Академический Документы
Профессиональный Документы
Культура Документы
12714
ABSTRACT
First-break picking of microseismic data is a significant step in microseismic moni-
toring. There is a great error in conventional first-break picking methods based on
time domain analysis in low signal to noise ratio. S-transform may provide a novel
approach, it can extract the time–frequency features of the signal and reduce the
picking error because of its high time–frequency resolution and good time–frequency
clustering; however, the S-transform is not well suited for microseismic data with high
noise. For applications to array data where the weak signal has spatial coherency as
well as some distinct temporal characteristics, we propose to combine the shearlet
transform with a time–frequency transform. In the proposed method, the shearlet
transform is used to capture spatial coherency features of the signal. The information
of the signal and noise in shearlet domain is represented by shearlet coefficients. We
use the correlation of signal coefficients at adjacent fine scales to give prominence to
signal features to accurately discriminate the signal from noise. The prominent sig-
nal coefficients make the signal better gathered in time–frequency spectrum of the
S-transform. Finally, we can get reliable and accurate first breaks based on the change
of energy. The performance of the proposed method was tested on synthetic and field
microseismic data. The experimental results indicated that our method is outstanding
in terms of both picking precision and adaptability to noise.
1 INTRODUCTION signal to noise ratio (SNR). However, for the data with high
noise, because of the small underlying distinction between mi-
First-break picking is a fundamental step in microseismic data
croseismic signal and noise, the auto-picks obtained by time
processing. Automatic and accurate picking of first arrivals is
domain analysis methods are unsatisfactory.
an important prerequisite for high-precision hypocentre loca-
First-arrival picking methods based on time–frequency
tion (Maxwell et al. 2010). Various first-break picking algo-
analysis improve the picking accuracy in high-noise back-
rithms have been proposed, such as short- and long-time av-
ground to some extent. Time–frequency analysis can simul-
erage method (Allen 1978; Earle and Shearer 1994), Akaike
taneously analyse signal characteristics in time and frequency.
information criterion (Leonard 2000; Zhang, Thurber and
The time–frequency spectrum reflects more comprehensive
Rowe 2003), higher order statistics (Saragiotis, Hadjileon-
signal features. Therefore, it is suitable for signal recogni-
tiadis and Panas 2002), polarization analysis (Deflandre and
tion and first-break picking. Wavelet and S-transforms are
Dubesset 1992) and fractal dimension (Boschetti, Dentith and
commonly used in time–frequency analysis methods. Wavelet
List 1996). These methods get satisfactory results for high
transform decomposes the signals into time-scale domain, not
the time–frequency domain in the strict sense, which makes it
∗ E-mail: cz1@ualberta.ca impossible to give an intuitive time–frequency representation.
C 2018 European Association of Geoscientists & Engineers 85
86 Y. Cheng, Y. Li and C. Zhang
S-transform is the phase correction of the continuous wavelet where the window width σ is assumed to be σ = 1/| f |.
transform and maintains a direct relationship with the Fourier The window width varies inversely with the frequency,
spectrum. S-transform provides frequency-dependent reso- which makes the S-transform to have the ability of multi-
lution, showing excellent time–frequency analysis capacity. resolution analysis. The time window is wide at low-frequency
All these features make sure that the S-transform possesses band to obtain high-frequency resolution and the time win-
some desirable characteristics that are absent in the continu- dow is narrow at high-frequency band, which can obtain high
ous wavelet transform. Pinnegar and Mansinha (2003) used time resolution. Based on (1) and (2), the S-transform is rede-
S-transform to determine P-wave time arrival in a noisy seis- fined as:
mogram. On the basis of the time–frequency difference be- +∞
| f | −(t−τ )2 f 2
tween the P-wave arrival and noise, the first break is well S (τ, f ) = h (t) √ e 2 e−i2π f t dt. (3)
−∞ 2π
indicated in time–frequency spectrum. Microseismic data are
The Gaussian window of S-transform satisfies (Pinnegar
characterized by high-signal frequency and usually disturbed
and Mansinha 2003)
by a lot of noise (Akram and Eaton 2016). Due to the weaker
+∞
amplitude and higher frequency of microseismic signals, there | f | −(t−τ )2 f 2
√ e 2 dτ = 1. (4)
is no significant time–frequency difference between signal and −∞ 2π
noise, which makes the S-transform difficult to pick reliable Therefore, the integral of time–frequency spectrum over
first break. time becomes the frequency spectrum.
