GEOPHYSICS, VOL. 71, NO. 5 共SEPTEMBER-OCTOBER 2006兲; P. E49–E55, 10 FIGS.

10.1190/1.2227617

Annual Meeting Selections

Linearized amplitude variation with offset „AVO… inversion

with supercritical angles

Jonathan E. Downton1 and Charles Ursenbach2

scure the overlying zone of interest. The traditional way of dealing

ABSTRACT
Contrary to popular belief, a linearized approximation of
the Zoeppritz equations may be used to estimate the reflec-
tion coefficient for angles of incidence up to and beyond the
critical angle. These supercritical reflection coefficients are
complex, implying a phase variation with offset in addition to
amplitude variation with offset 共AVO兲. This linearized ap-
proximation is then used as the basis for an AVO waveform
inversion. By incorporating this new approximation, wider
offset and angle data may be incorporated in the AVO inver-
sion, helping to stabilize the problem and leading to more ac-
curate estimates of reflectivity, including density reflectivity.
with the condition is to limit the range of angles used in the AVO in-
version. However, because the critical angle may be as small as 25°,
the reliability of the estimates of the AVO inversion is limited.
Limiting the angles used in AVO inversion is a result of the com-
mon and mistaken belief that the Aki and Richards 共1980兲 linearized
approximation 共their equation 5.44兲 of the Zoeppritz equation is val-
id only for subcritical angles of incidence. For example, de Nicolao
et al. 共1993兲 state that the linearized approximation is valid only for
precritical angles. Aki and Richards 共1980兲 state as a precondition
for using the linearized approximation 共their equation 5.44兲 that
none of the angles of incidence and transmission may be near 90°,
thus precluding the critical angle.
This seems to be supported if the Aki and Richards 共1980兲 equa-
tion for the PP reflection coefficient R共p兲 is written in terms of hori-
zontal slowness p:

INTRODUCTION
1 ⌬␳ 1 ⌬␣
R共p兲 = 共1 − 4␤2 p2兲 +
Supercritical reflection coefficients arise in situations of interest
2 ␳ 2共1 − ␣2 p2兲 ␣
to explorationists. In trying to determine density through amplitude ⌬␤
variation with offset 共AVO兲, large angles of incidence and offset are − 4␤2 p2 , 共1兲
used in the inversion 共Downton, 2005兲. Most AVO inversion ap- ␤
proaches using a linearized approximation require incidence angles
where ␳, ␣, and ␤ are the average density, P-wave, and S-wave ve-
significantly smaller than the critical angle. Estimate reliability is
proportional to the range of angles used in the inversion, so restrict- locities and where the differentials ⌬␳, ⌬␣, and ⌬␤ are the change
ing the maximum angle to be significantly tighter than the supercriti- in layer properties for the density and velocities. As written, the re-
cal angle reduces this reliability. flection coefficient calculated by equation 1 must be real for homo-
In addition, the critical angle plays an important role in heavy-oil geneous waves 共i.e., p is real兲. However, supercritical reflection co-
plays in northeastern Alberta. Reflections at the Paleozoic unconfor- efficients are complex, making use of the complex numbers to con-
mity often are critical because of the large velocity contrast between tain the phase information.
the overlying clastics and underlying carbonates. This becomes a If, however, equation 1 is parameterized in terms of the average
problem because the large supercritical reflection coefficients ob- angle ␪,

Manuscript received by the Editor September 22, 2005; revised manuscript received February 27, 2006; published online August 28, 2006.
1
Veritas GeoServices, 715 Fifth Ave. SW, Suite 2200, Calgary, Alberta, T2P 5A2 Canada. E-mail: jon_downton@veritasdgc.com.
2
University of Calgary, CREWES 共Consortium for Research in Elastic Wave Exploration Seismology兲, Department of Geology and Geophysics, 2500 Univer-
sity Dr. NW, Calgary, Alberta T2N 1N4, Canada. E-mail: ursenbach@crewes.org.
© 2006 Society of Exploration Geophysicists. All rights reserved.

