You are on page 1of 6

Can P-wave AVO be quantitative?

GUILLAUME CAMBOIS, CGG, Paris, France

Fundamentals of P-wave AVO. Under the small incidence


angle approximation, Shuey showed in 1985 that the
prestack behavior of P-wave reflection data follows the simple formula:

R( ) = A + B sin 2
where is the angle of incidence. A and B, respectively
called intercept and gradient, are related to elastic rock
properties:

A = Rp =

I p Vp
=
+
2 I p 2Vp 2

B = Rp 2 Rs =

I p Is

Is
2Ip

where RP and RS are respectively P- and S-reflectivity, IP and


IS P- and S-impedance, VP and VS P- and S-velocity, and
the density. Note that the second formula is only valid when
the VP/VS ratio is 2. Using these two formulae, we can rewrite
Shueys equation as:

R( ) = Rp (1 + sin 2 ) 2 Rs sin 2
R( ) = Rp cos2 + 2( Rp Rs )sin 2
The first expression is the small angle approximation of
Fattis equation, and the second expression is equivalent to
the formula derived by Verm and Hilterman in 1995. They
called the second term on the right-hand-side Poissons
reflectivity (PR) because:

2( Rp Rs ) =

I p Is (Vp / Vs )

=
Ip
Is
Vp / Vs

Castagna and Smith showed in 1994 that PR is identical to


Smith and Gidlows fluid factor when VP/VS is 2.
Using any of the above equations, it is theoretically possible to compute shear or Poissons reflectivity from P-wave
data alone. This information is of course invaluable for fluid
prediction and hydrocarbon detection. The alternative way
of getting access to such data is through multicomponent
acquisition, which is a lot more costly. It is therefore critical
to assess whether these reflectivity series can be extracted
faithfully (i.e., quantitatively) from the P-wave data.
1246

THE LEADING EDGE

NOVEMBER 2000

30
gradient
decibels

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

t SEGs 1998 Annual Meeting in New Orleans, Ali Tura


and I organized a workshop titled Can P-wave AVO be
quantitative or do we need multicomponent data? At the
end of the workshop, after an afternoon of talks and debates,
we asked the audience to vote for either approach. Not surprisingly, the raised hands showed an almost perfect 50-50
split. This controversial issue is still not settled and, if this
years SEG convention is any indication, the debate is alive
and well. The success of AVO and direct hydrocarbon indicators (DHIs) in finding oil and gas has been firmly established over the years and is well documented in numerous
papers. However, the jury is still out on the quantitative
aspects. I claim that achieving quantitative results from Pwave AVO may be possible, but it will take a lot more effort
than is currently being done. Thus, my answer to the title
question is a definite and resounding maybe.

15
0

intercept

-15
stack
0.0

1.5

3.0

4.5

6.0

time in seconds
Figure 1. Attribute response to noise (or equivalently
standard deviation) for a typical marine geometry. The
mute zone roughly follows the 30 incidence limit, and
ends at 2700 ms. The intercept response to noise
plateaus 3.5 dB above the stack maximum attenuation
while the gradient boosts noise, especially below the
mute zone.
Reliability of AVO attributes. The standard way to proceed is to process the seismic data with a preserved amplitude sequence (preferably involving prestack time
migration). Once the data are corrected for propagation
effects (spherical divergence, NMO), the velocity field is raytraced to derive a relationship between offset, arrival time,
and incidence angle. It is then simply a matter of fitting
Shueys equation to the prestack amplitude samples to
obtain the desired shear or Poissons reflectivity. Although
it may not be the most robust norm, I will use in this paper
a least-squares curve fitting method because it provides an
elegant formalism to assess attribute reliability. (For all the
mathematical derivations, please refer to a couple of
Expanded Abstracts I wrote for the 1998 and 2000 SEG conventions.)
The reliability of AVO attributes is only a function of
geometry (number of traces, range of incidence angles) and
does not depend on the input data. Figure 1 shows the
response to noise (also called standard deviation) of stack,
intercept, and gradient for a typical marine geometry. The
stack decreases noise level (thus increases signal-to-noise
ratio) following the square root of the fold. Thus it reaches
a maximum noise attenuation level (15 dB) associated with
maximum fold (30) below the mute zone. Intercept and gradient need at least two-fold to be computed, which explains
why they tend to increase noise for shallow times. As the
fold builds up, the intercept starts decreasing noise, and
eventually reaches a plateau 3.5 dB above the stack level.
The gradient on the other hand first reaches a plateau within
the mute zone and then steadily increases noise level. The
plateau is associated with the mute function, which roughly
follows a constant incidence angle (30 in this case). The
slight decrease corresponds to fold building within the mute
zone.
Figure 1 shows quite clearly that typical acquisition
geometry cannot yield reliable gradient estimates. This
attribute increases noise level (or decreases signal-to-noise
ratio) by at least 15 dB. This is hardly surprising considering that the gradient essentially corresponds to the differNOVEMBER 2000

