Professional Documents
Culture Documents
2012 - Estimation of Near-Surface Quality Factors by Constrained Inversion of Rayleigh-Wave Attenuation Coefficients - Xia Et Al.
2012 - Estimation of Near-Surface Quality Factors by Constrained Inversion of Rayleigh-Wave Attenuation Coefficients - Xia Et Al.
a r t i c l e i n f o a b s t r a c t
Article history: Quality factors (Q) of near-surface materials are as important as velocities of the materials in many applica-
Received 16 December 2011 tions. Only phase information of surface-wave data is utilized when high-frequency (≥ 2 Hz) surface-wave
Accepted 6 March 2012 data are routinely inverted to determine near-surface shear (S)-wave velocities. Amplitude information of
Available online 13 March 2012
high-frequency surface-wave data can be used to determine quality factors of near-surface materials.
Given S-wave velocity, compressional (P)-wave velocity, and Rayleigh-wave phase velocities, it is feasible
Keywords:
Rayleigh wave
to solve for S-wave quality factor QS and P-wave quality factor QP (for some specific velocity models)
Attenuation coefficient down to 30 m below the ground surface in many settings by inverting high-frequency Rayleigh-wave atten-
Quality factor uation coefficients in a layered earth model. Amplitude of seismic data is an exponential function of attenu-
Constrained inversion ation coefficients. When calculating attenuation coefficients from changes in amplitude, this nonlinear nature
would result in that small variations in amplitude cause huge changes in attenuation coefficients. This result
suggests data (attenuation coefficients) that normally possess large errors could eventually transfer to a
model (quality factors); therefore, constraints (or a priori information) on models are necessary. Because
an inversion system to solve this problem is unstable, a regularization parameter must be introduced into
an inversion algorithm to stabilize the inversion. These characteristics of the inversion problem allow us to
solve the problem as a constrained and regularized linear system. Usually, a set of models that meet the de-
fined constraints can be obtained by solving the system. Based on the linear nature of the inversion system, a
smooth model can be selected from the set of models as a solution of the inversion using the L-curve method.
This approach is a trade-off solution between data misfit and model length. Several real-world examples
demonstrate the importance of constraints in finding acceptable realistic quality factors from empirical data.
© 2012 Elsevier B.V. All rights reserved.
0926-9851/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.jappgeo.2012.03.003
138 J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144
the material damping ratio D (= 0.5Q − 1) (Lai and Rix, 1998; Rix et Marquardt, 1965). The regularization parameter acts as a constraint
al., 2000), is of fundamental interest in earthquake engineering on the model space (Tarantola, 1987) and can be determined by a
(Kramer, 1996), geotechnical engineering, ground-water, and envi- trial-and-error method, especially because the number of unknowns
ronmental studies, as well as in oil exploration and earthquake seis- is less than the number of observed data after discretizing an earth
mology. A desire to understand the attenuative properties of the model for inversion of dispersion curves (e.g., Xia et al., 1999, 2002,
earth is based on the observations that seismic-wave amplitudes are 2003, 2010, in press) or assuming a specific nonlayered earth model
reduced as waves propagate through a dissipative medium. (Xia et al., 2006). A trade-off solution, equivalently, is a smoothed so-
Modeling results that examined the relationship between lution calculated by a truncated inverse of a data kernel (Hansen,
Rayleigh-wave attenuation coefficients and compressional (P)-wave 1998; Parker, 1994) with the singular value decomposition (Golub
and S-wave quality factors (QP and QS) (Xia et al., 2002) suggested and Reinsch, 1970; Lanczos, 1961).
that it is feasible to invert attenuation coefficients of Rayleigh waves In this study, a constrained linear system was solved by introduc-
for quality factors. Modeling analysis also showed that QP may be ing a regularization parameter for quality factors in a layered earth
inverted when Vs/Vp is greater than 0.45, a situation which is com- model (Xia et al., 2002). After a set of feasible solutions is found, we
mon in oil industry and crust seismology studies, and which also is
occasionally encountered in near-surface materials. Modeling results
also suggested that most contributions to Rayleigh-wave attenuation
coefficients from QP are in a relatively higher frequency range while
contributions from QS are in a lower frequency range. Using different
weighting, therefore, on QP and QS in different frequency ranges may
increase the possibility of obtaining QP. Modeling results (Xia et al.,
2002), in addition, reveal that the stability of the inversion system
for quality factors is much worse than that for S-wave velocities.
