International Journal of Geosciences, 2011, 2, 164-171
doi:10.4236/ijg.2011.22017 Published Online May 2011 (
Copyright © 2011 SciRes. IJG
Errors of Estimating Near-Surface Q-A Statistical
Chih-Sung Chen, Yih Jeng*
Department of Earth Sciences, National Taiwan Normal University, Taipei
Received December 18, 2010; revised February 7, 2011; accepted March 16, 2011
Although various estimating methods have been developed for measuring Q from near-surface seismic data,
less thought has been given to the accuracy of Q obtained. The errors of Q depend on the ways of measuring
Q and the computation techniques used in estimating. The main purpose of this paper is to give a compre-
hensive evaluation for the accuracy of measuring near-surface Q. We discuss the possible origins from which
errors may develop, and provide a statistical guide to the accuracy that can be expected. A set of real data
based on the improved spectral ratio method for near-surface Q was used as an example of validation and
sensitivity analysis. The Bonferroni procedure was adopted for deriving the joint confidence intervals for k
and n of the power law model. The same approach with modest modification may be applied to analyze the
accuracy of Q estimated by other methods.
Keywords: Q, Near Surface, Spectral Ratio, Seismic, Statistical Errors
1. Introduction
Wave energy attenuation is an important physics issue
for many applications, and has been defined in geosci-
ences as the quality factor Q indicating the inverse of
intrinsic attenuation. In other words, a large Q implies
small energy dissipation when wave propagates through
the rock. By contrast, rocks with small Q values will be a
poor transmitter for wave propagating. In geophysical
investigation, Q is useful for describing rock properties,
implementing inverse filters [1], and indicating fluid
saturation better than the velocity ratio. In particular,
anisotropy of Q could give unique information on sub-
surface properties, such as the lithology and orientation
of fractures [2,3]. Although the importance of Q is fully
recognized by geophysicists, the estimation of Q is still
controversial perhaps due to the practical difficulties of
in-situ measurement using geophysical method and in-
sufficient information of physical mechanisms to estab-
lish appropriate models. In view of this, some geophysi-
cists suggest that directly measuring variations of wave
velocity, amplitude, and frequency content due to ab-
sorption effects is more practical than estimating Q [4].
However, in near surface seismic survey, an accurate Q
model is still exclusively needed for performing inverse
Q filtering to recover the deep reflection energy and to
improve the resolution of the subsurface image. Two
basic hypotheses are considered in the estimate of Q for
shallow earth. The conventional one is that the Q model
is independent of frequency; but some investigators sug-
gest that Q may be frequency dependent [5,6]. A theo-
retical derivation proposed by Jeng et al. [7] indicates
that the Q of shallow seismic waves is frequency de-
pendent. The frequency dependent Q has gained more
attentions in recent inversion studies. Li et al. [8] have
developed Q estimates using both t* and spectral ampli-
tudes for frequencies between 1 and 40 Hz, and yielded
consistent results that Q increases with depth and fre-
quency up to 20 Hz within the Seattle Basin. Although
the frequency dependent model is more practical in use,
the standard error in the estimation of Q for each given
frequency may be enormous. Most seismologists attri-
bute the limitations of the Q estimation to the source
factor, the data quality, the effect of time window length,
and the velocity model. However, some peculiar prob-
lems may not be so easy to answer; for instance, the
negative Q is noted in some cases. Therefore, we suspect
the assumption of energy dissipation of traveling plane
waves may not be true for the shallow earth model. In
addition to this, the analysis problems should also be a
crucial factor. In order to give a comprehensive evalua-
tion for the accuracy of Q, this paper examines the
C.-S. CHEN ET AL.165
commonly used definition of Q from the near-surface
point of view and the basic theory for Q estimation along
with a statistical analysis. We present a framework of
determining the errors in near-surface Q, and provide a
statistical guide to the accuracy that can be expected. A
field example of data acquisition and processing meth-
odology is also briefly described. The physics behind the
data scatter may give us a better understanding of many
systematic errors of near-surface Q measurements.
