GEOPHYSICS, VOL. 65, NO. 5 (SEPTEMBER-OCTOBER 2000); P. 1446–1454, 9 FIGS.

Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Shear-wave velocity and density estimation from PS -wave AVO

analysis: Application to an OBS dataset from the North Sea

Side Jin∗ , G. Cambois‡ , and C. Vuillermoz∗∗

ABSTRACT
S-wave velocity and density information is crucial for
hydrocarbon detection, because they help in the dis-
crimination of pore filling fluids. Unfortunately, these
two parameters cannot be accurately resolved from con-
ventional P-wave marine data. Recent developments in
ocean-bottom seismic (OBS) technology make it possi-
ble to acquire high quality S-wave data in marine envi-
ronments. The use of S-waves for amplitude variation
with offset (AVO) analysis can give better estimates of
S-wave velocity and density contrasts.
Like P-wave AVO, S-wave AVO is sensitive to various
types of noise. We investigate numerically and analyti-
cally the sensitivity of AVO inversion to random noise
and errors in angles of incidence. Synthetic examples
show that random noise and angle errors can strongly
bias the parameter estimation. The use of singular value
decomposition offers a simple stabilization scheme to
solve for the elastic parameters.
The AVO inversion is applied to an OBS data set from
the North Sea. Special prestack processing techniques
are required for the success of S-wave AVO inversion.
The derived S-wave velocity and density contrasts help
in detecting the fluid contacts and delineating the extent
of the reservoir sand.
INTRODUCTION
Seismically derived elastic parameters of subsurface mate-
rials play an important role in seismic reservoir detection and
characterization. These parameters can be related to lithol-
ogy and fluid content using empirically derived relationships.
For example, Castagna et al. (1985) show that water-saturated
sandstones, siltstones, and shales fall approximately along a
line in the crossplot of P-wave velocity versus S-wave velocity,
whereas gas-saturated sandstones fall in different places on the
crossplot and are separated from the other media. From this
observation, Smith and Gidlow (1987) introduced the “fluid
factor” concept to highlight the gas-bearing sandstones using
weighted stacks of seismic data.
Amplitude variation with offset (AVO) analysis can be used
to evaluate elastic rock properties from seismic data. It is
usually performed on P-wave reflections. To fully describe
an isotropic elastic earth, three parameters are needed. Al-
though there are a number of possible choices, P-wave velocity,
S-wave velocity, and density are the most geophysically signif-
icant. The amplitude of a P-wave reflection is a function of all
three parameters. Unfortunately, using offset limited data, the
two velocities and density cannot be unambiguously resolved
(Tarantola, 1986; Debski and Tarantola, 1995). In fact, many
theoretical and numerical studies (Ursin and Tjaland, 1992;
De Nicolao et al., 1993; Jin et al., 1993) have demonstrated
that only two parameters can be effectively determined from
P-wave AVO data. Thus, in practice, one usually chooses a
parameterization with two parameters which have the most
influence on recorded seismic amplitudes, while the third one
can be neglected within the range of recorded incidence an-
gles. The two parameters used are generally P-wave and S-
wave impedances (Fatti et al., 1994) or intercept and gradi-
ent (Shuey, 1985). However, in some circumstances, these pa-
rameters are not adequate for hydrocarbon detection. In such
cases, the use of velocity and density may provide a remedy be-
cause they have detectable polarity-reversed contrasts for zero
impedances. Unfortunately, even if the impedances are well re-
solved from P-wave data, it is not possible to accurately convert
them to velocities due to the poor resolution—or absence—of
density information.
S-waves have been used in seismic exploration for several
decades. Recent developments in ocean-bottom multicompo-
nent recording technology can provide high quality P- and
S-wave data. The use of the S-wave allows new opportuni-
ties for AVO analysis to resolve elastic parameters. Since the
amplitudes of S-wave reflections depend only on two parame-
ters for an isotropic elastic medium, the numerical singularity

Published on Geophysics Online March 23, 2000. Manuscript received by the Editor May 17, 1999; revised manuscript received January 12, 2000.
∗
CGG American Services, Inc., 16430 Park Ten Place, Houston, Texas 77084. E-mail: jin@us.cgg.com.
‡CGG, 1 Rue Leon Migaux, 91314, Massy, France. E-mail: gcambois@group.cgg.com.
∗∗
CGG American Services, Inc., 16430 Park Ten Place, Houston, Texas 77084. E-mail: cvuillermoz@cgg.com.
°c 2000 Society of Exploration Geophysicists. All rights reserved.

