Professional Documents
Culture Documents
10.1190/GEO2011-0260.1
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Manuscript received by the Editor 19 July 2011; revised manuscript received 1 November 2011; published online 16 February 2012.
1
University of Alberta, Department of Physics and Signal Analysis and Imaging Group (SAIG), Edmonton, Alberta, Canada. E-mail: verarodr@ualberta.ca;
bonar@ualberta.ca; msacchi@ualberta.ca.
© 2012 Society of Exploration Geophysicists. All rights reserved.
V21
V22 Vera Rodriguez et al.
2 3
crosscorrelations. In reflection seismology, Bonar and Sacchi a1
(2010) proposed a method for spectral decomposition to 6 7
6 a2 7
analyze seismic sections. The decomposition technique is based 6 7
6 7
upon the transform of the time domain data into a sparse time- s ¼ ½ W1 W2 W3 ::: WN 6 a 7
6 3 7 ¼ Ra: (3)
frequency map. When implemented for denoising purposes this 6 . 7
6 .. 7
method is equivalent to basis pursuit denoising (BPDN) (Chen 4 5
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
et al., 1998).
aN
In this study, we are interested in attenuating noncoherent or am-
bient noise. Following the exposition by Bonar and Sacchi (2010),
we seek a sparse representation of the microseismic signal in terms Equation 3 is an underdetermined system that accepts an infinite
of a dictionary of complex Ricker wavelets. By imposing sparsity number of solutions. In other words, there are multiple possible re-
and misfit constraints to the decomposition process, we find a de- presentations of s in terms of linear combinations of the columns
noised version of the original signal. An extension of the time- of R. A common approach to solve underdetermined systems is
frequency transform to 3C data is accomplished through the concept through the damped least-squares (DLS) method. Mathematically,
of group sparsity (Yuan and Lin, 2006; Eldar and Bolcskei, 2009). this entails solving the following problem
Group sparsity is a condition that arises naturally in problems such
as linear regression (Yuan and Lin, 2006), brain activity imaging
(Cotter et al., 2005), or the determination of the direction of arrival âl2 ¼ arg min½ks − Rak22 þ λl2 kak22 : (4)
a
in antennas (Hyder and Mahata, 2010). In the field of microseismi-
city analysis, group sparsity has been recently used in the joint de- The cost function inside the square brackets consists of two
termination of origin time, hypocenter location, and moment tensor terms. The first term is the misfit function, a measure of fidelity
of microseismic events (Vera Rodriguez et al., 2010a, 2010b). Con- of fit between the observations and the synthesized data, and the
sidering that the three components of a single receiver share the second term is the l2 regularization term. Minimization of this
same sparsity pattern in the time-frequency plane, a group sparsity cost function yields the DLS solution âl2 . In this case, the solu-
thresholding criteria (Fornasier and Rauhut, 2008) is implemented tion vector of coefficients can be easily found utilizing the meth-
with a fast iterative soft thresholding algorithm (FISTA) (Beck and od of conjugate gradients (CG) (Hestenes and Stiefel, 1952). The
Teboulle, 2009) to simultaneously recover denoised versions of the
trade-off parameter λl2 is used to balance the relative strengths of
three components. The resulting methodology has been tested with
the misfit and regularization terms. An alternative solution to our
synthetic and real microseismic data sets. In all cases, the improve-
problem entails using an l1 regularization term. In other words,
ment in S/N with respect to the input noisy traces is clearly visible in
we will impose a measure to promote sparsity on the coefficients
the reconstructions obtained through the 3C sparse time-frequency
needed to represent the seismogram. This can be accomplished by
transform.
solving
1C SPARSE TIME-FREQUENCY TRANSFORM âl1 ¼ arg min½ks − Rak22 þ λl1 kak1 : (5)
a
The decomposition of a discrete seismic trace s½n in terms of
dilations and translations of a reference complex wavelet w½n is Minimization of the cost function inside square brackets in equa-
given by (Bonar and Sacchi, 2010) tion 5 yields the sparse solution âl1 . The trade-off parameter λl1
balances the fidelity of fit versus the l1 regularization term.