2. Methodology and Background
The work discussed in this paper is part of a series of
experiments related to the near-surface seismic explora-
tion project that has been carried out at the geophysical
laboratory of National Taiwan Normal University. Seve-
ral test sites located mainly in northern and central Tai-
wan provide the real data for analysis. The test sites are
very similar in geology, and consist basically of Recent
alluvium and unconsolidated Quaternary sediments [9,10].
In the previous experiments, Jeng et al. [7,11] proposed
an improved field measurement of near-surface Q using
a frequency dependent spectral ratio method. The field
lay-out is shown in Figure 1. A multichannel seismo-
graph with 4.5 Hz three-component geophones and 28 Hz,
40 Hz, and 100 Hz vertical geophones were used to ac-
quire data of different dominant frequencies. A me-
chanical striking seismic source (Figure 2) and a 7-kg
sledgehammer were used to generate P and S waves of
different impacting energy. Figure 3 demonstrates the
way of generating shear wave using the sledgehammer.
A typical shot gather acquired in the field for picking the
Figure 1. Standard field layout of the experiment. Geo-
phone intervals vary from 0.5 m to 2.0 m based on the site
environments and sources applied. An engineering seismo-
graph was employed to record data.
Figure 2. VAKIMPAK, a mechanical striking seismic source
used in the field for generating P- and S-waves. The striking
energy is much stronger than the sledgehammer, about
2500 Nm for P-wave and 2000 Nm for SH-wave.
Figure 3. Shear wave generating by using a sledgehammer
striking against the end of a block of a wooden plank.
first arrivals is shown in Figure 4. This technique firstly
estimates the Q of any given frequency by fitting a
straight line to the slope of the spectral ratio over a finite
range of arrival-time difference of two receivers, and
then a frequency dependent equation of Q is determined
by an optimum power law model. The final Q of any
particular frequency is then evaluated from the equation.
In practice, similar to the problem of the conventional
spectral ratio method, the ratio values rarely fall on a
straight line that has been described by Patton [12].
Therefore the estimated Q is far from a definite value no
matter what method is applied. It is more like an ap-
proximate value with an estimating error.
The origin of the estimation errors of near-surface Q
values from spectral ratios can be considered from two
aspects, i.e., the inherent and processing errors. The in-
herent error can be interpreted as a trade-off error that
stems from the measurement of spectral ratios itself, and
places fundamental limits on the accuracy of estimates of
Q [13]. The processing error depends predominantly on
Copyright © 2011 SciRes. IJG
Figure 4. Typical raw shot gather data acquired in the ex-
perimental sites.
the model of measurement. Conventional estimating me-
thod is based upon the assumption of frequency inde-
pendence that may be problematic in shallow Q values
[7,14] due to heterogeneity, variations in saturation and
porosity of the near surface layers, thin bed scattering,
leaky modes and so on. Given this, we investigated the
accuracy of near-surface Q in a broad view by way of the
improved method, developed by Jeng et al. [7], which
assumes frequency dependence.
3. Theoretical Consideration
3.1. Errors of Small Dissipation Assumption
The concepts of the seismic attenuation and the estima-
tion from the amplitude spectral ratios in near-surface
seismic data will be reviewed briefly before we discuss
the accuracy of near-surface Q.
The fundamental definition of Q is based upon the in-
trinsic property of rock, and can be described by the
harmonic oscillation theory
, (1)
where represents the elastic energy stored at maxi-
mum stress and strain, and
is the energy loss per
harmonic cycle. Q is also defined as
is the exponential decay constant, V is the
wave velocity, and f is the frequency [6,15]. The above
definitions are valid if the attenuation is small, i.e., large
Q is assumed. A more appropriate equation is proposed
by Hamilton [16] for large-dissipation (small Q):
denote the estimated Q on the small-dissipa-
tion assumption (Q >10) [15], and S the estimated Q
on the large-dissipation assumption, then Equation (3)
 
Vf Q
Most Q values estimated in near surface are lower than
2π, thus the error on small dissipation assumption can be
determined by Equation (4).