1446
 S-velocity and Density Estimation
associated with parameter estimation may be less severe than
for the P-wave three-parameter case. Hence, S-wave velocity
and density may be more reliably estimated from S-wave AVO
analysis.
Multicomponent field records from an explosive source con-
tain converted and nonconverted P-wave and S-wave reflec-
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
tions. The converted PS-wave is one of the S-waves which are
generally of good quality and easy to acquire. Since we only
focus on the PS-wave in this paper, “S-wave” will from now
on refer to the converted PS-wave unless otherwise specified.
We present some particularities of S-wave AVO inversion and
propose a stable method for elastic parameter extraction. Syn-
thetic and real data examples are shown to illustrate the ben-
efits of S-wave AVO analysis.
REFLECTION COEFFICIENT OF CONVERTED PS-WAVE
Following Aki and Richards (1980), when the changes in
elastic properties at the boundary of two layers are small and
the angle of incidence does not approach 90◦ , the reflection co-
efficient, R, of the converted P-SV wave can be approximated
as
·
sin θ p 1Vs
R = 0.5 4(sin2 θ s − γ cos θ p cos θ s )
cos θ s Vs
¸ (4)
1ρ
− (cos 2θ s + 2γ cos θ p cos θ s ) , (1)
ρ
where θ p is the angle of incidence of the P-wave, and θ s is the
reflection angle of the S-wave (Figure 1); θ s can be calculated
from θ p using Snell’s law; V p , Vs , and ρ represent, respectively,
the average P-wave velocity, S-wave velocity, and density; 1Vs
and 1ρ are the changes in S-wave velocity and density at the
boundary of two layers; and γ is the velocity ratio Vs /V p . Unlike
the P-wave reflection coefficient, the P-S reflection coefficient
does not depend on the P-wave velocity contrast.
Figure 2 shows the amplitude versus angle (AVA) curves
for several 1Vs /Vs and 1ρ/ρ values. We can see that the
amplitudes behave differently for S-wave velocity variations
(Figure 2a) and density variations (Figure 2b). Generally, the
S-wave velocity contrasts contribute more to AVA variations
and produce more complicated curves. The Vs /V p ratio also
affects the AVA responses. Figure 2 was obtained for Vs /V p =
0.7.
From Figure 2, we can see that the reflection coefficient ver-
sus angle is nearly a linear function at small incidence angles
FIG. 1. P-S conversion diagram at an interface where the elastic
parameters change; θ p is the P-wave incidence angle, and θ s is
reflection angle of S-wave in SV mode.
1447
(<30◦ ). In fact, Zheng (1991) showed that equation (1) can be
simplified to
R = Csρ sin θ p (2)
when the incidence angles are small. Csρ is a combination of
velocity and density contrasts:
µ µ ¶¶
1 1ρ 1ρ 21Vs
Csρ = − + 2γ + . (3)
2 ρ ρ Vs
If γ = 0.5, this equation becomes Csρ = −1Is /Is , where Is is
S-wave impedance (Is = ρVs ).
AVO INVERSION OF CONVERTED PS-WAVE
Converted wave amplitude variation patterns reveal the
changes in S-wave velocity and density, and thus provide im-
portant lithologic information which is useful for hydrocarbon
discrimination. AVO inversion consists in finding the relative
variations 1Vs /Vs and 1ρ/ρ from seismic data. This can be
done by fitting equation (1) to PS-wave reflection amplitudes
of real common conversion point (CCP) gathers. Following the
procedure described by Smith and Gidlow (1987), the least-
squares fitting results in a system of two equations:
Am = D
where m = (1Vs /Vs , 1ρ/ρ)T . The expressions for matrix A and
vector D are given in Appendix A. Solving for 1Vs /Vs and
1ρ/ρ requires the inversion of matrix A. Before doing that,
we need to know how stable the system is. Seismic data are
always offset limited and noisy. The range of incidence angles
for deep points can become very small. Even though A is a
2 × 2 matrix, the inversion may suffer from numerical instabil-
ity. In fact, the result of AVO inversion is firstly affected by
the processing steps which transformed the recorded seismic
data into reflection coefficients. These steps may be inaccurate
and complicated by many factors including, for example, resid-
ual velocity errors, NMO stretch, thin bed tuning, geometrical
spreading, directivity patterns of sources and receivers, energy
dissipation, anisotropy, and so on. Ostrander (1984) discussed
some of these complications. Here, we address another im-
portant factor: the inherent instability of the linear system (4).
This instability persists even when the processing related errors
mentioned above are removed.
Sensitivity to random noise
Noise, even in small amounts, can be very troublesome when
solving an unstable system of equations. We now illustrate the
noise sensitivity with a numerical example. A gather of P-S
reflection coefficients (Figure 3a) was generated using equa-
tion (1), and some random noise was added. The true elastic
parameters 1Vs /Vs and 1ρ/ρ used to calculate the reflection
coefficients are represented by solid lines in Figures 3b and 3c.
V p , Vs , and ρ are constant and equal to 2000 m/s, 1000 m/s, and
2 g/cm3 , respectively. The offset-angle (θ p ) curves are plotted
on the reflection coefficient gather.
We now try to solve for 1Vs /Vs and 1ρ/ρ from the gener-
ated reflection coefficients, which is done by solving the linear
system (4). The results are represented by the thin lines in Fig-
ures 3b and 3c. To facilitate the comparison, the true model
(thick lines) and inverted model (thin lines) are drawn on the
 1448 Jin et al.
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FIG. 2. AVA curves for velocity changes (a) and density changes (b). (a) 1Vs /Vs varies from −0.2 to 0.16 with 1ρ/ρ = 0. (b) 1ρ/ρ
varies from −0.2 to 0.16 with 1Vs /Vs = 0. In both (a) and (b), γ = 0.7.