THE LEADING EDGE

0000

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 2. Intercept versus gradient crossplot for pure


random noise. The geometry is that of Figure 1, and
the noise is crossplotted between 2 and 3 s. Note the
difference of scale between the two attributes and the
clear negative correlation.
ence between minimum and maximum angle stacks, divided
by the sine squared of the maximum incidence angle.
Because this number is generally very small, the slightest
amount of noise between the two angle stacks (e.g., wavelet
variation) is boosted compared with the expected signal.
Another popular tool to analyze AVO results is the crossplot of intercept and gradient. However, this technique is
flawed in the presence of noise as illustrated by Figure 2.
The negative statistical correlation between the two attributes is actually indicative of crosstalk: A scaled version
of the noisy part of the intercept leaks into the gradient
and overpowers any genuine signal. Figure 3 shows a real
data crossplot from the Gulf of Mexico. The seismic samples corresponding to a known gas sand are red triangles;
surrounding shales are blue dots. The clear background
trend has the same characteristics as the noise trend in
Figure 2 (negative correlation, difference of scale). To assess
whether this trend is a lithologic or a statistical effect, one
can for example crossplot stack and gradient (which are statistically uncorrelated). Because the background trend has
disappeared from this crossplot (Figure 4), it was a statistical effect due to noise in the data.
According to the equations in the previous section,
Poissons reflectivity (or the fluid factor) is simply the sum
of intercept and gradient. In light of the amplitude difference, such a sum will mostly result in the gradient, or rather,
in yet another scaled version of the noisy intercept. Some
people have suggested scaling the gradient before adding
it to the intercept. However, if the amplitude difference is
due to statistical leakage, the resulting scaled sum is merely
a far-angle stack (which is a rather good hydrocarbon indicator but should definitely not be mistaken for a quantitative measure of Poissons reflectivity!). As a general rule,
any measurement that can be directly connected to noise
should not be trusted quantitatively.
Most damaging types of noise. The definition of noise
includes anything that does not fit the assumed model represented by Shueys equation. If random noise is always an
issue, the most damaging types of noise are wavelet variations with offset (amplitude and phase), inaccurate NMO
corrections, inaccurate estimation of incidence angles, and
0000

THE LEADING EDGE

NOVEMBER 2000

Figure 3. Intercept versus gradient crossplot for real


Gulf of Mexico data. Samples corresponding to a
known gas sand are red triangles; surrounding shales
are blue dots. Note again the difference of scale and
the negative trend which could be either lithologic or
statistical (as in Figure 2).

Figure 4. Stack versus gradient crossplot for the same


data as in Figure 3. Because these two attributes are
statistically uncorrelated, the disappearance of the
background trend shows that it was a noise-related
effect. Note that the samples corresponding to the gas
sand were also biased by noise as they show a much
improved correlation in this crossplot.
organized noise (multiples, converted waves).
In a recent issue of GEOPHYSICS, Paul Hatchell claims that
transmission distortions are the most damaging noise, but
I have not fully analyzed these effects yet. The main concern is that the leakage introduced by these types of noise
looks like valid seismic information. Consider the simple
case where the only source of noise is an inaccurate
prestack amplitude balancing. (This problem is absolutely
nontrivial and was singled out by Smith and Gidlow as the
main reason why they could not use their theoretical
weights.) In a clastic environment without the presence of
hydrocarbons, we expect PR to be close to zero. Thus,
prestack amplitudes should, on average, decay with the
NOVEMBER 2000