E49
E50 Downton and Ursenbach
Figure 1. Input velocity and density well-log information used to
generate the synthetic model in Figure 2.
Figure 2. Synthetic model generated with 共a兲 the Zoeppritz equation and 共b兲 equation 2.
共c兲 The difference between 共a兲 and 共b兲.
Figure 3. Comparison of extracted amplitude at selected times from Figures 2a and b for
共a兲 160 ms, 共b兲 520 ms, and 共c兲 684 ms. The x-axis is offset in meters.
R共␪兲 =
1
2
冉 ␤2
1 − 4 2 sin2␪
␣
⌬␳
␳
+ 冊 1 ⌬␣
2 cos2␪ ␣
␤2 2 ⌬␤
−4 sin ␪ , 共2兲
␣2 ␤
the reflection coefficient can become complex. This is a conse-
quence of the fact that ␪ is the average of incidence and transmitted
angles, 共 ␪r + ␪t兲/2. These are related by Snell’s law, sin ␪r /␣1
= sin ␪t /␣2, where ␣1 and ␣2 are the P-wave velocities above and be-
low the interface. The incidence angle ␪r is always real, but the trans-
mitted angle becomes complex for angles beyond the critical angle:
␪t =
␲
2
− i cosh−1
␣2
␣1
冉sin ␪r , 冊 ␣2
␣1
sin ␪r ⱖ 1. 共3兲
To illustrate this, a simple convolutional model was generated us-
ing both the Zoeppritz equation and equation 2 to generate the reflec-
tion coefficients. The density, P-wave, and S-wave velocity-layer in-
formation 共Figure 1兲 used to construct this model are based on a well
log from northeastern British Columbia, Canada 共Downton, 2005兲.
The seismic synthetic data are shown with a 10/
14–90/110-Hz trapezoidal zero-phase filter. Both
synthetics are shown without moveout, geomet-
ric spreading, or other losses, highlighting differ-
ences that arise solely from the linearized approx-
imation. The synthetic data generated using equa-
tion 2 共Figure 2b兲 closely approximate the model
generated by using the Zoeppritz equation 共Fig-
ure 2a兲, as evidenced by the difference display of
the two 共Figure 2c兲. Figure 3 shows amplitude ex-
tractions taken from the two models at time sam-
ples corresponding to interfaces that include su-
percritical angles. The synthetic model generated
using equation 2 closely matches the results gen-
erated using the Zoeppritz equation.
METHOD
The above discussion suggests two issues that
must be dealt with if supercritical angles are to be
included in the AVO inversion. It is evident that
the phase changes as a function of angle of inci-
dence. This is a problem for traditional AVO in-
version schemes, which work on a sample-by-
sample basis. These AVO inversions ignore the
wavelet and so cannot handle these phase chang-
es. This suggests that some sort of AVO wave-
form inversion 共Simmons and Backus, 1996; Bu-
land and Omre, 2003; Downton and Lines, 2004兲
should be used.
Second, the amplitudes of the synthetic data
are very large at the far offsets compared to the
near offsets. Real seismic data do not behave like
this because of geometric spreading, attenuation,
and other losses. To address this imbalance, the
seismic data are normally preconditioned to cor-
rect for these losses prior to AVO inversion. How-
ever, including angles close to critical makes
 Linearized AVO inversion E51

these corrections unstable. Figure 4 shows the geometric spreading ent noise. Prior Gaussian reflectivity variables are assumed that may
losses calculated following Červeny 共2001兲. Note the large change be parameterized compactly by the model covariance matrix
in scalars close to the critical angle. Applying the inverse loss func- 共Downton and Lines, 2001兲:
tion as a gain correction to the seismic data introduces high-frequen-