FIG. 3. The reflection coefficients (a) are generated using the band-limited model shown by the thick lines in (b) and (c). The
offset-angle curves are also shown in (a). The results of AVO inversion of the reflection coefficients are shown by the thin lines in
(b) and (c).
 S-velocity and Density Estimation
same panels. We can see that, for a uniform data noise level,
the quality of the result degrades rapidly with the decrease of
the range of incidence angles. In the deeper part, the inver-
sion fails to resolve the model parameters because data noise
is amplified to the level of signal. This example shows that the
straightforward resolution of system (4) is not stable when the
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
data contain noise, and angular aperture is restricted. Cambois
(1998) shows that the presence of random noise can even bias
fluid factor calculations and mislead AVO interpretation be-
cause the noise statistically biases the crossplot trend.
Sensitivity to errors in incidence angles
For the two-parameter P-wave inversion, the first param-
eter (P-wave impedance or AVO intercept) is determined
by the zero angle reflection coefficient, the second (S-wave
impedance or AVO gradient) by the variation of the reflec-
tion coefficient with incidence angle. Hence, the errors in the
incidence angles primarily affect the second parameter. This
becomes more complicated for S-wave AVO inversion, since
the intercept is always zero. Both S-wave velocity and density
are determined from the amplitude variation with incidence
angle. Errors in incidence angle determination can therefore
severely bias both parameters.
Unlike errors in the data, which only affect the vector D
of system (4), angle errors result in both inaccurate matrix A
and vector D. We now show how these errors bias the AVO
inversion. Figure 4 was obtained by inverting synthetic reflec-
tion coefficients using incidence angles containing systematic
errors. The changes in 1Vs /Vs and 1ρ/ρ were both set to 0.1
for the calculation of the reflection coefficients. The maximum
incidence angle is 40◦ . Figure 4 illustrates the relative errors
|m ε − m|/m where m ε is the inverted parameters in the pres-
ence of angle errors; m stands for true 1Vs /Vs and 1ρ/ρ. The
errors in incidence angles are highly amplified after inversion
(the thick lines are errors resulting from the stabilized inversion
which will be described in the next section).
Stabilization
Since the resolution of system (4) is very sensitive to small
errors, some sort of stabilization is required. Jin et al. (1993)
showed that singular value decomposition (SVD) can be effec-
FIG. 4. Influence of incidence angle errors on the AVO inver-
sion. Thin lines: conventional inversion. Thick lines: SVD sta-
bilized inversion.
1449
tively used for AVO stabilization. The main benefits of SVD is
to provide a precise way of analyzing a matrix and to yield a
stable but approximate inverse. When a matrix is numerically
unstable for inversion, it has at least one very small or zero
eigenvalue (see Appendix B). Figure 5 shows the two eigen-
values for matrix A in equation (4) as a function of the P-wave
incidence angle. The matrix is a 2 × 2 matrix, so it has only two
eigenvalues. The smaller eigenvalue rapidly decreases to zero
when the maximum incidence angle falls below 40◦ . The ratio
of largest and smallest eigenvalues is the condition number of
the matrix, which is a measure of the singularity of a matrix
(bold line in Figure 5). An error (noise) can be amplified by a
factor of the condition number (see Appendix C). The noise
level in Figures 3b and 3c is proportional to the condition num-
ber shown in Figure 5. A matrix is well-conditioned when its
condition number is not far from 1. Figure 5 shows that ma-
trix A is ill-conditioned for incidence angles smaller than 40◦
because the condition number becomes very high (>400). This
explains why noise increases rapidly when the maximum inci-
dence angle falls below 40◦ .
The ill-conditioned nature of S-wave AVO inversion for
small incidence angles data can readily be understood from
the AVA curves in Figure 2. These curves are nearly straight
lines passing through the origin for small incidence angles. The
slopes of the lines are uniquely determined by Csρ [see equation
(2)]. In this case, we can only determine Csρ , but not 1Vs /Vs
and 1ρ/ρ individually, because any pair of 1Vs /Vs and 1ρ/ρ
that gives the same Csρ value has the same AVA response.
The SVD stabilization method consists of adding a small
positive number to the smaller eigenvalues (see Appendix D).
This technique is widely used in geophysical inversions and is
also called SVD damping. Figure 6 shows the AVO inversion