Therefore, regularization and/or constraints on the quality factors
are/is essential.
When no a priori information on the earth model is available, com-
mon practice is to seek a regularized solution that minimizes data
misfit and the length of an earth model (roughness of a model), for
example, a regularized least-squares solution (Levenberg, 1944;
Fig. 1. (a) Because of the linear nature of the system (Eq. (1)), a plot of model lengths Fig. 2. (a) A 24-channel streamer was used to record Rayleigh-wave data. (b) A seismic
vs. data misfits normally shows the shape of an L-curve. (b) An example of an L-curve source used in data acquisition was a rubber-band assisted weight drop vertically
generated with different damping factors. impacting on a plate and the streamer in the back of the photo.
J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144 139
use the L-curve method (Hansen, 1992, 1998; Lawson and Hanson, We build on Kudo and Shima's (1970) work to calculate the atten-
1974; Tikhonov and Arsenin, 1977) to determine an optimal regular- uation coefficients between two traces, U ðx þ dx; t Þ ¼ U ðx; t Þe−αdx ,
ization parameter and obtain a trade-off solution from the set of fea- where U is Rayleigh-wave amplitude, α is a Rayleigh-wave attenua-
sible solutions. tion coefficient, and x and dx are the nearest source-geophone offset
and a geophone interval, respectively. After the Fourier transform
2. Basic equations with respect to time t, we obtain
Fig. 3. (a) Raw Rayleigh-wave data (shot 1732) used for the first example. (b) Attenuation coefficients. Those labeled “Measured” were calculated by Eq. (2) and those labeled
“Final” were calculated by Eq. (1) with inverted quality factors shown in panel d. (c) The L-curve used to determine the trade-off value of a damping factor (a solid square). Figures
next to the symbols are values of damping factors associated with the solutions. (d) A trade-off QS model (in the logarithmic scale) under the constraint 0 b QS b 100.
140 J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144
velocities by inverting Rayleigh-wave phase velocities (Xia et al., → →
diagonal matrix, ϕd ¼
S A X − B is the data misfit, and ϕm ¼
1999) and measuring near-surface P-wave velocities by other seismic → 2
methods, such as reflection (Hunter et al., 1984; Steeples and Miller, →
X is the model length, which is the maximum component of X .
1990), refraction (Palmer, 1980), and/or tomography methods ∞
(Ivanov et al., 2006; Zhang and Toksöz, 1998), the dissipation factors Because of the linear nature of the system (Eq. (1)), a plot of (ϕm,
(Q −1
P and Q −1
S ) can be inverted directly for noise-free data using Eq. ϕd) normally shows the shape of an L-curve although interpreting
(1). Because Eq. (1) is a linear system, the same method used in Xia an L-curve usually requires experience.
et al. (1999) can be directly employed to solve QS and/or QP from Models corresponding to small damping factors usually possess
Rayleigh-wave attenuation coefficients. Eq. (1) can be written as a large errors and result in very long model lengths with small data
matrix form misfits. This is because the models try to represent errors in data
(solutions on the right side of a trade-off curve in Fig. 1b). On the
→ →
ð3Þ other hand, models corresponding to large damping factors are
A X ¼B ðxi > 0Þ;
usually stable and result in large data misfits with short model
→
lengths (solutions on the left side of a trade-off curve in Fig. 1b).
where X is a vector of inverse quality →
factors (a model vector 1/Q) Trade-off models could be found in the trade-off zone on the plot of
with xi as the ith component, the B is an attenuation coefficient (a (ϕm, ϕd).
data vector α R ðf Þ), and A is a data kernel matrix (Menke, 1984) deter-
mined by Eq. (1). Menke (1984) discussed an algorithm to solve Eq.
