GEOPHYSICS, VOL. 67, NO. 5 (SEPTEMBER-OCTOBER 2002); P. 15421547, 5 FIGS.

10.1190/1.1512799
Estimation of quality factors from CMP records
Changjun Zhang

and Tadeusz J. Ulrych

ABSTRACT
Estimates of the quality, Q, factor are commonly ob-
tained fromvertical seismic data or stacked surface seis-
mic data. This paper describes a method that allows
Q-factor tobeestimateddirectlyfromcommonmidpoint
(CMP) gathers. Absorption of the waveeld is depen-
dent on three parameters: frequency, traveltime in the
medium, and medium Q-factor. Assuming that the am-
plitude spectrum of the seismic source signature may be
modeled by that of a Ricker wavelet, we derive an ana-
lytical relation between Q-factor and seismic data peak
frequency variation both along offset and vertical time
direction. The Q-factor is estimated from CMP gathers
using a layer-stripping approach.
INTRODUCTION
Seismic waves traveling through the earth experience
absorptionmeaning attenuation and dispersionbecause of
theanelasticityandheterogeneityof themedium(Ricker, 1953;
Futterman, 1962; White, 1983; Kneib and Shapiro, 1995). Un-
derstanding, estimating, and compensating for absorption of
seismic waves are important in the quest to improve the res-
olution of seismic images, better interpret the effects of AVO,
and invert for material properties.
To compensate for absorption, we require an estimate of the
quality (Q) factor. Methods for estimating Q-factor from sur-
face seismic data are not well developed (Dasgupta and Clark,
1998). However, some research has been published concerning
the estimation of Q-factor from vertical seismic prole (VSP)
and crosswell data (Tonn, 1991). Almost all of these methods
use the amplitude of received signals, but this information is
often inaccurate because of noise, geometrical spreading, scat-
tering, and other effects.
Quan and Harris (1997) present a method for estimating
seismic absorption based on the frequency shift observed in
VSP data space. Since the centroid of the signals spectrum
Manuscript received by the Editor December 7, 1999; revised manuscript received December 21, 2001.

Formerly Department of Earth and Ocean Sciences, University of British Columbia; presently Kelman Technologies Inc., 600, 540 Fifth Ave. SW,
Calgary, Alberta T2P 0M2, Canada. E-mail: albert.zhang@kelman.com.
Department of Earth and Ocean Sciences, University of British Columbia, 2219 Main Mall, Vancouver, British Columbia V6T 1Z4, Canada.
E-mail: ulrych@eos.ubc.ca.
c 2002 Society of Exploration Geophysicists. All rights reserved.
experiences a shift to lower frequencies during propagation,
they develop a relationship between Q-factor and the centroid
of an amplitude spectrum which is represented by a Gaussian,
boxcar, or triangular shape.
We showhowto estimate the Q-factor fromprestack gathers
based on two observations. First, a common midpoint (CMP)
section represents multiple observations of an underground
structure. It provides information in the time and offset do-
mains, allowing for the extraction of information concerning
structure, lithology, and material properties such as velocity
and Q-factor. Second, reection arrival times are determined
by interval velocities and the geometric structure of the subsur-
face. Absorption of the received signals is only determined by
the interval Q-factors and traveltimes in each layer. If the am-
plitude spectrumof a seismic wavelet is assumed to be Ricker-
like, interval Q-factors can be computed solely fromthe varia-
tion of the peak frequency of a spectrum as a function of time.
We begin by developing an equation that relates absorption
tospectral peakfrequency variation. Fromthis relationship, we
estimate interval Q-factors using a layer-stripping approach.
Tests show that the relationship determines interval Q-factor
values with reasonable accuracy.
ABSORPTION AND PEAK FREQUENCY VARIATION
In seismic data processing, a recorded trace is commonly
modeled as the convolution of a seismic source signature with
a reectivity series. The seismic source signature is generally
unknown, although in some cases it can be measured or is as-
sumed to be minimum phase. The effects of anelasticity can
be incorporated into the model by convolving it with an earth
lter. This lter is causal, minimum phase, and depends on
the Q-factor (Aki and Richards, 1980). In routine seismic pro-
cessing, inverse Q-factor ltering is often used to remove this
absorption effect (Hargreaves and Calvert, 1991; Varela et al.,
1993). Instead of studying the details of the absorption lter
response, we consider only the relationship between Q-factor
and peak frequency translation.
We begin by assuming that the amplitude spectrum of the
source wavelet can be well represented by that of a Ricker
1542
Quality Factor Estimation 1543
wavelet (Ricker, 1953). The frequency spectrum of a Ricker
wavelet is expressed by the equation
B( f ) =
2