To overcome this shortcoming, this paper proposes a
+∞ +∞ +∞
| f | −(t−τ )2 f 2
first-break picking method based on cascading use of shearlet S (τ, f )dτ = h (t) √ e 2 dτ e−i2π f t dt
and S-transforms. The method takes the advantage of shear- −∞ −∞ −∞ 2π
+∞
let transform in capturing the geometric structures for two- = h (t)e−i2π f t dt = H ( f ) , (5)
dimensional signal. By the correlation of signal coefficients −∞
at adjacent scales, the signal produces features that are more where H( f ) represents the Fourier transform of h(t). Hence,
obvious. Therefore, the time–frequency performance of valid the original signal can be recovered by inverse Fourier trans-
signal is distinct from noise after the S-transform. Based on form:
the energy difference between signal and noise, we realize the +∞ +∞ +∞
first-break picking automatically. The effectiveness and the h (t) = H ( f )ei2π f t d f = S (τ, f ) dτ ei2π f t d f.
−∞ −∞ −∞
accuracy of the proposed method are demonstrated by exper- (6)
iments on the synthetic and field microseismic data.
Based on the above relationships, there is a link be-
tween S-transform and Fourier transform, which ensures
2 THEORY
that S-transform is invertible. The S-transform realizes the
2.1 S-transform transformation of the signal from time domain to time–
frequency domain, then the inverse S-transform can be used
Stockwell, Mansinha and Lowe (1996) proposed a non-
to revert the signal to the time domain without any loss of
stationary signal analysis method, S-transform, which is an
information.
extension of the short time Fourier and wavelet transforms.
The S-transform of a time series h(t) is defined as:
+∞ 2.2 Shearlet transform
S (τ, f ) = h (t)ω (t − τ, f ) e−i2π f t dt, (1)
−∞ Shearlet transform is an almost optimal sparse method, which
applies the framework of affine system and has the ability
where S(·) represents S-transform, t is the time and f is the
to capture the geometry of multidimensional data effectively
frequency, and τ denotes the position of the window function
(Lim 2010; Merouane, Yilmaz and Baysal 2015; Hosseini
on the time axis. S-transform adopts a normalized Gaussian
et al. 2015; Zhang et al. 2015). It has a simpler mathematical
window with the following representation:
structure and is able to achieve the optimal approximation
1 −t 2 and multiresolution analysis of the signals. The construction
ω (t) = √ e 2σ 2 , (2)
σ 2π of shearlet system involves three important parameters, which
control the resolution, orientation and the position of the
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
First-break picking by Shearlet and Stockwell transforms 87
shearlet, respectively. Shearlet transform shows directional between adjacent fine scales (Zhao et al. 2016), which is de-
sensitivity, and it can analyse the anisotropic information scribed as:
of the data at multiple scales. More mathematical details of
shearlet transform can be found in the Appendix. ECorr j, j+1 (k) = E ( j, k) E ( j + 1, k) . (8)
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
88 Y. Cheng, Y. Li and C. Zhang
Figure 1 Comparison of the proposed method and shearlet noise attenuation (select the sixth trace for detailed comparison). (a) Synthetic
microseismic record. (b) Time–frequency spectrum of the sixth trace before noise suppression. (c) Time–frequency spectrum of the sixth trace
after noise suppression. (d) Time–frequency spectrum of the shearlet coefficients in the sixth trace after the multiplication of adjacent fine scales
(third and fourth).
sampling point of the first and second local maxima that are pression, respectively. Figure 1d shows the time–frequency
larger than 0.05, respectively. spectrum of the shearlet coefficients after the multiplication
The S-transform cannot get satisfactory first break picks of adjacent fine scales. Through contrastive analysis, we can
in high noise background, because there is no remarkable see that most of the noise is effectively suppressed by shear-
time–frequency difference between the signal and noise. Com- let noise attenuation. However, the time–frequency values in
bining random noise attenuation with the S-transform is a Fig. 1c change around 0.175 s, which seriously deviates from
representative method of trying to improve the signal to noise the true first break time. The accuracy of S-transform picking
ratio (SNR) before picking. Next, we compare our method based on noise attenuation depends on good reconstruction
versus a combination of noise attenuation using shearlets fol- quality and denoising effect. However, the coefficients related
lowed by S-transform picking. We select a trace randomly to the signals and noise are very close in high noise condition,
(6th) from the synthetic record shown in Fig. 1a for detailed which makes it difficult to distinguish the signal from noise via
comparison (the true first break is 0.187 s). Figure 1b and c the threshold and leads to unsatisfactory denoising results. By
shows the time–frequency spectra before and after noise sup- contrast, the proposed method can provide more accurate and
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
First-break picking by Shearlet and Stockwell transforms 89
Figure 2 (a) Pure synthetic record. (b) Enlarged view of (a) (the red lines represent the vertical components for the geophones and the blue lines
represent the horizontal components).