THE LEADING EDGE

1247

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

PR

noise-free

PR

balancing

PR

NMO-stretch

PR

residual NMO

Figure 5. Effect of various noise types on synthetic data. From left to right: A noise-free estimation of A and PR
gives the expected result (no energy in PR); after prestack amplitude balancing (now PR = A); after simulated NMO
stretch (PR is a lot noisier); after residual NMO correction (PR is a negatively scaled version of A).
cosine squared of the incidence angle. Any statistical attempt
at prestack amplitude balancing will result in a constant
amplitude level for all offsets, in which case we have:

Rbal ( ) = Rp = Rp cos2 + Rp sin 2


In other words, we obtain an estimated Poissons reflectivity equal to the normal incidence reflectivity. This is what
is meant by statistical leakage.
Figure 5 illustrates the influence of some of these noise
types using synthetic data. The input data is a gather of 31
angle traces sampling the range of 0-30 at 1 intervals. The
P-wave impedance model is a random series, and the S-wave
impedance model is exactly its half (simulating a hydrocarbon-free clastic sequence). The prestack reflectivity series
are generated using a broadband zero-phase wavelet and
Shueys equation. At first, the gather is noise free and the
inverted intercept and Poissons reflectivity are exactly what
is expected (PR = A + B is almost zero), which is only natural because the same equation is used to model and to
invert the data. Next, the data undergo prestack amplitude
balancing to compensate for average amplitude variation
with offset (angle). As expected from the previous paragraph, PR is equal to A. To simulate NMO stretching, the
wavelet gradually changes with angle: At 30 the wavelet
has lost 10 Hz in high frequencies and gained 5 Hz in low
frequencies. The resulting PR is extremely noisy. Finally,
residual NMO corrections are introduced as a parabolic
time shift that reaches 5 ms at 30. Now PR becomes a negatively scaled version of A, with a large scale difference, as
predicted by the theory. The estimated PR is not just a
scaled-up version of its correct value; it is heavily contaminated by intercept leakage. On the other hand, A is always
1248

THE LEADING EDGE

NOVEMBER 2000

accurately estimated, as expected from Figure 1.


The presence of a hydrocarbon anomaly is simulated in
Figure 6. The crossplot shows that it is a typical Class 2 AVO
anomaly. After simulation of NMO-stretching, the anomaly
is still visible but buried within noise-related artifacts. A
stronger anomaly would manage to stand out from the
background, but it would have to be really strong to beat
more realistic noise leakage (as is the case in Figures 2 and
3). Again, large hydrocarbon anomalies can be identified
qualitatively with AVO analysis, but quantitative measures
or detection of smaller anomalies are heavily contaminated
by noise. Because real data never exactly conform to
Shueys equation (the wavelet, for one, always varies with
offset), the shear and Poissons reflectivity series extracted
from P-wave data using the standard AVO sequence cannot be trusted quantitatively.
The sources of noise described above have been recognized in the past, and there has been a lot of research (most
notably by Herb Swan) to address them. However, all types
of noise have to be eliminated (not just some of them) for
the results to be quantitatively meaningful. In fact, the only
way for P-wave AVO to be quantitative is to reach a level
of noise elimination such that the prestack samples exactly
conform to Shueys equationa gigantic task indeed but
perhaps not impossible.
Elastic impedance. In a 1999 TLE article, Patrick Connolly
develops an alternative approach, called elastic impedance,
which offers an elegant solution to some of the problems
described above. The principle is to first perform stratigraphic inversion of some angle-limited stacks and then to
compute P- and S-impedance and instantaneous VP/VS using
NOVEMBER 2000