f
2
f
2
m
e

f
2
f
2
m
, (1)
where f
m
is the dominant frequency of the main frequency. For
convenience, we refer to the frequency of maximumamplitude
as the peak frequency, denoted as f
p
. For a wavelet at its initial
state, the peak frequency is the dominant frequency.
Theevolutionof theamplitudespectrumthroughtimeis now
modeled as a Ricker wavelet traveling in a viscoelastic medium
(geometric spreading and other factors are not considered).
After traveling for a time t , the amplitude spectrum is
B( f, t ) =
_
2

__
f
2
f
2
m
_
e

f
2
f
2
m
H( f, t ). (2)
The value H( f, t ) is the absorption lter (Varela et al., 1993)
whose frequency response is
H( f ) = exp
_

_
ray
a( f, l) dl
_
, (3)
where the integral is evaluated along the raypath l, and
a( f, l) =
f
Q(l)v(l)
. (4)
Here, Q(l) and v(l) are the Q-factor and velocity, respec-
tively, dened along each point of the raypath. The value
Q(l) is assumed to be independent of frequency (Ricker, 1953;
Kjartansson, 1979; White, 1983).
One-layer case
By considering the propagationof a wave ina half-space with
a Q-factor for t seconds, we determine the amplitude spectrum
of the received signal as
B( f, t ) = B( f )e

f t
Q
. (5)
As time increases, absorption increases with frequency and re-
sults in the peak frequency translating toward lower frequency.
This phenomenon is clearly illustrated in Figure 1. Because
of absorption, the time width of the source wavelet increases
and the amplitude spectrum narrows.
If the traveltime is known, we can obtain the Q-factor from
the spectral variation. Acommon method to do this is the spec-
tral ratio method. Assume that a second wavelet is received
at time t
1
with a spectrum of
B( f, t
1
) = B( f )e

f t1
Q
. (6)
Comparing the two amplitude spectra, we obtain
B( f, t
1
)
B( f, t )
=
e

f t1
Q
e

f t
Q
. (7)
Taking logarithms, equation (7) becomes
ln
_
B( f, t
1
)
B( f, t )
_
=
f
Q
(t
1
t ). (8)
Denoting
A
r
= ln
_
B( f, t
1
)
B( f, t )
_
, (9)
plotting the logarithm of the spectral ratio A
r
as a function of
frequency f yields a linear trend whose slope p is a function
of Q, where
Q =
(t
1
t )
p
. (10)
The spectral ratio method is simple in principle, but in prac-
tice determining the Q-factor is complicated by overlapping
wavelets which lead to amplitude spectra that do not reect
the wavelet spectrum. The magnitude of the amplitude spec-
trum of seismic waves is affected by many factors, such as
underground structures, geometric spreading, and automatic
gain control (AGC), which is applied to balance the trace am-
plitude in seismic data processing. It is almost exclusively the
Q-factor, however, that affects the shape of the wavelet spec-
trum. Based on this idea, we have developed a method to esti-
mate the Q-factor from the spectral variation of reections
in a CMP gather. Because it contains both time and offset
information, a CMP gather is well suited to determine the
Q-factor.
Including all Q-factor unrelated functions into an amplitude
term, we write the amplitude spectrum as
B( f, t ) = A(t )B( f )e

f t
Q
, (11)
where A(t ) is an amplitude factor independent of frequency
and absorption. The peak frequency f
p
can be determined by
equating the derivative of the spectrum, with respect to fre-
quency, to zero:
FIG. 1. (a) An event in a CMPgather is generated by a reector
in an absorbing medium. Traces have been normalized based
on their maximum amplitudes. (b) Amplitude spectra of the
original source signature[0], trace 1[1], trace 11[11], and trace
21[21].
1544 Zhang and Ulrych
B( f, t )
f
= A(t )
B( f )
f
e