using SVD stabilization of the same PS-wave data shown in
Figure 3a. The noise affecting the estimated parameters is dra-
matically reduced. In this example, 2% of the largest eigenvalue
is added to the smaller eigenvalue.
SVD damping is efficient in reducing the amplification of
noise with depth. In fact, it also helps in attenuating the devi-
ations resulting from errors in the incidence angles. Figure 4
shows that errors in 1Vs /Vs and 1ρ/ρ are reduced to a lower
level by SVD damping. However, even when stabilization is ap-
plied, errors in incidence angle remain an important factor of
FIG. 5. Eigenvalues (thin lines) and condition number (thick
line) for matrix A of equation (4). The eigenvalues are dis-
played in a logarithmic scale.
 1450 Jin et al.
inaccuracy for AVO inversion. For example, a 10% angle error
can produce a 40% error on the results. Hence, too simplified
angle estimation methods, such as the straight ray approxima-
tion, are not adequate for PS-wave AVO analysis.
As mentioned above, SVD damping only yields an approx-
imate solution. In other words, the damping compromises the
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
accuracy of AVO inversion, which is the price to pay for a re-
duced noise level. The tradeoff depends on the range of the
incidence angle. In general, when the maximum angle is larger
than 40◦ , the stabilized inversion always gives good estimates
for 1Vs /Vs and 1ρ/ρ (i.e., the bias introduced by the damp-
ing is small). For more limited ranges of incidence angle, the
inversion gives compensating values for 1Vs /Vs and 1ρ/ρ, so
that they yield correct slopes for the AVA curves.
REAL DATA EXAMPLE
The synthetic examples have made an optimistic case for
elastic parameter extraction. We now illustrate the AVO inver-
sion for S-wave velocity and density contrasts with a 2-D real
data example from the North Sea. The data were acquired by
placing multicomponent sensors on the bottom of the ocean.
The water depth is approximately 120 m. The reservoir is a
late Paleocene Rogaland oil-bearing sand which is hardly ob-
servable on conventional streamer data. The objective of this
S-wave AVO study is to detect fluid contacts and delineate the
extent of the oil-bearing sand using PS-wave–derived S-wave
velocity and density contrasts.
FIG. 6. SVD stabilized results (thin lines) compared to the true model (thick lines).
Prestack processing
The data were processed for AVO purposes. The preserved
amplitude processing sequence included mode-converted
wavefield separation, geometrical spreading correction, com-
mon conversion point gathering, overburden anisotropy com-
pensation, NMO, and wavelet calibration taking into account
well information. Refer to Vuillermoz and Granger (1998) for
more data acquisition and processing information. Here, we
just briefly describe the overburden anisotropy compensation
and residual NMO corrections, which are of the utmost impor-
tance for shear-wave AVO inversion.
Equation (1) is based on the assumption of an isotropic
earth; however, subsurface materials are usually anisotropic.
Direct inversion of the anisotropy parameters such as elas-
tic stiffness coefficients is impractical because it increases the
number of unknowns and thus makes the problem severely ill-
conditioned. In order to apply the proposed inversion method,
the overburden anisotropy effects must be removed. The most
prominent effect of anisotropy is shear-wave splitting. In the
processing sequence, shear-wave splitting was investigated by
rotation analysis of the multicomponent data, similar to the
technique proposed by Thomsen (1988). This analysis is able
to find the principal axes of anisotropy and the time delay of
splitting. The compensation for overburden anisotropy is then
achieved by applying the estimated time shift, amplitude, and
rotation corrections, which simulate the reverse of the shear-
wave splitting effect of the overburden.
 S-velocity and Density Estimation
The final prestack processing step is residual NMO correc-
tions. Like P-wave AVO analysis, slight NMO errors can dra-
matically contaminate the results. To prevent this from hap-
pening, a horizon-based statistical correction was applied on
the common conversion point (CCP) gathers to perfectly align
the strong reflection events at the reservoir level.
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
The measured shear-wave amplitudes are more sensitive to
noise and processing errors than P-wave amplitude. The suc-
cess of AVO inversion strongly depends on the success of pre-
served amplitude processing. Care has to be taken to preserve
amplitudes at every step of the processing sequence.