3.2. Errors of the Over-Parameterized Model
White [14] has proposed that it is essential to parameter-
ize the estimation model based on the principle of parsi-
mony. According to this principle, fundamental limits on
the accuracy of estimating Q can be set up. The error of
the over-parameterized model mainly originates from the
calculation of spectral ratios. If the estimation of spectral
ratios comes from multiple coherence analysis of two
time intervals in surface seismic data, then the equation
of the relative standard error is
.. Q
se Q 2
T (5)
where Q denotes the sample mean of Q, F is the band-
width over which useful measurements can be made, and
T is the duration of the data segment [14].
3.3. Noise Effect
The primary assumption of the spectral ratios is the line-
arity of the data to be fitted. This assumption is some-
times invalid due to the peaks occurring in the spectral
ratio. The spikes in the spectral ratio obtained from the
conventional method are the contributions from the
P-wave leaky modes and other noises. The leaky modes
are the energy of surface waves (normal modes) below
certain cutoff frequencies that leak through the half-
space as body waves, and the phase velocity may exceed
Copyright © 2011 SciRes. IJG
C.-S. CHEN ET AL.167
the body wave’s velocity. Because the energy of leaky
modes attenuate exponentially with distance, the spikes
occurring at the spectral ratio can be reduced by increas-
ing the source-receiver offset, but it is still a probable
cause for unstable spectral ratios. The improved spectral
ratio method reduces the effect of leaky modes by taking
an optimum linear model from a variety of geophone
pairs of different source-receiver offsets, then assuming a
simple nonlinear power law regression model to estimate
the Q as a function of frequency. The data uncertainties
due to other noises are poorly known; therefore, the
noises usually are assumed to be a zero-mean Gaussian
process, and ignored in the inversion algorithm [17]. This
is an oversimplified assumption about the noise, and
adds errors to the Q estimation. Therefore, the standard
error of the slope of the optimum linear regression model
may reflect the residual error originating from the leaky
modes and other noises.
11 21
ln ,
SZftt ftf
 
4. Processing Errors
4.1. High and Low Estimates of the Regression
Coefficient from the Least-Squares Fit
Equation (1) can be generalized for an arbitrarily small Q
by taking the differential form:
is the period of the wave or oscillation [18].
Integrating Equation (6) yields
. (7)
Since energy is proportional to the square of amplitude A,
Equation (7) can be expressed as
. (8)
A conventional equation of spectral ratios is derived
from Equation (8) as
112 1
ln SZfZZ f
where is the amplitude at frequency f of the
signal recorded at offset n
and C is a constant that
takes into account the source function, receiver function,
and geometric function.
