
velocity-depth models shown in Figure 9.0-1.
tion 9.5). A common application of inversion to estimate
In reality, we record wavefields that comprise not
the errors in the initial model parameters — layer veloc-
just traveltimes but also amplitudes. Pretend that the
ities and reflector depths, is reflection traveltime tomog-
zero-offset traveltime section which we retained in Fig-
raphy. Tomographic inversion involves perturbing the
ure 9.0-2 is equivalent to a zero-offset wavefield section
model parameters by a small amount so as to match the as in Figure 8.2-1, insofar as traveltimes are concerned.
modeled reflection traveltimes with the observed travel- Now, consider the inverse problem. Suppose that
times. Refraction traveltime tomography (Section C.9) the analyst is given the zero-offset section in Figure 8.2-
and reflection traveltime tomography (Section J.6) both 1 and is asked to estimate an earth model in depth from
are based on the assumption that the perturbation re- it. As an answer, the analyst may provide any one of the
quired to update the model parameters is very small earth models shown in Figure 9.0-1. They all are equally
compared to the spatial variations in the model param- valid answers since they all are consistent with the in-
eters themselves. In practice, tomography is best used put zero-offset section in Figure 8.2-1. In fact, there are
strictly to touch-up a carefully estimated earth model infinitely many possible earth models, not just the four
based on some plausable geologic constraints; it should models in Figure 9.0-1, which are consistent with the
never be used by itself to estimate the model. zero-offset section. While there are infinite number of
solutions, there should only be one valid answer — the
geologic model shown in Figure 8.2-1. We will never
Velocity-Depth Ambiguity know which one, if any, of the answers corresponds to
the true velocity-depth model.
In practice, the analyst deals with a stacked section
A fundamental problem with inversion is velocity-depth
and not a zero-offset section. Shown in Figure 9.0-3 are
ambiguity. This means that an error in layer velocity can
three velocity-depth models estimated from the seismic
be indistinguishable from an error in reflector geometry. data associated with the stacked section in the same fig-
To help resolve the velocity-depth ambiguity, we must ure by means of model estimation procedures described
use nonzero-offset data to estimate the layer velocities in this chapter. The zero-offset traveltimes associated
and reflector geometries (Section J.7). We can never with the layer boundaries included in each of the three
claim to completely resolve the velocity-depth ambigu- models are computed by normal-incidence ray tracing.
ity and obtain the one and the only solution — the true When modeled traveltimes are repeatedly overlayed on
velocity-depth model, from inversion of seismic data. the same stacked section, we note that they all coin-
But, we can hope to reduce the many possible solutions cide with the observed traveltimes associated with the
to a few geologically plausable velocity-depth models. reflection events that correspond to the layer bound-
Of course, if there is ample well control, we may be able aries included in the velocity-depth models. Which one
to discard all but one of the few solutions and declare of the three models best represents the true subsurface
that as the solution. structure — that is the question. And this question has
Consider the four geologically plausable velocity- a profound implication in practice; in the present case,
depth models of a salt diapir shown in Figure 9.0-1. an accurate delineation of the geometry for the target
Compare these models and note the differences in re- reflector represented by the deepest layer boundary has
flector geometries, which are subtle in some places and an impact on the exploration and development of a po-
distinctively obvious in others. There also are differ- tential reservoir.
ences in layer velocities from one model to the other. Mathematically, the nature of any inversion prob-
Compute the zero-offset reflection traveltime re- lem with a multiple number of solutions that are
sponses by way of normal-incidence rays and display all consistent with observed data is characterized as
1358 Seismic Data Analysis

FIG. 9.0-1. Four different earth models in depth associated with a diapiric structure. Note the differences in layer velocities
and reflector geometries.
Earth Modeling in Depth 1359

FIG. 9.0-2. The zero-offset traveltime sections associated with the earth models in depth shown in Figure 9.0-1.
1360 Seismic Data Analysis

nonuniqueness. This mathematical nonuniqueness is consistent with the surface data set itself but also with
known to the exploration seismologist as velocity-depth the borehole data.
ambiguity. Refer back to Figure 9.0-1 and note that Aside from consistency, model verification has to
in reality only Model 1 is the correct model — the include a test of flatness of events on image gathers de-
same as the true geologic model shown in Figure 8.2-1. rived from prestack depth migration. A correct model,
However, any combination of layer velocities and re- again, within the limitations of velocity-depth ambigu-
flector geometries represented by the other three earth ity and accuracy of inversion methods, would yield an
models is equally consistent with the input data. This accurate image from prestack depth migration irrespec-
means that errors in layer velocities and errors in reflec- tive of the source-receiver offset. Thus, with the cor-
tor geometries are indistinguishable — a restatement of rect model, the resulting image gathers would have flat
the velocity-depth ambiguity. A robust quantification of events. An erroneous earth model, on the other hand,
velocity-depth ambiguity is provided in Section J.7. would cause residual moveout on image gathers. The
Since we can never obtain the true representation use of image gathers in model estimation, verification,
of an earth model from inversion of seismic data, the and update is discussed in Section 9.3 and illustrated
plausable strategy is to estimate an initial model and by some of the case studies presented in Chapter 10.
then update it to get a final model that may be con-
sidered an acceptable approximation to the true model.
Model building strategies are discussed in Section 9.4, Model Representation and Visualization
and model updating by reflection tomography is pre-
sented in Section 9.5. In earth modeling, a surface corresponding to a layer
The important question in practice is, following the boundary is usually represented by a set of triangles, the
model update, how much velocity-depth ambiguity re- size and shape of which vary depending on the complex-
mains unresolved in a final velocity-depth model. As il- ity of the reflector geometry. Shown in Figure 9.0-4 are
lustrated in Figures 9.0-1 and 9.0-2, the ambiguity with four surfaces associated with layer boundaries included
zero-offset data is infinite. By using nonzero-offset data, in an earth model. Also included in Figure 9.0-4 are the
we can hope to resolve the velocity-depth ambiguity up representations of these surfaces by triangulation.
to a certain theoretical limit. An important rule to keep A velocity-depth model usually is represented ei-
in mind is that, for data with good quality, velocity- ther in the form of a gridded or tessellated volume. Grid-
depth ambiguity for a reflector can be resolved with an ding means dividing the whole volume into a set of 3-D
acceptable degree of accuracy if the data used in inver- cells of equal size with appropriate dimensions in the in-
sion have been recorded with offsets greater than the line, crossline, and depth directions. Tessellation means
reflector depth (Bickel, 1990; Lines, 1993). dividing the volume associated with each layer into a
While an earth model can only be estimated with set of tetrahedra, the size and shape of which depend
an accuracy that is within the threshold of velocity- on the geometry of layer boundaries (Figure 9.0-5). In a
depth ambiguity, it does have to be consistent with tessellated velocity-depth model, a velocity and a gra-
the input data used in inversion to estimate the model. dient, if available from sonic logs, are assigned to each
Consistency is a necessary condition for an earth model corner of the tetrahedra.
to be certified as an acceptable estimate of the true A velocity-depth model is represented either in
model. A quick way to check for consistency is by ray- gridded or tesselated form depending on the applica-
theoretical forward modeling of zero-offset traveltimes tion that needs it as input. For instance, ray tracing in
associated with the reflector boundaries that are in the coherency inversion and prestack depth migration may
earth model itself, and then comparing them with the be performed using gridded or tessellated models. On
actual traveltimes picked from the stacked data. Any the other hand, wave extrapolation in 3-D poststack
discrepancy between the modeled and actual travel- depth migration based on finite-difference schemes is
times is an indication of errors in the earth model pa- performed conveniently using gridded models.
rameters — layer velocities and/or reflector geometries. Images of layer boundaries included in an earth
By using nonzero-offset data, we can hope to reduce the model in depth can be converted to a physical model us-
many possible solutions to a few. Furthermore, by in- ing various image construction techniques. Figure 9.0-6
troducing constraints, we may be able to converge to a shows the four surfaces that represent the layer bound-
single solution provided the set of constraints are reli- aries from the earth model as in Figure 9.0-4 and the
able. One set of constraints is the depth information at slabs that represent the layers of the earth model cre-
well locations. This well-top information can be used to ated by laser lithography.
calibrate results of inversion of surface seismic data and In areas with complex structures, we often have to
obtain a single earth model in depth that not only is deal with mulitvalued depth surfaces. For instance, salt
Earth Modeling in Depth 1361

FIG. 9.0-3. Three velocity-depth models (left column) all of which are consistent with the observed traveltimes on the
stacked section (right column). Superimposed on the same stacked section are the modeled zero-offset traveltimes associated
with the velocity-depth models in the left column.
1362 Seismic Data Analysis

FIG. 9.0-4. Surfaces that represent the reflector geometries (left column) and the triangulated form of these surfaces (right
column).
Earth Modeling in Depth 1363

FIG. 9.0-5. (a) The model volume that includes the surfaces in Figure 9.0-4, (b) the tessellated form of the model.
1364 Seismic Data Analysis

FIG. 9.0-6. Physical modeling of an earth model — surfaces that represent layer boundaries (left) and slabs that represent
the layers themselves (right). The physical model dimensions have an aspect ratio of one.
Earth Modeling in Depth 1365

overhangs associated with diapirism and imbricate Table 9-2. Parameters of the model with horizontal
structures associated with overthrusting cause a sur- layers and constant-velocity near-surface layer.
face to fold onto itself. Shown in Figure 9.0-7 is an earth
model that comprises a complex diapiric structure. The Layer Velocity (m/s) Depth (m)
top-salt boundary is repesented by the multivalued yel-
low surface and the base-salt boundary is represented by H1 1500 100
the pink surface. The surfaces that represent the layer H2 2000 1000
boundaries within the overburden must be attached to H3 2400 − 2700 1500
the top-salt boundary surface without any gaps so as to H4 3000 − 3500 1800
form individual volumes associated with each layer. Fig- H5 4500 2250
ure 9.0-8a shows the solid-model representation of the H6 3000 2700
earth model in Figure 9.0-7. When expanded, the solid
model shows the individual volumes associated with the
layers included in the earth model (Figure 9.0-8b).
Isolate the salt mass from the model as shown in locity variations (Figure 9.1-1a). The stacking velocity
Figure 9.0-9 to examine the complexity of the salt di- section is shown in Figure 9.1-2b with the color bar
apir. The multivalued surface that represents the top- on the right-hand margin. The stacking velocity section
salt boundary is shown in Figure 9.0-10a. The triangu- was derived from the horizon-consistent stacking veloc-
lated mesh for this surface is shown in Figure 9.0-10b ity profiles shown in Figure 9.1-3.
with an enlarged view in Figure 9.0-11.

Dix Conversion
9.1 MODELS WITH HORIZONTAL LAYERS
The horizon-consistent stacking velocity profiles (Fig-
In this section, we shall examine the accuracy of Dix ure 9.1-3) at each of the layer boundaries are used to
conversion and coherency inversion to estimate layer ve- perform Dix conversion to derive the interval velocity
locities using two earth models with horizontal layers, profiles for each of the layers. Dix conversion is based
but with lateral velocity variations. The near-surface on the formula
layer is of constant velocity in the first model and its
velocity varies laterally in the second model. Otherwise,
2 τ
Vn2 τn − Vn−1
both models are identical. Shown in Figure 9.1-1a are n−1
vn = , (9 − 1)
the velocity profiles for the six layers in the model. We τn − τn−1
shall refer to the layers by the horizon names corre-
sponding to the base of each layer, H1 through H6. where vn is the interval velocity within the layer
Listed in Table 9-2 are the layer velocities and depths bounded by the (n − 1)st layer boundary above and
to the base of each layer. When the layer velocity is not the nth layer boundary below, τn and τn−1 are the cor-
constant, the range is given in Table 9-1. The lateral responding two-way zero-offset times, and Vn and Vn−1
velocity gradients in layers H3 and H4 are about 125 are the corresponding rms velocities. Derivation of equa-
m/s and 200 m/s over one cable length, respectively. tion (9-1) is provided in Section J.4.
Figure 9.1-1b shows the velocity-depth model with the Equation (9-1) is based on the assumptions that
color bar on the right-hand margin. the layer boundaries are flat and the offset range used
A total of 384 shot records was modeled using the in estimating the rms velocities Vn and Vn−1 corre-
two-way acoustic wave equation. The simulated record- sponds to a small spread. Additionally, keep in mind
ing geometry consists of an off-end cable with 96 re- that the rms velocities used in equation (9-1) are based
ceivers and offset range 25-2400 m. Shot and receiver on a straight-ray assumption; thus, ray bending at layer
intervals are both 25 m, and the CMP interval is 12.5 boundaries are not accounted for in Dix conversion.
m and CMP fold is 48. Figure 9.1-2a shows the CMP- The procedure for estimating the layer velocities
stacked section with the picked time horizons that corre- and reflector depths using Dix conversion of stacking
spond to the layer boundaries H1 through H6 in Figure velocities includes the following steps:
9.1-1b. Compare with the velocity-depth model (Figure
9.1-1b) and note that flat horizons in depth correspond (a) For each of the layers in the model, pick the time
to curved horizons in time because of the lateral ve- of horizon on the unmigrated CMP-stacked data
1366 Seismic Data Analysis

FIG. 9.0-7. Two different views of a complex earth model FIG. 9.0-8. (a) The solid-model representation of the earth
associated with a salt diapirism. (Modeling by Cyril Gre- model in Figure 9.0-7, (b) the exploded form of (a). (Mod-
gory; courtesy Paradigm Geophysical.) eling by Cyril Gregory; courtesy Paradigm Geophysical.)
Earth Modeling in Depth 1367

FIG. 9.0-9. Two different views of the salt mass (the pink FIG. 9.0-10. (a) The surface that represents the top-salt
solid) and the substratum isolated from the solid model boundary, and (b) its representation by triangulation. (Mod-
shown in Figure 9.0-8. (Modeling by Cyril Gregory; cour- eling by Cyril Gregory; courtesy Paradigm Geophysical.)
tesy Paradigm Geophysical.)
1368 Seismic Data Analysis

FIG. 9.0-11. A detailed portion of the triangulated surface shown in Figure 9.0-10b. The surface underneath the triangulated
mesh is the base-salt boundary. (Modeling by Cyril Gregory.)
Earth Modeling in Depth 1369

that corresponds to the base-layer boundary (Fig- make an attempt in Section 9.5 to update this result by
ure 9.1-2a). These times are used in lieu of the two- using tomography.
way zero-offset times in equation (9-1).
(b) Extract the rms velocities at horizon times (Figure
9.1-3).
(c) Use equation (9-1) to compute the interval veloci- Coherency Inversion
ties for each of the layers from the known quantities
— rms velocities and times at top- and base-layer The procedure for estimating layer velocities from co-
boundaries. herency inversion requires CMP gathers at analysis
(d) Use interval velocities and times at layer bound- locations and horizon times picked from unmigrated
aries to compute depths at layer boundaries. If the stacked data. Alternatively, time horizons picked from
input times are from an unmigrated stacked sec- time-migrated data can be forward-modeled to derive
tion as in Figure 9.1-2a, use normal-incidence rays the zero-offset traveltimes needed for coherency inver-
for depth conversion. If the input times are from a sion. As for Dix conversion, the velocity estimate from
migrated stacked section, use image rays for depth coherency inversion is local, independent of data away
conversion. from the analysis location.
A procedure for velocity-depth model estimation
Interval velocity profiles derived from Dix conver- that includes coherency inversion is conducted layer-
sion are shown in Figure 9.1-4a. The earth model can be by-layer starting from the surface. As for Dix conver-
constructed by combining the estimated interval veloc- sion, consider the synthetic data set associated with the
ity profiles and depth horizons (Figure 9.1-4b). Com- model with horizontal layers shown in Figure 9.1-1. We
parison with the true model shown in Figure 9.1-1b shall adopt the interval velocity profile for the first layer
clearly demonstrates that the interval velocity estima- H1 estimated from Dix conversion and start the appli-
tion based on Dix conversion is not completely accurate. cation of coherency inversion with layer H2. Assume
The interval velocity profiles derived from Dix conver- that the velocity-depth model for the first n − 1 layers
sion (Figure 9.1-4a) exhibit the sinusoidal oscillations already has been estimated. For the nth layer, follow
caused by the swings in the stacking velocity profiles the steps below for coherency inversion:
themselves (Figure 9.1-3).
The fundamental problem is that the stacking ve- (a) For a trial constant velocity assigned to the nth
locity estimation is based on fitting a hyperbola to layer, perform normal-incidence traveltime inver-
CMP traveltimes associated with a laterally homoge-
sion to convert the time horizon corresponding to
neous earth model. If there are lateral velocity varia-
the base-layer boundary to a trial depth horizon.
tions in layers above the layer under consideration, and
(b) Given the geometry of the CMP gather at the
if these variations are within a cable length, then stack-
analysis location, assign a trial velocity to the
ing velocities would oscillate in a physically implaus-
half space that includes the layer yet to be deter-
able manner (Lynn and Claerbout, 1982; Loinger, 1983;
mined and compute the CMP traveltimes using the
Rocca and Toldi, 1983). As a consequence, the result-
ing interval velocity estimation based on Dix conversion known overburden velocity-depth model. The mod-
is adversely affected. In the present case, Dix conver- eled CMP traveltime trajectory that corresponds
sion has produced fairly accurate estimates for the in- to the base of the layer under consideration is
terval velocities of the top three layers — H1, H2, and in general nonhyperbolic, for the ray tracing used
H3 as shown in Figure 9.1-4a. But the interval veloc- to compute the CMP traveltimes accounts for ray
ity estimates for layers H4 and H5 have been adversely bending at layer boundaries and incorporates ver-
affected by the laterally varying velocities within the tical velocity gradients within layers above. Shown
layer above, H3. in Figures 9.1-7a through 9.1-11a are the CMP ray-
The pragmatic approach would be to smooth out paths at an analysis location through the unknown
the oscillations in the stacking velocities before Dix con- and half space that includes the layer under con-
version and smooth out the oscillations in the velocity sideration and the known overburden part of the
profiles after Dix conversion (Figure 9.1-5a). Then, the model. The corresponding CMP traveltime trajec-
resulting earth model is expected to be free of the ad- tories are shown in color in Figures 9.1-7b through
verse effects of stacking velocity anomalies (Figure 9.1- 9.1-11b overlayed on the CMP gather at the anal-
5b). A closer look at the central portions of the esti- ysis location.
mated models using Dix conversion is shown in Figure (c) Extract data window along the modeled traveltime
9.1-6. Note that the model derived from the smoothed trajectory. Shown in Figures 9.1-7c through 9.1-
interval velocities is closer to the true model. We shall 11c are the data windows for seven trial velocities
(text continues on p. 1375)
1370 Seismic Data Analysis

FIG. 9.1-1. An earth model that comprises six flat layers: (a) the interval velocity profiles for the six horizons H1-H6; (b)
true velocity-depth model created from the profiles in (a).
Earth Modeling in Depth 1371

FIG. 9.1-2. (a) The CMP-stacked section derived from the modeled common-shot gathers using the earth model shown in
Figure 9.1-1b; (b) the stacking velocity section.
1372 Seismic Data Analysis

FIG. 9.1-3. Horizon-consistent stacking velocity semblance spectra computed from the CMP gathers of the synthetic data
as in Figure 9.1-2a along the time horizons H1-H6.
Earth Modeling in Depth 1373

FIG. 9.1-4. (a) The interval velocity profiles derived from Dix conversion of the horizon-consistent stacking velocity profiles
picked from the semblance spectra shown in Figure 9.1-3; (b) estimated velocity-depth model. Compare with the true velocity-
depth model shown in Figure 9.1-1b.
1374 Seismic Data Analysis

FIG. 9.1-5. (a) The interval velocity profiles as in Figure 9.1-4a displayed by the thick curves and smoothed interval velocity
profiles displayed by the thin curves; (b) estimated velocity-depth model using the smoothed interval velocity profiles in (a).
Compare with the true velocity-depth model shown in Figure 9.1-1b and the model derived from the unsmoothed interval
velocities shown in Figure 9.1-4b.
Earth Modeling in Depth 1375

FIG. 9.1-6. Central portions of (a) the true velocity-depth model shown in Figure 9.1-1b, (b) the velocity-depth model shown
in Figure 9.1-4b estimated using the unsmoothed interval velocity profiles shown in Figure 9.1-4a, and (c) the velocity-depth
model shown in Figure 9.1-5b estimated using the smoothed interval velocity profiles shown in Figure 9.1-5a.

around the optimum velocity that best flattens the 7c through 9.1-11c can be considered as events on
event within the data window. a moveout-corrected CMP gather. A flat event on a
(d) Compute semblance using the data within the se- moveout-corrected CMP gather suggests that the stack-
lected window as a measure of discrepancy between ing velocity associated with that event is optimum.
the modeled and the actual traveltimes. Figure 9.1- Likewise, a flat event in the data window panels in
12 shows the semblance spectra computed from Figures 9.1-7c through 9.1-11c suggests that the layer
the CMP gather in Figures 9.1-7b through 9.1-11b velocity from coherency inversion is optimum. An er-
from the reflection events associated with layers H2 roneously too low or too high velocity causes residual
through H6. moveout which can be observed on the event within the
(e) Repeat all the steps above for a range of constant data window.
velocities. The semblance spectra in Figure 9.1-12 were com-
(f) Pick the constant trial velocity as the layer velocity puted using three different maximum offsets — 2400
for which the semblance is the maximum. m, 1400 m, and 400 m. Note that, for a given layer, the
The results of coherency inversion are used to pick longer the cable length, the sharper the semblance peak
velocity nodes at analysis locations. Specifically, the — the higher the velocity resolution. Also, the deeper
layer velocity at a given location is selected based on the event and the higher the velocity, the broader the
the semblance curve and the data window along the semblance peak — the poorer the velocity resolution.
modeled traveltime trajectory making sure that the es- Figures 9.1-7 through 9.1-11 show the results of co-
timated velocities are geologically plausable. Data win- herency inversion at one CMP location. In practice, for
dows along modeled traveltime trajectories can be ex- 2-D data, coherency inversion often is applied contin-
amined for flatness criterion to pick an optimum ve- uously along the line. As for horizon-consistent stack-
locity node. Specifically, the events in Figures 9.1- ing velocity analysis, for each layer, a horizon-consistent
(text continues on p. 1382)
1376 Seismic Data Analysis

FIG. 9.1-7. Coherency inversion for layer H2 as in Figure 9.1-2a: (a) CMP raypaths, (b) the CMP gather at the analysis
location, (c) the data window that includes the color-marked event in the CMP gather in (b) with moveouts associated with
the different trial velocities assigned to layer H2. See text for details.
Earth Modeling in Depth 1377

FIG. 9.1-8. Coherency inversion for layer H3 as in Figure 9.1-2a: (a) CMP raypaths, (b) the CMP gather at the analysis
location, (c) the data window that includes the color-marked event in the CMP gather in (b) with moveouts associated with
the different trial velocities assigned to layer H3. See text for details.
1378 Seismic Data Analysis

FIG. 9.1-9. Coherency inversion for layer H4 as in Figure 9.1-2a: (a) CMP raypaths, (b) the CMP gather at the analysis
location, (c) the data window that includes the color-marked event in the CMP gather in (b) with moveouts associated with
the different trial velocities assigned to layer H4. See text for details.
Earth Modeling in Depth 1379

FIG. 9.1-10. Coherency inversion for layer H5 as in Figure 9.1-2a: (a) CMP raypaths, (b) the CMP gather at the analysis
location, (c) the data window that includes the color-marked event in the CMP gather in (b) with moveouts associated with
the different trial velocities assigned to layer H5. See text for details.
1380 Seismic Data Analysis

FIG. 9.1-11. Coherency inversion for layer H6 as in Figure 9.1-2a: (a) CMP raypaths, (b) the CMP gather at the analysis
location, (c) the data window that includes the color-marked event in the CMP gather in (b) with moveouts associated with
the different trial velocities assigned to layer H6. See text for details.
Earth Modeling in Depth 1381