AVO inversion
Figure 7 shows one of 225 CCP gathers of PS horizontal
component data used in this study. The Vs /V p ratio is found to
be 0.3 from velocity analysis of PP and PS data. A smoothed
interval velocity model was derived from PP data. The offset-
angle relationship is determined by ray tracing: a P-wave ray
is first traced to the surface from a depth point with an an-
gle of incidence, then a S-wave ray is traced from the same
depth point using a departure angle determined by Snell’s law.
The offset corresponding to the incidence angle is the distance
between the two emerging points of the rays at the surface.
The procedure is repeated several times with different inci-
dence angles to obtain the offset-angle relationship. The offset
angle curves are plotted on the CCP gather in Figure 7 after
depth/time conversion. The maximum angle of incidence at the
reservoir level is around 45◦ . The previous condition number
analysis has shown that it is then possible to extract S-wave
FIG. 7. One of the 225 processed CCP gathers of PS horizontal
component data. The offset-angle curves are overlaid (angles
are in degrees).
1451
velocity and density contrasts from the CCP gathers. Due to
asymmetrical ray paths, the incidence angle is larger for PS
wave reflection than PP reflection for a given offset. The same
offset-angle relationship is used for all CCP gathers.
Figure 8 shows the S-wave velocity and density contrasts
obtained from the SVD-stabilized AVO inversion of the 225
CCP gathers. A damping factor of 2% of the first eigenvalue
is applied to the second one (see Appendix D). An AVO
anomaly indicator is also shown in this figure, which is the
difference between the velocity contrast and scaled density
contrast. The concept of this indicator is similar to the fluid
factor of traditional P-wave AVO analysis. We will hereafter
also call it fluid factor.
Interpretation
The reservoir sand is identified from previous work (Vuiller-
moz and Granger, 1998). The bottom of the reservoir is indi-
cated by arrows in Figure 8. S-wave velocity and density indi-
cate rock properties by responding differently to lithology and
fluid content. We can observe an overall uniform amplitude
of S-wave velocity contrast along the reservoir sand, except
for some local variations due to remaining noise. However, we
do not have such result for the density contrast; CDP 476 (ar-
row position) separates the reservoir sand into two segments
of different amplitudes, marked by red and green dashed lines
in Figure 8. The density contrast is stronger on the left than on
the right. Does this indicate a fluid contact or thin bed interfer-
ence? The thin-bed tuning effect does exist, but it is unlikely to
be the cause of the amplitude difference because the structure
is similar on both sides of this CDP location. Besides, if tuning
caused the amplitude effect on the density contrasts, it should
have a similar effect on the S-wave velocity contrast, which we
do not observe. The interpretation that CDP 476 represents the
separation between hydrocarbon-bearing sand and brine sand
seems to be more reasonable. With this interpretation, we can
easily explain why the density changes across this CDP location
while S-wave velocity does not. In fact, S-waves are less sensi-
tive to pore fill because the shear modulus only depends on the
rock matrix (Ensley, 1984). In contrast, the density depends on
both the matrix and the material in the pore space of the rock.
As shown in Figure 8, the reservoir sand can also be high-
lighted by the fluid factor 1F in terms of S-wave velocity and
density contrasts: 1F = 1Vs /Vs − α1ρ/ρ. The coefficient α is
the slope of the background trend on the crossplot of 1Vs /Vs
and 1ρ/ρ (Figure 9). This trend is referred to as the fluid line
in conventional AVO analysis. If a point lies on the fluid line,
1F is equal to zero. Conversely, if a point lies off the line,
then 1F 6= 0. Therefore, the fluid factor is a measure of how
“anomalous” a reflection is to the background trend.
The fluid factor section (Figure 8) shows that the strong re-
flections above and below the reservoir sand lie on or near
the fluid line because they have small fluid factor values. For
example, the strong reflection, marked by the yellow dashed
line, disappears from the fluid factor section because it is part
of the background trend (yellow cluster in Figure 9). The fluid
factor highlights all events uncorrelated with the background
trend. For example, the sea bottom multiples around 3800 ms
appear on the fluid factor section (Figure 8). The integrated
analysis of fluid factor and original sections helps identifying
hydrocarbon-related AVO anomalies.
 1452 Jin et al.