f t
Q
+ A(t )B( f )e

f t
Q
_

t
Q
_
= 0. (12)
From equation (1), determining
B( f )
f
=
2

_
2 f
f
2
m
_
e

f
2
f
2
m
+
2

_
f
2
f
2
m
_
e

f
2
f
2
m
_
2 f
f
2
m
_
(13)
and inserting this expression into equation (12), we obtain the
peak frequency at time t :
f
p
= f
2
m
_
_
_
t
4Q
_
2
+
_
1
f
m
_
2

t
4Q
_
. (14)
The relationship between Q-factor and the shift of peak fre-
quency is now
Q =
t f
p
f
2
m
2
_
f
2
m
f
2
p
_. (15)
This shows that if the dominant frequency f
m
is known, the
Q-factor can be computed from the CMP gather using only
one offset.
In practice, of course, we do not know the initial f
m
. How-
ever, it can be estimated if we assume that the amplitude spec-
trum of the initial source wavelet is approximated by a Ricker
wavelet. Designating the peak frequencies at times t
1
and t
2
by
f
p1
and f
p2
, respectively,
Q =
t
1
f
p1
f
2
m
2
_
f
2
m
f
2
p1
_ =
t
2
f
p2
f
2
m
2
_
f
2
m
f
2
p2
_. (16)
Thus, we can derive the dominant frequency of the source
wavelet fromthe peak frequencies of a reection at two differ-
ent time points:
f
m
=
_
f
p1
f
p2
(t
2
f
p1
t
1
f
p2
)
t
2
f
p2
t
1
f
p1
. (17)
Equations (15) and (17) allow us to obtain an average
Q-factor by using the peakfrequency variationalong all offsets,
therebyallowingus toremovetheeffects of surfaceuctuations
and randomnoise, consequently improving the accuracy of the
Q-factor.
Although we assume a Ricker amplitude spectrum in our
computations, other models are also possible because of its
similarity with observed amplitude spectra. One example is
the Gaussian spectrum of standard deviation :
B( f ) = exp
_
( f f
m
)
2

2
_
. (18)
Following the above procedure, the relationship between the
Q-factor and the variation of peak frequencies is
Q =
t
2
f
m
f
p
. (19)
For real seismic data, the amplitude spectrum of a seismic
wavelet is not exactly that of a Ricker or a Gaussian wavelet.
In many situations, however, it can still be approximated by a
Ricker spectrum (Ricker, 1953).
Multilayer case
Consider rst the case of two layers with quality factors Q
1
and Q
2
and traveltimes t
1
and t
2
in each layer, respectively.
Applying equation (11), we get
B( f, t ) = A(t )B( f )e

f t1
Q1
e

f t2
Q2
, (20)
where t =t
1
+t
2
. Using the peak frequency f
p
associated with
B( f, t ), Q
2
can be estimated by knowing Q
1
and the dominant
frequency f
m
of the source wavelet:
Q
2
=
t
2
Q
1
Q
1
t
1
(21)
where
=
2 f
2
m
2 f
2
p
f
p
f
2
m
. (22)
Now use the concept of an equivalent Q. In other words, let
e

f t1
Q1
e

f t2
Q2
= e

f t
Q
, (23)
which allows an expression to be derived that relates the inter-
val Q
2
to the equivalent Q-factor by
Q
2
=
t
2
Q
1
Q
(t
1
+t
2
)Q
1
t
1
Q
. (24)
The equivalent Q-factor can be calculated from equation (15)
using the peak frequency at time t . Since the dominant fre-
quency f
m
of the initial wavelet and Q
1
have already been
determined fromupper-layer arrivals and since t
1
and t
2
can be
estimated, then Q
2
can be computed from equation (24).
Given the complexity of the subsurface, some approxima-
tions must be made to use all offset information. Assuming
a multilayer medium, write the amplitude attenuation equa-
tion (11) as
B( f, t ) = A(t )B( f ) exp
_
N

i =1
f t
i
Q
i
_
, (25)
where Q
i
and t
i
are the quality factor and the traveltime in
layer i , respectively. Taking an approach from velocity estima-
tion, assume straight raypaths andcompute the total traveltime
of a reection at a particular offset as
N

i =1
t
i
= t
N
. (26)
Now dene t
0(N)
as the zero-offset traveltime of reection N,
Q
N
as the Q-factor value for layer N, and t
0(i )
as the zero-offset
traveltime of a reectionabove layer N. Using similar triangles,
the traveltime in layer i is
t
i
=
t
N
t
0(N)
_
t
0(i )
t
0(i 1)
_
. (27)
Furthermore, we can split the amplitude attenuation operator
in equation (25) into two factors as follows:
Quality Factor Estimation 1545
B( f, t ) = A(t )B( f ) exp
_
N1

i =1
f t
i
Q
i
_
exp
_
f t
N
Q
N
_
.
(28)
This allows us to obtain the following equation for Q
N
:
Q
N
=
t
N