FIG. 9.1-12. Semblance spectra derived from coherency inversion applied to the central CMP gather associated with the
stacked data in Figure 9.1-1a using maximum offsets of 2400 m, 1400 m, and 400 m. See text for details.
1382 Seismic Data Analysis

semblance spectrum is computed using coherency inver- m/s. The interval velocity profiles are shown in Figure
sion. For 3-D data, as for conventional velocity analy- 9.1-18a and the velocity-depth model is shown in Figure
sis, coherency inversion normally is applied at uniformly 9.1-18b.
spaced grid points over the survey area. As for the model with a constant-velocity near-
Figure 9.1-13 shows the semblance spectra derived surface layer (Table 9-2), a total of 384 shot records
from coherency inversion for layers H2 through H6 of was modeled using the two-way acoustic wave equation
the model shown in Figure 9.1-1. The interval velocity with identical line geometry. Figure 9.1-19a shows the
profiles shown in Figure 9.1-14a have been picked by CMP-stacked section with the picked time horizons that
tracking the semblance peaks. Compare with the pro- correspond to the layer boundaries H1 through H6 in
files associated with the true model shown in Figure Figure 9.1-18b. The stacking velocity section is shown
9.1-1a, and note that by coherency inversion the inter- in Figure 9.1-19b with the color bar on the right-hand
margin.
val velocities for layers H2 and H3 have been estimated
The stacking velocity section was derived from the
accurately. But the semblance spectra in Figure 9.1-13
horizon-consistent stacking velocity profiles shown in
for the underlying layers H4, H5, and H6 exhibit oscil-
Figure 9.1-20. As a direct consequence of the lateral ve-
lations akin to those associated with the interval veloc-
locity variations in the near-surface layer H1, the stack-
ities derived from Dix conversion of stacking velocities ing velocities oscillate violently for the layers below.
(Figure 9.1-4). Specifically, when there are lateral ve- If Dix conversion is done using the unedited stacking
locity variations within a cable length in one layer, in velocity profiles, the resulting interval velocity profiles
this case layer H3, the interval velocity estimation for exhibit geologically implausable variations (Figure 9.1-
the layers below is adversely affected as shown in Figure 21a). It is imperative in practice to remove the oscilla-
9.1-13. This phenomenon also was observed by Sorin et tions from the stacking velocity profiles before Dix con-
al. (1996). version. The resulting interval velocity profiles shown in
Shown in Figure 9.1-14b is the velocity-depth Figure 9.1-21b are closer to the true profiles shown in
model derived from coherency inversion. The model was Figure 9.1-18a.
constructed layer-by-layer starting from the top. Co- By using the interval velocity profiles, convert
herency inversion was used to estimate the layer veloc- the time horizons shown in Figure 9.1-19a to depth
ities (Figure 9.1-14a), and normal-incidence traveltime horizons. Then, combine the interval velocity profiles
inversion of the time horizons picked from the stacked with the depth horizons to compose the velocity-depth
section (Figure 9.1-1) was used to obtain the depth hori- model shown in Figure 9.1-21c. A comparison with
zons. Note the distortions caused by the oscillations of the true model shown in Figure 9.1-19b, once again,
the interval velocity profiles in the estimated velocity- clearly demonstrates that the interval velocity estima-
depth model. As in Dix conversion, these oscillations tion based on Dix conversion is not completely accurate.
must be smoothed out for one layer before estimating We shall make an attempt in Section 9.5 to update the
the interval velocity for the next layer. The resulting model in Figure 9.1-21c by using tomography.
semblance spectra exhibit a reduced degree of oscilla- Figure 9.1-23 shows the semblance spectra derived
tions (Figure 9.1-15). The velocity-depth model based from coherency inversion for layers H2 through H6 of
the model shown in Figure 9.1-18. The analysis was con-
on the revised interval velocity profiles (Figure 9.1-16a),
ducted layer by layer starting from the top. Based on
which have been picked from the new set of semblance
the lessons learned from the model experiments shown
spectra, is shown in Figure 9.1-16b. A closer look at the
in Figures 9.1-13 and 9.1-15, the oscillations in the sem-
central portion of the estimated model using coherency blance spectra were rejected while tracking the interval
inversion is shown in Figure 9.1-17. Provided that the velocity profile from the semblance spectrum for the
oscillations shorter than a cable length are eliminated layer under consideration before moving down to the
from the interval velocity profiles, coherency inversion next layer. Despite the rejection of the oscillations, the
seems to produce an acceptable estimate of the layer semblance spectra still exhibit the influence of the near-
velocities. surface layer with lateral velocity variations (Figure 9.1-
23).
Shown in Figure 9.1-24b is the velocity-depth
Near-Surface Layer with model using the interval velocity profiles (Figure 9.1-
Laterally Varying Velocities 24a) derived from coherency inversion. A closer look at
the central portion of the estimated model using co-
We now examine the accuracy of Dix conversion and herency inversion is shown in Figure 9.1-25. As long as
coherency inversion for the model with horizontal lay- the oscillations shorter than a cable length are elim-
ers (Table 9-2) but with a near-surface layer H1 with inated from the interval velocity profiles, coherency
laterally varying velocities between 800 m/s and 1500 inversion seems to produce a better estimate of the
Earth Modeling in Depth 1383

FIG. 9.1-13. Horizon-consistent coherency inversion semblance spectra computed from the CMP gathers associated with the
stacked data in Figure 9.1-2a. No smoothing has been applied to the velocity profiles picked from these semblance spectra.
1384 Seismic Data Analysis

FIG. 9.1-14. (a) The interval velocity profiles derived from coherency inversion semblance spectra shown in Figure 9.1-13;
(b) estimated velocity-depth model. Compare with the true velocity-depth model shown in Figure 9.1-1b.
Earth Modeling in Depth 1385

FIG. 9.1-15. Horizon-consistent coherency inversion semblance spectra computed from the CMP gathers associated with the
stacked data in Figure 9.1-2a. When computing the semblance spectrum for a given layer, smoothing has been applied to the
velocity profiles picked from the semblance spectra associated with the layers above.
1386 Seismic Data Analysis

FIG. 9.1-16. (a) The interval velocity profiles derived from coherency inversion semblance spectra shown in Figure 9.1-15;
(b) estimated velocity-depth model. Compare with the true velocity-depth model shown in Figure 9.1-1b.
Earth Modeling in Depth 1387

FIG. 9.1-17. Central portions of (a) the true velocity-depth model shown in Figure 9.1-1b, and (b) the velocity-depth model
shown in Figure 9.1-14b estimated using the smoothed interval velocity profiles shown in Figure 9.1-16a.

velocity-depth model compared to Dix conversion (Fig- the presence of low-relief structures. Figure 9.2-1 shows
ure 9.1-22). The difference between the two estimates a velocity-depth model with such characteristics. The
is in the reflector geometries. Dix conversion has intro- model simulates a transgressive depositional sequence
duced spurious structures into the model (Figure 9.1- within the first 1-km depth, a deltaic sequence between
22), while coherency inversion has introduced a bulk 1.5-2 km, and a deeper depositional sequence between
shift in the reflector depths (Figure 9.1-25). In earth 2-2.5 km. Our goal is to detect the subtle lateral velocity
modeling, an error in the form of a distorted reflector variations within the individual sequences.
geometry is worse than an error in the form of a bulk A total of 154 shot records was modeled using the
shift in the reflector depth. While the error in the form two-way acoustic wave equation. The simulated record-
of a bulk shift can be corrected for by calibrating the ing geometry consists of a split-spread cable with 97
estimated model to well tops, the error in the form of a receivers and an offset range of 0-2350 m. Shot and re-
distorted reflector geometry may require a serious revi- ceiver intervals are both 50 m. Figure 9.2-2 shows the
sion of the estimated model. CMP-stacked section with and without the interpreted
time horizons. Assuming that the CMP-stacked section
is largely equivalent to a zero-offset section, these time
9.2 MODEL WITH LOW-RELIEF horizons correspond to two-way zero-offset time picks
STRUCTURE which are used in stacking velocity inversion and co-
herency inversion.
In this section, we shall review velocity estimation via The color-coded true velocity-depth model is
stacking velocity inversion and coherency inversion in shown in Figure 9.2-3. Using coherency inversion and
1388 Seismic Data Analysis

FIG. 9.1-18. An earth model that comprises six flat layers: (a) the interval velocity profiles for the six horizons H1-H6; (b)
true velocity-depth model created from the profiles in (a). This model is the same as the model in Figure 9.1-1 except that
H1 has laterally varying velocities.
Earth Modeling in Depth 1389

FIG. 9.1-19. (a) The CMP-stacked section associated with the earth model shown in Figure 9.1-18; (b) stacking velocity
section.
1390 Seismic Data Analysis

FIG. 9.1-20. Horizon-consistent stacking velocity semblance spectra computed from the CMP gathers of the synthetic data
as in Figure 9.1-19a.
Earth Modeling in Depth 1391

FIG. 9.1-21. (a) The interval velocity profiles derived from Dix conversion of the horizon-consistent stacking velocity profiles
picked from the semblance spectra shown in Figure 9.1-20, (b) the interval velocity profiles of (a) after lateral smoothing, (c)
estimated velocity-depth model. Compare with the true velocity-depth model shown in Figure 9.1-18b.
1392 Seismic Data Analysis

FIG. 9.1-22. Central portions of (a) the true velocity-depth model shown in Figure 9.1-18b, (b) the velocity-depth model
shown in Figure 9.1-21c estimated using the smoothed interval velocity profiles shown in Figure 9.1-21b.

stacking velocity inversion to estimate layer velocities, Stacking Velocity Inversion


the velocity-depth models also shown in Figure 9.2-3
are obtained. In both cases, normal-incidence traveltime The procedure for estimating layer velocities from
inversion is used to delineate the reflector geometries. stacking velocity inversion requires the horizon times
Note that both techniques are able to delineate the shal- picked from unmigrated stacked data (Figure 9.2-2) and
low transgressive sequence. The individual units within stacking velocities at time horizons that correspond to
the deltaic sequence are not included in the velocity-
layer boundaries in the model (Figure 9.2-4). Horizon-
depth model estimation. Instead, only the top (Horizon
consistent stacking velocities are estimated by comput-
5) and base (Horizon 6) of the sequence are included
ing semblance for a range of constant velocities contin-
in the model. Nevertheless, both techniques are able to
detect the existence of velocity variations from one unit uously along the time horizon picked from the stacked
to the next. The deltaic sequence, however, is delin- data (Figure 9.2-2). The velocity spectrum for each time
eated by coherency inversion more accurately. Finally, horizon (Figure 9.2-4) then is picked to derive a stacking
stacking velocity inversion fails to estimate the internal velocity curve along the midpoint axis. As for coherency
velocity distribution of the deepest sequence between inversion, the velocity estimate from stacking velocity
2-2.5 km, correctly. Coherency inversion, on the other inversion is local, independent of data away from the
hand, has at least been able to detect the relative mag- analysis location.
nitude of the velocity variations within this sequence A procedure for velocity-depth model estimation
with reasonable accuracy. In the following paragraphs, that includes stacking velocity estimation is conducted
we shall discuss how the two velocity-depth model esti- layer-by-layer starting from the surface. Assume that
mates in Figure 9.2-3 were made. the velocity-depth model for the first n−1 layers already
Earth Modeling in Depth 1393

FIG. 9.1-23. Horizon-consistent coherency inversion semblance spectra computed from the CMP gathers associated with
the stacked data in Figure 9.1-19a. When computing the semblance spectrum for a given layer, smoothing has been applied
to the velocity profiles picked from the semblance spectra associated with the layers above.
1394 Seismic Data Analysis

FIG. 9.1-24. (a) The interval velocity profiles derived from coherency inversion semblance spectra shown in Figure 9.1-23;
(b) estimated velocity-depth model. Compare with the true velocity-depth model shown in Figure 9.1-18b.
Earth Modeling in Depth 1395

FIG. 9.1-25. Central portions of (a) the true velocity-depth model shown in Figure 9.1-18b, and (b) the velocity-depth model
shown in Figure 9.1-24b estimated using the smoothed interval velocity profiles shown in Figure 9.1-24a.

has been estimated. For the nth layer, follow the steps tory, and thus determine the modeled stacking ve-
below for stacking velocity inversion: locity for the trial interval velocity.
(d) Measure the discrepancy between the modeled and
(a) For a trial constant velocity assigned to the nth the actual stacking velocities by way of semblance.
layer, perform normal-incidence traveltime inver- (e) Repeat all the steps above for a range of constant
sion to convert the time horizon corresponding to velocities.
the base-layer boundary to a trial depth horizon. (f) Pick the constant trial velocity as the layer velocity
(b) Given the geometry of the CMP gather at the anal- for which the difference between the modeled and
ysis location (not the CMP gather itself, but only the actual stacking velocities is minimum or the
the source-receiver geometry associated with it), semblance is maximum.
compute the CMP traveltimes. The modeled CMP
traveltime trajectory that corresponds to the base Following the steps described above, perform stack-
of the layer under consideration is in general nonhy- ing velocity inversion to obtain the layer-velocity sem-
perbolic, because the ray tracing used to compute blance spectra shown in Figure 9.2-5. Note that velocity
the CMP traveltimes accounts for ray bending at estimates down to Horizon 3 are fairly accurate. The es-
layer boundaries and incorporates vertical velocity timates for Horizons 4 and 5 show some departure from
gradients within layers above. the true velocity but are largely acceptable. Note the
(c) Compute the best-fit hyperbolic traveltime trajec- variations in the estimate for Horizon 6 — the method

(text continues on p. 1404)


1396 Seismic Data Analysis

FIG. 9.2-1. A velocity-depth model with low-relief structures.

FIG. 9.2-2. The CMP-stacked section associated with the velocity-depth model in Figure 9.2-1 with the interpreted time
horizons.
Earth Modeling in Depth 1397

FIG. 9.2-3. (a) True velocity-depth model associated with the stacked section in Figure 9.2-2; (b) result of stacking velocity
inversion; (c) result of coherency inversion.
1398 Seismic Data Analysis

FIG. 9.2-5. Stacking velocity inversion for the model in Fig-


FIG. 9.2-4. Horizon-consistent stacking velocity spectra for ure 9.2-1. Input to inversion are horizon-consistent stacking
the model in Figure 9.2-1. velocities (Figure 9.2-4) and the horizon times picked from
stacked data (Figure 9.2-2). Posted velocities are the true
layer velocities as in Figure 9.2-1.
Earth Modeling in Depth 1399

FIG. 9.2-6. Coherency inversion for the model in Figure 9.2-1. Picked velocity profiles are overlayed on the semblance spectra.
Input to inversion are CMP gathers and the horizon times picked from stacked data (Figure 9.2-2). Posted velocities are the
true layer velocities as in Figure 9.2-1.
1400 Seismic Data Analysis

FIG. 9.2-7. Coherency inversion for the model in Figure 9.2-1. (a) CMP gather at analysis location CMP 75 as in Figure
9.2-8; (b) semblance curve; (c) data window for three trial velocities. CMP raypaths that correspond to these trial velocities
are shown in Figure 9.2-8.
Earth Modeling in Depth 1401
1402 Seismic Data Analysis

FIG. 9.2-10. Data windows from coherency inversion at


CMP location 75 as in Figure 9.2-8 applied to the horizons
FIG. 9.2-9. Semblance curves from coherency inversion at as in Figure 9.2-2. The CMP gather is shown in Figure 9.2-
CMP location 75 as in Figure 9.2-8 applied to the horizons 7a and the semblance curves are shown in Figure 9.2-9. True
as in Figure 9.2-2. The CMP gather is shown in Figure 9.2-7a layer velocities are posted at the bottom of each panel.
and the data windows are shown in Figure 9.2-10.
Earth Modeling in Depth 1403
1404 Seismic Data Analysis

is attempting to distinguish the units within the deltaic tion event for Horizon 7a is displayed in Figure 9.2-7c
sequence with different velocities. The velocity estimate along the modeled traveltime trajectories. The flatness
for Horizon 7a has significant departures from the true criterion suggests that the optimum layer velocity at
velocity of 3500 m/s. Finally, the method has failed to the analysis location is 3500 m/s. The maximum of the
estimate the internal velocity variations within the deep semblance curve derived from coherency inversion co-
sequence, correctly. incides with the optimum choice for the layer velocity
(Figure 9.2-7b).

Coherency Inversion
Velocity Resolution
The results of coherency inversion for the model with
low-relief structures (Figure 9.2-1) are shown in Figure Figure 9.2-9 shows the semblance curves derived from
9.2-6. Note that, for shallow layers, the inversion yields coherency inversion at CMP location 75. (Horizon 3
accurate velocity estimates. It also has detected the ve- is missing at this location.) The CMP gather itself is
locity variations within the deltaic sequence over the shown in Figure 9.2-7a. Note that the sharpness of the
full-fold CMP range, correctly — note the increase in peak in the semblance curve, hence the velocity resolu-
velocity for Horizon 6 from around 3000 m/s on the left tion, depends upon the depth of the layer boundary and
to 3200 m/s in the middle and back to 3000 m/s on the magnitude of the layer velocity. Also, recall from
the right, and compare with Figure 9.2-1. Nevertheless, Figure 9.1-12 that the velocity resolution also depends
note the very short-wavelength variations in the veloc- on the effective cable length.
ity curves derived from picking the maxima of the sem- The sampling interval for the velocity axis in the
blance plots. As was demonstrated by the coherency in- semblance curves should be chosen by taking into con-
version tests applied to the model with horizontal layers sideration the velocity resolution that can be achieved.
(Section 9.1), this observation has an important practi- Figure 9.2-10 shows the CMP data windows with the re-
cal implication with regard to evaluation and use of the flections that correspond to the selected horizons. The
results of velocity estimation. Specifically, lateral veloc- horizontal axis spans the offset range used in coherency
ity variations of very short-wavelengths that are much inversion. For shallow horizons (1, 2, and 4), the offset
less than a cable length should not be incorporated into range is smaller because of muting. The CMP gather
a velocity-depth model. Instead, some lateral smoothing is shown in Figure 9.2-7a and the semblance curves are
of velocity estimates is almost always needed. shown in Figure 9.2-9. For a given horizon, the data win-
Shown in Figure 9.2-7a is a CMP gather used in dow with the flattest event is distinguishable when the
coherency inversion to obtain a velocity estimate for velocity sampling is appropriate. Specifically, for shal-
the layer above Horizon 7b. Following the procedure de- low horizons with low velocity (e.g., Horizons 1 and 2),
scribed in the previous section, use the normal-incidence velocity increment needs to be small enough to pick a
rays to convert the time horizon picked from the stacked layer velocity, accurately. However, for deeper events
data (Horizon 7a in Figure 9.2-2) at the analysis lo- with high velocity (e.g., Horizons 7a and 7e), the ve-
cation to a trial depth horizon. The normal-incidence locity increment does not have to be as small, since
depth conversions are shown in Figure 9.2-8 for the event curvature in the data windows becomes indistin-
three trial velocities as in Figure 9.2-7. The normal- guishable as seen in Figure 9.2-10. It is only with large
incidence rays in Figure 9.2-8 are overlayed on the true velocity increments that we observe a marked difference
velocity-depth model. Note that, in case of 3000 m/s, in event curvature as demonstrated in Figure 9.2-11. A
the trial depth horizon is shallower than the actual layer good rule of thumb in practice is that the sampling in-
boundary for Horizon 7a; it coincides with the actual terval in velocity used in stacking velocity inversion and
boundary in case of 3500 m/s (the true velocity of the coherency inversion needs to be specified as small as 25
layer above Horizon 7a); and, it is deeper in the case of m/s for velocities as low as 1500 m/s and can be spec-
4000 m/s velocity. ified as large as 200 m/s for velocities as high as 5000
For each trial constant velocity assigned to the m/s.
layer above Horizon 7a, perform modeling of CMP trav-
eltimes at the analysis location. The raypaths associ-
ated with the CMP traveltimes for the three trial ve- 9.3 MODEL WITH COMPLEX
locities are also displayed in Figure 9.2-8. Note that the OVERBURDEN STRUCTURE
velocity contrast at layer boundaries and the geometry
of these boundaries within the overburden has caused Figure 9.3-1 illustrates a complex overburden structure
reflection point smearing at Horizon 7a. associated with overthrust tectonics. Such structures
The CMP data window that includes the reflec- were formed as a result of the tectonic movements dur-
Earth Modeling in Depth 1405

FIG. 9.3-1. A velocity-depth model with complex overburden structure caused by overthrust tectonics.
1406 Seismic Data Analysis

ing the Lower Miocene and Upper Cretaceous, and are the depth image in Figure 9.3-3b, which was obtained by
common in North America, South America, and the stacking the image gathers as in Figure 9.3-6a, yields the
Middle East. The target horizon is the flat reflector at correct reflector geometry for the top of the carbonate
2.5 km below the imbricate structures. sequence. Interpret the top of the carbonate sequence
The velocity-depth model comprises a shallow se- from this section and insert it as a layer boundary into
quence with a relatively simple structure. Underneath the velocity-depth model (Figure 9.3-4a).
this shallow sequence is a shale-marl sequence with a Next, assign the velocity for the carbonate se-
strong vertical velocity gradient (0.5 m/s/m). Then, we quence (5700 m/s) to the half-space below the top-
have the imbricated fault structures of the carbonate carbonate boundary as shown in Figure 9.3-4a. Using
sequence, and finally the target level at 2.5 km charac- this model, perform prestack depth migration and ob-
terized as the detachment zone that separates the in- tain the depth image shown in Figure 9.3-4b. The image
competent rock layers above from the competent rock gathers indicate flat events for the top T and base B of
layers below. the carbonate sequence (Figure 9.3-6b), thus confirming
A total of 154 shot records was modeled using the that the velocity of the carbonate sequence is correct.
two-way acoustic wave equation. The simulated record- Additionally, the depth image in Figure 9.3-4b, which
ing geometry consists of a split-spread cable with 97 was obtained by stacking the image gathers as in Fig-
receivers and an offset range of 0-2350 m. Shot and re- ure 9.3-6b, yields the correct reflector geometry for the
ceiver intervals are both 50 m. Figure 9.3-2a shows the base of the carbonate sequence. Interpret the base of the
CMP-stacked section. Note the traveltime distortions carbonate sequence from this section and insert it as a
along the deepest reflection caused by the severe ray layer boundary into the velocity-depth model (Figure
bending within the overburden. Time migration (Fig-
9.3-5a).
ure 9.3-2b) — whether post- or prestack, will not re-
Finally, assign the substratum velocity (5000 m/s)
solve the deleterious effect of the complex overburden
to the half-space below the base-carbonate boundary
structure.
as shown in Figure 9.3-5a. Using this model, perform
prestack depth migration and obtain the depth image
shown in Figure 9.3-5b. The image gathers indicate flat
Image Gathers events for the top T and base B of the carbonate se-
quence, and the flat reflector F within the substratum
Suppose that we already have estimated the velocity-
(Figure 9.3-6c). The velocity-depth model in Figure 9.3-
depth model in Figure 9.3-1 down to 1-km depth.
5a is the same as the true velocity-depth model in Fig-
We shall construct the remaining part of the model
ure 9.3-1. Additionally, the section in Figure 9.3-5b,
layer-by-layer — the shale-marl sequence, the imbri-
which was obtained by stacking the image gathers as
cate structures and the substratum, by prestack depth
in Figure 9.3-6c, represents the correct image in depth.
migration. Our goal here is to examine image gath-
ers from prestack depth migration for layer velocity Interpret the flat reflector at 2.5 km from this section
estimation and the stack of the image gathers for re- and insert it as a layer boundary into the velocity-depth
flector geometry delineation. The results of the layer-by- model (Figure 9.3-5a).
layer velocity-depth model estimation based on image- Now examine further the image gathers in Figure
gather analysis are shown in Figures 9.3-3, 9.3-4, and 9.3-6. The flatness of an event on image gathers is an in-
9.3-5. dication of the accuracy of the velocity field associated
Figure 9.3-3a shows the velocity-depth model with with the layer above the layer boundary that is rep-
the known shallow sequence down to 1 km, and the un- resented by that event. Actually, to declare the layer
known part that is represented by the half-space below. velocity as correct and accurate, not only the event as-
The velocity assigned to the half-space is that of the sociated with the base-layer, but also all events above
shale-marl sequence (Figure 9.3-1). Note that this layer that event should be flat. Note that in Figure 9.3-6a
has a vertical velocity gradient. Figure 9.3-3b shows the all events down to and including the top-carbonate are
depth image from prestack depth migration using the fairly flat.
velocity-depth model in Figure 9.3-3a. Superimposed on The “nonflatness” of an event on image gathers is
this section are the layer boundaries in the velocity- detectable only if there is sufficient cable length. For
depth model. instance, the shallow events on image gathers in Figure
Figure 9.3-6a shows selected image gathers from 9.3-6a extend to a narrow range of offsets as a result of
prestack depth migration. The event T associated with muting. As a result, it is difficult, if not impossible, to
the top of the carbonate sequence exhibits a flat char- infer how flat these events are.
acter on the image gathers. This indicates that the ve- The events associated with the base-carbonate and
locity field for the layer above is correct. Consequently, the underlying flat reflector exhibit residual moveout
(text continues on p. 1415)
Earth Modeling in Depth 1407

FIG. 9.3-2. (a) CMP-stacked section associated with the velocity-depth model in Figure 9.3-1; (b) poststack time migration.
1408 Seismic Data Analysis

FIG. 9.3-3. (a) Velocity-depth model associated with the overthrust data in two parts — the known part down to 1 km
and the unknown part represented by the half-space below; (b) prestack depth migration using the model in (a). Half-space
velocity is that of the shale-marl sequence — 3200 m/s referenced to the top of the layer with 0.5 m/s/m gradient.
Earth Modeling in Depth 1409

FIG. 9.3-4. (a) Velocity-depth model associated with the overthrust data in two parts — the known part down to the top of
the imbricate structures and the unknown part represented by the half-space below; (b) prestack depth migration using the
model in (a). Half-space velocity is that of the carbonate sequence — 5700 m/s.
1410 Seismic Data Analysis

FIG. 9.3-5. (a) Velocity-depth model associated with the overthrust data in two parts — the known part down to the base
of the imbricate structures and the unknown part represented by the half-space below; (b) prestack depth migration using
the model in (a). Half-space velocity is that of the substratum — 5000 m/s.
Earth Modeling in Depth 1411
1412 Seismic Data Analysis

FIG. 9.3-7. (a) Velocity-depth model associated with the overthrust data in two parts — the known part down to the base
of the imbricate structures and the unknown part represented by the half-space below; (b) prestack depth migration using
the model in (a). Half-space velocity is 4500 m/s.
Earth Modeling in Depth 1413

FIG. 9.3-8. (a) Velocity-depth model associated with the overthrust data in two parts — the known part down to the base
of the imbricate structures and the unknown part represented by the half-space below; (b) prestack depth migration using
the model in (a). Half-space velocity is 5500 m/s.
1414 Seismic Data Analysis
Earth Modeling in Depth 1415

on image gathers in Figure 9.3-6a. The event curva- Suppose that the velocity-depth model has been es-
ture for both layer boundaries is upward; this suggests tablished for the first n − 1 layers, and that we want
that the velocity field assigned to the half-space which to estimate the layer velocity for the nth layer. Using
includes these two layer boundaries (Figure 9.3-3a) is the known overburden velocity-depth model for the first
erroneously low. n − 1 layers, assign a constant velocity to the half-space
The detectability of residual moveout on image below that includes the nth layer. Perform prestack
gathers is possible, again, only if there is sufficient ca- depth migration and output image gathers at some ap-
ble length. The smaller the effective cable length for propriate interval along the line. The image gathers
an event, the less detectable is the residual moveout, should exhibit flat character for the events associated
thus the poorer the velocity resolution. In the limit of with the n − 1 layers, but show a residual moveout
zero offset, velocity resolution becomes nill. The resid- for the event associated with the nth layer. Repeat the
ual moveout also is influenced by the magnitude of the analysis using the same overburden model and a range
layer velocity and the depth of the layer boundary. of constant velocities assigned to the half-space.
Using the correct velocity for the carbonate se- Consider the velocity-depth model in Figure 9.3-5a.
quence (Figure 9.3-4a), the event associated with the The substratum below the base-carbonate layer bound-
base-carbonate becomes flat on image gathers (Figure ary is the half-space that contains the flat reflector at
9.3-6b). Nevertheless, the event associated with the flat 2.5 km. The half-space velocity is 5000 m/s. Selected
reflector now exhibits a downward curvature, suggest- image gathers from prestack depth migration using this
ing that the half-space velocity (5700 m/s) in Figure model are shown in Figure 9.3-6c, and the depth image
9.3-4a is erroneously low for the substratum. is shown in Figure 9.3-5b. Keep the same overburden
Finally, using the correct velocity for the substra- model as in Figure 9.3-5a, but assign a constant velocity
tum (Figure 9.3-5a), the event associated with the flat of 4500 m/s to the half-space as shown in Figure 9.3-7a.
reflector F exhibits a flat character on image gathers The resulting image gathers are shown in Figure 9.3-9a
(Figure 9.3-6c). Not only is this event flat in Figure and the depth image in Figure 9.3-7b. The event on the
9.3-6c, but so are all the other events above. Suppose image gathers associated with the flat reflector exhibits
that an image gather contains 10 events, and that all a small, but still detectable residual moveout that sug-
are flat except the sixth from the top. This does not im- gests erroneously low velocity for the layer above the flat
ply that the velocity-depth model is correct for all the reflector. Repeat the analysis for a half-space velocity
layers except for the sixth layer. Instead, it implies that of 5500 m/s (Figure 9.3-8a). Selected image gathers are
the velocity-depth model is correct down to and includ- shown in Figure 9.3-9b and the stack of image gathers
ing the fifth layer, and the deeper part of the model is in Figure 9.3-8b.
incorrect. The velocity-depth models (Figure 9.3-5a, 9.3-7a,
Flatness of all the events on image gathers is a 9.3-8a) corresponding to the image gathers in Figures
means of verifying the accuracy of the velocity-depth 9.3-6c, 9.3-9a, and 9.3-9b, respectively, have the same
model used in prestack depth migration. This is a nec- overburden, but with three different constant half-space
essary but not sufficient condition for verifying the ac- velocities — 5000, 4500, and 5500 m/s, assigned to the
curacy of a model. As stated earlier in Section 9.0, by substratum below the carbonate sequence. First, exam-
using nonzero-offset data, we can hope to resolve the ine the residual moveout on the image gathers for the
velocity-depth ambiguity for a reflector if the data used event associated with the flat reflector F . Although the
in inversion have been recorded with offsets greater than moveout differences are subtle, the flatness is achieved
the reflector depth. The additional limitations in model with the 5000-m/s half-space velocity. By analyzing the
verification based on image-gather analysis are the re- image gathers from constant half-space velocity scans,
flector depths and the magnitude of the layer velocities. we can make optimum velocity picks at analysis loca-
Specifically, the deeper the reflector or the higher the tions that best satisfy the flatness criterion for the layer
velocity of the layer above, the less the detectability of under consideration.
a residual moveout on image gathers (Figure 9.3-6).