The crossplot also shows a clear average density difference of
the sand between the red and the green segments. This reveals
a pore fluid or lithology difference between the two segments.
From a well penetrating the green segment, we know that it is
an oil sand at the well location. No well exits in the area of the
red segment. Is it a brine sand? The AVO inversion results show
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
converted S-wave data in a marine environment and to use
these measurements to enhance extraction of rock properties.
We have presented a linearized AVO inversion of PS-wave
data. S-wave velocity and density are strongly coupled in
PS-wave AVO, so special care must be taken for parameter
resolution. Using synthetic data, we investigated the sensitiv-

that it is at least different from the sand in the green segment.
CONCLUSIONS
Numerous theoretical and numerical studies have reported
that neither velocity nor density can be well resolved from
conventional marine data where P-waves are dominant and the
angular aperture is small. One alternative is to use converted
S-wave. With the recent developments in ocean-bottom seismic
technology, it is now possible to acquire high-quality mode
ity of the PS-wave AVO inversion to random noise and errors
in angles of incidence. Resolution of S-wave velocity and den-
sity contrasts is stabilized using a SVD technique.
Under the assumption of an isotropic earth and small
changes of rock properties, only two parameters are respon-
sible for S-wave AVO, and they can be reasonably resolved
from high-quality data with greater than 40◦ angular aperture.
Application to a carefully processed OBS data set illustrates
the practical use of the S-wave AVO inversion and the benefit
of the combined use of S-wave velocity and density to seismic