, (29)
where is the same as that in equation (22) and
=
N1

i =1
t
i
Q
i
. (30)
The Q-factor values can now be calculated layer by layer
throughlayer stripping. Since a straight raypathapproximation
is used, the computed Q
N
is not the actual interval Q-factor.
Analogous to rms velocities, we refer to such Q-factor values
as rms Q-factor values. In a manner similar to interval velocity
analysis, we can determine the apparent interval Q-factor of
the i th layer, Q
i nt
i
, from the rms values using a relation similar
to the Dix formula:
Q
int
i
=
_
Q
2
i
t
0(i )
Q
2
i 1
t
0(i 1)
t
0(i )
t
0(i 1)
. (31)
We emphasize that Q
i
is the average of the calculated
Q-factor values at different offsets for layer i and t
0(i )
is the
zero-offset arrival time of reection i when the raypath is as-
sumed to be straight across interfaces. Since values of Q
i nt
are
derived from rms Q-factor values, they are apparent, rather
than real, interval Q-factors.
EXAMPLES AND DISCUSSION
To estimate the Q-factor fromCMP gathers, we assume that
the arrival times for the main reection events are known.
Fourier transforms are computed in the windowcontaining the
reection at each offset, each amplitude spectrumis tted with
a Ricker spectrum, and the peak frequency of the spectrum is
estimated. Using equations (17), (15), and (29), we can extract
the Q-factors layer by layer from peak frequency variation.
Figure 2 shows a simple test on a synthetic CMP gather with
two events. The original Ricker wavelet has a dominant fre-
quency of 60 Hz. Absorption is modeled by using low Q-factor
values to emphasize the effect; the values in the two layers
are 10 and 20, respectively. The actual and estimated Q-factor
curves are shown in Figures 2b and 2c, respectively. The cor-
responding dominant frequency is estimated as 60.67 Hz. The
agreement between the corresponding values shows that the
method works well for ideal synthetic data.
The variation of peak frequencies with offset is shown in
Figure 3. The remarkable decrease is, of course, from absorp-
tion. This example demonstrates clearly why absorption must
be compensated for if resolution of prestack data is an is-
sue. Theoretically, for large Q-factors or weak attenuation,
this method will work. But in practice, the errors in the mea-
sured variation of peak frequencies are relatively large and will
decrease the accuracy of estimated Q-factor values.
Once we determine the Q-factor, we are faced with the task
of inverse Q-ltering. The details of the inverse Q-lters are
not discussed in this paper. An example showing the effect of
inverse Q-ltering on a prestack CMP gather using obtained
Q-factor values is illustrated in Figure 4. A synthetic CMP
gather comprised of three reections is modeled in Figure 4a,
where signal amplitudes remain unbalanced. The rst two re-
ections merge at far offsets; we consider them as one to esti-
mate the peak frequency variation. The extracted Q-factors
are shown in Figure 4b at relative magnitude. Using these
values, inverse Q-ltering is performed. The compensated re-
sult is shown in Figure 4c, demonstrating that the reections
associated with the two upper layers are separated.
These two examples show the validity of our method for
Q-factor estimation and the resolution improvement of
prestack seismic records using accurate Q-factor values. Al-
though our method is illustrated using CMP geometry, it can
FIG. 2. (a) A synthetic CMP gather of two reections with ab-
sorptionand10%randomnoise. Traces have alsobeennormal-
ized based on their maximumamplitudes. (b) Real Q-factor in
layers 1 and 2. It is unknown in layer 3. (c) Computed Q-factor
values.
FIG. 3. Variation of peak frequencies of the two reection
events in Figure 2. Solid line is for the rst layer. Dashed line
is for the second layer.
1546 Zhang and Ulrych
FIG. 4. The effect of inverse Q-ltering. (a) A synthetic CMP
gather with three reections. (b) Extracted Q-factor values.
Two layers have been separated. (c) Result of absorption com-
pensation.
FIG. 5. A real data example. (a) The original common shot
gather. (b) The common shot gather after inverse Q-ltering.
The Q curve is not shown here.
also work on common source gathers (CSGs) if the geometry
difference can be neglected by considering the substructure as
a horizontally layered medium.
The next step is to test our extracted Q-factors in real data