cy noise into the seismic data. It is more stable to apply the geometric
spreading losses as a forward operator of the AVO model before ap-
plying the wavelet. This is done easily as part of the AVO waveform
inversion.
The Downton and Lines 共2004兲 approach is modified easily to in-
corporate both of these considerations. The ray theoretic convolu-
tional model is the basis for this forward model. The model assumes
plane waves and presupposes the earth is composed of a series of flat,
homogeneous, isotropic layers. Gain corrections such as spherical
divergence, directivity, and array corrections can be incorporated
easily into this model, whereas multiples, converted waves, head
waves, and transmission and absorption losses are not incorporated
and must be addressed through prior processing. The linear model is
parameterized in terms of the fractional perturbations, defined as the
change in layer properties divided by the average layer properties.
This definition is the same form as the zero-offset reflectivity for
both the P-wave and S-wave impedances. As such, for brevity and
conciseness we call these fractional perturbations reflectivity. Hav-
ing defined this terminology, the linear model of Downton and Lines
共2004; their equation 1兲 is
冤 冥
␴r2p ␴ r prs ␴ r prd
C m = ␴ r prs ␴r2s ␴ rsrd 共5兲
␴ r prd ␴ rsrd ␴r2d
The diagonal terms of this matrix describe the relative power of the
different reflectivity series, while its off-diagonal terms describe the
degree of correlation between these different reflectivity series. The
model covariance matrix can be constructed using empirical rock
physical relationships 共Downton, 2005兲, such as the mudrock rela-
tionship 共Castagna et al., 1985兲 and the Gardner velocity-density re-
lationship 共Gardner et al., 1974兲.
The inverse problem is solved using a preconditioned conjugate
gradient, with the number of conjugate gradient iterations solved
acting as the penalty term 共Hansen, 1998兲. In the conjugate gradient
algorithm, each iteration adds independent information. This addi-
tional information is associated with smaller and smaller eigenval-
ues, and as such there is greater noise amplification with these later
iterations. Thus, by selecting the number of conjugate gradient itera-
tions, the user controls the trade-off between resolution or the