Jeng et al. [7] have improved the conventional spectral
ratio method by measuring the two-way travel time dif-
ference rather than the difference of the offset. Accord-
ingly the equation of spectral ratios is given by
In order to obtain a stable Q, the improved spectral ratio
method assumes that Q is frequency dependent. For each
given frequency f, the natural logarithm of the amplitude
112 2 is taken; then plot it against
the arrival time difference Δt of the two receivers at off-
sets Z1 and Z2. A large number of different Δts and cor-
responding spectral ratios are estimated for the same
frequency (normally 30 to 60 different Δts were taken for
one frequency); then the Q of the given frequency is de-
termined by the slope of an optimum linear regression
,,SZfS Zf
The estimated standard error of the slope for the linear
regression model is the square root of our estimator of
variance, thus
.. s
se b
where b is the sample mean of the slope and x represents
the value of Δt. The numerator 2n
s is the square
root of the residual variance in which rs is the residual
sum of squares. The standard error in the estimate of the
slope can be related to the standard error in estimated Q
through Equation (10). The estimated limits for the slope
b of the regression model with a 95% confidence interval
0.025, 2tns
where t is the test for significance from the t table, 0.025
is one-half of the significance level of t test (usually de-
noted by α), and n 2 is the degrees of freedom of the
t-distribution [19-21]. These are also the estimated limits
for the Q of a given frequency with a 95% confidence
4.2. Standard Errors and Confidence of the
Optimum Power Law Model
Since near-surface seismic data show strong evidence of
frequency dependence [7], the true value of Q can only
be approximated by estimating an optimum regression
model related to the frequency. We calculated the error
by estimating the deviation of our calculated Q from the
power law model where f is the frequency, k
and n are constants to be determined. For a very small Q,
Equation (4) is suggested to correct the small-dissipation
(large Q) assumption error before calculating the uncer-
Copyright © 2011 SciRes. IJG
tainty at this stage. Because the nonlinear regression
model consists of two parameters, k and n, thus we use
confidence regions that involve separate intervals in each
parameter. The joint confidence intervals for multi-re-
gression parameters in nonlinear regression can be de-
rived by the Bonferroni procedure [21]. A [100(1 α)]%
joint confidence interval for k and n is given by
kp k
np n
where α is the significance level, 12
is the
100 percentile of the standard normal distribu-
tion, and
k and
n are the standard error of parame-
ters k and n, respectively.
5. Example of Real Data
We use some typical field data acquired at the Keelung
experimental site in northern Taiwan to demonstrate the
statistical approach of the previous section. Figure 5
shows a representative numerical plot of the spectral
ratios of P-wave versus arrival time difference at one
particular frequency (60 Hz) for any two geophones of
different offsets in one geophone spread (Table 1). The
linear regression model of the data is Y = 0.05442 +
0.09025X. The estimated standard error of the slope of
the regression model was
. .0.012149se b , and a
95% confidence interval for the slope of the regression
model was 0.09025 ± 0.0245 since t(0.025,n 2) =
t(0.025,37) = 2.02 in Equation (12). The standard error in
the estimate of Q can be related to the standard error of
the slope of the regression model.
Table 2 is the frequency dependent near-surface Q
values estimated for the P-wave data recorded at the
same experimental site [22]. Inspection of the examples
and Table 5 in Jeng et al. [7], these values are reasonable
for the near surface Q estimation and more stable than
those obtained by using the conventional spectral ratio
method. Let us take the data set of SHOT NO. P0430-2
for example. Given that Q = 5 for the representative Q
value in our case, data of T = 30 ms with Δt = 5 ms were
analyzed. The seismic bandwidth was from 60 to 300 Hz
normally. Then the error of the over-parameterized
model for our near surface Q in this case is about 1.2. A
desirable way of reducing the relatively large inherent
standard error is to expand the duration of the analyzed
data segment and their time separation because the
bandwidth is wide enough for seismic frequency ranges.
Figure 5. Representative numerical values of the spectral
ratios of P-wave versus arrival time difference at one par-
ticular frequency (60 Hz). Dashed lines define the confi-
dence bands. Data were recorded at the Keelung experi-
mental site in northern Taiwan (Table 1). The linear re-
gression model of the data is Y = 0.05442 + 0.09025X with a
correlation coefficient of 0.77371.
Table 1. Original representative 60 Hz P-wave spectral
ratio data acquired at the Keelung experimental site. The
arrival time difference in millisecond is calculated from any
two geophones of different offsets in one geophone spread.
The amplitude ratio is natural logarithmic transformed.