X
s½n ¼ ak ½iwk ½n − i; (1) Sparse solutions from l1 regularization are estimated using the
i;k
fast iterative soft thresholding algorithm (FISTA) (Beck and
Teboulle, 2009) (Appendix A). The denoised version of the input
where the indexes i and k are related, respectively, to the translations data is obtained via
and dilations of w½n, and the complex coefficients ak ½i control the
amplitude and phase of the wavelets. More specifically, the index k ŝ ¼ Râlp ; (6)
reflects a parameter/property of the wavelet that controls its dilation
or frequency content. In a Ricker wavelet, for example, k represents
the central frequency of the wavelet. In matrix-vector notation, where p can be either one or two depending on which regular-
equation 1 can be written as the convolution of a set of complex ization approach we are using. In Figure 1, we provide a syn-
coefficient vectors fak g with a set of dilated versions of the refer- thetic example where we examine the two denoising schemes.
ence wavelet fwk g, In this case, a complex Ricker wavelet is used as the reference
wavelet to create the dictionary R (Figure 2), where the dilation
X
s¼ W k ak ; (2) index k is related to the central frequency that describes each
k wavelet in R. Complex wavelets can be easily computed using
the Hilbert Transform. Notice that, although the construction
where s is the vector representation of the trace s½n and Wk is the of the dictionary requires a fixed reference wavelet, in this case
convolution matrix of the wavelet with dilation index k. To general- a complex Ricker wavelet, the dictionary is not adaptive and can
ize this model, the convolution matrices of a set of N dilated ver- be used to represent a wide range of source signatures in the
sions of the reference wavelet can be incorporated to create a input data through translations and changes in phase and fre-
redundant dictionary R (Mallat, 2008), quency content of the reference wavelet. To produce a fair
Microseismic data denoising V23
assessment of the denoising results, we have determined λl1 and operator (Fornasier and Rauhut, 2008) is introduced into the inver-
λl2 using the χ 2 criterion for the goodness of fit (see e.g., Aster sion routine (see Appendix A). The denoised version of the 3C data
et al., 2005). Representations using DLS tend to be “spread” in is finally recovered via
the solution (Figure 1c), whereas, sparse representations are well
localized and produce cleaner signal estimations using a smaller
number of nonzero coefficients (Figure 1e). For the purpose of
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
S ¼ ½x y z ; (7)
where the columns of Âl2;1 are the sparse representations in the time-
frequency plane of the input three components. The symbol jj · jjp;q
denotes the mixed lp;q -norm defined as (Kowalski and
Torrésani, 2009)
X X q 1
p q
p
kXkp;q ¼ n k
jxn;k j ; (9)
where xn;k is the nth row, kth column element of the matrix X. To Figure 2. Reference complex Ricker wavelet. Real part (dashed
apply FISTA to MMVs a multidimensional shrinkage-thresholding line), imaginary part (dotted line), and envelope (solid line).
V24 Vera Rodriguez et al.
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Figure 3. Comparison of denoising results between the 1C and 3C versions of the sparse time-frequency transform. (a) Synthetic 3C data. (b)
Synthetic data contaminated with random noise. (c) Denoised data using the 1C sparse time-frequency transform. (d) Denoised data using the
3C sparse time-frequency transform. H1 and H2 stand for horizontal components.
APPLICATION TO DENOISE
MICROSEISMIC DATA
The 3C sparse transform is first tested with a synthetic microseis-
mic data set. The synthetic data consists of direct P and S arrivals
computed in a 1D velocity model (Figure 4). The amplitude varia-
tions due to the source radiation pattern are included in the model-
ing, corresponding to a pure double couple source mechanism.
Wave arrivals are computed in a vertical array of eight 3C receivers
located at a horizontal distance of ∼400 m from the source location.