冤冥冤 冥冤 冥
d1 WN1F1 WN1G1 WN1H1 rP amount of independent information and the noise level.
To speed up the convergence of the problem, preconditioning is
⯗ = ⯗ ⯗ ⯗ rS , 共4兲
performed. This consists of solving for a set of transformed parame-
dN WNNFN WNNGN WNNHN rd ters that orthogonalizes the model covariance matrix Cm. To a good
approximation, this involves solving for the transformed variables:
where r P, rS, and rd are the P- and S-wave impedances and density
P-impedance reflectivity, fluid stack 共Smith and Gidlow, 1987兲, and
reflectivity. These are vectors whose elements correspond to differ-
the difference between the density and P-wave velocity reflectivity
ent time samples. Likewise, the elements of the data vector dn repre-
共density-difference reflectivity兲. The original parameters from equa-
sent the processed seismic data for the nth offset for the correspond-
tion 4 are obtained by the inverse transformation of the results of the
ing time samples. The block matrices describe the physics of the
problem. The diagonal matrices F, G, and H contain weights that de-
scribe how amplitude changes as a function of offset. These complex
weights follow from equation 2. Following Claerbout 共1992兲, the
block matrix Nn describes the kinematics or the normal moveout
共NMO兲. This operator can be constructed using whatever offset trav-
eltime relationship one desires, including the traveltimes generated
from ray tracing. To invert data at large angles of incidence, it is im-
portant to position the event correctly without introducing residual
NMO. Implicit in this derivation is that the velocity is known before-
hand and that static corrections are applied. Last, W is a convolution
matrix that contains the source wavelet. To ensure the output of this
operation is real, Hermitian symmetry is imposed in the frequency
domain using a Hilbert transform. Gain corrections such as geomet-
ric spreading may be implemented by appropriately scaling F, G,
and H individually.
Applying these three operators in series, the block matrices Fn,
Gn, and Hn model the offset-dependent reflectivity from the zero-
offset reflectivity, Nn applies NMO, and W convolves the offset-de-
pendent reflectivity with the source wavelet modeling the band-lim-
ited seismic data with NMO. The inversion of equation 4 can be
thought of as three separate inversion problems: deconvolution, in-
verse NMO, and AVO inversion.
The actual inverse problem is formulated following a Bayesian
methodology 共Downton, 2005兲, where the likelihood function is
modeled with a variety of long-tailed distributions to give the algo- Figure 4. Geometric spreading losses shown with a log scale. The
rithm better noise-rejection characteristics in the presence of coher- critical angle generates a poorly illuminated zone shown in black.
E52 Downton and Ursenbach
inversion. The transformed reflectivity is used only to speed up the
convergence of the problem. In addition, some insight may be
gained by observing the behavior of the transformed variables as a
function of the iteration number. The geophysicist gains an under-
standing of the relative importance of the fluid stack and the density-
difference reflectivity to the overall solution as a function of noise
level in the solution. It is possible to see whether it is the seismic data
or the constraints that drive the density solution.
Figure 5. The input P-wave velocity and density curves used to gen-
erate the synthetic gathers along with their reflectivity. Note the den- higher-order NMO corrections. The residual NMO issue may be
sity reflectivity Rd is uncorrelated with the P-wave velocity reflectiv- dealt with using some automatic higher-order NMO correction such
ity RVP at a number of significant interfaces, e.g., 0.34, 0.42, and
0.67 s.
Figure 6. 共a兲 Synthetic seismic data input to the AVO waveform inversion along with 共b兲
the estimated data and 共c兲 the difference between 共a兲 and 共b兲.
RESULTS
The synthetic model based on the well logs from Figure 1 is used
to test theAVO waveform inversion algorithm. Figure 5 shows the P-
wave velocity and density curves along with their corresponding re-
flectivity series. There are a number of interfaces for which the den-
sity reflectivity is uncorrelated with the P-wave velocity reflectivity.
The constraints assume a relationship between these parameters.
Therefore, to estimate the density at these locations accurately, the
AVO inversion must rely solely on the seismic data rather than on the
constraints.
The synthetic model shown in Figure 2b was rerun with moveout,
geometric spreading, and other losses. Coherent and random noise
were added to the result 共Figure 6a兲 to give a S/N ratio of 10:1. This
was used as input into the AVO waveform inversion. Figure 6b
shows the estimated data generated by theAVO inversion, while Fig-
ure 6c shows the difference between the estimated and actual data.
TheAVO waveform inversion estimated the original data quite accu-
rately. Most of the difference arises from the inversion not being able
to estimate the random and coherent noise, although some small the-
oretical error is evident for two supercritical reflections.
Figure 7 compares the estimated reflectivity for the new algorithm
and the reference zero-offset reflectivity. The estimated reflectivity
is almost a direct overlay of the ideal zero-offset reflectivity. Note
that the density reflectivity is estimated accurately even in areas
where the density reflectivity follows a different trend than the
P-wave velocity reflectivity. This same model was run with S/N ra-
tios of 4:1, 2:1, and 1:1. Density was estimated accurately for S/N ra-
tios down to 2:1. The results are superior to those obtained from an
equivalent AVO inversion performed sample by sample.
The robustness of theAVO inversion in the presence of substantial
noise is attributable to the extra stability afforded by incorporating
wide-angle seismic reflectivity information. However, in practice
other obstacles impede incorporating this information. Processing
this wide-angle information is a problem. The potential exists for
significant NMO stretch and residual NMO, even when applying
as offered by Swan 共2001兲 or Ratcliffe and Roberts 共2003兲. Our ap-
proach builds the NMO stretch into the forward-modeling operator
共Downton, 2005兲.
To illustrate the significance of addressing
NMO stretch, Figure 8 shows the results of per-
forming a time sample-by-sample AVO inversion
on NMO-corrected gathers with a S/N ratio of
10:1. The inversion was performed using angles
up to 45° to avoid supercritical reflections. The
resultant reflectivity estimates are a poorer match
to the reference zero-offset reflectivity than those
obtained using the AVO waveform inversion
共Figure 7兲. This poorer result is mostly the result
of ignoring the NMO stretch and offset-depen-
dent tuning, introducing theoretical error. This er-
ror or noise is amplified by the AVO inversion,
creating the ringing apparent in the estimates in
Figure 8.
The AVO waveform inversion was also tested
on a small 3D seismic survey shot over the Nisku
pinnacle reef in Alberta, Canada. Supercritical re-
flections occur at the interface between the Banff
 Linearized AVO inversion E53