processing. A CSG gather obtained from a shallow seismic
survey is shown in Figure 5a. Absorption phenomena can be
observed from the broadened wave shape along the time axis.
The inverse Q-ltering result, showninFigure 5b, was obtained
using the Q-compensationlter basedonthe methodproposed
by Varela et al. (1993). Generally, aninverse Q-lter is unstable
for high-frequency noise. To stabilize the process, before atten-
uation compensation was performed, the data were processed
by low-pass frequency ltering to attenuate the frequency
content above 40 Hz.
CONCLUSION
Absorption effects as a function of offset may be mistakenly
interpreted as AVO phenomena. Without prestack Q-factor
compensation, the reection amplitudes may appear to de-
crease with offset. Compensating for absorption in prestack
data can improve the resolution of seismic data and be impor-
tant for AVO analysis.
Figure 4 clearly shows that signals attenuate as a function of
bothtraveltimeandoffset. Absorptioncompensation, basedon
theunderlyingphysics of theabsorbingprocess, lets us compute
a section with increased resolution. Subsequently, a stationary
source wavelet can be used to improve the resolution by means
of deconvolution. It is superior to conventional deconvolution
techniques which rely on an adaptive formulation of wavelet
estimation.
Prestack inverse Q-ltering is well suited to compensate
for attenuation. It is superior to poststack compensation since
stacking distorts the frequency information of the original seis-
mic data. It is advisable to extract Q-factors fromprestack seis-
mic data that have not been processed by frequency ltering
methods that violate the assumptions of our method.
Prestack processing provides another dimensionoffset
that increases the information available for signal and noise
separation. Inverse Q-factor compensation performed in the
prestack domain can provide a balanced signal spectrum over
offset and over time as well, improving the vertical and lateral
resolution of the nal image.
Underground lithology is characterized by its velocity, den-
sity, and Q-factor. While the Q-factor does not affect the ar-
rival times of reections, it does affect amplitude and the sig-
nals frequency content. Extracting velocity information from
CMP gathers is a common practice. We have devised an anal-
ogous approach to estimate the absorption character from the
variation of spectrum with both time and offset. We derived
a relationship that computes Q-factors from observed peak
frequency variation, based on the reasonable assumption of a
Ricker-like amplitude spectrum of the source signature.
ACKNOWLEDGMENTS
This research was carried out with grants from the compa-
nies who support CDSST, a consortium for the development
of specialized seismic technology at the University of British
Columbia, Canada. Our gratitude goes to the three review-
ers for their constructive comments and suggestions. C. Z.
also extends his thanks to the R&D department of Kelman
Technologies Inc.
REFERENCES
Aki, K., and Richards, P. G., 1980, Quantitative seismologyTheory
and methods: W. H. Freeman & Co., 167185.
Dasgupta, R., and Clark, R. A., 1998, Estimation of Q from surface
seismic reection data: Geophysics, 63, 21202128.
Quality Factor Estimation 1547
Futterman, W. I., 1962, Dispersive body waves: J. Geophys. Res., 67,
52795291.
Hargreaves, N. D., and Calvert, A. J., 1991, Inverse Q-ltering by
Fourier transform: Geophysics, 56, 519527.
Kjartansson, E., 1979, Constant Q-wave propagation and attenuation:
J. Geophys. Res., 84, 47374748.
Kneib, G., and Shapiro, S. A., 1995, Viscoacoustic wave propagation
in 2-D random media and separation of absorption and scattering
attenuation: Geophysics, 60, 459467.
Quan, Y., and Harris, J. M., 1997, Seismic attenuation tomography
using the frequency shift method: Geophysics, 62, 895905.
Ricker, N., 1953, The formand laws of propagation of seismic wavelets:
Geophysics, 18, 1040.
Tonn, R., 1991, The determination of the seismic quality factor Q
from VSP data: A comparison of different computational method:
Geophys. Prosp., 39, 127.
Varela, C. L., Rosa, A. L., and Ulrych, T. J., 1993, Modeling of attenu-
ation and dispersion: Geophysics, 58, 11671173.
White, J. E., 1983, Underground sound application of seismic waves:
Elsevier Science Publ. Co., Inc., 83137.