The resulting synthetic data is dominated by the amplitudes of the S
arrivals (Figure 5, first row). The S/N is defined by
μs
S∕N ¼ ; (11)
σn
visible improvement in the S/N, however, the signal of interest is Notice also that low-pass frequency filters introduce apparent de-
also negatively affected in its frequency content (Figure 5, third lays in the wave arrivals, which are not present in the sparse
row). The denoised version of the data using the 3C sparse trans- transform.
form on the other hand, preserves the frequency content of the wave Figures 7 through 11 present the results of applying the 3C sparse
arrivals while removing also the random noise (Figure 5, fourth transform to a real microseismic data set. The time-frequency maps
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
row). The advantage of denoising using the 3C sparse transform obtained from the sparse transform provide an alternative domain to
is particularly evident in the vertical component of this example, analyze the microseismic data (Figure 7). Furthermore, the time-
where the weak S arrivals in the first six receivers are clearly below frequency analysis of the microseismic traces can reveal variations
the noise level in the noisy data, but are partially recovered by the in the dominant frequency of the wave arrivals in different receivers.
transform (Figure 5, column c). Another important property of In the example shown in Figure 7, the arrivals in the deeper three
the transform is its ability to preserve the amplitude information receivers display a lower dominant frequency than in the rest of the
of the wave arrivals. Amplitude information is important for ana- array of receivers. If time picks were to be obtained from these time-
lyses of, for example, polarization angles and source mechanisms. frequency maps, different frequency slides can be analyzed to
Polarization angles are usually estimated through hodogram determine where the onset of the arrivals is more evident. In this
analysis (Figure 6). Random noise introduces uncertainty in the es- example, the maximum dominant frequency observed in all recei-
timation of polarization angles by causing the hodogram curves to vers for the arrivals around 60 ms is used to low-pass filter the raw
deviate from the actual wave arrival angles (Figure 6, column b). microseismic data (Figures 8b, 9b, 10b, and Figure 11, column b).
The effect of reducing the frequency content in the input data is This frequency filter is not optimum however for the receivers
to smooth the hodogram curves, which can allow a more robust where the dominant frequency is lower. Additionally, the compo-
estimation of polarization angles (Figure 6, column c). Similar ef- nent of the random noise that is below the cut-off frequency of
fects are observed when the traces are denoised using the 3C sparse the filter is preserved in the data. As a consequence, the low-pass
transform (Figure 6, column d), with the advantage that the ampli- filtered traces display higher noise content before the first arrivals
tude and frequency content of the arrivals of interest are preserved. when compared with the traces that were denoised with the sparse
transform. The presence of cleaner wave arrivals in the denoised between ∼0.9 and ∼1.3 in synthetic and real data examples ana-
traces is of advantage for processes such as automatic time-picking, lyzed in this study. The best λl2;1 value for a particular data set
selection of hodogram windows, and amplitude stacking. After the can be found by constructing a Pareto curve (e.g., Berg and
first arrival, the sparse transform retains the oscillations of the signal Friedlander, 2008) or by simple trial and error. The preservation
that represent potential information. For example, consider the of the amplitude and phase properties of the arrivals in the input
arrival around ∼110 ms in the H1 component of receiver 12
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Figure 6. Comparison of hodogram analysis for receiver three of the synthetic microseismic data set of Figure 5. Column (a) synthetic data.
Column (b) synthetic data contaminated with random noise. Column (c) noisy data filtered with a low-pass frequency filter. Column (d)
denoised data using the 3C sparse transform. Vertical lines in the first row delimit the hodogram window. Hodogram curves in column
(a) are repeated with dotted lines in columns (b), (c), and (d) for comparison. H1 and H2 stand for horizontal components.
Microseismic data denoising V27
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Figure 8. Denoising results in the horizontal H1 component of the Figure 10. Denoising results in the vertical component of the real
real microseismic data. (a) Raw microseismic data. (b) Raw data microseismic data. (a) Raw microseismic data. (b) Raw data filtered
filtered with a low-pass frequency filter. (c) Denoised data using with a low-pass frequency filter. (c) Denoised data using the 3C
the 3C sparse transform. sparse transform.