shales and the Wabamun carbonates about 80 ms above the zone of duces a slight bias to the wide-angle reflectivity response 共Haase and
interest. The AVO waveform inversion was performed using aver- Ursenbach, 2005兲, especially for supercritical reflections. Subse-
age angles up to and including 50° with the objectives of 共1兲 stabiliz- quently ignoring this difference will bias AVO estimates. This issue
ing the inversion in the presence of coherent noise
and 共2兲 estimating the density reflectivity section.
Figure 9a shows a representative input gather in
the AVO waveform inversion, Figure 9b shows
the forward-modeled data from the AVO parame-
ter estimates for the same gather, and Figure 9c
shows the misfit. By using the large range of an-
gles, the AVO waveform inversion could identify
and reject significant coherent noise. Figure 10
shows the estimated P- and S-wave impedance
reflectivities and density reflectivity. The Nisku
pinnacle reef is the structural high in the middle
of the section at 1.12 s. By including a greater
range of angles in the inversion, the AVO wave-
form inversion can estimate a more accurate
S-impedance reflectivity than that which could be
obtained using a traditional time sample-by-sam-
ple AVO inversion.
As discussed in the Method section, the inver-
sion is solved using a preconditioned conjugate
gradient for the transformed variables: P-wave
impedance reflectivity, fluid stack, and differ-
ence-density reflectivity. In observing the solu-
tion as a function of iteration, most of the early it-
erations add information first to the P-wave im-
pedance reflectivity, then to the fluid stack. Only Figure 7. AVO waveform inversion estimate 共dashed gray line兲 of the P-wave impedance
in later iterations is any information added to the reflectivity R P, S-wave impedance reflectivity RS, density reflectivity Rd, and fluid stack
density-difference reflectivity. Further, angles reflectivity RFl compared to reference zero-offset reflectivity 共solid black line兲. Note the
beyond 45° must be included for this to occur for good agreement.
solutions with reasonable noise levels. In Figure
10c most of the independent information con-
tained in the density reflectivity is between 1.0
and 1.2 s. The rest of the density reflectivity gen-
erally tracks the P- and S-wave impedance reflec-
tivity. Unfortunately density logs were unavail-
able for the well locations within this survey, but
the density response was consistent with nearby
logs and known well control.