Figure 3 Noisy synthetic records. (a) Noisy record added with white Gaussian noise. (b) Noisy record added with coherent noise.
reliable first break for noisy microseismic data. It is worth em- containing 24 traces, which is composed of three components
phasizing that the shearlet transform is not used as a denoising with the dominant frequency of 200 Hz (see Fig. 2). The sam-
pre-processing in our method, we only consider the character- pling rate is 1 ms, and the spatial sampling of the array of
istic difference between signal and noise in shearlet domain. 24 receivers is 10 m. The vertical components for the geo-
Based on the correlation of the signal coefficients at adjacent phones are represented by the red lines, and the horizontal
scales, the weak signal shows significant property that differs components are represented by the blue lines. Figure 3a and b
from noise. As a result, the signal is better gathered in time- shows noisy records added with white Gaussian noise (WGN)
frequency spectrum of the S-transform. Our method shows and coherent noise for a signal to noise ratio (SNR) of –9 dB,
better performance than the picking method based on shear- respectively.
let noise attenuation. This is due to using a standard noise We take the noisy record contaminated by WGN as an ex-
attenuation method without the enhancements for process- ample to illustrate the detailed implementation process of our
ing in the shearlet transform domain discussed in this paper. method. We use shearlet transform to decompose the record
However, these enhancements are essential to overcome the for four scales and each scale is divided into six directions.
large noise. The third and the fourth scales are fine scales that contain
effective information. Figure 4a–f shows the coefficients in
different directions of the fourth scale from the first to the
4 APPLICATIONS sixth directions, respectively. The correlation energy between
the two adjacent fine scales is shown in Fig. 5. We can see
4.1 Synthetic microseismic record that the energy of the first two directions is larger than other
To illustrate the performance of the proposed method, we use directions. It also shows that the valid signal is mainly mapped
Ricker wavelet to establish a synthetic microseismic record to the first two directions, as demonstrated by Fig. 4a and b.
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
90 Y. Cheng, Y. Li and C. Zhang
Figure 4 (a–f) Shearlet coefficients of the fourth temporal frequency scale from the first to the sixth directions.
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
First-break picking by Shearlet and Stockwell transforms 91
Figure 6 Comparison waveform and time–frequency spectrum of the sixth trace in Fig. 3a. (a) Waveform of the sixth trace. (b) Time–frequency
spectrum of the sixth trace. (c) Waveform of the sixth trace after the multiplication of the coefficients for the same position at adjacent fine scales
(third and fourth). (d) Time–frequency spectrum of the sixth trace after the scale multiplication. (e) Normalized energy. (f) Energy difference
between adjacent sampling points.
To further verify the reliability of the proposed method, shows the waveform data examples added with different types
we simulate a Ricker wavelet that contains P-wave and of noise (SNR = –8 dB).
S-wave, as shown in Fig. 8a. The dominant frequency is 200 Tables 1 and 2 show the statistical results of P- and S-
Hz and the sampling rate is 1 ms. We add WGN and real noise wave time arrivals in two groups of tests, respectively. In our
to the pure wavelet, respectively, and generate noisy wavelets paper, when the absolute error between the auto-pick and the
with different SNRs. The SNR is defined as: true arrival is within 3 ms, the pick is considered as an ac-
curate pick. The two tables also list the statistics of the picks
|s (t)|2 whose absolute errors are zero (pick with no error) and not
SNR = 10log10 t
, (11)
t |y (t) − s (t)|2 more than 1 ms (pick within 1 ms). It can be seen that our
method can detect the P- and S-wave arrivals with high ac-
where y(t) is the noisy signal and s(t) is the original signal. curacy. It provides reliable results for noisy data, most of the
We conduct 1000 experiments in each group. Figure 8b and c picks are satisfactory, even at the SNR of –10 dB.