FIG. 8. Density contrast (top), S-wave velocity contrast (middle), and AVO anomaly indicator (bottom) obtained from the inversion
of PS-wave data. The bottom of the reservoir sand is indicated by arrows. Note the amplitude difference in density contrast between
the two sides of the arrow. This difference is not so clear for S-wave velocity.
 S-velocity and Density Estimation 1453

REFERENCES
Aki, K., and Richards, P. G., 1980, Quantitative seismology: W. H.
Freeman and Co.
Cambois, G., 1998, AVO attributes and noise: pitfalls of crossploting:
68th Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded Abstracts.
Castagna, J. P., Bazle, M. L., and Eastwood, R. L., 1985, Relationships
between compressional-wave and shear-wave velocities in clastic sil-
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

icate rocks: Geophysics, 50, 571–581.

FIG. 9. Crossplot of 1Vs /Vs and 1ρ/ρ for the data shown in
Figure 8. A general trend is clearly observed. The slope of
the trend is 0.8 and is used for fluid factor calculation. The
red, green, and yellow clusters are the points of the horizons
marked by red, green, and yellow dashed lines, respectively, in
Figure 8. One possible interpretation for the clusters is shown.
reservoir characterization, especially for fluid contact detection
and changes in pore fluids.
ACKNOWLEGMENTS
We thank Renaud Daures and Pierre-Yves Granger for the
excellent processing of the converted wave data. We also thank
John Castagna, Keith Wrolstad, and one anonymous reviewer
for their valuable comments on the manuscript.
APPENDIX A
LEAST-SQUARES CURVE FITTING
De Nicolao, A., Drufuca, G., and Rocca, F., 1993, Eigenvalues and
eigenvectors of linearized elastic inversion: Geophysics, 58, 670–679.
Debski, W., and Tarantola, A., 1995, Information on elastic parameters
obtained from the amplitudes of reflected waves: Geophysics, 60,
1426–1436.
Ensley, R. A., 1984, Comparison of P-wave and S-wave seismic data—
A new method for detecting gas reservoir: Geophysics, 49, 1420–
1431.
Fatti, J. L., Vail, P. J., Smith, G. C., Strauss, P. J., and Levitt, P. R., 1994,
Detection of gas in sandstone reservoirs using AVO analysis: A 3-D
seismic case history using the Geostack technique: Geophysics, 59,
1362–1376.
Golub, G. H., and Reinsch, C., 1970, Singular value decomposition and
least squares solutions: Springer-Verlag.
Golub, G. H., and Van Loan, C. F., 1989, Matrix computations: Johns
Hopkins University Press.
Jin, S., Beydoun, W., and Madariaga, R., 1993, A stable elastic inver-
sion for marine data: 63rd Ann. Internat. Mtg., Soc. Expl. Geophys.,
Expanded Abstracts, 665–668.
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.
Ostrander, W. J., 1984, Plane-wave reflection coefficients for gas sands
at nonnormal angles of incidence: Geophysics, 49, 1637–1648.
Smith, G. C., and Gidlow, P. M., 1987, Weighted stacking for rock prop-
erty estimation and detection of gas: Geophys. Prosp., 35, 993–1014.
Shuey, R., 1985, A simplification of Zoeppritz equations: Geophysics,
50, 609–614.
Tarantola, A., 1986, A strategy for nonlinear elastic inversion for seis-
mic reflection data: Geophysics, 51, 1983–1903.
Thomsen, L., 1988, Reflection seismology over azimuthally anisotropic
media: Geophysics, 53, 304–313.
Ursin, B., and Tjaland, E., 1992, Information content of the elastic
reflection matrix: 62nd Ann., Mtg., Soc. Expl. Geophys., Expanded
Abstracts, 796–799.
Vuillermoz, C., and Granger, P. Y., 1998, Multicomponent seismic en-
hances characterization of offshore reservoirs: Word Oil, 219, April,
141–142.
Zheng, X., 1991, Approximation to P-P, P-SV, SV-SV, and SV-P reflec-
tions and transmissions: 61st Ann. Internat. Mtg., Soc. Expl. Geo-
phys., Expanded Abstracts, 1102–1105.

p " #2
1X
Consider N sampled reflection coefficients Rs(θi ) with inci-
sin θi ¡ ¢
N p
p
dence angle θi , i = 1, 2, . . . , N . The objective is to find a set of A22 = s p
cos 2θi + 2γ cos θi cos θis
,
p
1Vs /Vs and 1ρ/ρ which best fits Rs(θi ) via equation (1). The 4 i=1 cos θis
solution in the least-squares sense is obtained by minimizing
"Ã !2
the following expression: X
N p
sin θi ¡
X ¡ ¡ A12 = A21 = − sin2 θis
p¢ ¡ p ¢¢2
N
Rs θi −R θi , (A-1) i=1
cos θis
i=1 #
p p ¢¡ p ¢
where R(θi )is the theoretical reflection coefficient given by −γ cos θi cos θi
s
cos 2θis + 2γ cos θi cos θi
s
,
equation (1). Taking the derivatives of expression (A-1) with
respect to 1Vs /Vs and 1ρ/ρ and setting them to zero gives " #
system (4): Am = D. Explicitly, it can be written as XN
sin θi ¡ 2 s
p
¢ ¡ ¢
 D1 = 2
p p
sin θi − γ cos θi cos θiS Rs θi ,
1V cos θis
à ! s
 Ã ! i=1
A11 A12  Vs  D1 " #
 