9.4 MODEL BUILDING


Constant Half-Space Velocity Analysis
Although doing it right the first time is most desirable,
Image gathers from prestack depth migration are used there is never a situation where this is possible when es-
to estimate layer velocities in two ways — constant half- timating an earth model in depth. The velocity-depth
space velocity analysis and residual moveout analysis. ambiguity that is inherent to inversion makes it very
1416 Seismic Data Analysis

difficult getting the right answer — the true geologi- (d) Perform vertical-ray or image-ray depth conversion
cal model, let alone the first time. Limitations in the of the time horizons from step (a) using the interval
resolving power of the methods to estimate layer ve- velocity maps from step (c).
locities that arise from the band-limited nature of the
recorded data and finite cable length used in record- The combination of the interval velocity maps from step
ing further compound the problem. Finally, traveltime (c) with the depth horizons from step (d) constitutes
picking that is needed for most velocity estimation tech- the initial model derived from time-to-depth conver-
niques and time-to-depth conversion as well as picking sion. This initial model may then be calibrated to well
depth horizons from depth-migrated data to delineate data if the depth horizon maps are to be used as depth
reflector geometries are all adversely affected by noise structure maps for well positioning. Alternatively, the
present in the data. estimated initial model may be updated and used to
All things considered, we can only expect to do our perform 3-D post- or prestack depth migration to de-
best in estimating what may be called an initial model, rive an image volume in depth.
and update this model to get an acceptable final model. In reference to step (c) of the procedure outlined
In this section, we shall discuss ways to estimate an above, interval velocities may be estimated by methods
initial model, and in the next section, we shall discuss other than Dix conversion (Table 9-1). Nevertheless, as
application of residual moveout analysis and reflection part of the common strategy for time-to-depth conver-
traveltime tomography to update the initial model. sion, interval velocities are derived from Dix conversion
We shall discuss two strategies applicable to both of rms velocities. In reference to step (d) of the proce-
2-D and 3-D seismic data for initial model building: dure outlined above, depth conversion of time horizons
may be performed by one of the following three strate-
(a) A time-to-depth conversion strategy based on in- gies:
terpretation in the time domain, and
(b) A layer-by-layer inversion strategy based on inter- (a) Most commonly applied strategy is based on a com-
pretation in the depth domain. bination of Dix conversion of rms velocities to in-
terval velocities and image-ray depth conversion of
Practical methods to estimate layer velocities and de- time horizons interpreted from the time-migrated
lineate reflector geometries used in implementing the volume of data. This is the usual implementation
two strategies are listed in Table 9-1. A widely used of map migration. Stacking velocity inversion some-
combination for time-to-depth conversion is Dix con- times may be substituted for Dix conversion to es-
version to estimate layer velocities and image-ray depth timate interval velocities.
conversion to delineate reflector geometries. Whereas (b) Alternatively, depth conversion may be performed
for layer-by-layer inversion, a widely used combination using vertical rays. This is acceptable only if lat-
is coherency inversion to estimate layer velocities and eral mispositioning because of lateral velocity vari-
poststack depth migration to delineate reflector geome- ations is negligible (Section 8.0). Again, the interval
tries. velocities are estimated by Dix conversion.
(c) Albeit rarely, a third option is to use normal-
Time-to-Depth Conversion incidence rays for depth conversion. Time horizons
intepreted from the time-migrated volume of data
The time-to-depth conversion strategy involves the fol-
may first be forward-modeled to derive 3-D zero-
lowing steps:
offset traveltimes, which are then depth-converted
(a) Interpret a set of time horizons from an image vol- using normal-incidence rays. Dix conversion still is
ume derived from time migration; these time hori- the robust method for interval velocity estimation.
zons are usually associated with layer boundaries
with velocity contrast or geological formations of We now describe the details of the steps involved
interest. in the usual implementation of map migration based on
(b) Intersect rms velocity functions picked at speci- image-ray depth conversion to derive depth structure
fied analysis locations over the survey area with maps.
the time horizons from step (a) to derive horizon-
consistent rms velocity maps. The rms velocity
functions are preferably picked from gathers de- Time Structure Maps
rived from prestack time migration.
(c) Perform Dix conversion of the rms velocity maps Time horizons are picked from image volumes obtained
from step (b) to derive interval velocity maps. from either 3-D post- or 3-D prestack time migration.
Earth Modeling in Depth 1417

Aside from improved imaging of conflicting dips with Note the fault signatures especially evident on hori-
different stacking velocities, the latter offers the advan- zons H1, H2, and H3. Actually, the sharp boundary be-
tage of providing a 3-D rms velocity field that is asso- tween the green-blues and the yellow-reds observed on
ciated with events in their migrated positions (Section all the maps is clear evidence of a major fault through
7.4). Figure 9.4-1 shows a 3-D view of the image volume the middle of the survey area parallel to the longer
derived from 3-D prestack time migration of a marine dimension. Adjoining this major fault are the oblique
3-D data set. faults that can be detected on H1, H2, and H3. Since
The interpreter identifies the time horizons that are we use rays to perform time-to-depth conversion, it is
associated with depositional sequence boundaries and important to ensure that ray tracing is made stable by
geologically and lithologically significant layer bound- applying a carefully measured amount of smoothing to
aries within some of the depositional units. Then, re-
the gridded surfaces (Figure 9.4-4). This smoothing also
flection times are picked by combining seed detection
is needed to edit outliers among the control points that
and line-based interpretation strategies (Section 7.5).
have inevitably corrupted the grid points. Finally, the
A simplified form of an interpretation session without
gridded surfaces are usually displayed in the form of
explicit fault identification is given below.
contour maps as shown in Figure 9.4-5.
(a) Seed points are placed on the event that is being
picked at locations with good signal-to-noise ratio.
These are then used to drive a seed detection al- Interval Velocity Maps
gorithm to pick patches of the time surface around
each seed point. Figure 9.4-2 shows picks from six The Dix equation (9-1), which relates rms velocities
time horizons from the shallowest (H1) to the deep- to interval velocities, is used to derive interval velocity
est (H6). Depending on the signal-to-noise ratio maps. RMS velocities, in principle, are most appropri-
and complexity of geometry of the time horizon, ately estimated from prestack time-migrated data (Sec-
the extent of the surface patches varies from hori- tion 7.4). Recall from Section 3.1 that the type of veloc-
zon to horizon. Some horizons are almost entirely ity that can be most reliably estimated from CMP data
picked by seed detection (H2), some are covered by is the velocity used to apply normal-moveout correction.
limited amounts of seed-detected surface patches To stack the data we also substitute NMO velocities for
(H4), and some are not eligible for seed detection stacking velocities. The use of NMO velocities as stack-
(H6). Note that seed detection has failed especially ing velocities is based on the small-spread hyperbola as-
in intensively faulted areas. sumption. To further substitute stacking velocities for
(b) To ensure structural control and adequate coverage rms velocities is only allowed if the CMP data are as-
of the time horizons, additional picking along in- sociated with horizontally layered earth. To justify the
lines and crosslines is required. Figure 9.4-2 shows use of stacking velocities as rms velocities, we first need
the horizon strands derived from line-based pick- to correct for the dip effect on stacking velocities by
ing. way of dip-movoeut (DMO) correction (Section 5.1). In
(c) The surface patches derived from seed detection the case of a 3-D survey, we also need to correct for
and horizon strands derived from line-based pick- the source-receiver azimuthal effects on stacking veloc-
ing are then combined to form the complete set of ities by way of 3-D DMO correction (Section 7.2). This
control points for each horizon as shown in Fig-
means that it is the stacking velocity field derived from
ure 9.4-2. At this stage, a comprehensive editing
3-D DMO-corrected data that should be considered as a
and repicking are required to ensure consistency in
plausable substitute for the rms velocity field. But then
picking. The edited control points are then input
the DMO velocities are associated with CMP gathers in
to a surface fitting algorithm ( Section J.5) to cre-
ate grid points that define the surface by a map their unmigrated positions. Strictly, we need the move-
function tn (x, y) at every inline and crossline in- out velocities not only corrected for dip and azimuth ef-
tersection, where the function value tn represents fects but also estimated from gathers in their migrated
the reflection time at the (x, y) location on the nth positions. This is because the rms velocities used in
surface. Figure 9.4-3 shows the gridded surfaces de- Dix equation (9-1) are defined for a horizontally layered
rived from the control points shown in Figure 9.4-2 earth model (Section J.4). Thus the desired strategy is
for the six horizons. The red corresponds to the that velocities derived from 3-D prestack time migra-
highs and the blue corresponds to the lows on each tion should be substituted for rms velocities.
horizon. Each horizon has been color-coded inde- Although prestack time migration velocities are
pendently. most desired to substitute for rms velocities, the in-
(text continues on p. 1425)
1418 Seismic Data Analysis

FIG. 9.4-1. An image volume of data derived from 3-D prestack time migration. The color-coded top surface is the water
bottom with the acquisition footprint exhibited by the striations along the inline direction.
Earth Modeling in Depth 1419

FIG. 9.4-2. Results of interpretation of six time horizons from the image volume shown in Figure 9.4-1.
1420 Seismic Data Analysis

FIG. 9.4-3. Gridded surfaces derived from the horizon picks shown in Figure 9.4-2.
Earth Modeling in Depth 1421

FIG. 9.4-4. Smoothed versions of the gridded surfaces shown in Figure 9.4-3.
1422 Seismic Data Analysis

FIG. 9.4-5. Time contour maps associated with the smoothed surfaces shown in Figure 9.4-4.
Earth Modeling in Depth 1423

FIG. 9.4-6. Horizon-consistent rms velocity maps associated with the time horizons shown in Figure 9.4-5.
1424 Seismic Data Analysis

FIG. 9.4-7. Interval velocity maps derived from the horizon-consistent rms velocity maps shown in Figure 9.4-6.
Earth Modeling in Depth 1425

tepreter may be compelled to use whatever veloc- this goal, we want to trace the image ray from the point
ity functions that may be available. These may have of emergence (x0 , y0 , 0) back to its point of departure
been derived from velocity analysis applied to DMO- (xn , yn , zn ).
corrected data or even, although hardly desirable, to Suppose that the first n − 1 horizons have already
CMP data without DMO correction. Under those cir- been converted to depth, and that next we want to con-
cumstances, the velocity functions picked at analysis vert the nth horizon to depth. Since the earth model is
locations need to be edited for any dip effect by ei- known for the first n − 1 layers, then we know the co-
ther eliminating the suspect functions altogether or by ordinates of the intersection point S1 of the image ray
smoothing. with the first layer, (x1 , y1 , z1 ), where by definiton of
Whatever the source of information, the interpreter the image ray, x1 = x0 and y1 = y0 . By using Snell’s
starts with a set of velocity functions, each made up of law, we can determine the direction of the ray as it de-
parts point S1 reaching point S2 on the next surface.
a set of time-velocity pairs and associated with analysis
As the image ray moves from one surface to the next,
locations over the survey area. The analysis grid typi-
we add up the time it takes to travel. When it reaches
cally varies from 500 × 500 m to 2 × 2 km; hence, there
the (n − 1)st layer, the elapsed two-way time tn−1 is
may be as many as 400 velocity functions per 100 km-
n−1
squared of the survey area. The gridded time horizons ∆sk
are intersected with the velocity functions and, for each tn−1 = 2 , (9 − 2)
vk
horizon, velocity nodes are extracted from the veloc- k

ity functions coincident with the horizon times at the where vk is the interval velocity of the kth layer, and
locations of the velocity functions themselves. These ve- ∆sk is the distance between the intersection points of
locity nodes are then used as control points input to a the image ray, Sk−1 and Sk , on the (k − 1)st and kth
gridding algorithm to create horizon-consistent rms ve- surfaces given by
locity maps (Figure 9.4-6). There may be a need for 1/2
further editing and smoothing of the rms velocity grids. ∆sk = (xk − xk−1 )2 + (yk − yk−1 )2 + (zk − zk−1 )2 .
Finally, the horizon-consistent rms velocity values
(9 − 3)
and the horizon times at each grid point are used in Dix
equation (9-1) to compute the interval velocity values, Now, we examine the situation when the image ray
which are then used to create the horizon-consistent in- departs the (n − 1)st layer at point Sn−1 on the way
terval velocity maps (Figure 9.4-7). Once again, there to the nth surface. Again, by Snell’s law we know the
may be further need for editing and smoothing of the direction of the ray. We also know the elapsed time tn −
interval velocity maps to remove any geologically im- tn−1 from the (n − 1)st surface to the nth surface since
plausable velocity variations. we know the total elapsed time tn−1 from equation (9-
2) and the total elapsed time tn from the input time
horizon read at point (x0 , y0 , z = 0). Finally, we know
Depth Structure Maps the interval velocity vn of the nth layer from the interval
velocity map. Therefore, we can calculate the elapsed
We now have a set of time horizon maps (Figure 9.4-4) distance ∆sn along the raypath as it departs the point
and a corresponding set of interval velocity maps (Fig- Sk−1 on the (k − 1)st surface in the direction dictated
by Snell’s law. The quantity ∆sn is given by
ure 9.4-7). Recall from Section 8.0 that, in principle, a
time-migrated image can be converted to a depth sec- vn
∆sn = tn − tn−1 . (9 − 4)
tion by mapping the amplitudes along image rays. This 2
notion also can be employed to convert time horizons Finally, the coordinates of the point Sn that we
into depth horizons. The process is done layer by layer need to know to perform the time-to-depth conversion
starting with the shallowest horizon. are given by
A comprehensive mathematical discussion on xn = xn−1 + ∆sn cos α, (9 − 5a)
image-ray tracing is given by Hubral and Krey (1980).
yn = yn−1 + ∆sn cos β, (9 − 5b)
Refer to Figure 9.4-8 for a brief description of image-
ray tracing. Consider an image ray that departs the nth zn = zn−1 + ∆sn cos γ, (9 − 5c)
surface at point Sn with coordinates (xn , yn , zn ) and where α, β, and γ are the directional cosines of the ray
emerges at the right angle at the earth’s surface at point at point Sn−1 . The directional cosines are known by
S0 with coordinates (x0 , y0 , z = 0). Our goal is to de- the application of Snell’s law at point Sn−1 with known
termine the coordinates of the output point (xn , yn , zn ) coordinates (xn−1 , yn−1 , zn−1 ).
on the depth map from image-ray depth conversion of To summarize, given the depth and interval veloc-
the input point (x0 , y0 , tn ) on the time map. To achieve ity maps for the first n − 1 horizons, and the time and
1426 Seismic Data Analysis

FIG. 9.4-8. Principles of image-ray depth conversion. See text for details.

interval velocity maps for the nth horizon, we can trace The displacement vector also has a directional az-
an image ray associated with the time tn (x0 , y0 ) on the imuth φn which is given by
time map and derive the depth value zn (xn , yn ) on the yn − y0
depth map. φn = tan−1 (9 − 6b)
xn − x0
Figure 9.4-9 shows the depth maps derived from
image-ray depth conversion of the time maps shown in as measured from the inline x direction. The displace-
Figure 9.4-4 using the interval velocity maps shown in ment azimuth maps are shown in Figure 9.4-12. Again,
Figure 9.4-7. As for the time maps, the usual way to note that the most significant azimuthal variations are
post the depth maps is to contour them (Figure 9.4- along the fault zones.
10).
The depth maps are compatible with the time
maps; nevertheless, there can be subtle differences be- Calibration to Well Tops
cause of velocity variations that would give rise to the
departure of image rays from the vertical. To quantify The depth structure maps derived from time-to-depth
the differences between vertical-ray and image-ray tra- conversion or layer-by-layer inversion invariably will not
jectories, and thus to quantify the differences between match the well tops. The sources of discrepancy between
the time maps and depth maps, we can calculate the the estimated reflector depths and the well tops include
modulus ∆dn of the lateral displacement vector between limitations in the methods for interval velocity estima-
the points S0 and Sn as tion, mispicking of time horizons input to depth con-
version, and limitations in the actual depth conversion
∆dn = (xn − x0 )2 + (yn − y0 )2 , (9 − 6a)
itself within the context of ray tracing through an earth
and create the displacement modulus maps shown in model that includes complex layer boundaries. For the
Figure 9.4-11. Note that the most significant displace- depth structure maps to be usable in subsequent reser-
ment between the vertical rays and image rays is at the voir modeling and simulation, it is imperative to cali-
fault zones. brate them to well tops.
(text continues on p. 1433)
Earth Modeling in Depth 1427

FIG. 9.4-9. Depth horizons derived from time-to-depth conversion of the time horizons shown in Figure 9.4-4 using the
interval velocity maps shown in Figure 9.4-7.
1428 Seismic Data Analysis

FIG. 9.4-10. Contour maps of the depth horizons shown in Figure 9.4-9.
Earth Modeling in Depth 1429

FIG. 9.4-11. Modulus maps for the displacement vectors that describe the lateral departure between vertical rays and image
rays. See text for details.
1430 Seismic Data Analysis

FIG. 9.4-12. Azimuth maps for the displacement vectors that describe the directional difference between vertical rays and
image rays. See text for details.
Earth Modeling in Depth 1431

FIG. 9.4-13. (a) Depth horizons as in Figure 9.4-9 derived from image-ray depth conversion of the time horizons shown in
Figure 9.4-4; (b) solid model created from the depth horizons in (a).
1432 Seismic Data Analysis

FIG. 9.4-14. Explosion of the solid model shown in Figure 9.4-13b.


Earth Modeling in Depth 1433

Consider a seismically derived depth structure map (n−1)st layer should not be calibrated before estimating
zs (x, y) based on time-to-depth conversion, say, for the the model for the next layer n. This is because seismi-
layer boundary associated with the top-reservoir. Also cally derived layer velocities almost never match with
consider Nw well tops zw (xi , yi ) for this horizon at loca- well velocities. The discrepancy between the two is at-
tions (xi , yi ), i = 1, 2, . . . , Nw . Since the velocity-depth tributable to several factors, including the limited res-
model derived from time-to-depth conversion is sup- olution in velocities estimated from seismic data (Sec-
posed to be consistent with the input data — the time tions 9.1, 9.2, and 9.3) and seismic anisotropy (Section
structure map τ (x, y) created from the interpretation of 11.7). Additionally, the high-frequency variations in the
the time-migrated volume of data, we have well velocities are absent from the seismically derived
velocities.
zs (x, y) The calibrated depth maps can be used to create
τ (x, y) = 2 , (9 − 7a)
Vs (x, y) a solid model of the earth as illustrated in Figure 9.4-
where, for the purpose of calibration, Vs (x, y) can be 13. Each layer is represented by a solid (Figure 9.4-14)
with its interior populated by specific layer parameters.
either the average or rms velocity map associated with
These may include compressional- and shear-wave ve-
the horizon zs (x, y). Also, for simplicity in calibration,
locities, densities, and rock physics parameters such as
we consider vertical rays rather than image rays as in
porosity, permeability, pore pressure, and fluid satura-
equation (9-7a). tion. When populated by the petrophysical parameters,
There exists a calibration velocity Vc (x, y) such the solid associated with the reservoir layer represents a
that, at a well location (xi , yi ), it satisfies the relation reservoir model. For the purpose of reservoir modeling,
zw (xi , yi ) the solid for the reservoir layer usually is downscaled
τ (xi , yi ) = 2 . (9 − 7b) in the vertical direction by dividing it into thin slices
Vc (xi , yi )
with a thickness as small as 1 m — much less than the
Combine equations (9-7a) and (9-7b) to get a relation threshold for vertical seismic resolution (Section 11.1).
that is satisfied at the well locations Additionally, the solid for the reservoir layer is upscaled
Vc (xi , yi ) zw (xi , yi ) in the lateral direction by dividing each thin slice into
= . (9 − 8) finite elements with a varying size of up to 250 m on
Vs (xi , yi ) zs (xi , yi )
one side. The reservoir model is eventually fed into a
From the knowledge of the well tops zw (xi , yi ) and reservoir simulation scheme to predict the geometry of
the seismically derived reflector depths zs (xi , yi ) at the the fluid flow from the given reservoir parameters.
well locations, equation (9-8) gives a calibration factor
c(xi , yi ) = Vc (xi , yi )/Vs (xi , yi ) computed at each of the
well locations. Next, apply kriging or some other inter-
Layer-by-Layer Inversion
polation technique to the sparsely defined calibration
factors c(xi , yi ), i = 1, 2, . . . , Nw to derive a calibration
factor map c(x, y) specified at all grid locations (x, y). The second strategy for initial model building involves
Kriging is a statistical method of determining the best a layer-by-layer application of an appropriate combina-
tion of inversion methods listed in Table 9-1. Layer-by-
estimate for an unknown quantity such as c(x, y) at
layer inversion involves the following steps:
some location (x, y) using a sparse set of values such
as c(xi , yi ) specified at locations (xi , yi ) (Sheriff, 1991;
David, 1987). (a) Interpret a set of time horizons from unmigrated
The final step in calibration is to scale the depth data to be used in lieu of zero-offset reflection trav-
eltimes required by coherency inversion. Alterna-
structure map zs (x, y) by the calibration factor map
tively, interpret a set of time horizons from time-
c(x, y)
migrated data and perform forward modeling to
zc (x, y) = c(x, y) zs (x, y), (9 − 9) obtain the required zero-offset traveltimes.
(b) Assume that interval velocities and reflector ge-
where zc (x, y) is the calibrated depth structure map. ometries for the first n − 1 layers have been esti-
Note that, by way of equations (9-8) and (9-9), the cal- mated, and that we want to estimate the same for
ibrated depth zc coincides with the well top zw at well the nth layer. By using the time horizons from step
location (xi , yi ). (a), first apply one of the inversion methods listed
Calibration to well tops is done only after the com- in the left-hand column of Table 9-1 to estimate
pletion of model building, and just before well plan- the interval velocity field.
ning and reservoir modeling. When estimating an earth (c) Assign the estimated interval velocity field for the
model by following a layer-by-layer inversion procedure nth layer to the half space that includes the un-
(next subsection), depth horizon associated with the known part of the model — the nth layer and the
1434 Seismic Data Analysis

layers below, and perform depth migration. Often, (e) Interpret the depth horizon that corresponds to the
it is sufficient to do poststack depth migration. base boundary of the layer under consideration.
(d) Interpret the depth horizon associated with the (f) Incorporate the interpreted depth horizon into the
base of the nth layer from the depth image and intermediate velocity-depth model (panel (d) in
incorporate it into the velocity-depth model. Figures 9.4-16 through 9.4-21).
(g) Proceed to the next layer below and repeat the
The layer-by-layer inversion strategy alternates be- above steps.
tween layer velocity estimation and reflector geometry
delineation for each layer starting from the earth’s sur- The coherency semblance spectra shown in panel
face and moving down one layer at a time. The layer-by- (a) of Figures 9.4-16 through 9.4-21 have been compiled
layer estimation strategy facilitates checking the results into a single panel as shown in Figure 9.4-22. First, note
of inversion for one layer before moving down to the the gradual decrease in resolution as we move down
next. As such model updating can be interleaved with to the deeper layers. Much like conventional stacking
model estimation to circumvent accumulation of errors velocity analysis, the resolving power of coherency in-
in layer velocities and reflector geometries as the anal- version is governed by the cable length, the reflector
ysis proceeds from the top down. depth, the layer velocity, and bandwidth of the data.
We shall demonstrate the layer-by-layer inversion Second, observe the short-wavelength variations in the
strategy by applying a combination of coherency inver- velocity profiles defined by the semblance peaks. As was
sion to estimate layer velocities and poststack depth mi- demonstrated by the synthetic data experiments in Sec-
gration to delineate reflector geometries to the marine tion 9.1, these are caused by lateral velocity variations
data set shown in Figure 9.4-15. Superimposed on the
in the overlying layers that are much less than a cable
stacked section are the segments of interpreted reflec-
length (Figures 9.1-13 and 9.1-20). In the present case
tion traveltime horizons. Starting from the top, H1 is
shown in Figure 9.4-22, note that the interval velocity
the water bottom whereas H6 is the top-salt boundary.
profiles for layers H6 and H7 exhibit much more pro-
The abundance of diffractions within the salt layer is
nounced fluctuations compared to the layers above. This
associated with high-velocity anhydrite-dolomite rafts.
is caused by the strong lateral variations in the interval
Input to coherency inversion is the zero-offset reflec-
velocity profile for layer H5. In practice, the interval ve-
tion traveltimes associated with normal-incidence rays
and unmigrated data as shown in Figure 9.4-15. Diffrac- locity profile picked from the semblance spectrum must
tion flanks present in the unmigrated data must be exclude such rapid fluctuations. Otherwise, the reflec-
avoided when interpreting the reflection traveltime seg- tor geometry associated with the base of the layer under
ments from unmigrated data. consideration will be corrupted by geologically implaus-
Details of coherency inversion for interval veloc- able variations. Also note from the semblance spectra
ity estimation are provided in Section 9.1. Results of that we have to interpolate the interval velocity profiles
the layer-by-layer analysis have been compiled in Fig- through the zones with missing reflection events.
ures 9.4-16 through 9.4-21. Begin with the trivial task The accuracy in velocity estimation by coherency
of modeling the water layer by normal-incidence time- inversion can be monitored closely by examining the
to-depth conversion of the time horizon H1 in Figure modeled CMP traveltimes and the associated raypaths.
9.4-15 using a constant layer velocity of 1500 m/s. For Shown in Figure 9.4-23 are the CMP gathers at a se-
each layer H2 through H6, and one layer at a time, the lected analysis location for each of the layers H2 through
analysis includes the following steps: H7. Superimposed on these gathers are the modeled
CMP traveltime trajectories that correspond to the op-
(a) Apply coherency inversion to compute the horizon- timum layer velocities picked from the semblance spec-
consistent semblance spectrum (panel (a) in Fig- tra shown in Figure 9.4-24. The raypaths associated
ures 9.4-16 through 9.4-21), and with the modeled CMP traveltimes shown in Figure
(b) pick the interval velocity profile from the semblance 9.4-25 illustrate the ray bending at layer boundaries.
spectrum by tracking the semblance peaks. As such, the modeled traveltimes in Figure 9.4-23 are
(c) Assign the interval velocity profile to the half space in general nonhyperbolic. From the surface lateral ex-
that includes the unknown layer itself and use the tent of the raypath family for each layer in Figure 9.4-25
known overburden layer velocities and reflector ge- and the offset range of the modeled traveltimes in Fig-
ometries to create the intermediate velocity-depth ure 9.4-23, we note the resolution that can be attained
model (panel (b) in Figures 9.4-16 through 9.4-21). from the semblance spectra. Specifically, note that the
(d) Perform poststack depth migration using the inter- sharpness of the spectra in Figure 9.4-24 decreases with
mediate velocity-depth model (panel (c) in Figures increasing layer velocity, decreasing cable length, and
9.4-16 through 9.4-21). increasing reflector depth.
(text continues on p. 1440)
Earth Modeling in Depth 1435