Time Difference
(ms) Amplitude Ratio Time Difference
(ms) Amplitude Ratio
4.0 0.15 20.4 1.40
4.0 0.45 20.4 2.87
4.1 0.90 21.5 1.85
4.0 1.40 21.5 3.03
6.0 0.15 22.4 0.73
7.3 0.45 22.6 2.60
8.0 1.63 24.8 1.14
9.0 0.45 25.2 1.90
10.8 0.60 24.5 3.06
10.7 1.03 27.8 0.94
10.7 2.06 27.8 1.95
12.7 0.30 28.4 2.45
14.0 0.88 27.4 2.80
14.4 2.45 31.3 3.45
15.2 0.27 32.3 3.00
15.1 1.98 34.5 3.20
16.0 2.80 38.0 3.45
17.5 0.70 41.3 3.75
19.3 2.27 45.0 5.50
20.3 0.74
Copyright © 2011 SciRes. IJG
C.-S. CHEN ET AL.169
Table 2. Qp of the Keelung experimental site versus
frequency. An improved spectral ratio method is used to
estimate Qp.
FREQ.(Hz) P0430-1 P0430-2 P0430-3
60.00 2.03 2.71 2.17
90.00 2.95 3.16 3.04
120.00 3.23 3.37 3.31
150.00 3.60 4.03 4.13
180.00 3.41 5.66 4.04
210.00 4.63 5.33 4.58
240.00 6.00 6.18 5.37
270.00 2.97 6.38 7.01
300.00 6.33 7.46 6.89
Correction for the error of small dissipation assumption
is another way to improve the accuracy of near surface Q.
However, it produces only an insignificantly small dif-
ference (about 1% of difference for Q with a value of 5)
when compared with the standard errors from other ori-
For estimating the optimum power law model we used
the Simplex estimation algorithm to do the nonlinear
regression (Figure 6). From Equation (13) and Equation
(14), the estimated parameters k and n of the power law
model are
and the standard errors of parameters k and n are
Figure 6 shows the power law model for the Q values
obtained at the Keelung experimental site. Because the
0.04314 0.22491k
0.57108 0.824n
0.57108 0.824n
100 percentile of the standard normal distribu-
tion for a 95% confidence interval is 1.960 in Equation
(13) and Equation (14), the approximated 95% confi-
dence interval for k is
and for n is
The joint confidence interval for the family coefficient
approximates 90% for
0.04314 0.22491k
to be valid simultaneously. The standard error of the es-
timated Q at this stage can readily be obtained by substi-
tuting the numerical values of ks and ns into the power
law model. The error of the individual Q values of each
given frequency can be inferred from the plot of the re-
siduals against the fitted values (Figure 7). However,
different estimation algorithms and the convergence cri-
teria in nonlinear estimation may change the standard
error of the power law model dramatically; therefore, a
more accurate treatment to optimize the residual variance
around the regression line is probably to be important.
6. Discussions
The apparent Q estimated using the field data is the total
attenuation Qt that includes the intrinsic Qi and scattering
Qs induced by effects such as thin-bed tuning and scat-
tering. Thus, the error in Q is inversely proportional to
interval two-way time thickness, but increases with depth
of the interval [23]. By use of the parallel circuit model,
Lerche and Menke [24] prove that the two attenuations
are additive, i.e.,
Figure 6. Power law model for the Q values obtained at the
Keelung experimental site. The data used is P0430-2
shown in Table 2. The nonlinear regression model of the
data is Q = 0.1340208 × f0.6975394 with a correlation coeffi-
cient of 0.97445.
Figure 7. Plot of the residuals against the model predicted
values showing the significant relationship. Dashed lines de-
fine the confidence bands. The data used is P0430-2 listed in
Table 1. The linear regression model for residual values (R)
to predicted values (P) of the data is R = 0.08052 0.01390 P
with a correlation coefficient of 0.0610.
Copyright © 2011 SciRes. IJG
Neep et al. [25] propose that this equation is valid if at
least one full wavelength propagates through the medium,
and the relationship appears to be valid in the spectral
ratio method despite the scattering Q having both posi-
tive and negative values. The scattering Q in the near-
surface is more significant due to thin bed scattering, and
therefore, it may contribute more errors to the total Q
because the geometrical spreading factor is insignificant
in the near-surface Q estimation [7].