DISCUSSION
The ability to model supercritical reflection co-
efficients accurately removes one obstacle in us-
ing wide-angle reflectivity data for AVO inver-
sion. However, other issues remain. Our ap-
proach assumes isotropy and plane-wave rather
than spherical-wave reflection coefficients and
ignores multiples, converted waves, and head
waves. Ideally, the multiples, converted waves,
and head waves are removed by processing be-
fore AVO inversion; but if residual energy re-
mains, the robust inversion is reasonably insensi-
Figure 8. Three-term AVO inversion estimate 共dashed gray line兲 of the P-wave velocity
tive to coherent noise 共Figure 6兲. reflectivity RVP, S-wave velocity reflectivity RVS, density reflectivity Rd, and fluid stack
Using plane-wave reflection coefficients rather reflectivity RFl compared to reference zero-offset reflectivity 共solid black line兲. The esti-
than spherical-wave reflection coefficients intro- mate is markedly poorer because NMO stretch is ignored.
E54 Downton and Ursenbach
might be dealt with using an approach of Ursenbach et al. 共2005兲,
which demonstrates an efficient approximation to calculate spheri-
Figure 9. 共a兲 Real seismic data input to the AVO waveform inversion along with 共b兲 the
estimated data. 共c兲 The difference between 共a兲 and 共b兲.
Figure 10. AVO inversion estimates of 共a兲 P-wave impedance reflectivity, 共b兲 S-wave im-
pedance reflectivity, and 共c兲 density reflectivity.
cal wave reflection coefficients from plane-wave reflection coeffi-
cients.
Last, a linearized version of the anisotropic
Zoeppritz equations may be used rather than
equation 2. Note, however, the vertical TI param-
eterization of Rüger 共2002兲 leads to an underre-
solved inverse problem where only three of the
five unknown parameters may be solved for. In
addition to the modified reflectivity expression,
anisotropy must be accounted for as part of the
angle-of-incidence calculation and geometric-
spreading correction. Both of these might be ac-
counted for in a fashion similar to Xu et al.
共2005兲.
CONCLUSIONS
By performing a slight change in parameteriza-
tion, the Aki and Richards linearized approxima-
tion of the Zoeppritz equations can be used to
model supercritical reflection coefficients accu-
rately. This approximation is valid for weak frac-
tional perturbations in velocity and density. How-
ever, we do not believe this assumption is too re-
strictive because most well logs digitized at the
sample rate of the seismic data meet this criterion.
Further, this approximation closely matches the
actual values calculated by the Zoeppritz equa-
tions for a synthetic model based on a well log
from Canada.
This linearized approximation is the basis for
our AVO inversion. The more accurate approxi-
mation permits the use of wider-offset seismic
data, improving the stability of the inversion
problem. In the synthetic example we obtained
accurate density reflectivity estimates even with
prestack S/N ratios as low as 2:1. Likewise, we
obtained more accurate estimates of the S-wave
impedance reflectivity for the real seismic data
example in the presence of coherent noise and de-
termined a geologically reasonable density re-
flectivity section.
This linearized approximation removes one
obstacle from using wide-angle seismic data in
AVO inversion. Other practical and theoretical is-
sues limit the application of this approach. How-
ever, incorporating this new approximation in
AVO inversion should provide more precise re-
flectivity estimates with less bias than merely ig-
noring or muting the supercritical reflections.
ACKNOWLEDGMENTS
J. D. would like to thank Veritas for permitting
him to publish this paper, while C. U. would like
to thank the sponsors of CREWES for their sup-
port of his research. In addition the authors ac-
knowledge valuable discussions and insight from
Dave Gray, and they thank Jan Dewar for many
helpful editorial suggestions.
 Linearized AVO inversion
REFERENCES
Aki, K., and P. G. Richards, 1980, Quantitative seismology: Theory and
methods: W. H. Freeman and Company.
Buland, A., and H. Omre, 2003, Bayesian linearized AVO inversion: Geo-
physics, 68, 185–198.
Castagna, J. P., M. L. Batzle, and R. L. Eastwood, 1985, Relationships be-
tween compressional-wave and shear-wave velocities in clastic silicate
rocks: Geophysics, 50, 571–581.
Červeny, V., 2001, Seismic ray theory: Cambridge University Press.
Claerbout, J. F., 1992, Earth soundings analysis: Processing versus inver-
sion: Blackwell Scientific Publications.
de Nicolao, A., G. Drufuca, and F. Rocca, 1993, Eigenvectors and eigenval-
ues of linearized elastic inversion: Geophysics, 58, 670–679.
Downton, J. E., 2005, Seismic parameter estimation from AVO inversion:
Ph.D. dissertation, University of Calgary.
Downton, J., and L. Lines, 2001, Constrained three parameter AVO inversion
and uncertainty analysis: 71st Annual International Meeting, SEG, Ex-
panded Abstracts, 251–254.
——–, 2004, Three-termAVO waveform inversion: 74thAnnual Internation-
al Meeting, SEG, Expanded Abstracts, 215–218.
Gardner, G. H. F., L. W. Gardner, and A. R. Gregory, 1974, Formation veloc-
E55
ity and density — The diagnostic basics for stratigraphic traps: Geophys-
ics, 39, 770–780.
Haase, A. B., and C. P. Ursenbach, 2005, Spherical-wave AVO modeling in
elastic and anelastic transversely isotropic 共VTI兲 media: 75th Annual In-
ternational Meeting, SEG, Expanded Abstracts, 316–319.
Hansen, P. H., 1998, Rank-deficient and discrete ill-posed problems: Society
for Industrial and Applied Mathematics.
Ratcliffe, A., and G. Roberts, 2003, Robust, automatic, continuous velocity
analysis: 73rd Annual International Meeting, SEG, Expanded Abstracts,
2080–2083.
Rüger, A., 2002, Reflection coefficients and azimuthal AVO analysis in an-
isotropic media: Geophysical monograph series 10: SEG.
Simmons, J. L., Jr., and M. M. Backus, 1996, Waveform-based AVO inver-
sion and AVO prediction error: Geophysics, 61, 1575–1588.
Smith, G. C., and P. M. Gidlow, 1987, Weighted stacking for rock property
estimation and detection of gas: Geophysical Prospecting, 35, 993–1014.
Swan, H. W., 2001, Velocities from amplitude variations with offset: Geo-
physics, 66, 1735–1743.
Ursenbach, C. P., A. B. Haase, and J. E. Downton, 2005, An efficient method
for AVO modeling of reflected spherical waves: 75th Annual International
Meeting, SEG, Expanded Abstracts, 202–205.
Xu, X., I. Tsvankin, and A. Pech, 2005, Geometrical spreading of P-waves in
horizontally layered, azimuthally anisotropic media: Geophysics, 70,
D43–D53.