THE LEADING EDGE

0000

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

PR

A
PR

PR

PR

Figure 6. Same synthetic data but after introduction of a PR anomaly. The noise-free estimates give the expected
Class 2 response (as illustrated by the crossplot). The NMO-stretched results also give the expected anomaly but
buried within noise-related artifacts. The anomaly (red dots) does not stand out in the crossplot.
Shueys equation (or an equivalent). The prestack stratigraphic inversion removes the effects of the angle-dependent wavelet and, depending on the inversion algorithm,
may help identify residual NMO corrections. I will use a
real data set to illustrate this approach. The test data consist of three angle-cubes (Figure 7). The AVO anomaly is
obvious as the near- and far-angle sections are completely
different. Although there is clearly no need for a sophisticated DHI to identify the presence of hydrocarbons, can we
extract quantitative shear information from these data?
Figure 8 shows sonic, shear and density logs from a well
that penetrated the reservoir. The density drop at the top of
the reservoir exactly compensates the P-velocity increase so
that the P-impedance log does not show any contrast. The
top of the reservoir is therefore invisible to normal-incidence P-reflectivity. In contrast, Poissons reflectivity should
be extremely sensitive to the top-reservoir as the VP/VS ratio
undergoes a significant drop. This is a classic Class 2 AVO
anomaly. An angle gather is modeled using Zoeppritz equation. As expected, the near-angle trace does not show any
top reservoir reflection, which only becomes visible with
increasing angle. This confirms what we observe in the
angle cubes (Figure 7). However, it is important to remember, as originally noted by Allen and Peddy in their AVO
case studies book, that it is extremely difficult to pick accurate velocities in the presence of Class 2 anomalies. The synthetic gather illustrates the risk of picking the wrong seismic
phase and undercorrecting (or overcorrecting) the top-reservoir reflection.
Next, the angle-dependent synthetic seismograms are
used to extract a wavelet from each angle-cube (Figure 9).
The match is quite good on all three cubes, but the derived
wavelets are significantly different. Between 10 and 30, the
wavelet peak amplitude has increased 50%, the central frequency has decreased 25% and the phase has rotated 45.
Bandwidth and amplitude changes with angle are hardly
surprising and, in fact, are expected. The phase rotation is
probably due to the impact of NMO stretch on a nonzero0000

THE LEADING EDGE

NOVEMBER 2000

Figure 7. Near-, mid-, and far-angle sections from the


3-D test data. Their obvious difference at the reservoir
level (red rectangle) indicates a strong AVO anomaly.
phase wavelet. This simple analysis shows that Shueyderived shear and Poissons reflectivity would be totally
unreliable (at least quantitatively) if they were extracted from
these data.
Now that the wavelets are known, it is possible to invert
each angle-cube with its own wavelet. The stratigraphic
inversion of choice is the 3-D layer-based nonlinear algorithm described by Gluck, Juve, and Lafet in their 1997 TLE
article. The top-reservoir layer is not readily identifiable on
the 10 impedance volume but appears quite clearly on the
30 impedance volume (Figure 10). Intriguingly, the shape
of the top-reservoir layer in the far-angle volume does not
match the shape of the corresponding layer in the near-angle
volume. This is most likely a consequence of residual NMO
corrections. Again, this can be expected from a Class 2 anomaly, especially when comparing the angle-cubes in Figure
7. They look so different that it is very unlikely accurate
velocities can be picked. This is yet another reason why shear
and Poissons reflectivity cannot be quantitatively estimated
using standard AVO analysis. Note that other, nonstandard
AVO methods, like near- and far-angle crossplotting, will
NOVEMBER 2000

THE LEADING EDGE

1249

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 8. S-velocity (red), P-velocity (blue), and density (green) logs for the previous data set. The top-reservoir (red
ellipse) shows a drop in density and an increase in P- and S-velocities. The resulting acoustic impedance (black)
does not show a contrast. However, the VS/VP ratio squared (red to the right) does show a significant jump. The
prestack behavior of this typical Class 2 sand is shown by the Zoeppritz-derived angle-gather. The red line is
indicative of the risk of picking an incorrect phase during velocity analysis: The apparent upward shift of the
trough is due to waveform interference, not NMO errors.
also fail in the presence of residual NMO corrections. This
problem has to be resolved to stand any chance of achieving quantitative results.
Once these elastic impedance volumes have been computed and the layer mismatches have been resolved, it is a
matter of simple algebra to reconstruct a shear impedance
volume and a VP/VS volume using Shueys equation.
Although it is the same equation as that used for standard
AVO, using it after stratigraphic inversion is a lot more reli-