FIG. 9.4-15. An unmigrated CMP-stacked section with seven time horizons that correspond to layer boundaries with
significant velocity contrast.
1436 Seismic Data Analysis
Earth Modeling in Depth 1437
1438 Seismic Data Analysis
Earth Modeling in Depth 1439

FIG. 9.4-22. Semblance spectra from coherency inversion with picked interval velocity profiles for layers H2 through H7 as
denoted in Figure 9.4-15.
1440 Seismic Data Analysis

The results of the alternating steps of layer velocity (Figure 9.4-15), in the present case of a complex struc-
estimation and reflector geometry delineation involved ture it is easier and safer to interpret the time-migrated
in layer-by-layer model building have been compiled in data.
Figure 9.4-26. The left and the right columns corre- Results of the alternating steps of layer velocity
spond to the compilations of panels (b) and (c) of Fig- estimation using coherency inversion and reflector ge-
ures 9.4-16 through 9.4-21, respectively. The analysis ometry delineation using poststack depth migration in-
sequence illustrated by Figure 9.4-26 begins with panel volved in layer-by-layer model building have been com-
(a) and ends with panel (l). As outlined above, per- piled in Figure 9.4-29. The left column shows the in-
form coherency inversion to generate the interval veloc-
termediate models for layers H2 to H5 (H1 is the wa-
ity profile and, hence, the intermediate velocity-depth
ter layer), and the right column shows the intermedi-
model for layer H2 (panel (a)). Next, perform poststack
depth migration and delineate the reflector geometry ate depth images from poststack depth migration. Start
associated with the base of layer H2 (panel (b)). Then with the intermediate velocity-depth model in panel (a)
move down to the next layer and repeat the same anal- and perform poststack depth migration to get the depth
ysis (panels (c) and (d)), until you reach the last layer image in panel (b). Interpret the depth horizon associ-
in the sequence (panels (k) and (l)). ated with the base of layer H2 and insert this horizon
The final form of the velocity-depth model esti- into the velocity-depth model in panel (c). Then, move
mated by applying the layer-by-layer inversion strategy down to the next layer and perform coherency inversion
is shown in Figure 9.4-27a. Check the consistency of to derive the interval velocity profile for it. Assign this
this velocity-depth model with the depth image derived velocity profile to the half space below the base of layer
from poststack depth migration (Figure 9.4-27b) and H2. This constitutes the updated intermediate velocity-
the stacked section (Figure 9.4-27c). Specifically, over- depth model (panel (c)) which is used to obtain the
lay the depth horizons from the velocity-depth model depth image in panel (d). Continue in this alternating
onto the depth image and note that the reflector ge- manner for estimating the layer velocities and delineat-
ometries implied by the latter are in agreement with ing the reflector geometries until you reach horizon H5
the depth horizons. Then, perform zero-offset normal- (panel (h)).
incident modeling of the traveltimes associated with the
We do not intend to proceed further and complete
layer boundaries included in the velocity-depth model
the model building for all the layers. Instead, we shall
and overlay them onto the stacked section. Again, note
examine the accuracy of velocity estimation for the next
that the actual reflection traveltimes on the stacked
section are in good agreement with the modeled trav- layer H6 just above the salt layer. Figure 9.4-30 shows
eltimes. This exercise confirms the consistency of the the coherency semblance spectra for layers within the
estimated velocity-depth model with the input stacked overburden from H2 to H6. We observe that the qual-
data. Yet, the estimated model is but one of the many ity of the semblance peaks is very good for layer H2,
possible solutions. It can only be considered an initial while it begins to degrade almost immediately for the
model; thus, it needs to be verified by checking its con- layers below. Specifically, in the case of layer H3, the
sistency with prestack data and, finally, it needs to be semblance peaks lose their sharpness at the center por-
updated (Section 9.5). tion of the line where layer H3 becomes very thin (Fig-
We now examine the performance of the layer-by- ure 9.4-28a) causing instability in ray tracing. Next, the
layer inversion strategy applied to structurally complex lateral velocity variation in layer H4 has caused rapid
data. Figure 9.4-28a shows a time-migrated stacked sec- ondulations in the semblance spectrum for layer H5. Fi-
tion with a complex overburden structure associated nally, note that the quality of the semblance peaks has
with salt tectonics. The top-salt boundary is repre- deteriorated significantly for layer H6.
sented by horizon H6. The section has been interpreted To closely examine this rapid degradation in the
to identify the layer boundaries with significant velocity quality of semblance, we shall conduct velocity estima-
contrast. For coherency inversion to estimate interval
tion for layer H6 at four selected locations. For each lo-
velocities layer by layer, the required zero-offset travel-
cation, Figures 9.4-31 to 9.4-34 show the modeled CMP
times were modeled from the time horizons interpreted
from the time-migrated section. The modeled zero-offset raypaths that correspond to the modeled CMP travel-
traveltimes are shown in Figure 9.4-28 superimposed on time trajectory computed by assigning the velocity as-
the unmigrated stacked section; note that they are con- sociated with the semblance peak to the half space be-
sistent with the observed reflection times. The alterna- low horizon H5. Where the overburden is relatively less
tive to using the modeled zero-offset times for coherency complex, we note that the semblance peak is distinctive
inversion is clearly the reflection times picked directly and thus the estimate velocity is not ambiguous (Fig-
from the unmigrated stacked data. While this alterna- ure 9.4-31). At CMP locations where the overburden is
tive may be appropriate for cases of simple structures sufficiently complex to cause significant ray bending at
(text continues on p. 1450)
Earth Modeling in Depth 1441

FIG. 9.4-23. CMP gathers associated with the stacked section in Figure 9.4-15 displayed with the modeled traveltimes
overlayed. The corresponding CMP raypaths are shown in Figure 9.4-25.
1442 Seismic Data Analysis

FIG. 9.4-24. Semblance spectra derived from coherency inversion to estimate velocities for layers H2 through H7 as denoted
in Figure 9.4-15.
Earth Modeling in Depth 1443

FIG. 9.4-25. Part 1: The CMP raypaths associated with the modeled traveltimes shown in Figures 9.4-23.
1444 Seismic Data Analysis

FIG. 9.4-25. Part 2: The CMP raypaths associated with the modeled traveltimes shown in Figures 9.4-23.
Earth Modeling in Depth 1445

FIG. 9.4-26. Model building layer by layer starting from the surface. Left column shows the velocity-depth models and the
right column shows the image sections from poststack depth migration using the stack shown in Figure 9.4-15.
1446 Seismic Data Analysis

FIG. 9.4-27. (a) Final velocity-depth model derived from the layer-by-layer application of coherency inversion and poststack
depth migration, (b) depth migration using the velocity-depth model in (a), (c) modeled zero-offset traveltimes using the
velocity-depth model in (a) overlayed on the unmigrated stacked section as in Figure 9.4-15.
Earth Modeling in Depth 1447

FIG. 9.4-28. (a) A time-migrated stacked section with a set of interpreted horizons, (b) the unmigrated stacked section with
the zero-offset traveltimes modeled from the time horizons in (a).
1448 Seismic Data Analysis

FIG. 9.4-29. Model building layer by layer starting from the surface. Left column shows the velocity-depth models and the
right column shows the image sections from poststack depth migration using the stack shown in Figure 9.4-28b. Layers are
labeled as the horizons that correspond to the base labeled as in Figure 9.4-28a.
Earth Modeling in Depth 1449

FIG. 9.4-30. Semblance spectra derived from coherency inversion to estimate velocities for layers H2 through H6 as denoted
in Figure 9.4-28a.
1450 Seismic Data Analysis

layer boundaries, however, the semblance curves have the flat time horizons (Figure 9.4-36a) to transform
poor quality (Figures 9.4-32 to 9.4-34). Note that, at to nonflat depth horizons.
these locations, reflection point dispersal at the vicin- (e) Perform poststack depth migration (Figure 9.4-
ity of the normal-incidence point is large and raypaths 38a) using the initial interval velocity field (Figure
are complex. As a result, the complex layer boundaries 9.4-37b) and overlay the depth horizons derived in
adversely affect the quality of interval velocity estima- step (c) onto the depth section. Note that the depth
tion using methods that rely on ray tracing. Also recall horizons do not conform to the geometry of the re-
from the model experiments in Section 9.1 that lateral flectors inferred by depth migration; thus, the term
velocity variations in the layers above cause rapid fluc- structure-independent model estimation.
tuations in the semblance profile for the layer below. (f) Discard the structure-independent depth horizons
We conclude that all known practical velocity estima- and replace them with the depth horizons inter-
tion techniques based on ray theory alone suffer from a preted from the depth-migrated section (Figure
degradation of lateral resolution in areas with complex 9.4-38b).
overburden structures. (g) Overlay the depth horizons from step (f) onto the
interval velocity section from step (d) (Figure 9.4-
39a).
Structure-Independent Inversion (h) Extract the interval velocity profiles along the
depth horizons from the interval velocity section
(Figure 9.4-39b).
In areas with low-relief structures and moderate lat-
(i) Eliminate the oscillations from these profiles and
eral velocity variations, a structure-independent inver-
combine them with the depth horizons from step
sion strategy can be used to circumvent interpretation
(f) to build a structurally consistent earth model
of time horizons when deriving an initial estimate of
in depth (Figure 9.4-40a).
the earth model. Compared to a layer-by-layer inver-
(j) Perform prestack depth migration and obtain the
sion strategy, it can prove to be robust and less labor-
image section shown in Figure 9.4-40b from the
intensive. We shall outline a procedure for structure-
independent earth model estimation using a field data image gathers as in Figure 9.4-41.
example.
Shown in Figure 9.4-35a is a time-migrated CMP- Events on image gathers, except for the multiples, are
stacked section that exhibits a highly developed deltaic mostly flat. This means that the estimated earth model
depositional sequence. Note from the stacking velocity in depth (Figure 9.4-40a) is fairly accurate. In practice,
field shown in Figure 9.4-35b that the lateral velocity to attain consistency of the estimated model with the
variations are mild to moderate. Because the dips are input data, depth migration may have to be iterated a
gentle and the structures have low reliefs, we may sub- few times (Section 8.2). This then is followed by model
stitute the stacking velocity field in Figure 9.4-35b for updating with reflection tomography, which is discussed
the rms velocity field that we need for Dix conversion. in the next section.

(a) Consider a set of fictitious, flat time horizons as


displayed in Figure 9.4-36a, and extract the rms
9.5 MODEL UPDATING
velocity profiles along these horizons from the rms
velocity section (Figure 9.4-36b).
(b) Perform Dix conversion to generate interval veloc- Limitations in the techniques for velocity estimation
ity profiles from the rms velocity profiles (Figure and velocity-depth ambiguity inherent to seismic inver-
9.4-37a). Note that for deeper horizons, lateral ve- sion are compelling reasons for the need to update an
locity variations in the layers above have caused estimated earth model in depth, however it has been
oscillations in the interval velocity profiles. constructed. Unfortunately, model updating tools them-
(c) Apply lateral smoothing to remove these oscilla- selves also have limitations in terms of their ability to
tions and use the edited interval velocity profiles resolve lateral velocity variations and refine reflector ge-
to convert the flat time horizons (Figure 9.4-36a) ometries. Again, the cable length and reflector depth
to depth horizons. dictate the extent that model updating techniques can
(d) Combine the interval velocity profiles (Figure 9.4- resolve the velocity-depth ambiguity. In this section, we
37a) with the depth horizons to build an initial shall review residual moveout corrections applied to im-
interval velocity field as shown in Figure 9.4-37b. age gathers as a local method and reflection tomography
Note that lateral velocity variations have caused as a global method for model updating.
(text continues on p. 1462)
Earth Modeling in Depth 1451

FIG. 9.4-31. (a) Raypaths associated with the optimum layer velocity estimated from coherency inversion applied to the
gather at CMP location 216, (b) the coherency semblance spectrum with its peak corresponding to the velocity assigned to
the layer under consideration for computing the raypaths in (a), (c) enlarged view of the raypaths as in (a).
1452 Seismic Data Analysis

FIG. 9.4-32. (a) Raypaths associated with the optimum layer velocity estimated from coherency inversion applied to the
gather at CMP location 479, (b) the coherency semblance spectrum with its peak corresponding to the velocity assigned to
the layer under consideration for computing the raypaths in (a), (c) enlarged view of the raypaths as in (a).
Earth Modeling in Depth 1453

FIG. 9.4-33. (a) Raypaths associated with the optimum layer velocity estimated from coherency inversion applied to the
gather at CMP location 596, (b) the coherency semblance spectrum with its peak corresponding to the velocity assigned to
the layer under consideration for computing the raypaths in (a), (c) enlarged view of the raypaths as in (a).
1454 Seismic Data Analysis

FIG. 9.4-34. (a) Raypaths associated with the optimum layer velocity estimated from coherency inversion applied to the
gather at CMP location 803, (b) the coherency semblance spectrum with its peak corresponding to the velocity assigned to
the layer under consideration for computing the raypaths in (a), (c) enlarged view of the raypaths as in (a).
Earth Modeling in Depth 1455

FIG. 9.4-35. (a) A CMP-stacked section, and (b) the stacking velocity field.
1456 Seismic Data Analysis

FIG. 9.4-36. (a) The stacking velocity field as in Figure 9.4-35b with fictitious flat time horizons superimposed, and (b) the
rms velocity profiles extracted from the section in (a) along the fictitious flat time horizons.
Earth Modeling in Depth 1457

FIG. 9.4-37. (a) The interval velocity profiles derived from the rms velocity profiles in Figure 9.4-36b by way of Dix
conversion, and (b) a structure-independent velocity-depth model created by combining the interval velocity profiles as in (a)
with the fictitious depth horizons that correspond to the flat time horizons as in Figure 9.4-36a.
1458 Seismic Data Analysis

FIG. 9.4-38. (a) Depth migration of the stacked section in Figure 9.4-35 using the velocity-depth model in Figure 9.4-37b
with its associated structure-independent depth horizons superimposed, and (b) the same section as in (a) with a set of
interpreted structurally consistent depth horizons.
Earth Modeling in Depth 1459

FIG. 9.4-39. (a) The velocity-depth model as in Figure 9.4-37b with a set of interpreted structurally consistent depth
horizons as in Figure 9.4-38b, and (b) the interval velocity profiles extracted from the section in (a) along the structurally
consistent depth horizons superimposed onto the same section.
1460 Seismic Data Analysis

FIG. 9.4-40. (a) A structurally consistent velocity-depth model created by combining the interval velocity profiles as in
Figure 9.4-39b with the depth horizons as in Figure 9.4-38b, and (b) prestack depth migration of the data associated with
the stacked section in Figure 9.4-35.
Earth Modeling in Depth 1461

FIG. 9.4-41. Selected image gathers associated with the image section derived from prestack depth migration shown in
Figure 9.4-40.
1462 Seismic Data Analysis

Residual Moveout Analysis incorrectly, then you would observe events with resid-
ual moveout. In principle, this residual moveout can be
We begin with building an initial velocity-depth model computed and used to update the initially picked veloc-
from the data as in Figure 9.4-15 using the time-to- ity function.
depth conversion strategy described in Section 9.4. Fig- Analogous to the conventional stacking velocity
ure 9.5-1a shows the time-migrated stacked section with analysis, if the initial velocity-depth model has been es-
a set of interpreted horizons that correspond to layer timated with sufficient accuracy, then the image gath-
boundaries with significant velocity contrast. Perform ers derived from prestack depth migration using this
Dix conversion of the horizon-consistent rms velocity model should exhibit flat events associated with the
profiles (Figure 9.5-1b) to derive the interval velocity layer boundaries included in the model. Any errors
profiles as shown in Figure 9.5-1c. Clearly, you often in layer velocities and/or reflector geometries, on the
are compelled to apply some smoothing to the rms other hand, should give rise to residual moveout along
velocity profiles before Dix conversion and even ap- those events on the image gathers. Again, in principle,
ply additional smoothing to the interval velocity pro- this residual moveout can be determined and used for
files afterwards. Then, perform image-ray depth con- model updating as illustrated in Figure 9.5-5. The im-
version of the time horizons interpreted from the time- age gather in Figure 9.5-5a exhibits events with residual
migrated section (Figure 9.5-1a) to generate the depth moveout represented by the purple trajectories. Assume
horizons. Finally, combine the interval velocity profiles that the residual moveout is parabolic and compute the
with the depth horizons to create the velocity-depth semblance spectrum as shown in Figure 9.5-5b. The hor-
model shown in Figure 9.5-2a. izontal axis of the semblance plane represents the depth
Complete the analysis by checking for consistency error and the vertical axis represents the depth of the
of this estimated initial velocity-depth model with the event. The semblance spectrum has two quadrants that
depth image derived from poststack depth migration correspond to positive and negative residual moveouts,
and the stacked section. Note that the reflector geome- or equivalently, to positive and negative depth errors. A
tries inferred by the depth image shown in Figure 9.5- flat event would yield a semblance peak that coincides
2b are in agreement with the depth horizons. Addition- with the vertical axis with zero depth error, whereas an
ally, observe that the actual reflection traveltimes on the event with residual moveout would yield a semblance
stacked section are in good agreement with the modeled peak situated either in the left or right quadrant de-
zero-offset traveltimes as shown in Figure 9.5-2c. pending on the sign of the depth error.
We now check for consistency of the initial velocity- Much like picking a velocity function from a con-
depth model in Figure 9.5-2a with prestack data and up- ventional stacking velocity semblance spectrum, a verti-
date it by correcting for the residual moveout observed cal function that represents the depth-dependent resid-
on image gathers derived from prestack depth migra- ual moveout can be picked from the semblance spec-
tion. Image gathers are like moveout-corrected CMP trum in Figure 9.5-5b. This function can then be used
gathers with vertical axis in depth. Unlike in CMP gath- to correct for the residual moveout as shown in Figure
ers, however, events in image gathers are in their mi- 9.5-5c. The actual steps to make the residual moveout
grated positions. Shown in Figure 9.5-3a is an image correction are as follows:
section derived from prestack depth migration of the
data associated with the stacked section in Figure 9.5- (a) Extract the interval velocity function from the
2c using the initial velocity-depth model in Figure 9.5- velocity-depth model at the image-gather location
2a. Superimpose the depth horizons from the velocity- where the residual moveout analysis is to be done.
depth model onto the image section and make some mi- (b) Convert the image gather from depth to time using
nor adjustments where necessary by re-interpreting the the interval velocity function.
depth horizons (Figure 9.5-3b). Selected image gathers (c) Assume that the residual moveout of events on the
associated with the depth image from prestack depth image gather in time is parabolic and compute the
migration are shown in Figure 9.5-4. semblance spectrum for a range of negative and
How do we make use of the image gathers to up- positive moveouts.
date the initial velocity-depth model? First, consider (d) Apply the residual moveout correction to the image
applying to a CMP gather conventional stacking veloc- gather.
ity analysis. Compute the velocity spectrum (Section (e) Compute a new rms velocity function from the re-
3.2) and pick a velocity function. Following the normal- sults of the residual moveout analysis.
moveout correction of the CMP gather using this veloc- (f) Compute a new interval velocity function (equation
ity function, events should look flat if the velocity func- 9-1) from the updated rms velocity function at the
tion had been picked correctly. If the picking was done image-gather location.
(text continues on p. 1469)
Earth Modeling in Depth 1463

FIG. 9.5-1. (a) A poststack time-migrated stacked section with a set of interpreted horizons, (b) the horizon-consistent rms
velocity profiles, (c) the interval profiles derived from Dix conversion of the rms velocity profiles.
1464 Seismic Data Analysis

FIG. 9.5-2. (a) The velocity-depth model created by using a combination of Dix conversion to estimate layer velocities
(Figure 9.5-1c) and image-ray depth conversion of the time horizons interpreted from the time-migrated stacked section
(Figure 9.5-1a), (b) poststack depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on
the unmigrated stacked section.
Earth Modeling in Depth 1465

FIG. 9.5-3. (a) Prestack depth migration using the velocity-depth model in Figure 9.5-2a, (b) the same image section with
depth horizons as in Figure 9.5-2a.
1466 Seismic Data Analysis

FIG. 9.5-4. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-3b and residual moveout
semblance spectra computed from these gathers.
Earth Modeling in Depth 1467

FIG. 9.5-5. (a) An image gather associated with the prestack depth-migrated section in Figure 9.5-3b, (b) residual moveout
spectrum computed from (a), (c) the image gather in (a) after residual moveout correction, (d) the residual moveout section
derived from the semblance spectra shown in Figure 9.5-6.
1468 Seismic Data Analysis

FIG. 9.5-6. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-4 along the depth horizons
shown in Figure 9.5-3b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-2a.
Earth Modeling in Depth 1469