(3). Eq. (3) will provide accurate QP and QS if the attenuation coeffi- 5. Real-world examples
cients contain no error as a synthetic example (Xia et al., 2002). Solu-
tions of Eq. (3) are not guaranteed to exist or solutions may possess A shallow Rayleigh-wave survey was conducted in an arid region
an unacceptable error when attenuation coefficients possess errors. of the Southwestern United States during the spring of 2010. The
The nonlinear nature of amplitude as an exponential function of at- objective of this survey was to determine the seismic properties of
tenuation coefficients means that small variations in amplitude near-surface sediments. Rayleigh-wave data were acquired using a
cause large changes in attenuation coefficients, especially for large Q towed streamer that consisted of 24 4.5-Hz vertical component geo-
values. To constrain the model space, we introduce a constraint for phones with a nearest-offset of 41 m and a geophone interval of
quality factors, that is 0 b Qi b ai, where ai is a constant. This constraint 1.2 m (Fig. 2a). The seismic source was an accelerated weight drop
is equivalent to xi > ci > 0, where ci is 1/ai. Because of the instability of
Eq. (3), Xia et al. (2002) introduced a damping factor into the system.
Therefore our inversion problem can be described by the following
constrained system:
→ →
ðA þ λI Þ X ¼B ðxi > ci > 0Þ; ð4Þ
4. L-curve method
→
→ → →
L X;λ ¼ S A X − B þ λ
X ¼ ϕd þ λϕm ; ð5Þ
∞ Fig. 4. (a) Attenuation coefficients calculated from shot 1732. Those labeled “Mea-
2
sured” were calculated by Eq. (2), those labeled “Constrained” were the same as
those labeled “Final” in Fig. 3b, and those labeled “Unconstrained” were calculated by
where S is related to a weighting matrix W (= S TS) that is deter- Eq. (1) with inverted quality factors shown in panel b. (b) An inverted QS model (in
mined by errors in attenuation coefficients. Since the weighting the logarithmic scale) without the constraint 0 b QS b 100 produced a quality factor of
matrix W is both diagonal and positive (Xia et al., 1999), S is also a 290 at a depth of around 10 m.
J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144 141
vertically impacting on a plate pressure coupled to the ground surface damping factor (0.055) possesses the shortest model length (78.1)
(Fig. 2b). and the largest data misfit (0.0086).
Although even Rayleigh-wave data (Fig. 3a) possess some noise, Models in the ranges at the both ends are obviously not the best
dominant Rayleigh wave energy in the time window of between solution. We need to find a model in the trade-off zone as indicated
200 and 400 ms allowed us to directly calculate attenuation coeffi- in Fig. 1b. A trade-off model associated with the trade-off damping
cients (triangles in Fig. 3b) using Eq. (2). After estimating S-wave factor (0.036) is indicated by a solid square in Fig. 3c, which produced
velocities for a layered model by inverting Rayleigh-wave phase a model with a model length of 86 and data misfit of 0.0068 (Fig. 3c).
velocities (Xia et al., 1999), we are able to estimate QS with the algo- The attenuation coefficients calculated from the trade-off model
rithm discussed in the previous sections. (Fig. 3d) are shown by solid squares in Fig. 3b. Inverted S-wave qual-
In solving for quality factors, we assumed QP = 2QS in inversion. ity factors (QS) are within the constrained range (Fig. 3d).
Because QS for near-surface sediments are rarely over 100 (Lai and Constraining xi > 0.01 (i.e., 0 b QSi b 100) is necessary for a well
Rix, 1998; Sheriff and Geldart, 1985), we set a constraint of behaved inversion (Eq. (3)). If this constraint is removed, using the
0 b QSi b 100 for all i, which results in constant ci in Eq. (4) of 0.01. L-curve method we may end up an improved fit of the attenuation
We found that the smallest feasible damping factor λ in Eq. (4) was coefficients (Fig. 4a) compared to what was shown in Fig. 3b but
0.026. With the damping factor changing up to 0.055 at an interval have to contend with wild variation in solutions (Fig. 4b). Clearly, a
of 0.001, we identified possible solutions (Fig. 3c) that produced a quality factor of nearly 290 at a depth of around 10 m is too large
nearly perfect L-curve. The first model on the right in Fig. 3c associat- and unrealistic for the near-surface sediments at this site.
ed with the smallest damping factor (0.026) possesses the longest Quality (S/N and uniformity in energy arrivals) of Rayleigh-wave
model length of 98.2 (b100) and the smallest data misfit (0.0056). data in the second example (Fig. 5a) is higher than the first example.