1250

THE LEADING EDGE

NOVEMBER 2000

able because two of the most damaging types of noise have


been resolved (wavelet variations and residual NMO corrections). Also, the 3-D aspect of the stratigraphic inversion
effectively removes a lot of random noise. However, other
types of noise remain, which are still damaging for both
methods.
Conclusions. P-wave seismic data do contain significant
shear information that has been used over the years successfully for the detection of hydrocarbons. However, the quantitative
extraction of the shear reflectivity from
P-wave data is so unreliable that the
slightest amount of noise biases the measurements. By noise I mean anything that
does not fit Shueys simple equation
in particular, wavelet variations with offset (which always occur) and residual
NMO corrections, often expected in the
presence of Class 2 AVO anomalies.
Noise, at large, introduces P-reflectivity
leakage in the other prestack attributes.
This leakage biases most DHIs in such a
way that they are still valid hydrocarbon
indicators but do not represent what they
are supposed to. In particular, the popNOVEMBER 2000

THE LEADING EDGE

0000

Downloaded 03/08/15 to 178.183.194.65. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(GEOPHYSICS 1985); Weighted stacking for rock


property estimation and detection of gas by
Smith and Gidlow (Geophysical Prospecting 1987);
Comparison of AVO indicators: A modeling
study by Castagna and Smith (GEOPHYSICS
1994); Detection of gas in sandstone reservoirs
using AVO analysis: a 3-D seismic case history
using the Geostack technique by Fatti, Vail,
Smith, Strauss and Levitt, (GEOPHYSICS 1994);
Lithology color-coded seismic sections: The
calibration of AVO crossplotting to rock properties by Verm and Hilterman, (TLE 1995); and
Elastic impedance by Connolly (TLE 1999).
The stratigraphic inversion algorithm is
described by Gluck, Juve and Lafet in Highresolution impedance layering through 3-D
stratigraphic inversion of the poststack seismic
Figure 9. Angle sections after well matching using the angle-dependata (TLE 1997). Everything I claim in this
dent synthetic seismograms derived in Figure 8. The matches are quite paper is rigorously derived in two SEG
good, but the derived wavelets are significantly different.
expanded abstracts: AVO attributes and noise:
pitfalls of crossplotting (1998) and AVO inversion and elastic impedance (2000). Finally, another perspective on noise is given by Hatchell in Fault whispers:
Transmission distortions on prestack seismic reflection data
E
(GEOPHYSICS 2000). L
Acknowledgment: This paper benefited considerably from numerous (and
animated) discussions with colleagues at CGG and across the industry.
Special thanks go to Yves Lafet for the inversion of the real data example.
I also thank Norsk Hydro, Conoco, BP, and Statoil for permission to show
their proprietary data.
Corresponding author: G. Cambois, gcambois@cgg.com

Figure 10. Layer-based stratigraphic inversion of two


angle cubes. Reservoir layers (blue) are clearly visible
at 30 but not at 10. The top-reservoir layer at 30
overlaid on the 10 impedance section (black dotted
line) shows a significantly different structure. This is
probably due to inaccurate NMO corrections.
ular fluid factor obtained by scaling and adding intercept
and gradient is in fact a far-angle stack. It is therefore highly
unlikely that any quantitative information can be extracted
from standard P-wave AVO.
Elastic impedance offers an interesting alternative
because it is not affected by wavelet variations with offset
and residual NMO corrections. However, other sources of
noise, like inaccurate incidence angles or residual multiple energy, still adversely affect both methods. Elastic impedance is a giant step forward in terms of gaining access to
more reliable quantitative estimates of shear reflectivity
from P-wave data, but there are still some hurdles to be
cleared. This is why I claim that P-wave AVO may, one day,
become quantitative, but it will require a lot more effort than
what we are currently doing.
Suggested reading. The two 1993 SEG books, Offset-Dependent
ReflectivityTheory and Practice of AVO Analysis by Castagna
and Backus and Amplitude Variations With Offset: Gulf Coast
Case Studies by Allen and Peddy, are a must. Classic papers
include: A simplification of the Zoeppritz equations by Shuey
0000

THE LEADING EDGE

NOVEMBER 2000

NOVEMBER 2000

THE LEADING EDGE

1251

You might also like