(g) Convert the image gathers back to depth using the shown in Figure 9.5-10. Superimposed on these spectra
new interval velocity functions. are the residual moveout profiles that were picked from
(h) Finally, update the velocity-depth model using the the spectra before the update (Figure 9.5-6). Note that
new interval velocity functions. the residual moveout errors have been reduced signifi-
cantly after the model update.
In principle, residual moveout analysis can be car- The residual moveout analysis of image gathers
ried out for image gathers at some spatial interval. and the update of velocity-depth model should be per-
The residual moveout spectra computed from the im- formed iteratively until the velocity-depth model and
age gathers displayed in Figure 9.5-4 indicate varying the depth image are consistent (Deregowski, 1990).
degrees of errors in the initial velocity-depth model Consistency may be achieved with only a few iterations
shown in Figure 9.5-2a. An alternative way to mea- for cases where residual moveouts are small. Sometimes,
sure the residual moveout is by computing it along consistency is never achieved even after several itera-
the depth horizons themselves. Shown in Figure 9.5-6 tions. This often occurs when the initial residual move-
are the horizon-consistent residual moveout semblance outs are caused largely by significant errors in the ini-
spectra. The vertical axis of the semblance plane repre- tial velocity-depth model. An erroneous initial estimate
sents the depth error which corresponds to either posi- most likely is a result of rapid lateral velocity variations
tive or negative residual moveout. less than a spread length (Section 9.1). In general, up-
Pick the residual moveout profiles from the sem- dating velocity-depth models based on residual move-
blance spectra as shown in Figure 9.5-6 and combine out analysis of image gathers yields acceptable results
them with the depth horizons to create the residual for moderately complex structures associated with com-
movoeut section shown in Figure 9.5-5d. This section pressional and extensional tectonics. However, it may
gives an impression of how much residual moveout, thus not be suitable for complex overburden structures asso-
the range of errors in the initial velocity-depth model, ciated with overthrust or salt tectonics.
is present on image gathers as a function of depth and
distance along the line.
Figure 9.5-7a shows the interval velocity profiles Reflection Traveltime Tomography
before (as in Figure 9.5-1c) and after the residual move-
out update, and Figure 9.57b shows the updated depth Reflection traveltime tomography is based on perturb-
horizons displayed on top of the initial velocity-depth ing the initial model parameters by a small amount and
model as in Figure 9.5-2a. When combined, the updated then matching the change in traveltimes to the travel-
interval velocity profiles and depth horizons yield the time measurements made from residual moveout anal-
updated velocity-depth model shown in Figure 9.5-8a. ysis of image gathers (Sherwood et al., 1986; Kosloff et
Following the update, the new model needs to be al., 1996). A mathematical treatment of the subject is
checked for consistency with the input seismic data. given in Section J.6. Here, we remind ourselves of the
Figure 9.5-8b shows the image section derived from underlying assumptions, outline the theory, and exam-
prestack depth migration using the updated model in ine the performance of the method with model experi-
Figure 9.5-8a. Note that the depth horizons associated ments. We shall complement the discussion on tomog-
with the updated model, when superimposed on the raphy with a field data example.
image section, coincide with the reflectors that corre- We must do the best we can in building an accu-
spond to the layer boundaries. The modeled zero-offset rate earth model in depth (Section 9.4) so that only
reflection traveltimes using the updated model also are small changes remain to be made to the model by to-
in good agreement with the observed traveltimes of mography. Specifically, a tomographic update can be
the events on the unmigrated stacked section that are expected to work provided the changes to be made to
associated with the layer boundaries included in the the initial earth model parameters in terms of slownness
velocity-depth model of Figure 9.5-8a. and depths at layer boundaries are small compared to
To further verify the accuracy of the updated the model parameters themselves. Additionally, we shall
model, compare the residual moveout semblance spectra assume that the initial model is made up of horizontal
computed from the image gathers at selected locations layers with laterally invariant model parameters.
along the line after the model update (Figure 9.5-9) In the usual implementation of reflection traveltime
with those before the update (Figure 9.5-4). Note that tomography, the model parameters are perturbed while
most of the semblance peaks are now positioned along preserving the offset values of the seismic data (Sec-
the vertical axis of the semblance spectra that corre- tion J.6). The tomographic update ∆p to the model
sponds to zero residual moveout. The horizon-consistent parameters that comprise the changes in the slowness
residual moveout spectra after the model update are and depths to layer boundaries is given by the gener-
1470 Seismic Data Analysis

FIG. 9.5-7. Model updating by residual moveout corrections: (a) The interval velocity profiles as in Figure 9.5-1c before
(thin curves) and after (thick curves) residual moveout corrections, (b) the depth horizons in Figure 9.5-2a after an update
with the velocity-depth model before update as in the same figure shown in the background.
Earth Modeling in Depth 1471

FIG. 9.5-8. (a) The velocity-depth model in Figure 9.5-2a after model updating by residual moveout corrections, (b) prestack
depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated stacked section.
1472 Seismic Data Analysis

FIG. 9.5-9. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-8b and residual moveout
semblance spectra computed from these gathers.
Earth Modeling in Depth 1473

FIG. 9.5-10. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-9 along the depth
horizons shown in Figure 9.5-8b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-8a.
1474 Seismic Data Analysis

and depths to layer boundaries is given by the gener- (h) Update the paramater vector p + ∆p. Figure 9.5-
alized linear inversion (GLI) solution (equation J-88 in 11a shows the interval velocity profiles before and
Section J.6) after the tomographic update and Figure 9.5-11b
shows the updated depth horizons superimposed
∆p = (LT L)−1 LT ∆t, (9 − 10)
on the initial velocity-depth model of Figure 9.5-2a.
where ∆t denotes the column vector that represents In a tomographic update, you may wish to perturb
the residual moveout times measured from the image a subset of layer velocities and/or reflector geome-
gathers, L is a sparse matrix — its elements are in terms tries. This depends on your confidence in the initial
of the slowness and depth parameters associated with model parameters for the layers and the quality of
the initial model, and T denotes matrix transposition. the residual moveout profiles to be used in inver-
Consider the earth model with horizontal layers sion.
shown in Figure 9.1-1. We shall make an attempt to
update the initial estimate shown in Figure 9.1-5 using
reflection tomography. By combining the updated interval velocity pro-
files (Figure 9.5-11a) with the new depth horizons (Fig-
(a) Generate a set of image gathers (Figure 9.5-4) from ure 9.5-11b), we obtain the velocity-depth model af-
prestack depth migration (Figure 9.5-3) using an ter the first iteration of tomographic update (Figure
initial velocity-depth model (Figure 9.5-2a). 9.5-12a). Following the update, we check for consis-
(b) Convert the image gathers from depth to time us- tency of the new model with the input seismic data.
ing the interval velocity functions extracted from Overlay the depth horizons from the updated model
the initial velocity-depth model at the image- in Figure 9.5-12a onto the image section derived from
gather locations. prestack depth migration shown in Figure 9.5-12b and
(c) Compute the horizon-consistent residual moveout note that they coincide with the reflectors associated
for all offsets along events on image gathers that with the layer boundaries included in the model. Also,
correspond to the layer boundaries included in the the modeled zero-offset reflection traveltimes using the
model. Figure 9.5-6 shows the residual moveout updated model coincide with the observed traveltimes
spectra along the six horizons of the model in Fig- of the events on the unmigrated stacked section (Figure
ure 9.5-2a. The vertical axis represents the residual 9.5-12c) that are associated with the layer boundaries
moveout measured at a reference offset, usually the included in the velocity-depth model of Figure 9.5-12a.
maximum offset. The horizontal axis represents the Next, compute the residual moveout semblance
CMP locations along the line. Since residual move- spectra from the image gathers at selected locations
out can be either negative or positive, the vertical along the line after the model update (Figure 9.5-13)
axis is in both the positive and negative directions. and compare them with those before the update (Fig-
(d) Pick the residual moveout profiles for all the hori- ure 9.5-4). While most of the semblance peaks are now
zons by tracking the semblance peaks. Any depar- aligned with the vertical axis of the spectra, the to-
ture from the horizontal axis indicates a nonzero mographic update may be repeated to further remove
value for residual moveout. any remaining residual moveout errors. The horizon-
(e) Build the traveltime error vector ∆t using the consistent residual moveout semblance spectra after the
residual moveout times. As an example, you may first iteration of the tomographic model update are
have 10 layers, 1000 CMPs with a fold of 30. This shown in Figure 9.5-14. The residual moveout profiles
means that the length of the traveltime error vector that were picked from the spectra before the update
∆t is 300 000. (Figure 9.5-6) have been overlayed on the spectra af-
(f) Define the initial model by a set of slowness and ter the update to observe the extent of the removal of
depth parameters and construct the coefficient ma- residual moveout errors by the tomographic update.
trix L in equation (9-10) (Section J.6). As an ex- Proceed to a second iteration of tomographic up-
ample, you may have 10 layers and each layer may date by picking a new set of residual moveout profiles
be defined by 50 pairs of slowness and depth val- from the spectra shown in Figure 9.5-14. Then, follow
ues in the lateral direction. This means you would the updating procedure described above to obtain a
have 1000 parameters in your model space. It also new set of interval velocity profiles (Figure 9.5-15a) and
means that for the example given in step (b), you depth horizons (Figure 9.5-15b). Note that the changes
have 300 000 equations to solve for 1000 parame- in interval velocities and depth horizons that result from
ters. The solution for this overdetermined system the second iteration are smaller compared to those from
is given by equation (9-10). the first iteration (Figure 9.5-11).
(g) Estimate the change in parameters vector ∆p, by Combine the new set of interval velocity profiles
way of the GLI solution given by equation (9-10). and depth horizons to create the next update of the
Earth Modeling in Depth 1475

FIG. 9.5-11. First iteration of model updating by reflection tomography: (a) The interval velocity profiles as in Figure 9.5-1c
before (thin curves) and after (thick curves) tomographic updating, (b) the depth horizons in Figure 9.5-2a after the update,
with the velocity-depth model before the update as in Figure 9.5-2a shown in the background.
1476 Seismic Data Analysis

FIG. 9.5-12. (a) The velocity-depth model in Figure 9.5-2a after model updating by reflection tomography, (b) prestack
depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated stacked section.
Earth Modeling in Depth 1477

FIG. 9.5-13. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-12b and residual
moveout semblance spectra computed from these gathers.
1478 Seismic Data Analysis

FIG. 9.5-14. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-13 along the depth
horizons shown in Figure 9.5-12b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-12a.
Earth Modeling in Depth 1479

velocity-depth model shown in Figure 9.5-16a. Verify Limitations in Resolving Velocity-Depth


the updated model by performing prestack depth migra- Ambiguity by Tomography
tion to generate an image section (Figure 9.5-16b) and
zero-offset modeling of traveltimes associated with the Reflection traveltime tomography can be surprisingly
layer boundaries included in the model (Figure 9.5-16c). successful in updating an initial model with significant
Next, compute the residual moveout semblance spec- errors. Consider, for instance, the velocity-depth model
tra to examine any remaining moveout errors (Figure in Figure 9.5-25 with constant layer velocities. This
9.5-17). Finally, compute the horizon-consistent resid- model is indeed practically consistent with the input
ual moveout semblance spectra and overlay the residual seismic data as verified by the depth image in Figure
moveout profiles from the first iteration (Figure 9.5-18). 9.5-25b and the modeled zero-offset traveltimes over-
Note that the changes from the first to the second iter- layed on the unmigrated stacked section in Figure 9.5-
ation are marginal. At this point, you may wish to end 25c.
the iterations for tomographic update. Perform prestack depth migration using the initial
The extent to which an initial velocity-depth model model to derive the image section in Figure 9.5-26a and
is perturbed by a tomographic update depends on the compute the residual moveout semblance spectra of the
accuracy of that initial model, which in turn, depends selected image gathers in Figure 9.5-27. Note the signfi-
on how it has been estimated. Consider the two initial cant moveout errors manifested by the semblance peaks
velocity-depth models shown in Figures 9.4-27a and 9.5- which are off the zero moveout centerline on the spectra.
2a, which, for convenience, will be referred to as Models Now, compute the horizon-consistent residual moveout
1 and 2. Model 1 in Figure 9.4-27a was built by applying semblance spectra along the depth horizons shown in
the layer-by-layer inversion strategy with a combination Figure 9.5-26b and note the significant departures from
of coherency inversion and poststack depth migration the zero moveout line (Figure 9.5-28). Pick the resid-
to estimate the layer velocities and delineate the reflec- ual moveout profiles from these spectra and perform a
tor geometries. Model 2 in Figure 9.5-2a was built by tomographic update to obtain a new set of interval ve-
applying the time-to-depth conversion strategy with a locity profiles and depth horizons as shown in Figure
combination of Dix conversion to estimate the layer ve- 9.5-29.
locities and image-ray depth conversion to delineate the Combine the updated interval velocities and re-
reflector geometries. flector geometries to create the updated velocity-depth
As with the model in Figure 9.5-2a, perform model shown in Figure 9.5-30a. Note that this model is
prestack depth migration using Model 1 to obtain the fairly close to the updated models derived from the ap-
image section shown in Figure 9.5-19a. Then, compute plications of time-to-depth conversion (Figure 9.5-16a)
the residual moveout semblance spectra for selected im- and layer-by-layer inversion (Figure 9.5-23a) strategies.
age gathers (Figure 9.5-20). By comparing these spectra Model verification tests (Figures 9.5-30b,c) and residual
with those in Figure 9.5-4, it may not be obvious as to moveout semblance spectra (Figures 9.5-31 and 9.5-32)
whether Model 1 or Model 2 requires more perturba- demonstrate the feasibility and accuracy of the updated
tion. However, a comparison of the horizon-consistent model in Figure 9.5-30a.
residual moveout semblance spectra shown in Figures As demonstrated by the extreme case of an erro-
9.5-6 and 9.5-21 reveals that Model 1 gives rise to less neous initial model (Figure 9.5-25), a tomographic up-
residual moveout on image gathers than does Model date can steer the model toward an acceptable final
2. As such, the tomographic update of Model 1 yields model. How far, though, can we go with tomography to
changes in interval velocities and reflector geometries obtain a finely accurate model? In the following model
(Figure 9.5-22) that are smaller than those for Model 2 experiments, we shall examine this fundamentally im-
(Figure 9.5-7). portant question regarding tomography.
Irrespective of the strategy followed to derive the Figure 9.5-33 shows horizon-consistent residual
initial model, the results of model updating need to be moveout semblance spectra derived from image gath-
verified for consistency with the input seismic data (Fig- ers that were created by prestack depth migration of
ure 9.5-23) and examined for any remaining residual the synthetic data as in Figure 9.1-2a using the initial
moveouts (Figure 9.5-24) to decide whether or not to velocity-depth model shown in Figure 9.1-5b. Shown in
continue with the iterations of tomographic update. Figure 9.5-34 are the results of tomographic updating
(text continues on p. 1494)
1480 Seismic Data Analysis

FIG. 9.5-15. Second iteration of model updating by reflection tomography: (a) The interval velocity profiles as in Figure
9.5-11a before (thin curves) and after (thick curves) tomographic updating, (b) the depth horizons in Figure 9.5-12a after
update, with the velocity-depth model before the update as in Figure 9.5-12a shown in the background.
Earth Modeling in Depth 1481

FIG. 9.5-16. (a) The velocity-depth model in Figure 9.5-12a after model updating by reflection tomography, (b) prestack
depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated stacked section.
1482 Seismic Data Analysis

FIG. 9.5-17. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-16b and residual
moveout semblance spectra computed from these gathers.
Earth Modeling in Depth 1483

FIG. 9.5-18. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-17 along the depth
horizons shown in Figure 9.5-16b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-16a.
1484 Seismic Data Analysis

FIG. 9.5-19. (a) Prestack depth migration using the velocity-depth model in Figure 9.4-27a, (b) the same image section
with depth horizons as in Figure 9.4-27a.
Earth Modeling in Depth 1485

FIG. 9.5-20. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-19b and residual
moveout semblance spectra computed from these gathers.
1486 Seismic Data Analysis

FIG. 9.5-21. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-20 along the depth
horizons shown in Figure 9.5-19b. The velocity-depth model used in prestack depth migration is shown in Figure 9.4-27a.
Earth Modeling in Depth 1487

FIG. 9.5-22. Model updating by reflection tomography: (a) The interval velocity profiles as in Figure 9.4-22 before (thin
curves) and after (thick curves) tomographic update, (b) the depth horizons in Figure 9.4-27a after, with the velocity-depth
model before the update as in Figure 9.5-27a shown in the background.
1488 Seismic Data Analysis

FIG. 9.5-23. (a) The velocity-depth model in Figure 9.4-27a after model updating by reflection tomography, (b) prestack
depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated stacked section.
Earth Modeling in Depth 1489

FIG. 9.5-24. Residual moveout semblance spectra computed from the image gathers associated with the prestack depth-
migrated section in Figure 9.5-23b along the depth horizons shown in the same figure. The velocity-depth model used in
prestack depth migration is shown in Figure 9.5-23a.
1490 Seismic Data Analysis

FIG. 9.5-25. (a) A velocity-depth model created by iterative poststack depth migration using constant layer velocities, (b)
poststack depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated
stacked section.
Earth Modeling in Depth 1491

FIG. 9.5-26. (a) Prestack depth migration using the velocity-depth model in Figure 9.5-25a, (b) the same image section
with depth horizons as in Figure 9.5-25a.
1492 Seismic Data Analysis

FIG. 9.5-27. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-26b and residual
moveout semblance spectra computed from these gathers.
Earth Modeling in Depth 1493

FIG. 9.5-28. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-27 along the depth
horizons shown in Figure 9.5-26b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-26a.
1494 Seismic Data Analysis

— the interval velocity profiles before (thin curves) and The fact that the tomographic update for one layer
after (thick curves) the tomographic update, and the is influenced by the errors in the parameters for the
updated velocity-depth model itself. Compare with the layers above is best illustrated by the model in Figure
initial velocity-depth model shown in Figure 9.1-5b and 9.1-18. This is the model with horizontal layers (Ta-
the true velocity-depth model shown in Figure 9.1-1b. ble 9-2) but with a near-surface layer H1 with laterally
Compare with the true model (Figure 9.1-1) and the varying velocities between 800 m/s and 1500 m/s. Re-
initial model (Figure 9.1-5) and note that tomographic sults of earth model estimation using Dix conversion
update has produced a result fairly close to the true are shown in Figure 9.1-21c. Consider this model as
model. A close-up of the center portions of the initial the initial velocity-depth model and perform prestack
model, the final model, and the true model after the depth migration of the data associated with the stacked
tomographic update is shown in Figure 9.5-35. section in Figure 9.1-19a, generate image gathers, and
Iterate the updating process until the residual compute the residual moveout semblance spectra shown
moveout for all the layers is minimized. Specifically, per- in Figure 9.5-41. Pick the spectra to obtain the resid-
form prestack depth migration using the tomographi- ual moveout profiles by tracking the semblance peaks,
cally updated velocity-depth model in Figure 9.5-34 and then use these profiles in a tomographic update to get
generate a new set of image gathers. Compute a new set the new model shown in Figure 9.5-42. Use the result-
of residual moveout semblance spectra from the image ing model as input to a second iteration for model up-
gathers as shown in Figure 9.5-36. Compare with the dating. Compare the resulting residual moveout spectra
residual moveout spectra in Figure 9.5-33 before the to- shown in Figure 9.5-43 with those from the first itera-
mographic update and note that the residual moveout tion shown in Figure 9.5-41 and note that the second it-
after the update has been reduced significantly. eration has reduced the residual moveout, significantly.
We now test the tomographic update in the pres- The updated model from the second iteration is shown
ence of large errors in the initial velocity-depth model. in Figure 9.5-44. A close-up of the central portions of the
Consider the velocity-depth model shown in Figure 9.1-
intermediate velocity-depth model (Figure 9.5-42b), the
1 based on the parameters listed in Table 9-2. We shall
velocity-depth model updated for the second time (Fig-
deliberately introduce errors to this model by setting
ure 9.5-44b), and the true velocity-depth model (Figure
the velocities for layer H3 to a constant value of 2550
9.1-18b) is shown in Figure 9.5-45. Compare these es-
m/s and for layer H4 to a constant value of 3250 m/s.
timates with the close-up view (Figure 9.1-22b) of the
The erronenous velocity-depth model is shown in Figure
initial velocity-depth model shown in Figure 9.1-21c.
9.5-37.
The lesson to learn from this experiment is not to start
By adopting the distorted model in Figure 9.5-37
with an initial model (Figure 9.1-21c) with large errors
as the initial velocity-depth model, perform prestack
depth migration, generate image gathers and compute and expect a tomographic update to correct for these
residual moveout semblance spectra as shown in Figure errors.
9.5-38. Note the large moveouts caused by the velocity Finally, we shall update the model shown in Fig-
errors associated with layers H3 and H4. The tomo- ure 9.1-24 which was estimated by coherency inversion.
graphic update shown in Figure 9.5-39a has modified Consider this model as the initial velocity-depth model
the velocities for layers H3 and H4, and for the underly- and perform prestack depth migration, generate im-
ing layers. Specifically, the constant-velocities assigned age gathers, and compute the residual moveout spec-
to layers H3 and H4 have been replaced with laterally tra shown in Figure 9.5-46. Pick the spectra, again,
varying velocity profiles that are consistent with the ac- by tracking the semblance peaks to obtain the resid-
tual velocity profiles shown in Figure 9.1-1. Neverthe- ual moveout profiles. Then use these profiles in a tomo-
less, compare the profiles in Figures 9.1-1a and 9.5-39a graphic update to get the new model shown in Fig-
and note that the velocity estimate for the underlying ure 9.5-47. A close-up of the central portions of the
layer H5 has been adversely influenced by the tomo- initial velocity-depth model (Figure 9.1-24b), the up-
graphic update applied to the layers above. dated velocity-depth model (Figure 9.5-47b), and the
A close-up of the central portions of true velocity- true velocity-depth model (Figure 9.1-18b) is shown in
depth model (Figure 9.1-1b), the initial velocity-depth Figure 9.5-48.
model (Figure 9.5-37b), and the updated velocity-depth Based on the results of the experiments using Dix
model (Figure 9.5-39b) is shown in Figure 9.5-40. Al- conversion and coherency inversion applied to the model
though the errors in the initial velocity-depth model for with a near-surface layer that has lateral velocity varia-
the shallow layers have been reduced, note that, in gen- tions (Figure 9.1-18), again, the lesson to learn is that,
eral, starting with an initial velocity-depth model with whatever the strategy used in the model building, if you
large errors, tomographic updating does not always pro- start with an initial model with large errors, do not ex-
duce satisfactory results. pect tomographic updating to correct for these errors.
(text continues on p. 1512)
Earth Modeling in Depth 1495

FIG. 9.5-29. Model updating by reflection tomography: (a) The interval velocity profiles before (thin curves) and after (thick
curves) tomographic update, (b) the depth horizons in Figure 9.5-25a after update, with the velocity-depth model before the
update as in Figure 9.5-25a shown in the background.
1496 Seismic Data Analysis

FIG. 9.5-30. (a) The velocity-depth model in Figure 9.5-25a after model updating by reflection tomography, (b) prestack
depth migration with depth horizons as in (a), (c) modeled zero-offset traveltimes overlayed on the unmigrated stacked section.
Earth Modeling in Depth 1497

FIG. 9.5-31. Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-30b and residual
moveout semblance spectra computed from these gathers.
1498 Seismic Data Analysis

FIG. 9.5-32. Residual moveout semblance spectra computed from the image gathers as in Figure 9.5-31 along the depth
horizons shown in Figure 9.5-30b. The velocity-depth model used in prestack depth migration is shown in Figure 9.5-30a.
Earth Modeling in Depth 1499

FIG. 9.5-33. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-2a using the initial velocity-depth model shown in Figure 9.1-5b.
1500 Seismic Data Analysis

FIG. 9.5-34. (a) The interval velocity profiles before (thin curves) and after (thick curves) tomographic update; (b) the
updated velocity-depth model. Compare with the initial velocity-depth model shown in Figure 9.1-5b and the true velocity-
depth model shown in Figure 9.1-1b.
Earth Modeling in Depth 1501

FIG. 9.5-35. Central portions of (a) the estimated initial velocity-depth model shown in Figure 9.1-5b using the smoothed
interval velocity profiles shown in Figure 9.1-5a, (b) the updated velocity-depth model shown in Figure 9.5-34b, and (c) the
true velocity-depth model shown in Figure 9.1-1b.
1502 Seismic Data Analysis

FIG. 9.5-36. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-2a using the updated velocity-depth model shown in Figure 9.5-34b.
Earth Modeling in Depth 1503

FIG. 9.5-37. (a) The velocity profiles as in Figure 9.1-1a except for horizons H3 and H4 which deliberately have been assigned
wrong velocities; (b) the velocity-depth model created from the interval velocity profiles in (a).
1504 Seismic Data Analysis

FIG. 9.5-38. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-2a using the initial velocity-depth model shown in Figure 9.5-37b.
Earth Modeling in Depth 1505

FIG. 9.5-39. (a) The interval velocity profiles before (thin curves) and after (thick curves) tomographic update; (b) the
updated velocity-depth model. Compare with the initial velocity-depth model shown in Figure 9.5-37b and the true velocity-
depth model shown in Figure 9.1-1b.
1506 Seismic Data Analysis

FIG. 9.5-40. Central portions of (a) the true velocity-depth model shown in Figure 9.1-1b, (b) the initial velocity-depth
model shown in Figure 9.5-37b, and (c) the updated velocity-depth model shown in Figure 9.5-39b.
Earth Modeling in Depth 1507

FIG. 9.5-41. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-19a using the initial velocity-depth model shown in Figure 9.1-21c.
1508 Seismic Data Analysis

FIG. 9.5-42. (a) The interval velocity profiles before (thin curves) and after (thick curves) the first tomographic update;
(b) the updated velocity-depth model. Compare with the initial velocity-depth model shown in Figure 9.1-21c and the true
velocity-depth model shown in Figure 9.1-18b.
Earth Modeling in Depth 1509

FIG. 9.5-43. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-19a using the updated velocity-depth model shown in Figure 9.5-42b.
1510 Seismic Data Analysis

FIG. 9.5-44. (a) The interval velocity profiles before (thin curves) and after (thick curves) the second tomographic update;
(b) the velocity-depth model updated for the second time. Compare with the initial velocity-depth model shown in Figure
9.1-21c, the intermediate velocity-depth model shown in Figure 9.5-42b, and the true velocity-depth model shown in Figure
9.1-18b.
Earth Modeling in Depth 1511

FIG.9.5-45. Central portions of (a) the intermediate velocity-depth model shown in Figure 9.5-42b, (b) the velocity-depth
model updated for the second time shown in Figure 9.5-44b, and (c) the true velocity-depth model shown in Figure 9.1-18b.
1512 Seismic Data Analysis