The first model on the left (Fig. 3c) associated with the largest We calculated attenuation coefficients (triangles in Fig. 5b) using
Fig. 5. (a) Raw Rayleigh-wave data (shot 2420) used for the second example. (b) Attenuation coefficients. Those labeled “Measured” were calculated using Eq. (2) and those labeled
“Final” were calculated using Eq. (1) with inverted quality factors shown in panel d. (c) The L-curve used to determine the trade-off value of a damping factor (a solid square).
Figures next to the symbols are values of damping factors associated with the solutions. (d) A trade-off QS model (in the logarithmic scale) under the constraint 0 b QS b 100.
142 J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144
Eq. (2). Consistent with the last example, we assumed QP = 2QS in in- first two examples, we calculated attenuation coefficients (triangles
version. Also as in the first example, a constraint of 0 b QSi b 100 is set in Fig. 6b) and determined the trade-off damping factor from the
for all i. We found that the smallest feasible damping factor was L-curve (Fig. 6c). A trade-off damping factor 0.16 was chosen
0.023. The damping factor changed up to 0.052 at an interval of (a solid square in Fig. 6c), which produced a model with a model
0.001 resulting in possible solutions that produced a nearly perfect length of 97.8 (b100) and data misfit of 0.0029. The attenuation coef-
L-curve (Fig. 5c). ficients calculated from the trade-off model (Fig. 6d) are shown by
The first model on the right of Fig. 5c associated with the smallest solid squares in Fig. 6b. Inverted S-wave quality factors are within
damping factor (0.023) possesses the longest model length of 98.9 the constrained range. If the constraint xi > 0.01 is not imposed in
(b100) and the smallest data misfit (0.0048). The first model on the the inversion system, the L-curve method can produce a model with
left of Fig. 5c associated with the largest damping factor (0.052) pos- QS of over 700 at a depth of 8 m, which is way outside realistic values
sesses the shortest model length (79.3) and the largest data misfit for near-surface sediments.
(0.0063). A trade-off damping factor of 0.031 was chosen (a solid A constraint xi > 0.01 (equivalent to 0 b QS b 100) limits the search
square in Fig. 5c), which produced a model with a model length of range in the model space. We expect fittingness in the data space to
87.3 and data misfit of 0.0052 (Fig. 5c). The attenuation coefficients decrease as the price we have to pay for no a priori information on
calculated from the trade-off model (Fig. 5d) are shown by solid quality factors at an investigation site. We applied the same proce-
squares in Fig. 5b. Inverted S-wave quality factors (Fig. 5d) are within dure and constraint as the previous examples to data shown in
the constrained range. If the constraint xi > 0.01 is not imposed during Fig. 7a. The attenuation coefficients calculated from the data
the inversion, the L-curve method may produce a model with an (Fig. 7a) are indicated by triangles in Fig. 7b. A smooth model was
unrealistic QS of nearly 300 at a depth of 7 m. obtained for QS under the constraint xi > 0.01 (diamonds in Fig. 7c).
Data quality of the third example (Fig. 6a) is virtually the same as Necessity of constraints was demonstrated by the unrealistic quality
the second example. Using the same procedure and constraint as the factor at a depth of 9 m (over 300) when no constraints were applied
Fig. 6. (a) Raw Rayleigh-wave data (shot 2350) used for the third example. (b) Attenuation coefficients. Those labeled “Measured” were calculated using Eq. (2) and those labeled
“Final” were calculated using Eq. (1) with inverted quality factors shown in panel d. (c) The L-curve used to determine the trade-off value of a damping factor (a solid square).
Figures next to the symbols are values of damping factors associated with the solutions. (d) A trade-off QS model (in the logarithmic scale) under the constraint 0 b QS b 100.
J. Xia et al. / Journal of Applied Geophysics 82 (2012) 137–144 143
Fig. 7. (a) Raw Rayleigh-wave data (shot 1778). (b) Attenuation coefficients. Those labeled “Measured” were calculated using Eq. (2), those labeled “Constrained” were calculated
using Eq. (1) with inverted quality factors labeled “Inverted Qs with constraint” in panel c, and those labeled “Unconstrained” were calculated using Eq. (1) with inverted quality
factors labeled “Inverted Qs without constraint” in panel c. (c) Inverted quality factors. Those labeled “Inverted Qs with constraint” were trade-off models under the constraint
0 b QS b 100 and those labeled “Inverted Qs without constraint” were trade-off models without the constraint 0 b QS b 100.