The real data example in this paper demonstrates that
the standard error of the estimated Q is strongly related
to the numerical values of ks and ns of the power law
model; thus the accuracy of the estimated Q depends on
the frequency bandwidth of data. This is also a coinci-
dent result proposed by White [14].
The total error associated with the near-surface Q es-
timation can be obtained by summing up all the relative
errors if they are independent and that the change of Q is
linear. Because these conditions may not be true in
reality, the total error only describes the maximum or the
worst-case scenario.
7. Conclusions
The attributes of uncertainties in estimating Q depend on
the methods used. The approach discussed in this study
suggests that the accuracy of near surface Q is affected
by a variety of factors resulting from errors of small dis-
sipation assumption, over-parameterized model, noise
effect, slope of the optimum linear regression model, and
optimum power law model.
Except under exceptionally favorable conditions, the
value of Q is always unstable and far from a definite
value no matter what method is applied. Our statistical
analysis can help understand the stability of the Q esti-
mated. However, the noise effect is the most uncertain
factor that may affect the accuracy of near surface Q.
Since no any filtering technique may remove noises
completely, it is suggested that every endeavor should be
made to acquire accurate estimates of error in the field.
The errors in estimating Q can be handled well by the
proposed method if the data acquisition errors have zero
mean, are approximately Gaussian, and have been well
8. Acknowledgements
The authors thank Jeffrey C. Green for his assistance in
manuscript preparation. Special thanks go to the editor
and one anonymous reviewer for their interest in this
work and constructive comments. This research was par-
tially funded by the National Science Council of Taiwan,
ROC, Grant NSC 98-2116-M-003-007.
9. References
[1] Y. H. Wang, “Inverse Q-Filter for Seismic Resolution
Enhancement,” Geophysics, Vol. 71, No. 3, May-June
2006, pp. V51-V60. doi:10.1190/1.2192912
[2] A. I. Best, J. Sothcott and C. McCann, “A Laboratory
Study of Seismic Velocity and Attenuation Anisotropy in
Near-Surface Sedimentary Rocks,” Geophysical Prospect-
ing, Vol. 55, No. 5, September 2007, pp. 609-625.
[3] J. H. Xia, R. D. Miller, C. B. Parka and G. Tian, “Deter-
mining Q of Near-Surface Materials from Rayleigh Waves,”
Journal of Applied Geophysics, Vol. 51, No. 2-4, De-
cember 2002, pp. 121-129.
[4] A. I. Best, “Introduction to Special Section – Seismic
Quality Factor,” Geophysical Prospecting, Vol. 55, No. 5,
September 2007, pp. 607-608.
[5] J. E. White, “Computed Seismic Speeds and Attenuation
in Rocks with Partial Gas Saturation,” Geophysics, Vol.
40, No. 2, April 1975, pp. 224-232.
[6] R. J. O'Connell and B. Budiansky, “Viscoelastic Pro-
perties of Fluid-Saturated Cracked Solid,” Journal of Ge-
ophysical Research, Vol. 82, No. 36, December 1977, pp.
5719-5735. doi:10.1029/JB082i036p05719
[7] Y. Jeng, J. Y. Tsai and S. H. Chen, “An Improved
Method of Determining Near-Surface Q,” Geophysics,
Vol. 64, No. 6, September-October 1999, pp. 1608-1617.
[8] Q. Li, W. S. D. Wilcock, T. L. Pratt, C. M. Snelson and T.
M. Brocher, “Seismic Attenuation Structure of the Seattle
Basin, Washington State, from Explosive-Source Refrac-
tion Data,” Bulletin of the Seismological Society of Ame-
rica, Vol. 96, No. 2, April 2006, pp. 553-571.