We now examine reflection traveltime tomography There is of course some mid-range of depths and
for its ability to update an initial model in the presence velocities for which residual moveout semblance spec-
of a complex overburden. Figure 9.5-49a shows the rms tra can be used for model updating. In the present ex-
velocity profiles along the time horizons in Figure 9.4- ample, residual moveout semblance spectra associated
28a. Dix conversion yields the interval velocity profiles with horizons H5 to H8 (as denoted in Figure 9.5-50a)
shown in Figure 9.5-49b. Admittedly, some editing and were computed and picked as input to the reflection
smoothing before and after Dix conversion were done to traveltime tomography solver (Section J.6). When pick-
tame the interval velocities to behave in a structurally ing the residual moveout profiles from the semblance
sensible manner. spectra in Figure 9.5-52, keep in mind not to track all
By using the interval velocity profiles shown in the rapid fluctuations that are much less than a cable
Figure 9.5-49b, the time horizons interpreted from the length. These are the manifestations of the lateral veloc-
time-migrated section shown in Figure 9.4-28a were ity variations in the layers above as was demonstrated
converted to depth by way of image rays. Then, these by the experiments conducted using synthetic data in
depth horizons were combined with the interval veloc- Section 9.1.
ity profiles (Figure 9.5-49b) to create an initial velocity- The workflow for tomographic updating —
depth model. This was followed by depth migration of prestack depth migration to generate image gathers,
the stacked section in Figure 9.4-28b using the initial computing residual moveout along events associated
velocity-depth model, which in turn, was revised based with the layer boundaries, picking residual moveout pro-
on the interpretation of the depth-migrated section. A files and using them as input to the tomography to
second iteration of poststack depth migration (Figure derive the updates for layer velocities (Figure 9.5-53a)
9.5-50a) produced a velocity-depth model (Figure 9.5- and reflector geometries (Figure 9.5-53b), were repeated
49c) that is consistent with the reflection times observed twice to produce the final model and image shown Fig-
on the unmigrated stacked section (Figure 9.5-50b). ure 9.5-54. In each iteration, the first four layers were
Actually, the now consistent velocity-depth model kept unchanged because of poor estimates for resid-
in Figure 9.5-49c is what we shall update by way of re- ual moveout along the horizons that correspond to the
flection tomography. To begin with, perform prestack base of these layers (H1 through H4). Compare the im-
depth migration and generate image gathers as in Fig- ages from prestack depth migrations shown in Figures
ure 9.5-51. Compute the residual moveout semblance 9.5-52a and 9.5-54a obtained from the initial and final
spectra and observe that there are semblance peaks that velocity-depth models of Figures 9.5-52b and 9.5-54b,
are off the center line, indicating the presence of residual respectively, and note that the differences are marginal.
moveout. Stacking of the image gathers yields the image Iterative tomographic updating, as demonstrated
section from prestack depth migration shown in Figure by the model experiments in this section, does not nec-
9.5-52. The significant improvement in imaging the di- essarily result in convergence of an initial model to a
apiric structure and the base-salt reflector (horizon H8 final model with zero residual moveout. Instead, the so-
as in Figure 9.5-50a) becomes obvious when this section lution may just wobble and never converge. This sce-
is compared with the poststack depth-migrated section nario is especially likely for cases of complex overbur-
in Figure 9.5-50a. Nevertheless, the fact that the sem- den structures as in the present example. In such cases,
blance spectra (Figure 9.5-51) exhibit residual moveout additional information; such as well control, check-shot
along the events associated with the layer boundaries velocities, and geologic constraints, is required to derive
included in the velocity-depth model suggests that the an acceptable final model.
image from prestack depth migration may be improved
by updating the initial velocity-depth model.
Now here is a dilemma that we have to live with Turning-Ray Tomography
when updating velocity-depth models. As a result of
the inevitable muting of the image gathers (Figure 9.5- Rapid discharge of the sediments by the Mississippi
51), there is often insufficient offset range to accurately River has given rise to an accummulation of gas-charged
measure residual moveouts for shallow horizons. Be- mudflows within complex, meandering channels down
cause of the very poor residual moveout spectra, shallow the slopes of its delta. Velocities within these mudflows
horizons can be unreliable in model updating. Errors are significantly lower than the surrounding deltaic sed-
in the model parameters for the shallow layers will of iments and can be as low as 300 m/s. The underlying
course adversely affect the estimation and updating of predeltaic Holocene sequence also is characterized by
the model parameters for the deeper layers. At greater sediments with lateral velocity variations (Coleman et
depths, although there is sufficient cable length, faster al., 1980; May et al., 1988).
velocities prevent recognition of velocity or depth errors Because of the absence of a near-surface refractor
with adequate resolution. with a strong velocity contrast, first arrivals on shot
(text continues on p. 1524)
Earth Modeling in Depth 1513

FIG. 9.5-46. Horizon-consistent residual moveout semblance spectra derived from image gathers that were created by prestack
depth migration of the synthetic data as in Figure 9.1-19a using the initial velocity-depth model shown in Figure 9.1-24b.
1514 Seismic Data Analysis

FIG. 9.5-47. (a) The interval velocity profiles before (thin curves) and after (thick curves) the first tomographic update;
(b) the updated velocity-depth model. Compare with the initial velocity-depth model shown in Figure 9.1-24b and the true
velocity-depth model shown in Figure 9.1-18b.
Earth Modeling in Depth 1515

FIG. 9.5-48. Central portions of (a) the initial velocity-depth model shown in Figure 9.1-24b using the interval velocity
profiles shown in Figure 9.1-24a, (b) the updated velocity-depth model shown in Figure 9.5-47b using the interval velocity
profiles shown in Figure 9.5-47a, and (c) the true velocity-depth model shown in Figure 9.1-18b.
1516 Seismic Data Analysis

FIG. 9.5-49. (a) RMS velocity profiles associated with the data shown in Figure 9.4-28a, (b) the interval velocity profiles
computed from (a) by way of Dix conversion, and (c) an initial velocity-depth model derived from the combination of the
interval velocity profiles in (b) with the depth horizons that were created by image-ray depth conversion of the time horizons
shown in Figure 9.4-28a.
Earth Modeling in Depth 1517

FIG. 9.5-50. (a) Poststack depth migration of the stacked section in (b) using the initial velocity-depth model in Figure
9.4-49c, (b) zero-offset traveltimes modeled from the depth horizons in (a) and superimposed on the stacked section.
1518 Seismic Data Analysis

FIG. 9.5-51. Part 1: Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-52 and residual
moveout semblance spectra computed from these gathers.
Earth Modeling in Depth 1519

FIG. 9.5-51. Part 2: Selected image gathers associated with the prestack depth-migrated section in Figure 9.5-52 and residual
moveout semblance spectra computed from these gathers.
1520 Seismic Data Analysis

FIG. 9.5-52. Prestack depth migration of the data associated with the stacked section shown in Figure 9.5-50b using the
initial velocity-depth model shown in Figure 9.5-49c, and (b) the residual moveout semblance spectra computed from the
image gathers associated with the prestack depth-migrated section in (a) along Horizons H5-H8 labeled as in Figure 9.5-50a.
Earth Modeling in Depth 1521

FIG. 9.5-53. (a) The interval velocity profiles before (thin curves) and after (thick curves) the second tomographic update,
and (b) the depth horizons after the update superimposed on the intermediate velocity-depth model derived from the first
iteration of tomographic update.
1522 Seismic Data Analysis

FIG. 9.5-54. (a) The final velocity-depth model after the second tomographic update, and (b) prestack depth migration of
the data associated with the stacked section shown in Figure 9.5-50b using the final velocity-depth model in (a).
Earth Modeling in Depth 1523

FIG. 9.5-55. (a) A map of statics derived from turning-ray tomography applied to a 3-D seismic data set, (b) an inline stack
along the traverse indicated by the horizontal line in (a) with refraction statics corrections, and (c) the same inline stack with
turning-ray tomographic statics corrections. (Kim and Bell, 2000; courtesy Geosignal, Western Geophysical.)
1524 Seismic Data Analysis

gathers recorded over the Mississippi Delta often do not (f) Iterate steps (a) through (e) as necessary to mini-
represent the typical refracted arrivals associated with mize the discrepancy between the modeled and ac-
a head wave. Therefore, neither refraction traveltime tual first-arrival times.
tomography (Section C.9) is applicable to estimate a
near-surface model nor refraction statics (Section 3.4) The final velocity-depth model resulting from the
may be a rigorous solution to resolving the near-surface iterative application of turning-ray tomograpy is then
complexity in the Mississippi Delta. used to compute the one-way traveltimes through the
Instead, the first arrivals are often associated near-surface model along vertical raypaths. These are
with diving waves through the deltaic and predeltaic then used to apply the necessary source and receiver
Holocene sediments (Zhou et al., 1992). Because of the statics corrections to the prestack data. Figure 9.5-
unusual velocity gradients within the near-surface lay- 55a shows the statics solution derived from turning-
ers, the downgoing incident wave rapidly turns around ray tomography applied to a 3-D offshore seismic data
before being reflected and is recorded by the receiver as set from the Mississippi Delta (Kim and Bell, 2000).
the first arrival. Note the mudflows characterized by the strings of neg-
Just as reflection traveltime tomography (Section ative statics shifts. Figure 9.5-55b shows a stacked sec-
J.6) can be used to update an initial estimate of a sub- tion along an inline traverse with refraction statics cor-
surface velocity-depth model, turning-ray tomograpy rections (Section 3.4), and Figure 9.5-55c shows the
may be used to update an initial estimate of a near- stacked section along the same traverse with statics
surface velocity-depth model (Zhou et al., 1992; Bell et corrections as in (a) based on turning-ray tomography.
al., 1994). Note the singnificant improvement of event continuity
in the central part of the section.
(a) Begin with the picking of the first arrivals that rep-
resent the diving waves through the near surface.
(b) Define an initial velocity-depth model by a set EXERCISES
of near-surface layers with constant velocities and
thicknesses, and model the first-arrival times by ray Exercise 9-1. Consider the hypothetical case of a
tracing. pure strike-slip fault with the same horizontally situated
(c) Compute the difference between the modeled and formation on both sides. Can such an earth model be
observed first-arrival times. identified by seismic imaging?
(d) Estimate the change in parameters vector ∆p of Exercise 9-2. There can be a multiple number
equaiton (9-10), by way of the GLI solution given of velocity-depth models such as shown in Figure 9.0-
by equation (J-88) by perturbing the velocities of 1 consistent with a single zero-offset traveltime section
the near-surface layers only. (Figure 9.0-2). Can these velocity-depth models also be
(e) Update the paramater vector p + ∆p to obtain a consistent with a single nonzero-offset traveltime sec-
new near-surface velocity-depth model. tion?
Appendix J
DATA MODELING BY INVERSION

J.1 The Generalized Linear Inversion

We shall start by making the following formal statement:

Given an observed data set d, estimate a set of parameters p which are used to construct a
model d of the observed data set d, such that the difference between the observed data set d and
the modeled data set d is minimum based on a specific norm.

For most geophysical applications, the norm for minimization usually is chosen to be L2 , in which
case the error energy in data modeling is made minimum in the least-squares sense. Another
norm that is appropriate in some applications is L1 , in which case the magnitude of the error
in data modeling is made minimum.
To form a mathematical framework for the objective stated above, we need a model equation
that relates the modeled data with the model parameters to be estimated as
d = Lp, (J − 1)
where d is the modeled data vector, p is the model parameter vector, and L is the matrix that
relates the modeled data vector to the model parameter vector.
The error vector e is defined as the difference between the modeled data vector and the
observed data vector
e=d−d. (J − 2)
Substitute equation (J-1) into equation (J-2) to obtain
e = d − Lp. (J − 3)
Following Lines and Treitel (1984), the least-squares solution for equation (J-3) can be
determined. First, the cumulative squared error S is expressed as
S = eT e, (J − 4a)
where T is for transpose. By substituting for e from equation (J-3), we get
S = (d − Lp)T (d − Lp). (J − 4b)
Expand the right-hand side to get
S = dT d − pT LT d − dT Lp + pT LT Lp. (J − 4c)
Now differentiate both sides of equation (J-4c) with respect to p and observe the requirement
for least-squares minimization that ∂S/∂p = 0
−dT L + pT LT L = 0. (J − 5a)
Apply matrix transpose and rearrange the terms
(LT L)p = LT d, (J − 5b)
1526 Seismic Data Analysis

which yields the desired least-squares solution


p = (LT L)−1 LT d, (J − 6a)
where LT L is the covariance matrix and (LT L)−1 LT is the least-squares (also called generalized
linear) inverse of L.
Equation (J-6a) represents the generalized linear inverse (GLI) solution to the parameter
vector p. This solution is widely used in many stages of seismic data analysis. Examples include
deconvolution (Section 2.3), residual statics corrections (Section 3.3), refraction statics correc-
tions (Section 3.4), and the discrete Radon transform (Section 6.4). In most applications, the
solution needs to be constrained since the covariance matrix LT L needs to be inverted. The
constrained solution is given by
p = (LT L + βI)−1 LT d, (J − 6b)
where β is called the damping factor and I is the identity matrix.
In some applications, the generalized linear inverse problem is formulated in the frequency
domain. Then, the unconstrained solution is given by
p = (LT∗ L)−1 LT∗ d, (J − 7a)
and the constrained solution is given by
p = (LT∗ L + βI)−1 LT∗ d, (J − 7b)
where the asterisk denotes complex conjugate, and p, d, and L are complex.
The computationally suitable technique to solve for the parameter vector p is dictated by
the properties of the coefficient matrix L; that in turn is influenced by how the parameter vector
p is defined. The coefficient matrix L may be sparse with many of its elements being zero as
for the residual and refraction statics. It may be near singular with the elements of two rows
being very similar in size as for the discrete Radon transform. The covariance matrix LT L may
have the Toeplitz structure with its elements being diagonally symmetric as for deconvolution
and the discrete Radon transform. It may also be diagonally dominant as for deconvolution.
In geophysical applications, techniques to solve for the parameter vector p in equations (J-
6a,b) or (J-7a,b) include Levinson recursion, conjugate gradient, Gauss-Seidel and singular-value
decomposition.

J.2 The GLI Formalism of Deconvolution

Deconvolution is fundamentally a data modeling technique. Specifically, we model a 1-D seismo-


gram for a minimum-phase estimate of the source wavelet, to predict multiples, and ultimately
obtain an estimate of white reflectivity series (Section 2.3). The mathematical foundation of
deconvolution is provided in Appendix B. Here, we shall apply the theory of the generalized
linear inversion outlined in the previous section to describe deconvolution within the framework
of 1-D seismic inversion. Inversion techniques at higher dimensions will be summarized in the
next section.
We shall consider designing a least-squares inverse filter f (t) that converts a wavelet w(t) to
a desired form d(t) such that the difference e(t) between the actual output y(t) and the desired
output d(t) is minimum in the least-squares sense. The zero-delay unit spike is a special case
of the desired output d(t). Other forms of d(t) can also be considered, such as a zero-phase
band-limited wavelet. The model equation for deconvolution is given by
y(t) = w(t) ∗ f (t) (J − 8)
Consider the discrete form of equation (J-8), with w(t) represented by the m−length time se-
ries (w0 , w1 , w2 , ..., wm−1 ), and f (t) represented by the n−length time series (f0 , f1 , f2 , ..., fn−1 ).
Data Modeling by Inversion 1527

Equation (J-8) can then be expressed in matrix form


 
w0 0 ... 0
 w1 w0 
    w2 w1 w0

  
y0  . f0
 .. 
 y1   .
 
. w2 w1 . 
 

f1 

 y2  =  . ..   f2 . (J − 9)
   wm−1 .. w2 .   
 ..   

  .. 
.  ..  .
ym+n−1  0 wm−1 .  fn−1
 . 
 . 
. w m−1
0 wm−1
This is the equation of complete transient convolution. Define the output vector on the left-hand
side by y, the coefficient matrix on the right-hand side by L, and the filter vector by f . Equation
(J-9) takes the compact form
y = Lf . (J − 10)
Follow the steps from equations (J-2) through (J-6) to obtain the least-squares solution for the
filter vector f :
(LT L) f = LT d. (J − 11)
Consider the special case of a three-point wavelet (w0 , w1 , w2 ). Set up the L matrix of
equation (J-9) for this special case
 
w0 0 0
 w1 w0 0 
 
L =  w2 w1 w0  , (J − 12a)
 
0 w2 w1
0 0 w2
with its transpose LT given by
 
w0 w1 w2 0 0
LT =  0 w0 w1 w2 0 . (J − 12b)
0 0 w0 w1 w2
Now multiply the two matrices to get

 
w02 + w12 + w22 w1 w0 + w2 w1 w2 w0
T 
L L = w0 w1 + w1 w2 w02 + w12 + w22 w1 w0 + w2 w1  . (J − 12c)
w0 w2 w0 w1 + w1 w2 w02 + w12 + w22
Compute the first three lags of the autocorrelation (r0 , r1 , r2 ) of the wavelet (w0 , w1 , w2 ):
r0 = w02 + w12 + w22
r1 = w0 w1 + w1 w2 (J − 13)
r2 = w0 w2 ,

and note that the elements of the covariance matrix LT L given by equation (J-12c) are the first
three autocorrelation lags of the wavelet (w0 , w1 , w2 ) given by equations (J-13). Thus, for the
general case, we note that
 r r r ... r 
0 1 2 n−1
 r1 r0 r1 . . . rn−2 
 
LT L =  r2 r1 r0 . . . rn−3  , (J − 14)
 . .. .. .. .. 
 . . 
. . . .
rn−1 rn−2 rn−3 ... r0
1528 Seismic Data Analysis

where (r0 , r1 , r2 , ..., rn−1 ) are the first n autocorrelation lags of the input wavelet series
(w0 , w1 , w2 , ..., wm−1 ).
For the special case of a desired output vector d

 
d0
 d1 
 
d =  d2  , (J − 15a)
 
d3
d4
obtain the matrix product LT d
 
w0 d0 + w1 d1 + w2 d2
LT d =  w0 d1 + w1 d2 + w2 d3  . (J − 15b)
w0 d2 + w1 d3 + w2 d4
Now compute the first three lags of the crosscorrelation (g0 , g1 , g2 ) of the desired output
(d0 , d1 , d2 , d3 , d4 ) with the input wavelet (w0 , w1 , w2 )

g0 = w0 d0 + w1 d1 + w2 d2 , (J − 16a)
g1 = w0 d1 + w1 d2 + w2 d3 , (J − 16b)
g2 = w0 d2 + w1 d3 + w2 d4 , (J − 16c)
and note that the elements of the matrix LT d given by equation (J-15b) are the first three lags of
the crosscorrelation of the desired output (d0 , d1 , d2 , d3 , d4 ) with the input wavelet (w0 , w1 , w2 )
given by equations (J-16a,b,c). Thus, for the general case, we note that
 
g0
 g1 
 
L d=
T g 
 .2  , (J − 17)
 . 
.
gn−1
where (g0 , g1 , g2 , ..., gn−1 ) are the first n lags of the crosscorrelation of the desired output d with
the input wavelet w.
By substituting equations (J-14) and (J-17) into equation (J-11), we get

 r r1 r2 . . . rn−1     
0 f0 g0
 r1 r0 r1 . . . rn−2   f1   g1 
     
 r2 r1 r0 . . . rn−3   f2 = g2 . (J − 18)
 . .. .. .. ..   ..   .. 
 . .     
. . . . . .
rn−1 rn−2 rn−3 ... r0 fn−1 gn−1
This is the discrete form of the classic Wiener-Hopf integral equation to estimate the least-
squares shaping filter f that converts an input wavelet w into a desired form d.
In general, the autocorrelation matrix LT L in equation (J-11) is diagonally dominant since
r0 is the largest value of the autocorrelation lags (r0 , r1 , r2 , ..., rn−1 ). Nevertheless, practical
implementations of equation (J-11) often require adding a small fraction of the zero-lag of the
autocorrelation to the diagonal elements of the matrix LT L:

(LT L + I) f = LT d, (J − 19)
where is called the prewhitening factor and I is the identity matrix.
Note that the autocorrelation matrix LT L given by equation (J-14) is of Toeplitz form. A
typical length for the deconvolution filter f in equation (J-18) is between 40 and 80 samples.
Data Modeling by Inversion 1529

This makes the size of the autocorrelation matrix LT L to be between 40 × 40 and 80 × 80. If the
autocorrelogram is computed from an input seismogram represented by a single trace, then the
length of the input data vector is typically 1000 samples. The ratio of the length of the input
seismogram used in computing the autocorrelation lags to the filter length should be no less
than 8.
Equation (J-11) may be solved for the filter vector f using Levinson recursion (Section B.6).
It may also be solved by the conjugate gradient method (Wang and Treitel, 1973; Koehler and
Taner, 1985; Claerbout, 1992). The conjugate gradient method by Hestenes (1956) described
and exemplified by Wang and Treitel (1973) is outlined below.
Refer to equation (J-11) and define

A = LT L (J − 20a)
and
B = LT d, (J − 20b)
so that
Af = B. (J − 21)
Our goal is to estimate the filter vector f recursively starting with an initial estimate X0
which may be defined as a null vector. The recursive estimate is terminated when the residual
vector Ri defined by

Ri = B − AXi (J − 22)
becomes a null vector itself. Here, Xi is the parameter vector estimate after the ith iteration. If
the autocorrelation matrix A has dimensions n × n, the conjugate gradient method yields the
solution f after m < n iterations.
The recursive scheme requires the following initial values

c−1 = 1, (J − 23a)

P−1 = 0, (J − 23b)
and
R0 = B − A X0 . (J − 23c)
The recursive scheme computes the energy of the residual after the ith iteration

ci = RTi Ri , i = 0, 1, 2, · · · , m. (J − 24a)
To begin the recursion, set i = 0 and compute the residual error c0 using equation (J-24a).
Then, insert the values for c−1 from equation (J-23a) and c0 from equation (J-24a) into (Wang
and Treitel, 1973)

bi−1 = ci /ci−1 (J − 24b)


to get the value for b−1 . Next insert the values for P−1 from equation (J-23b), R0 from equation
(J-23c), and b−1 from equation (J-24b) into

Pi = Ri + bi−1 Pi−1 (J − 24c)


1530 Seismic Data Analysis

to get a value for P0 . Then, apply the recursions given by

Qi = A Pi , m ≤ n, (J − 24d)

di = PTi Qi , (J − 24e)

and

ai = ci /di (J − 24f )

to get an estimate X1 of the filter vector f given by

Xi+1 = Xi + ai Pi (J − 24g)

and the associated new residual vector R1 given by

Ri+1 = Ri − ai Qi . (J − 24h)

Return to the beginning of the recursion and repeat the computations given by equations (J-24a)
through (J-24h). Stop the iteration after a specified value of m < n.
I have tested this algorithm and compared the results with those obtained from Levinson
recursion. For a filter length n, conjugate gradient gives the exact solution derived from Levin-
son recursion after m = n iterations. Nevertheless, an approximate solution that is acceptable
without much compromise can be achieved by m < n. This means that for a large filter length
n, conjugate gradient can be more efficient than Levinson recursion.

J.3 Applications of the GLI Technique

In this section, we shall review selected applications of generalized linear inversion (GLI) for
data modeling as part of a seismic data processing sequence. Specifically, we shall compile a
summary of the theoretical discussions in previous appendixes on deconvolution (Sections B.5
and B.8), residual statics corrections (Section C.4), refraction statics corrections (Section C.8),
and the discrete Radon transform (Section F.3) within the context of the GLI solution given by
equations (J-6a,b) or (J-7a,b).
For each application, we shall make reference to the model equation, input data vector d,
the coefficient matrix L, the parameter vector p and their specific sizes, and the technique to
solve for the parameter vector.
The model equation for surface-consistent deconvolution expressed in the frequency domain
is given by (Section B.8):

X̃ij (ω) = S̃j (ω) + H̃l (ω) + Ẽk (ω) + G̃i (ω), (J − 25)

where X̃ij (ω) is the logarithm of the amplitude spectrum of the modeled seismogram xij (t), and
S̃j (ω), H̃l (ω), Ẽk (ω) and G̃i (ω) are the log-amplitude spectra of the individual components in
the time domain — sj (t), gi (t), hl (t), and ek (t), respectively. The term sj (t) is the waveform
component associated with source location j, the term gi (t) is the component associated with
receiver location i, and the term hl (t) is the component associated with offset dependency of
the waveform defined for each offset index l = |i − j|. The fourth component ek (t) represents
the earth’s impulse response at the source-receiver midpoint location, k = (i + j)/2.
Consider a data set with ns shot locations and nc channels, so that the total number of
traces is ns × nc . For ns × nc values of the actual spectral components X̃ij at frequency ω,
and ns shot locations, nr receiver locations, ne midpoint locations, and nh offsets, we have the
Data Modeling by Inversion 1531

following set of model equations:


 . 
..
 
 S̃j 
   . 
  .. 

   .. 
    
    . 

 .     H̃l  
 ..   
     .. 
 X̃
 ij
  
 = ···1··· 1··· 1··· 1··· . . (J − 26)
 .     .. 

 .    .  
 .   
     Ẽk 
    .  
   .. 
 . 
 . 
 . 
 
 G̃i 
..
.