 1ρ  = , (A-2) 1X N
sin θi ¡
p
¢ ¡ ¢
A21 A22   D2 D2 =
p p
cos 2θis + 2γ cos θi cos θis Rs θi ,
ρ 2 i=1 cos θis

where and
" #2
X
N
sin θi ¡ 2 s
p
p ¢ Vs
A11 = 4 sin θi − γ cos θi cos θis , γ = .
i=1
cos θis Vp
 1454 Jin et al.

APPENDIX B
SVD ANALYSIS

The matrix A of equation (A-2) is symmetrical and is a func- where the vectors V1 = (ν11 , ν12 )T and V2 = (ν21 , ν22 )T are the
tion of incidence angles and the ratio γ . The sensitivity of A orthonormal eigenvectors of the system
determines how accurately we can resolve the elastic parame-
Downloaded 04/08/19 to 140.115.21.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ters 1Vs /Vs and 1ρ/ρ in presence of noise. SVD is a valuable AAT Vi = λi2 Vi , i = 1, 2, (B-2)
tool to address the numerical difficulties encountered in solv-
ing equation (A-2). The SVD for general matrices can be found and λi the corresponding eigenvalues. Eigenvalues are sorted
in many publications (e.g., Golub and Reinsch, 1970; Lines and as λ1 ≥ λ2 . The solution of equation (4) is then
Treitel, 1984). We only consider here the special case of ma-
trix A. X
2
VT D
SVD factors A into a product of three matrix, A = V3VT , m = A−1 D = V3−1 VT D = i
Vi . (B-3)
or explicitly, i=1
λi
à ! à !à !à !
A11 A12 ν11 ν21 λ1 0 ν11 ν12 We see that the solution m is the weighted sum of the two
= ,
A21 A22 ν12 ν22 0 λ2 ν21 ν22 eigenvectors with weights ViT D/λi . It is clear that the solution
(B-1) is not stable if λi is zero or very small.

APPENDIX C
CONDITION ANALYSIS

The system contains errors of different origins, which con- κ(A) is defined as
taminate both A and D of equation (2). A precise measure of
λ1
the sensitivity can be obtained by considering the perturbed κ(A) = kAk kA−1 k = . (C-2)
linear system λ2

(A + δA)mε = D + δD, (C-1) As demonstrated by Golub and Van Loan (1989), the rela-
tive error of the solution is bounded by
where δA is the perturbation of matrix A and δD is the per- µ ¶
turbation of the weighted stack data; mε is the solution of the kmε − mk kδAk kδDk
≤ κ(A) + . (C-3)
perturbed system. For real seismic data application, we always kmk kAk kDk
have a perturbed system (C-1) instead of the exact system (4).
For example, errors in incidence angle and velocity estimations Thus, the relative errors in m can be κ(A) times the rela-
contribute to δA and δD. Random noise, multiples, residual tives error in A and D. In this sense, the condition number
NMO, and other unaccounted for factors in true amplitude κ(A) quantifies the sensitivity of the system Am = D. If κ(A)
recovery contribute to δD. is large, then A is said to be an ill-conditioned matrix. In-
To describe the sensitivity of the solution m to perturbations, versely, matrices with small condition number are said to be
we need to introduce the condition number κ(A). For L2-norm, well-conditioned.

APPENDIX D
SVD STABILIZATION

From the previous section, we see that the solution m X2

ViT D
is increasingly sensitive to errors in A and D if the ma- mη = V3−1 VT D = Vi .
trix A is ill-conditioned. The SDV stabilization consists in
λ +η
i=1 i
adding a small number η to λi . The stabilized solution The number η is called the damping factor and can be added
is only to λ2 .