[9] Y. Jeng, “Shallow Seismic Investigation of a Site with
Poor Reflection Quality,” Geophysics, Vol. 60, No. 6,
November-December 1995, pp. 1715-1726.
[10] Y. Jeng, C.-S. Chen, H.-M. Yu, A. S.-R. Jeng, C.-Y.
Tang and M.-J. Lin, “Ultrashallow Seismic Experiment
on a Trenched Section of the Chelunpu Fault Zone, Tai-
wan,” Tectonophysics, Vol. 443, No. 3-4, October 2007,
pp. 255-270. doi:10.1016/j.tecto.2007.01.021
[11] Y. Jeng, S. H. Chen and K. J. Chen, “Effects of Uncon-
solidated Layers on Shallow Seismic Wave Propagation,”
EGS Annales Geophysical, Vol. 14, Supplement 1, April
1996, p. C168.
[12] S. W. Patton, “Robust and Least-Squares Estimation of
Acoustic Attenuation from Well-Log Data,” Geophysics,
Vol. 53, No. 9, September 1988, pp. 1225-1232.
Copyright © 2011 SciRes. IJG
Copyright © 2011 SciRes. IJG
[13] M. Sams and D. Goldberg, “The Validity of Q Estimates
from Borehole Data Using Spectral Ratios,” Geophysics,
Vol. 55, No. 1, January 1990, pp. 97-101.
[14] R. E. White, “The Accuracy of Estimating Q from Seis-
mic Data,” Geophysics, Vol. 57, No. 11, November 1992,
pp. 1508-1511. doi:10.1190/1.1443218
[15] D. H. Johnston and M. N. Toksöz, “Definitions and Ter-
minology,” In: M. N. Toksöz and D. H. Johnston, Eds,
Seismic Wave Attenuation, Geophysics Reprints Series,
Society of Exploration Geophysics, Tulsa, 1981, pp. 1-5.
[16] E. L. Hamilton, “Compressional-Wave Attenuation in Ma-
rine Sediments,” Geophysics, Vol. 37, No. 4, August
1972, pp. 620-646. doi:10.1190/1.1440287
[17] S. C. Constable, R. L. Parker and C. G. Constable, “Oc-
cam’s Inversion: A Practical Algorithm for Generating
Smooth Models from Electromagnetic Sounding Data,”
Geophysics, Vol. 52, No. 3, March 1987, pp. 289-300.
[18] F. D. Stacey, “Physics of the Earth,” John Wiley & Sons,
Inc., Hoboken, 1977.
[19] D. Rees, “Foundations of Statistics,” Chapman and Hall,
New York, 1987.
[20] J. F. Gibbs and E. F. Roth, “Seismic Velocities and At-
tenuation from Borehole Measurements near the Park-
field Prediction Zone, Central California,” Earthquake
Spectral, Vol. 5, No. 3, August 1989, pp. 513-537.
[21] R. H. Myers, “Classical and Modern Regression with Ap-
plications,” PWS-KENT Publishing Company, Boston,
[22] J. Y. Tsai, “Analysis of Shallow 3-Component Q Val-
ues,” Master’s Thesis, National Taiwan Normal Univer-
sity, Taiwan, 1994.
[23] R. Dasgupta and R. A. Clark, “Estimation of Q from Sur-
face Seismic Reflection Data,” Geophysics, Vol. 63, No.
6, November-December 1998, pp. 2120-2128.
[24] I. Lerche and W. Menke, “An Inversion Method for
Separating Apparent and Intrinsic Attenuation in Layered
Media,” Geophysical Journal of the Royal Astronomical
Society, Vol. 87, No. 2, November 1986, pp. 333-347.
[25] J. P. Neep, M. S. Sams, M. H. Worthington and K. A.
O’Hara-Dhand, “Measurement of Seismic Attenuation
from High-Resolution Crosshole Data,” Geophysics, Vol.
61, No. 4, July-August 1996, pp. 1175-1188.