Write equation (J-26) in matrix notation, X̃ = Lp as in equation (J-1). Here, X̃ is the


column vector of (ns × nc )-length on the left-hand side in equation (J-26), L is the sparse matrix
with dimensions (ns ×nc )×(ns +nh +ne +nr ), and p is the column vector of (ns +nh +ne +nr )-
length on the right-hand side of the same equation. Except the four elements in each row, the
L matrix contains zeros. Consider an example of ns = 1000, nc = 240, nh = 60, ne = 2000,
and nr = 1000. You would then have 240 000 equations to estimate 4060 parameters — more
equations than unknowns.
We want to estimate for each frequency ω the model parameters p such that the difference
between the actual spectral component X̃ and the modeled spectral component X̃ is minimum
in the least-squares sense. Follow the steps from equation (J-2) to (J-7a) for the complex case
and derive the GLI solution for surface-consistent deconvolution, p = (LT∗ L)−1 LT∗ X̃. The pa-
rameter vector p that contains the spectral components S̃j , G̃i , H̃l and Ẽk , which are associated
with the source and receiver locations, offset dependency, and earth’s impulse response, is solved
for each frequency component ω. Results from all frequency components are then combined to
obtain the terms in equation (J-25). The surface-consistent spiking deconvolution operator to
be applied to each trace in the data set is then the minimum-phase inverse of sj (t) ∗ gi (t) ∗ hl (t).
The size of the matrix LT L is (ns + nh + ne + nr ) × (ns + nh + ne + nr ). Consider 500
frequency components for which equation (J-28) is to be solved for. For the example specified
above, you would have to solve a matrix equation p = (LT∗ L)−1 LT∗ X̃ of the size 4060 × 4060
for each of the 500 frequency components. A practical scheme for solving this equation is based
on the Gauss-Seidel iterative scheme described in Section B.8.
To estimate surface-consistent shot and receiver residual statics, we model traveltime devi-
ations on moveout-corrected CMP gathers. The model equation for residual statics estimation
is given by (Section C.4)

tij = sj + ri + Gk + Mk x2ij , (J − 27)

where tij are the modeled traveltime deviations associated with a reflection event on moveout-
corrected CMP gathers, sj is the residual statics shift at the jth source location, ri is the residual
statics shift at the ith receiver location, Gk is the structure term at the kth midpoint location,
and Mk x2ij is the residual moveout at the kth midpoint location.
Consider a data set with ns shot locations and nc channels, so that the maximum number
of picks is ns × nc . For ns × nc picks of tij , and ns shot locations, nr receiver locations, nG
1532 Seismic Data Analysis

midpoint locations, we have the following set of equations:


 . 
.
 .  
..  . 
 sj 
 .    . 
 ..    
     .. 
 ..   
 r 
 .    i 
   .. 
 tij  =  · · · 1 · · · 1 · · · 1 · · · x2 · · · 
 , (J − 28)
   ij 
 . 
 .    
 ..    G 
     .k 
 ..    
 .   .. 
..  
 Mk 
. ..
.
which in compact matrix notation is given by t = Lp as in equation (J-1). Here t is the column
vector of (ns × nc )-length in equation (J-28), L is the sparse matrix in the same equation with
dimensions (ns × nc ) × (ns + nr + nG + nG ), and p is the column vector of (ns + nr + nG + nG )-
length on the right-hand side of the same equation. Except for the three elements in each row,
the L matrix contains zeros. Consider an example of ns = 1000, nc = 240, nr = 1000, and
nG = 2000. You would then have 240 000 equations to estimate 6000 parameters.
Follow the steps from equation (J-2) to (J-6) and derive the GLI solution for surface-
consistent residual statics, p = (LT L)−1 LT t. Here t denotes the column vector of (ns × nc )-
length that represents the traveltime deviations picked from moveout-corrected CMP gathers.
The size of the matrix LT L is (ns +nr +nG +nG )×(ns +nr +nG +nG ). For the example specified
above, you would have to solve a matrix equation p = (LT L)−1 LT t of the size 6000 × 6000. A
practical scheme for solving this equation is based on the Gauss-Seidel method (Section C.4).
To estimate surface-consistent shot and receiver intercept time anomalies, and thus obtain
shot and receiver refraction statics, we model refracted arrival times derived from picking first
breaks (Section 3.4). The model equation for the refracted arrivals is given by (Section C.8)
tij = Tj + Ti + sb xij , (J − 29)
where Tj and Ti are the intercept time anomalies at shot and receiver locations, respectively,
and sb is the bedrock slowness. Here, we shall consider the special case of surface shots. Hence,
for n shot/receiver stations the parameter vector is p : (T1 , T2 , . . . , Tn ; sb ).
Consider a data set with n shot locations and nc channels, so that the total number of
traces is n × nc . Assume that you will pick the first breaks from half the number of the traces.
For n × nc /2 picks of tij , and n + 1 parameters p : (T1 , T2 , . . . , Tn ; sb ), we have the following set
of equations:
 . 
..  
 .   ..
 ..   . 
  
 ..    Tj 
 .    
    . 
 tij  =    .. 
   · · · 1 · · · 1 · · · xij   , (J − 30)
 .     Ti 
 ..   
   . 
 ..   .. 
 . 
.. sb
.
which in matrix notation is t = Lp as in equation (J-1). Here t is the column vector of
(n×nc /2)-length in equation (J-30), L is the sparse matrix in the same equation with dimensions
(n × nc /2) × (n + 1), and p is the column vector of (n + 1)-length on the right-hand side of the
same equation. Except for the three elements in each row, the L matrix contains zeros. Consider
an example of n = 1000 and nc = 240. You would then have 120 000 equations to estimate 1001
parameters.
Data Modeling by Inversion 1533

Follow the steps from equation (J-2) to (J-6) and derive the GLI solution for surface-
consistent refraction statics, p = (LT L)−1 LT t. Here t denotes the column vector of (n × nc /2)-
length that represents the observed (picked) refracted arrival times in equation (J-30). The size
of the matrix LT L is (n + 1) × (n + 1). For the example specified above, you would have to solve
a matrix equation p = (LT L)−1 LT t of the size 1001 × 1001. A practical scheme for solving this
equation is based on the Gauss-Seidel method (Section C.4).
In some applications, instead of directly modeling the observed data, such as refracted
arrivals, it may be preferable to perturb the parameter vector p by a small amount ∆p and
compute the change in the model vector ∆t . Perturbation of the parameter vector associated
with the earth is the basis for tomographic inversion. Since it is the change in the parameter
vector and not the data vector that we estimate, tomographic application of the GLI technique
may be appropriately considered an example of earth modeling rather than data modeling. In
Section J.6, we review reflection traveltime tomography for updating earth model parameters.
Here, we will refer to the model equation for refraction tomography (Section C.9)
∆tij = Zj ∆swj + Zi ∆swi + Xij ∆sb , (J − 31)
where ∆tij is the amount of change in the difference between the observed traveltimes tij and
the initial estimate of the modeled traveltimes tij as a result of the change in the parameter
∆p; ∆swj and ∆swi are the slowness perturbations within the weathering layer at shot and
receiver locations, respectively; and ∆sb is the slowness perturbation within the bedrock layer.
The coefficient terms Zj and Zi and Xij are functions of the weathering layer thickness, critical
angle of refraction at shot and receiver locations and shot-receiver separation (Section C.9).
Examine the structure of equation (J-31) and note that, instead of modeling refraction
traveltimes by way of equation (J-29), we model the change in traveltimes by way of equa-
tion (J-31) and thus estimate the near-surface parameters. Hence, for n shot/receiver stations,
the parameter vector is ∆p : (∆sw1 , ∆sw2 , · · · , ∆swn , ∆sb ). We refer to the scheme based on
equation (J-31) as the iterative GLI solution.
Consider a data set with n shot locations and nc channels, so that the total number of
traces is n × nc . Again, assume that you will pick the first breaks from half the number of the
traces. For n × nc /2 picks of tij and n + 1 parameters ∆p : (∆sw1 , ∆sw2 , · · · , ∆swn , ∆sb ), we
have the following set of equations:
 . 
..  
 .   ..
 ..   .
   
 ..    ∆swj 
 .     
    . 
 ∆tij  =    . 
   · · · Zj · · · Zi · · · Xij   . . (J − 32)
 .     
 ..    ∆swi 
   . 
 ..   .. 
 . 
.. ∆sb
.
We write these equations in matrix notation as in equation (J-1) as
∆t = L ∆p, (J − 33)
where ∆t is the column vector of (n + nc /2)-length on the left-hand side of equation (J-32), L
is the sparse matrix in the same equation with dimensions (n + nc /2) × (n + 1), and p is the
column vector of (n + 1)-length on the right-hand side of equation (J-32). Except for the three
elements in each row, the L matrix contains zeros. For the example of n = 1000 and nc = 240,
you would have 120 000 equations to estimate 1001 parameters.
The GLI solution to equation (J-34) satisfies the requirement that the energy of the error
vector e = ∆t − ∆t is minimum and is given by
∆p = (LT L)−1 LT ∆t, (J − 34)
1534 Seismic Data Analysis

where ∆t denotes the column vector of n + nc /2-length that represents the difference between
the observed (picked) refracted arrival times and the initial estimate of the modeled times in
equation (J-32). The size of the matrix LT L is (n + 1) × (n + 1). For the example specified
above, you would have to solve a matrix equation p = (LT L)−1 LT t of the size 1001 × 1001. A
practical scheme for solving this equation is based on the Gauss-Seidel method (Section C.4).
We now review data modeling by the discrete Radon transform. Let d(h, t) represent a CMP
gather in the offset space (the plane of offset versus two-way traveltime) and u(v, τ ) be the
transformed data in the velocity space (the plane of stacking velocity versus two-way zero-offset
time). The model equation for the hyperbolic Radon transform in the Fourier transform domain
is given by (Section F.3)

d (h, ω ) = u(v, ω ) exp (−iω 4h2 /v 2 ), (J − 35)


v

where ω is the Fourier dual of t , with t = t2 , t being the two-way traveltime. Also, v is the
hyperbolic moveout velocity and h is the (half) offset.
Equation (J-33) can be written in the matrix form d = Lu as in equation (J-1) for each
ω component of d (h, ω ) and u(v, ω ), where L now is a complex matrix of dimensions m × p:
 −iω 4h2 /v2 2 2 2 2 
e 1 1 e−iω 4h1 /v2 . . . e−iω 4h12 /vp2
2 2 2 2
 e−iω 4h2 /v1 e−iω 4h2 /v2 . . . e−iω 4h2 /vp 
L=  . .. .. .. ,
 (J − 36)
.. . . .
2 2 2 2 2 2
e−iω 4hm /v1 e−iω 4hm /v2 ... e−iω 4hm /vp
and d and u are complex vectors of lengths m and p, respectively, where m and p are the
number of offsets and the number of velocities used in computing the discrete Radon transform.
Note that the elements of the L matrix only depend on the geometry of the input CMP gather
and the range of velocities used in constructing the velocity-stack gather u(v, τ ).
Follow the steps described by equations (J-2) through (J-7a) to obtain the GLI solution for
each Fourier component of u(v, ω ) given by u = (LT∗ L)−1 LT∗ d, where the asterisk denotes
complex conjugate. The size of the matrix LT L is m × p. For m = 120, p = 120, and 500 Fourier
components, you would have to solve a matrix equation p = (LT L)−1 LT t of the size 120 × 120
for each of the 500 Fourier components. The elements of the matrix L given by equation (J-36)
may have similar magnitudes in two different rows because of the small separation of offsets
associated with the input CMP gather. As such, the matrix can be near singular and requires a
solution based on singular-value decomposition (Section F.3).
Finally, slant-stack transform is a special case of the generalized discrete Radon transform
based on linear moveout equation. For details of the generalized linear inversion formulation of
the slant-stack transform, see Section F.3.

J.4 Dix Conversion

We shall develop the traveltime equation for a CMP recording geometry and an earth model
that comprises horizontal isovelocity layers (Figure J-1). Then, we shall derive the Dix equation
for interval velocities computed from rms velocities associated with the horizontal layers. From
the geometry of Figure J-1, note that the half-offset h is given by
n
h= ∆xk
k=1
or
n
h= tan θk ∆zk . (J − 37a)
k=1
Data Modeling by Inversion 1535

FIG. J-1. Geometry of a horizontally layered earth model with constant velocities to develop the theory
for the normal moveout and Dix equations (Section J.4).

From Snell’s law, the ray paramater associated with the raypath in Figure J-1 is
sin θk
p= , (J − 37b)
vk
where vk is the interval velocity for the kth layer. It follows that
pvk
tan θk = . (J − 37c)
1 − p2 vk2
Also note that
vk ∆τk
∆zk = , (J − 37d)
2
where ∆τk is the two-way zero-offset time along the vertical raypath within the kth layer.
Substitute equations (J-37c) and (J-37d) into equation (J-37a), and for convenience, define the
full offset x = 2h
n
pvk2 ∆τk
x= . (J − 38)
k=1
1 − p2 vk2

From the geometry of Figure J-1, note that the two-way traveltime tn (x) along the raypath
from the source S to the reflection point R on the nth interface back to the reflector G is given
1536 Seismic Data Analysis

by the sum of the traveltimes ∆tk along the raypath segment within each layer
n
tn (x) = 2 ∆tk ,
k=1
or
n
∆zk
tn (x) = 2 . (J − 39a)
vk cos θk
k=1

From Snell’s law given by equation (J-37b), it follows that

cos θk = 1 − p2 vk2 . (J − 39b)

Substitute equations (J-37d) and (J-39b) into equation (J-39a)


n
∆τk
tn (x) = . (J − 40)
k=1
1 − p2 vk2

Equations (J-38) and (J-40) are the raypath equations for an earth model with horizontal
isovelocity layers. Based on source-receiver reciprocity, the traveltime t2n (x) can be represented
by a curve that is symmetric with respect to offset x (Taner and Koehler, 1969)
t2n (x) = C0 + C1 x2 + C2 x4 + . . . , (J − 41)
which also is referred to in Section C.1. We need to expand the square-root term in equations
(J-38) and (J-40) into the Taylor series to derive the expressions for t2 (x) and the even powers
of x displayed in equation (J-41) (Hubral and Krey, 1980). The series expansions are
n
1 3
x= pvk2 ∆τk 1 + p2 vk2 + p4 vk4 + . . . (J − 42a)
2 8
k=1

and
n
1 3
tn (x) = ∆τk 1 + p2 vk2 + p4 vk4 + . . . . (J − 42b)
2 8
k=1

Now, write the infinite series in equations (J-42a) and (J-42b) as a summation
n ∞
x= p2j−1 aj vk2j ∆τk (J − 43a)
k=1 j=1

and
n ∞
tn (x) = p2j bj vk2j ∆τk . (J − 43b)
k=1 j=0

In equation (J-43a), the terms (a1 , a2 , a3 , . . .) are the fractional coefficients (1, 1/2, 3/8, . . .) dis-
played in equation (J-42a). Similarly, in equation (J-43b), the terms (b0 , b1 , b2 , . . .) are the frac-
tional coefficients (1, 1/2, 3/8, . . .) displayed in equation (J-42b). Interchange the order of the
summations in equations (J-43a) and (J-43b) to obtain
∞ n
x= p2j−1 aj vk2j ∆τk (J − 44a)
j=1 k=1

and
∞ n
tn (x) = p2j bj vk2j ∆τk . (J − 44b)
j=0 k=1
Data Modeling by Inversion 1537

Define two terms, Aj and Bj


n
Aj = aj vk2j ∆τk , (J − 45a)
k=1

n
Bj = bj vk2j ∆τk , (J − 45b)
k=1

and rewrite equations (J-44a) and (J-44b) as



x= p2j−1 Aj (J − 46a)
j=1

and

tn (x) = p2j Bj . (J − 46b)
j=0

The even powers of x that we need to substitute into equation (J-41) can also be expressed
as a double summation of the form given by equation (J-46a)

x2m = p2(j+m−1) Aj,2m , (J − 47a)
j=1

where m = 1, 2, 3, . . ., and Aj,2m is a recursive combination of Aj of equation (J-45a) (Hubral


and Krey, 1980).
Similarly, t2n (x) that we need to substitute into equation (J-41) can be expressed as a double
summation of the form given by equation (J-46b)

t2n (x) = p2j Bj,2 , (J − 47b)
j=0

where Bj,2 is a recursive combination of Bj of equation (J-45b) (Hubral and Krey, 1980).
By using equations (J-47a) and (J-47b) in equation (J-41), and setting the terms on both
sides of the equation with like powers of p equal to another, the coefficients, C0 , C1 , C2 , . . ., can
be determined in terms of the model parameters — interval velocities vk and layer thicknesses
defined in terms of two-way vertical times ∆τk . While details can be found in Hubral and Krey
(1980), here, we shall consider a truncated form of equation (J-41) to derive the Dix equation.
Specifically, we shall consider the following form of equation (J-41)

t2n (x) ≈ C0 + C1 x2 . (J − 48)

Expand the series in equations (J-44a) and (J-44b) up to the second power of the ray parameter
p and set a1 = 1, b0 = 1, and b1 = 1/2 to obtain
n
x≈p vk2 ∆τk (J − 49a)
k=1

and
n n
p2
tn (x) ≈ ∆τk + vk2 ∆τk . (J − 49b)
2
k=1 k=1

Refer to equation (J-37b) and note that by retaining only the terms up to the second power of
p, we are making a small-spread approximation as in equation (J-48).
1538 Seismic Data Analysis

Define
n
τn = ∆τk (J − 50a)
k=1

and
n
1
Vn2 = vk2 ∆τk , (J − 50b)
τn
k=1

where τn is the total two-way zero-offset elapsed time along the vertical raypath from the surface
down to the nth interface, and Vn is the rms velocity at the nth interface measured along the
vertical raypath (Figure J-1).
Use the definitions given by equations (J-50a) and (J-50b) to rewrite equations (J-49a) and
(J-49b) to yield
x ≈ pVn2 τn (J − 51a)
and
p2 Vn2 τn
tn (x) ≈ τn + . (J − 51b)
2
Now, square both sides of equations (J-51a) and (J-51b), and again, retain only the terms up
to the second power of p
x2 ≈ p2 Vn4 τn2 (J − 52a)
and
t2n (x) ≈ τn2 + p2 Vn2 τn2 . (J − 52b)
Finally, substitute equations (J-52a) and (J-52b) into equation (J-48)
τn2 + p2 Vn2 τn2 = C0 + C1 p2 Vn4 τn2 , (J − 53)
and set the terms with like powers of p on both sides of the equation equal to one another to
determine the coefficients C0 and C1
C0 = τn2 (J − 54a)
and
1
C1 = . (J − 54b)
Vn2
Back substitution into equation (J-48) yields the familiar hyperbolic normal moveout equation
x2
t2n (x) ≈ τn2 + . (J − 55)
Vn2
We now return to equation (J-50b) and split it in the following manner:
n−1
1
Vn2 = vk2 ∆τk + vn2 ∆τn . (J − 56a)
τn
k=1

By the definition of the rms velocity given by equation (J-50b), note that
n−1
2 1
Vn−1 = vk2 ∆τk . (J − 56b)
τn−1
k=1

Substitute equation (J-56b) into equation (J-56a)


2
Vn−1 τn−1 + vn2 ∆τn
Vn2 = . (J − 56c)
τn
Data Modeling by Inversion 1539

Finally, rearrange the terms and note that ∆τn = τn − τn−1 to obtain the Dix equation
2 τ
Vn2 τn − Vn−1 n−1
vn = , (J − 57)
τn − τn−1
which is equation (9-1) of the main text. Equation (J-57) is the basis for Dix conversion of rms
velocities to interval velocities. Given the rms velocities measured from the surface down to the
(n − 1) and nth layer boundaries along the vertical raypaths and the associated two-way zero-
offset times tn−1 and tn , the layer velocity within the interval between the two layer boundaries
can be computed using equation (J-57). Keep in mind the underlying assumption that the earth
model comprises horizontal isovelocity layers. Although reflector dip can be incorporated into
the Dix equation (Hubral and Krey, 1980), in practice, it is most desirable to work with rms
velocities estimated from gathers generated by prestack time migration. Additionally, keep in
mind that the rms velocities used in equation (J-57) are based on straight-ray assumption; thus,
ray bending at layer boundaries is not accounted for in Dix conversion. Finally, since equation (J-
48) is an approximation to equation (J-41), the offset range used in estimating the rms velocities
Vn and Vn−1 corresponds to a small spread.

J.5 Map Processing

A map is defined as a 2-D surface g(x, y). Depending on the quantity being mapped, g(x, y)
may have many types of units; for example, gravitational attraction (mGal), magnetic intensity
(gamma), elevation, or even times picked along marker horizons from seismic data. Here, we set
the convention that the positive x-axis points eastward and the positive y-axis points northward.
A discrete map function is represented by a grid of mesh points over the x − y plane. These
mesh points are spaced commonly at equal intervals in the x and y directions. For many types of
mapping, g(x, y) is a smooth function and such a map can be analyzed in the Fourier transform
domain. However, there are situations (as for isochron and structure maps) in which the map
function has discontinuities that represent faulting.
Maps usually are created from irregularly spaced observation values. Thus, the map function
at a particular grid point must be computed by some fitting procedure. A local plane surface
fitting procedure is described later in this section.
Figure J-2 shows the contour map of the test surface used in this study. Void grid points have
been filled with the arithmetic mean of the map function. Before map creation, some correction
may be applied to observed data, followed by various types of editing. The main topic of this
section is the next processing stage, which involves various 2-D processing techniques, each of
which is aimed at achieving a particular interpretive goal. To clarify the objectives in digital
processing, consider the properties of the quantity represented by a map function. In particular,
consider the gravity problem. Ultra-long wavelength anomalies generally are associated with
variations in crustal thickness. Moderately long wavelengths usually are caused by variations in
the basement topography. Medium to short wavelength anomalies are related closely to local
tectonic disturbances such as folding. Finally, very short wavelength anomalies have a variety of
sources; some are because of small, near-surface features, while others are just noise of various
types. Note that gradation in wavelength variations is related somewhat to the depth of the
source that causes the anomaly. The longer the wavelength, the deeper the anomaly; the shorter
the wavelength, the shallower the anomaly.
Similar physical interpretations are possible with other map functions, such as the seis-
mic (time) horizon map in Figure J-2. Some of the very short wavelengths here result from
near-surface effects that are manifested as residual statics. Also note in Figure J-2 that some
moderately long wavelengths correspond to structural undulations that exist in the area. The
most important observation made from this map is that most of the features with different
wavelengths are not spatially isolated, but are superimposed. This characteristic is common for
1540 Seismic Data Analysis

FIG. J-2. A time map of a seismic horizon. The faults have been smoothed out.

all types of map functions, whether gravity, magnetic, elevation, or time. To separate the effects of
different features from each other, we must analyze them in terms of wavelength. This wavelength
analysis also is a way to discriminate the depth of the source for some types of data such as
potential-field measurements.
The basic motivation behind digital processing of maps is to separate anomalies. Simple 2-D
smoothing and wavelength filtering are techniques for separating anomalies. Vertical derivatives
and analytic continuation also are useful for enhancing certain anomalies so that they appear
more pronounced on the map. Some techniques for separating anomalies by their wavelengths
are described next.
Before applying digital processing techniques, it is best to study the map in terms of wave-
length composition. The 2-D amplitude spectrum of a map is an excellent tool for recognizing
not only the wavelength content, but also the orientation of various components. The most use-
ful display is the color contour plot of the amplitude spectrum (Figure J-3) from which various
bands of wavelengths are distinguished clearly. Pink represents long, beige represents moderate,
and yellow represents short wavelength anomalies. For a 2-D real function, such as a map, the
amplitude spectrum is antisymmetric. Thus, only a pair of quadrants (the first and second) of
the amplitude spectrum needs to be displayed.
A simple 2-D smoothing operation is the easiest way to obtain a map that represents the
regional anomaly. Figure J-4 shows the map in Figure J-2 after 2-D smoothing. Most of the
very short wavelength features present in the center of the original map have been eliminated.
Depending on the interpreter’s goal, this output might be a satisfactory regional map. (However,
a smoother regional map often is desired.)
Data Modeling by Inversion 1541

FIG. J-3. 2-D amplitude spectrum of the map in Figure J-2.

Smoothing basically is done by computing the average value of the grid points that fall onto
a ring of some desired radius. The center of the ring coincides with the output point. There is
no reason why more than one ring should not be considered. For n concentric rings with mi
points over the ith ring, the average value gi of the quantity gij that is being mapped is
mi
1
ḡi = gij . (J − 58a)
mi j=1

Thus, the cumulative average g over n rings is


n
1
ḡ = ḡi . (J − 58b)
n i=1

Weighting factors, which depend on the distance from the center of the rings, often are used in
smoothing algorithms. In these cases, equation (J-58b) becomes
n
ḡ = wi ḡi , (J − 58c)
i=1

where wi are the weights. The residual anomaly is defined by


g = g0 − ḡ, (J − 59)
where g0 is the grid value at the center of the rings. Figure J-5 shows the regional anomaly
obtained by using 15 rings. In general, the more rings, the more smoothing of the data. Note
that the output of the ring method is relative. Depending on the nature of the objectives of
1542 Seismic Data Analysis

FIG. J-4. A smoothed version of the contour map in Figure J-2.

map processing, Figure J-5 may or may not be a good estimate of the regional anomaly. In most
cases, determining what is regional and what is residual is a matter of interpretive judgment.
The hypothetical map in Figure J-6a has ten different anomalies of various shapes and
orientations. Figure J-6b is the 2-D amplitude spectrum of this map with (kx , ky ) as the Fourier
duals of (x, y). Anomaly 3, which has infinite wavelength in the north-south direction (y-axis)
and a wavelength of 2AA in the east-west direction (x-axis), lies on the kx -axis in the transform
domain. Anomaly 7 also lies on the kx -axis, but is farther from the origin since it has a shorter
wavelength component (2BB ) in the x direction. Anomaly 8 tilted counterclockwise by an angle
θ degrees from the y direction has an apparent wavelength component 2CC in the y direction
and 2DD in the x direction.
Note the following relations:
λx
tan θ = , (J − 60a)
λy
where λx = DD and λy = CC are the wavelength components. By definition

λx =
kx
and

λy = ,
ky
which, after substitution into equation (J-60a), become
ky
tan θ = . (J − 60b)
kx
Data Modeling by Inversion 1543

FIG. J-5. A regional anomaly map based on equation (J-58c). The input is the contour map in Figure
J-2.

Equations (J-60a) and (J-60b) imply that map trend M M is orthogonal to transform trend
TT .
Unlike the simple, linear anomalies 3, 7, and 8, rounded anomalies 1, 2, and 10 map onto
the entire transform plane. Moreover, because of their equal size, anomalies 1 and 2 cannot be
distinguished on the plane of 2-D amplitude spectrum, even though they are isolated in the space
domain. However, the spatial extent of these two anomalies is much wider than that of anomaly
10. Thus, they are confined to lower wavenumbers. Finally, elongated features like anomalies 4,
5, 6, and 9 map onto the transform plane at certain dominant directions that are orthogonal to
their respective spatial trends. The dots on the kx −ky plane in Figure J-6b depict the maximum
energy associated with the anomalies.
Gridding involves fitting a locally plane surface to a set of control points around each grid
output point. Consider a plane-surface fit in the least-squares sense:
g̃(x, y) = a0 + a1 x + a2 y. (J − 61)
The least-squares error is
M
2
L= gi − g̃i , (J − 62)
i=1

where g is the observed value at the grid point (x, y), and M is the number of observations at
and around that grid point. For local plane fitting, M usually is set to 9 points. We want to find
a set of (a0 , a1 , a2 ) for which L is minimum;
∂L ∂L ∂L
= = = 0. (J − 63)
∂a0 ∂a1 ∂a2
1544 Seismic Data Analysis

FIG. J-6. Anomalies of various shapes in (a) the space and (b) the wavenumber domains.