V28 Vera Rodriguez et al.
ACKNOWLEDGMENTS
This work was partially supported by the
sponsors of the Signal Analysis and Imaging
Group (SAIG) and Pinnacle/Halliburton. We also
acknowledge Fairborne Energy Ltd. for provid-
ing the real microseismic data used in this study.
APPENDIX A
FISTA FOR GROUP SPARSE
SOLUTIONS
Figure 11. Comparison of hodogram analysis with real microseismic data (receiver 7). To incorporate the mixed l2;1 -norm within the
Column (a) raw data. Column (b) data filtered with a low-pass frequency filter. Column FISTA algorithm for the purposes of promoting
(c) denoised data using the 3C sparse transform. Vertical lines in the first row delimit the group sparsity, a brief introduction to the iterative
hodogram window. Hodogram curves in column (a) are repeated with dotted lines in soft thresholding function (ISTA) is presented.
columns (b) and (c) for comparison.
The mathematical framework for this algorithm
was first presented in Daubechies et al. (2004).
CONCLUSIONS Minimizing the cost function defined in equa-
tion 5 through a gradient based approach, the ISTA algorithm iter-
Processing of microseismcity records is a challenge due to, ates toward the solution, from an initial estimate a0 , by:
among other problems, the low S/N displayed by the signals of in-
terest. Application of techniques and processing tools developed in 1 T
other knowledge areas can be of great advantage to deal with akþ1 ¼ T λl ∕2α ak − R ðs − Rak Þ : (A-1)
1 α
microseismicity records. The method presented in this study is a
time-frequency transform constrained to impose sparsity into the The constant α is chosen to be greater than or equal to the maximum
time-frequency representation of the signal. When compared eigenvalue of RT R to prevent the argument, i.e., gradient step, with-
against the equivalent l2 -norm transform, the sparsity constrained in the operator T τ from becoming a negative function. The operator
transform using the l1 -norm displays improved results for denoising T τ is known as the soft thresholding function and is defined by
purposes. This transform can be implemented component by
component (1C algorithm) or by simultaneously utilizing the three T τ ðgÞ ¼ ðjgj − τÞþ signðgÞ; (A-2)
components of a receiver (3C algorithm). In particular, the 3C ver-
sion of the sparse transform is advantageous in cases where the where g is a scalar value, τ is a value greater than zero, and
wave arrival exists in the three components of a receiver but the ðjgj − τÞþ is the maximum between ðjgj − τÞ and zero. To accelerate
amplitude of the arrival is below the noise level in one component. the convergence rate of ISTA, the FISTA algorithm, developed in
In such cases, the 3C algorithm uses information from the compo- Beck and Teboulle (2009), uses a very specific linear combination
nents where the arrival amplitudes are above the noise level to re- of the previous two points or iterations. Specifically, this combina-
cover the amplitude information in the other component. The results tion at the kth iteration is
Microseismic data denoising V29
tk−2 Berg, E., and M. Friedlander, 2009, Probing the pareto frontier for basis
pursuit solutions: SIAM Journal on Scientific Computing, 31, no. 2,
yk ¼ ak−1 þ ðak−1 − ak−2 Þ; (A-3)
tk 890–912, doi: 10.1137/080714488.
Bonar, D., and M. Sacchi, 2010, Complex spectral decomposition via inver-
sion strategies: 80th Annual International Meeting, SEG, Expanded
where the variable tk is defined by Abstracts, 1408–1412.
Chen, S., D. Donoho, and M. Saunders, 1998, Atomic decomposition
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi by basis pursuit: SIAM Journal on Scientific Computing, 20, no. 1,
1þ 1 þ 4t2k−1
Downloaded 02/09/15 to 129.96.252.188. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/