Li, Y., Oldenburg, D.W., 2000. Incorporation geological dip information into geophysical Tikhonov, A.N., Arsenin, V.Y., 1977. Solution of Ill-posed Problems. W.H. Winston and
inversion. Geophysics 65, 148–157. Sons, Washington DC.
Marquardt, D.W., 1965. An algorithm for least squares estimation of nonlinear parameters. Toksöz, M.N., Johnston, D.H., 1981. Preface. In: Toksöz, M.N., Johnston, D.H. (Eds.), Seismic
Journal of the Society for Industrial and Applied Mathematics 2, 431–441. Wave Attenuation. Society of Exploration Geophysicists, Tulsa, OK, pp. v–vi.
Menke, W., 1984. Geophysical Data Analysis—Discrete Inversion Theory. Academic Xia, J., Miller, R.D., Park, C.B., 1999. Estimation of near-surface shear-wave velocity by
Press, Inc, New York. inversion of Rayleigh wave. Geophysics 64 (3), 691–700.
Mitchell, B.J., 1975. Regional Rayleigh wave attenuation in North America. Journal of Xia, J., Miller, R.D., Park, C.B., Tian, G., 2002. Determining Q of near-surface materials
Geophysical Research 80, 4904–4916. from Rayleigh waves. Journal of Applied Geophysics 51 (2–4), 121–129.
Oldenburg, D.W., Li, Y., 2005. Inversion for geophysics, a tutorial. In: Butler, Dwain K. Xia, J., Miller, R.D., Park, C.B., Tian, G., 2003. Inversion of high frequency surface waves
(Ed.), Near-surface Geophysics. Society of Exploration Geophysicists, pp. 89–150. with fundamental and higher modes. Journal of Applied Geophysics 52 (1), 45–57.
Palmer, D., 1980. The Generalized Reciprocal Method of Seismic Refraction Interpretation. Xia, J., Xu, Y., Miller, R.D., Chen, C., 2006. Estimation of elastic moduli in a compressible
Society of Exploration Geophysicists, Tulsa, Oklahoma. Gibson half-space by inverting Rayleigh wave phase velocity. Surveys in Geophysics
Parker, R.L., 1994. Geophysical Inverse Theory. Princeton University Press. 27 (1), 1–17.
Rix, G.J., Lai, C.D., Spang Jr., A.W., 2000. In situ measurement of damping ratio using Xia, J., Xu, Y., Miller, R.D., Zeng, C., 2010. A trade-off solution between model resolution
surface waves. Journal of Geotechnical and Geoenvironmental Engineering 126 and covariance in surface-wave inversion. Pure and Applied Geophysics 167 (12),
(5), 472–480. 1537–1547.
Sheriff, R.E., Geldart, L.P., 1985. Exploration Seismology (Volume 1): History, Theory, Xia, J., Xu, Y., Luo, Y., Miller, R.D., Cakir, R., and Zeng, C., in press, Advantages of using
and Data Acquisition. Cambridge University Press, New York. multichannel analysis of Love waves (MALW) to estimate near-surface shear-
Steeples, D.W., Miller, R.D., 1990. Seismic-reflection methods applied to engineering, wave velocity: Surveys in Geophysics. doi:10.1007/s10712-012-9174-2.
environmental, and ground-water problems. In: Ward, S.H. (Ed.), Geotechnical Zhang, J., Toksöz, M.N., 1998. Nonlinear refraction traveltime tomography. Geophysics
and Environmental Geophysics 1. Society of Exploration Geophysicists, Tulsa, OK, 63, 1726–1737.
pp. 1–30. Zhdanov, M.S., 2002. Geophysical Inverse Theory and Regularization Problems. Elsevier
Tarantola, A., 1987. Inverse Problem Theory. Elsevier Science B.V., Amsterdam, The Science B.V., Amsterdam, The Netherlands.
Netherlands.