By substituting equation (J-61) into equation (J-62), we have


M
2
L= gi − a0 − a1 xi − a2 yi . (J − 64)
i=1

Then, by doing the differentiations expressed by equation (J-63), we get the following set of
simultaneous equations
a0 + a1 x + a2 y = g,
a0 x + a1 x2 + a2 xy = xg,
a0 y + a1 xy + a2 y 2 = yg.
When put into matrix form, we obtain
    
M x y a0 g
 x x2 xy   a1  =  xg  . (J − 65)
y xy y2 a2 yg
Equation (J-65) is solved for the set of coefficients (a0 , a1 , a2 ).
Unlike the hypothetical anomalies in Figure J-6a, real data consist of anomalies of various
shapes and orientations that are superimposed on each other. However, in the transform domain,
the real anomalies can be separated in terms of their wavelength contents and orientations;
something that cannot be achieved in the space domain. Thus, the transform domain provides
a way to apply various filtering operations to a map. A band of wavelengths can be passed
Data Modeling by Inversion 1545

regardless of orientation, as shown by the radial filter transfer function in Figure J-7a. Anomalies
that are oriented in the range (θ1 , θ2 ) from the northerly direction can be passed regardless of
their size, using a directional filter whose transfer function is given by Figure J-7b. Finally, a
band of wavelengths with a directional orientation can be passed. This is achieved by a filter
with a transfer function as shown in Figure J-7c. In practice, as in other filter design techniques,
the transfer functions shown must be tapered in the neighborhood of cutoff wavelengths for
optimal performance.
Once the transfer function is designed to suit the purpose, the filter itself can be applied to
the map in the transform or the space domain. In the transform domain, the transfer function is
multiplied with the 2-D Fourier transform of the map. Subsequent inverse transformation yields
the filtered map. To apply the filtering in the space domain, first inverse Fourier transform the
filter’s transfer function to obtain its 2-D impulse response. Two-dimensional convolution of this
impulse response with the map yields the filtered map. Note that filters whose transfer functions
are depicted in Figure J-7 do not cause any phase shift; i.e., anomalies are not displaced in the
space domain after filtering.
Before applying the 2-D filters, the possible bands of the wavelengths present on the am-
plitude spectrum of the map to be filtered must be determined. Four regions were distinguished
from the color plot (Figure J-3) of the amplitude spectrum of the test surface. We have the
regional component down to 21-km wavelengths. A subregional part of the spectrum between
the 21- and 12-km wavelengths is defined. A moderate range of wavelengths is between 12 and
8 km, and the residual components are beyond the 8-km cutoff wavelength. This residual region
can be subdivided; anything less than 4 km corresponds to very short wavelength anomalies.
These anomalies may be caused by various types of noise.
The bands that were just determined now are used as cutoff wavelengths of the radial filter
transfer function. A given map can be scanned with a suite of low-pass filters and several filtered
maps can be produced, each with a potentially unique interpretational value. A result from such
a filtering operation is shown in Figure J-8. Note the resemblance between Figure J-8 and the
regional map (Figure J-5) that was obtained from the ring method described in this section.
Although not demonstrated here, note that as the bandwidth of the filter is increased, more and
more of the short wavelength anomalies are included, making the output less and less regional
in character.
A filter scan is not limited to low-pass filtering only. High-pass, band-pass, and band-reject
filters can be applied to maps to achieve a particular interpretational goal. For example, we may
want to obtain a residual map that is free of the very short wavelength anomalies that usually
are considered noise. This is accomplished by properly selecting the band-pass filter. A high-pass
filter is the proper choice to retain all of the short wavelength region of the spectrum.
Directional filters scan the map of interest at various angles to emphasize a particular
trend that may exist in the data. Figure J-9 shows the result of one of the several directional
filters applied to the test surface in Figure J-2. A low-pass radial filter was cascaded with each
directional filter so that any regional trend could be delineated in the area. The results of
directional filtering indicate the existence of a prominent regional trend approximately N300 W
(Figure J-9). In some cases, a certain band of wavelengths may have one dominant trend that
is different from that of another band of wavelengths. This situation may imply a change in the
tectonic setting over the geologic history in the area.

J.6 Reflection Traveltime Tomography

We want to update the parameters of an earth model that has already been estimated by
a suitable combination of inversion methods discussed in this chapter. Reflection traveltime
tomography is an inversion method for estimating the earth model parameters from the reflection
traveltimes associated with the observed seismic data. The reflection traveltime from a source
at the surface to a reflection point at the subsurface and back to a receiver at the surface is
1546 Seismic Data Analysis

FIG. J-7. Transfer functions for (a) band-pass, (b) directional, and (c) directional band-pass wavenum-
ber filters.

FIG. J-8. The contour map in Figure J-2 after low-pass filtering.

represented by an integral of the traveltime segments along the raypath that depend on the earth
model parameters themselves. This makes the direct inversion of the traveltimes to estimate the
earth model parameters a nonlinear problem. Nevertheless, based on a formal linearization of the
eikonal equation, which is a ray-theoretical approximation to the scalar wave equation (Section
H.2), it can be shown that small changes in reflection traveltimes are linearly related to small
changes in earth model parameters (Aldridge, 1994). We are thus motivated to use reflection
traveltime tomography not to estimate an initial earth model, but to update an already estimated
model.
Data Modeling by Inversion 1547

FIG. J-9. The contour map in Figure J-2 after directional low-pass filtering.

To develop a theory for the model update, we shall set the following framework for our
strategy:

(a) Define the earth model parameters in terms of slowness functions and depths at the bound-
aries of the layers included in the initial model.
(b) Assume that the initial model is made up of horizontal layers with laterally invariant model
parameters.
(c) Use the reflection times from CMP data associated with the layer boundaries included in
the model and update the earth model such that the discrepancy between the modeled
reflection times and the actual reflection times is minimum in the least-squares sense.
(d) Estimate the changes in the model parameters, rather than the parameters themselves.
(e) Perturb the initial model in two parts — slowness perturbation and depth perturbation.
(f) Perturb the model parameters while preserving the offset values of the seismic data.

Consider the raypath geometry in a horizontally layered earth model shown in Figure J-10.
We want to derive the reflection traveltime equation for the raypath from the reflection point
Ri at the interface zn in the subsurface to the receiver location Gj at the surface z = 0. This
one-way raypath is associated with a CMP gather at midpoint location y and half-offset h. Our
earth model consists of n layers above the reflection point Ri .
1548 Seismic Data Analysis

FIG. J-10. Geometry of a nonzero-offset ray to develop the theory for the reflection traveltime tomog-
raphy (Section J.6).

The modeled traveltime segment tAC from A to C along the raypath Ri Gj within the kth
layer is
AC
tAC = (J − 66a)
vk
or
tAC = zk − zk−1 sk sec θk , (J − 66b)
since AB = zk − zk−1 and AC = AB sec θk . In equation (J-66a), vk is the interval velocity
for the kth layer and x is the lateral distance from the midpoint location y. Also, in equation
(J-66b), sk = 1/vk is the slowness of the kth layer and θk is the angle of incidence at the kth
layer boundary.
The modeled traveltime tn from the reflection point Ri on the nth interface to the receiver
Gj is then the sum of the traveltime segments given by equation (J-66b) within n layers between
the points Ri and Gj :
n
tn = (zk − zk−1 )sk sec θk . (J − 67)
k=1

We now derive the expression for the half-offset hn between the midpoint location y and
the receiver location Gj . Refer to Figure J-10 once again and note that the offset segment BC
is given by
BC = AB tan θk (J − 68a)
or
BC = zk − zk−1 tan θk . (J − 68b)
Data Modeling by Inversion 1549

The half-offset hn is the sum of the lateral segments within the n layers above the reflection
point Ri given by equation (J-68b):
n
hn = (zk − zk−1 ) tan θk . (J − 69)
k=1

The raypath used in deriving the traveltime and offset equations (J-67) and (J-69), respectively,
is described by the ray parameter s as
s = sk sin θk , (J − 70)
which is the horizontal component of the slowness.
Consider an initial estimate of the parameter vector p : (· · · , sm , · · · , zm , · · ·) along the
raypath from Ri to Gj in Figure J-10, where 1 ≤ m ≤ n. We want to minimize the difference
between the observed times tn and the modeled times tn by iteratively perturbing the initial
estimate of the parameter vector. A change ∆p in the paramater vector will change the modeled
times as follows:
∂tn
tn = tn + ∆p. (J − 71)
modeled initial ∂p
modeled

The error in modeling the traveltimes is given by

en = tn − tn . (J − 72)
observed modeled

Substitute equation (J-71) into equation (J-72)

∂tn
en = tn − tn − ∆p. (J − 73)
observed initial ∂p
modeled

Now define the difference ∆tn between the observed traveltimes tn and the initial estimate of
the modeled traveltimes tn , and rewrite equation (J-73) as
∂tn
en = ∆tn − ∆p. (J − 74a)
∂p
The second term on the right is the amount of change in ∆tn as a result of the change in the
parameter ∆p. Define this term as ∆tn , and rewrite equation (J-74a) once more as
en = ∆tn − ∆tn , (J − 74b)
where
∂tn
∆tn = ∆p. (J − 75a)
∂p
The parameter vector p : (· · · , sm , · · · , zm , · · ·), with 1 ≤ m ≤ n, is composed of the slowness
variable sm and the depth variable zm , ∆tn in equation (J-75a) actually has two parts — one part
associated with the slowness perturbation and the other associated with the depth perturbation,
given by
n n
∆tn = ∆tn (sm ) + ∆tn (zm ), (J − 75b)
m=1 m=1

where ∆tn (sm ) and ∆tn (zm ) are the contributions of the slowness perturbation and depth
perturbation, respectively.
To compute the traveltime perturbations caused by the slowness and depth perturbations,
we shall follow the elegant formulation by Sherwood et al. (1986). The slowness perturbation
1550 Seismic Data Analysis

∆sm in layer m, where 1 ≤ m ≤ n, yields a traveltime perturbation ∆tn (sm ) given by


n
∂tn ∂tn
∆tn (sm ) = ∆sm + ∆θk . (J − 76a)
∂sm ∂θk
k=1

Since the offset of the input data is preserved in the perturbation, note from Figure J-10 and
equation (J-70) that, while perturbing the slowness, the propagation angle also is perturbed.
That is why we include the second partial differential term on the right-hand side of equation (J-
76a). In fact, slowness perturbation of a single layer m causes perturbation of propagation angles
within all of the n layers along the traveltime path. And that is why we have the summation in
the second term on the right-hand side of equation (J-76a).
Compute the partial derivatives in equation (J-76a) using equation (J-67) to yield
n
∆tn (sm ) = zm − zm−1 sec θm ∆sm + zk − zk−1 sk sec2k θk sin θk ∆θk . (J − 76b)
k=1

Substitute equation (J-70) into the second term on the right-hand side of this equation:
n
∆tn (sm ) = zm − zm−1 sec θm ∆sm + s zk − zk−1 sec2k θk ∆θk . (J − 76c)
k=1

Apply the same perturbation given by equation (J-76a) as for the traveltime equation (J-67)
to the offset equation (J-69) to obtain
n
∂hn ∂hn
∆hn (sm ) = ∆sm + ∆θk . (J − 77a)
∂sm ∂θk
k=1

Compute the partial derivatives in equation (J-77a) using equation (J-69) to yield
n
∆hn (sm ) = zk − zk−1 sec2k θk ∆θk . (J − 77b)
k=1

Note that this summation is the same as the summation in the second term on the right-hand
side of equation (J-76c). A direct consequence of the rule for preserving the offset value during
perturbation leads to setting ∆hn (sm ) = 0 in equation (J-77b). This means that the second
term on the right-hand side of equation (J-76c) vanishes, giving us the final expression for the
slowness perturbation:
∆tn (sm ) = zm − zm−1 sec θm ∆sm . (J − 78)
We now evaluate the traveltime perturbation ∆tn (zm ) caused by a depth perturbation ∆zm
in layer m, where 1 ≤ m ≤ n. This perturbation is given by
n
∂tn ∂tn
∆tn (zm ) = ∆zm + ∆θk . (J − 79a)
∂zm ∂θk
k=1

Since the offset of the input data is preserved in the perturbation, note from Figure J-1 that,
while perturbing the depth, the propagation angle also is perturbed. Again, that is why we
include the second partial differential term on the right-hand side of equation (J-79a). Since
depth perturbation of a single layer m causes perturbation of propagation angles within all
of the n layers along the traveltime path, we have the summation in the second term on the
right-hand side of equation (J-79a).
Compute the partial derivatives in equation (J-79a) using equation (J-67) to yield
n
∆tn (zm ) = sm sec θm − sm+1 sec θm+1 ∆zm + zk − zk−1 sk sec2k θk sin θk ∆θk .
k=1
(J − 79b)
Data Modeling by Inversion 1551

A perturbation in the thickness defined by (zm − zm−1 ) also causes a change in the thickness
(zm+1 −zm ), but in the opposite direction. That is why we have the two parts in the first term of
the right-hand side of equation (J-79b) — (sm sec θm ) and (sm+1 sec θm+1 ). Substitute equation
(J-70) into the second term on the right-hand side of this equation to obtain
n
∆tn (zm ) = sm sec θm − sm+1 sec θm+1 ∆zm + s zk − zk−1 sec2k θk ∆θk . (J − 79c)
k=1

Apply the same perturbation given by equation (J-79a) as for the traveltime equation (J-67)
to the offset equation (J-69):
n
∂hn ∂hn
∆hn (zm ) = ∆zm + ∆θk . (J − 80a)
∂zm ∂θk
k=1

Compute the partial derivatives in equation (J-80a) using equation (J-69) to yield
n
∆hn (zm ) = tan θm − tan θm+1 ∆zm + zk − zk−1 sec2k θk ∆θk . (J − 80b)
k=1

The two parts of the first term of the right-hand side of this equation (tan θm ) and (tan θm+1 )
are justified as for equation (J-79b).
A direct consequence of the rule for preserving the offset value during perturbation leads
to setting ∆hn (zm ) = 0 in equation (J-80b). This then leads to
n
zk − zk−1 sec2k θk ∆θk = − tan θm − tan θm+1 ∆zm , (J − 81a)
k=1

which, upon substitution into equation (J-79c) yields

∆tn (zm ) = sm sec θm − sm+1 sec θm+1 ∆zm − s tan θm − tan θm+1 ∆zm . (J − 81b)

A further substitution of equation (J-70) then is made to get

∆tn (zm ) = sm sec θm − sm+1 sec θm+1 ∆zm


(J − 81c)
− sm sin2 θm sec θm − sm+1 sin2 θm+1 sec θm+1 ∆zm .

Simplification of the terms gives us the final expression for the depth perturbation as

∆tn (zm ) = sm cos θm − sm+1 cos θm+1 ∆zm . (J − 82)

We now combine, using equation (J-75b), the slowness perturbation given by equation (J-
78) and the depth perturbation given by equation (J-82) and obtain
n n
∆tn = zm − zm−1 sec θm ∆sm + sm cos θm − sm+1 cos θm+1 ∆zm . (J − 83)
m=1 m=1

Examine the structure of equation (J-83) and note that, instead of modeling reflection travel-
times, we model the change in traveltimes, and thus estimate the change in the model parameters.
The parameter vector is ∆p : (∆sm , · · · , ∆zm , · · ·), where 1 ≤ m ≤ n.
1552 Seismic Data Analysis

Finally, write equation (J-83) in matrix form for all the traveltime perturbations to obtain
 . 
..
 .   . 
 ..    ..
 
 ..   
 .    ∆sm 
    .. 
 ∆tn  =  · · · Zm · · · Sm · · ·   , (J − 84)

 . 
   . 


 ..   ∆zm 
  ..
 .. 
 .  .
..
.
where
Zm = zm − zm−1 sec θm (J − 85a)
and
Sm = sm cos θm − sm+1 cos θm+1 . (J − 85b)
Write equation (J-84) in compact matrix notation:
∆t = L ∆p, (J − 86)
where ∆t is the column vector on the left-hand side of equation (J-84), L is the sparse matrix
with nonzero elements Zm and Sm given by equations (J-85a,b), and ∆p is the column vector
on the right-hand side of equation (J-84). Except for the two elements in each row, the L matrix
contains zeros.
The generalized linear inversion (GLI) solution to equation (J-86) satisfies the requirement
that the energy of the error vector
e = ∆t − ∆t (J − 87)
is minimum and is given by (Section J.1)
∆p = (LT L)−1 LT ∆t, (J − 88)
where ∆t denotes the column vector that represents the difference between the observed reflec-
tion traveltimes and the initial estimate of the modeled times, and T denotes matrix transposi-
tion.
We now outline the procedure for a tomographic model update based on the theory de-
scribed in this section:

(a) Perform prestack depth migration using the initial model and generate a set of image
gathers.
(b) Compute the residual moveout for all offsets along events on image gathers that correspond
to the layer boundaries included in the model, and thus build the traveltime error vector
∆t. For instance, you may have 10 layers, 1000 CMPs with a fold of 30. This means that
the length of the traveltime error vector ∆t is 300 000.
(c) Define the initial model by a set of slowness and depth parameters, and construct the
coefficient matrix L in equation (J-88) by computing the nonzero matrix elements Zm and
Sm given by equations (J-85a) and (J-85b). For instance, you may have 10 layers and each
layer may be defined by 50 pairs of slowness and depth values in the lateral direction. This
means you would have 1000 parameters in your model space. It also means that for the
example given in step (b), you have 300 000 equations to solve for 1000 parameters. The
solution for this overdetermined system is given by equation (J-88).
(d) Estimate the change in parameters vector ∆p, by way of the GLI solution given by equation
(J-88).
(e) Update the paramater vector p + ∆p.
Data Modeling by Inversion 1553

(f) Iterate steps (a) through (e) as necessary to minimize the discrepancy between the modeled
and actual traveltimes.

J.7 Threshold for Velocity-Depth Ambiguity

Velocity-depth ambiguity states that an error in depth is indistinguishable from an error in


velocity. Specifically, we shall investigate under what circumstances two earth models defined
by layer velocities and reflector geometries in depth are indistinguishable. We shall adopt the
derivation by Lines (1993). Albeit the underlying theory to be presented is somewhat simplistic
compared to the comprehensive analysis by Bickel (1990), it does provide a practical insight to
important aspects of velocity-depth ambiguity.
Consider a single, horizontal reflector at depth z in a medium with constant velocity v. The
reflection traveltime associated with a CMP recording geometry of a fixed offset 2h is given by
the hyperbolic moveout equation

2 z 2 + h2
t(v, z) = . (J − 89)
v
Perturbation of velocity v by ∆v and depth z by ∆z causes a change in traveltime t(v, z)
by ∆t expressed as
∂t ∂t
∆t = ∆v + ∆z. (J − 90)
∂v ∂z
First, differentiate both sides of equation (J-89) with respect to velocity v
∂t 1
= −2 z 2 + h2 2 , (J − 91a)
∂v v
and simplify by way of equation (J-89)
∂t t
=− . (J − 91b)
∂v v
Next, differentiate both sides of equation (J-89) with respect to depth z
∂t 2z
= √ , (J − 92a)
∂z v z 2 + h2
and simplify, again, by way of equation (J-89)
∂t 4z
= 2 . (J − 92b)
∂z v t
Substitute the partial differentials given by equations (J-91b) and (J-92b) into equation (J-90),
then normalize with respect to traveltime t
∆t ∆v 4z∆z
=− + 2 2 . (J − 93)
t v v t
For the case of zero offset, t = 2z/v; hence, equation (J-93) takes the special form
∆t ∆v ∆z
=− + . (J − 94)
t v z
Equation (J-94) states a rule of fundamental importance: for a zero-offset case, when the
perturbation in velocity ∆v/v is the same as the perturbation in depth ∆z/z, no change occurs
in traveltime; thus, velocity-depth ambiguity is infinite.
Now consider two earth models — an unperturbed model defined by the variables (z, v)
and a perturbed model defined by the variables (z + ∆z, v + ∆v), such that the two models are
indistinguishable at zero offset. This means that the zero-offset times t0 associated with the two
1554 Seismic Data Analysis

models are identical. Rewrite equation (J-89) in terms of the constant zero-offset time t0 = 2z/v

4h2
t(v) = t0 1+ . (J − 95)
t20 v 2
Perturb the velocity v by ∆v and evaluate the change in time ∆t to get
∂t
∆t = ∆v. (J − 96)
∂v
Evaluate the partial differential by way of equation (J-95) to obtain
4h2 ∆v 1
∆t = − . (J − 96)
v 3 t0 4h2
1+ 2 2
t0 v
Expand the square-root term by the Taylor series up to the second order to get
4h2 ∆v 2h2
∆t = − 3
1− 2 2 .
v t0 t0 v
Finally, retain only the term with power of two in h, normalize with respect to traveltime t0 and
rearrange the variables to obtain the expression for the absolute value of the fractional change
in velocity ∆v/v
∆v z 2 ∆t
= 2 , (J − 97)
v h t0
where z = vt0 /2. Equation (J-97) states another rule for velocity-depth ambiguity: for a reflector
at depth z, velocity ambiguity defined by ∆v/v can be made smaller by increasing the offset h and
decreasing the error in the time pick defined by ∆t/t0 at offset h. Note that for the zero-offset
case, h = 0, the velocity ambiguity is infinite. This conclusion also was reached earlier by way
of equation (J-94).

REFERENCES

Aldridge, D. F., 1994, Linearization of the eikonal equation: Geophysics, 59, 1631-1632.
Al-Yahya, K., 1989, Velocity analysis by iterative profile migration: Geophysics, 54, 718-729.
Bell, M. L., Lara, R., and Gray, W. C., 1994, Application of turning-ray tomography to the offshore
Mississippi Delta: 64th Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, 1509-
1512.
Bickel, S. H., 1990, Velocity-depth ambiguity of reflection traveltimes: Geophysics, 55, 266-276.
Claerbout, J. F., 1992, Earth soundings analysis — processing versus inversion: Blackwell Scientific
Publications.
Coleman, J. M., Prior, D. B., and Garrison, L. E., 1980, Subaqueous sediment instabilities in the
offshore Mississippi River Delta: USGS Open-File Report 80-01, 60.
David, M., 1987, Geostatistics: in Encyclopedia of Science and Technology, 6, 141-144, Academic
Press.
Deregowski, 1990, Common-offset migrations and velocity analysis: First Break, 8, 225-234.
Dix, C. H., 1955, Seismic velocities from surface measurements: Geophysics, 20, 68-86.
Faye, J-P. and Jeannaut, J-P., 1986, Prestack migration velocities from focusing depth analysis:
56th Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts, 438-440.
Hestenes, M. R., 1956, The conjugate gradient method for solving linear systems: Proc. Sympo. in
Applied Math., VI, McGraw-Hill Book Co.
Hubral, P. and Krey, T., 1980, Interval velocities from seismic reflection time measurements: Soc.
Expl. Geophys.
Data Modeling by Inversion 1555

Kim, H. S. and Bell, M. L., 2000, 3-D turning-ray tomography and its application to Mississippi
Delta: 70th Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts 593-596.
Koehler, F. and M. T. Taner, 1985, The use of the conjugate-gradient algorithm in the computation
of predictive deconvolution operators: Geophysics, 50, 2752-2758.
Kosloff, D., Sherwood, J. W. C., Koren, Z., Machet, E., and Falkovitz, Y., 1996, Velocity and
interface depth determination by tomography of depth migrated gathers: Geophysics, 61, 1511-
1523.
Landa, E., Kosloff, D., Keydar, S., Koren, Z., and Reshef, M., 1988, A method for determination
of velocity and depth from seismic data: Geophys. Prosp., 36, 223-243.
Landa, E., Thore, P., Sorin, V., and Koren, Z., 1991, Interpretation of velocity estimates from
coherency inversion: Geophysics, 56, 1377-1383.
Lines, L., 1993, Ambiguity in analysis of velocity and depth: Geophysics, 58. 596-597.
Lines, L. R. and Treitel, S., 1984, Tutorial: A review of least-squares inversion and its application
to geophysical problems: Geophys. Prosp., 32, 159-186.
Loinger, E., 1983, A linear model for velocity anomalies: Geophys. Prosp., 31, 98-118.
Lynn, W. S. and Claerbout, J. F., 1982, Velocity estimation in laterally varying media: Geophysics,
47, 884-897.
May, J. A., Meeder, C. A., Tinkle, A. R., and Wener, K. R., 1988, Seismic no-data zone, Offshore
Mississippi Delta: Part II: using geologic information to predict acoustic properites: Proc.
Offshore Tech. Conf., 75-84.
Reshef, M., 1997, The use of 3-D prestack depth imaging to estimate layer velocities and reflector
positions: Geophysics, 62, 206-210.
Reshef, M., 2001, Some aspects of interval velocity analysis using 3-D depth migrated gathers:
Geophysics (Scheduled for publication in Jan.-Feb. 2001 issue).
Reshef, M. and Kessler, D., 1989, Practical implementation of three-dimensional poststack depth
migration: Geophysics, 54, 309-318.
Rocca, F. and Toldi, J., 1983, Lateral velocity anomalies: 53rd Ann. Internat. Mtg., Soc. Expl.
Geophys., Expanded Abstracts, 572-574.
Sheriff, R. E., 1991, Encyclopedic dictionary of exploration geophysics: Soc. Expl. Geophys.
Sherwood, J. W. C., Chen, K. C., and Wood, M., 1986, Depths and interval velocities from seismic
reflection data for low-relief structures: Proc. Offshore Tech. Conf., 103-110.
Sorin, V., 1996, Velocity estimation in homogenous media: 56th Ann. Internat. Mtg., Soc. Expl.
Geophys., Expanded Abstracts, 1515-1517.
Thorson, J. R., Gever, D. H., Swanger, H. J., Hadley, D. M., and Apsel, R. J., 1987, A model-based
approach to interval velocity analysis: 57th Ann. internat. Mtg., Expanded Abstracts, 458-460.
Wang, R. J. and S. Treitel, 1973, The determination of digital Wiener filters by means of gradient
methods: Geophysics, 38, 310-326.
Zhou, X., Sixta, D. P. and Angstman, B. G., 1992, Tomostatics: turning-ray tomography and statics
corrections: The Leading Edge, 11, No. 12, 15-23.

You might also like