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
92 Y. Cheng, Y. Li and C. Zhang
Figure 7 First-break picking results for synthetic microseismic records. (a and b) Comparison of the proposed method (black dots) and the
S-transform method (green dots) for noisy records with white Gaussian noise and coherent noise, respectively. (c and d) Enlarged views of (a)
and (b).
Figure 8 (a) Pure Ricker wavelet. (b) Noisy wavelet added with white Gaussian noise. (c) Noisy wavelet added with real noise.
In order to show the superiority of the proposed method and the histograms of the absolute picking errors, it is clear
more intuitively, we give the histogram comparison of our to see that the proposed method can provide accurate P-wave
method and the S-transform method at the SNR of –8 dB, and S-wave time arrivals in high noise conditions. When the
which is shown in Fig. 9. We can observe that the S-transform SNR is lower, the picking accuracy of P-wave is slightly worse
method loses the ability for accurate picking when dealing than that of S-wave. This is due to the weaker amplitude of
with the data containing strong noise, only a small part of the P-wave, when the signal is disturbed by strong noise, the
the auto-picks is acceptable. In comparison, our method re- result of the P-wave arrival is probable to be false or missed.
duces the picking errors significantly, the picking results of
the P- and S-wave arrivals under different noise conditions
4.2 Field microseismic record
are convincing. More than 95% of the arrival times are de-
tected within 2 ms, among which nearly 60% are consistent To further validate the effectiveness of our proposed method,
with the real arrival times. From the data given in two tables we process a field microseismic data set containing 15 traces as
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
First-break picking by Shearlet and Stockwell transforms 93
SNR Number Accurate Pick with Pick within Accurate Pick with Pick within
(dB) of Tests Pick No Error 1 ms Pick No Error 1 ms
SNR Number Accurate Pick with Pick within Accurate Pick with Pick within
(dB) of Tests Pick No Error 1 ms Pick No Error 1 ms
Figure 9 Histograms of picking absolute error (s). The pick is considered as a valid pick when the absolute error is less than 3 ms. Note that a
large proportion of picks for the S-transform are outside this range of valid picks and the corresponding histogram classes are not shown on
the plots. (a and b) Statistics of P- and S-wave time arrivals in white Gaussian noise test, respectively. (c and d) Statistics of P- and S-wave time
arrivals in real noise test, respectively.
shown in Fig. 10a, where the vertical components are shown breaks is shown in Fig. 10b. We apply our method and the
in red and the horizontal components are shown in blue. The S-transform method to pick the arrival times. The picks are
sampling rate is 1 ms, and the spatial sampling of the array marked by black and green dots, respectively. Results show
of 15 receivers is 8 m. A magnified view around the first that most of the picks obtained by the S-transform method are
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
94 Y. Cheng, Y. Li and C. Zhang
Figure 10 First-break picking results for field microseismic record. (a) Field microseismic record. (b) Picking results of the proposed method
(black dots) and the S-transform method (green dots).
acceptable. However, for the third, fifth and the 14th traces, our method. The synthetic models in this paper only contain
whose valid signals are difficult to identify, the method gives P- and S-waves. The influence of other waveforms received
improper results. In contrast, the accuracy and consistency by the geophone in microseismic monitoring is neglected. In
of the arrival time picking are improved by our method; the future research, we will explore the picking performance of
picking times are reasonable and satisfactory. the proposed method in dealing with complex microseismic
data, and try to find effective countermeasures to resolve its
deficiencies.
5 CONCLUSIONS
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
First-break picking by Shearlet and Stockwell transforms 95
Hosseini S.A., Javaherian A., Hassani H., Torabi S. and Sadri M. of real number, positive real number and two-dimensional real
2015. Adaptive attenuation of aliased ground roll using the shearlet number, respectively. Mas is a matrix that can be factorized as
transform. Journal of Applied Geophysics 112, 190–205.
Kutyniok G. and Labate D. 2009. Resolution of the wavefront set Mas = Bs Aa = (1 s )(a √0 ), where Aa is the anisotropic dilation
01 0 a
using continuous shearlets. Transactions of the American Mathe- matrix and Bs is the shear matrix. The expression shows that
matical Society 361, 2719–2754.
the collection of basis functions ψa,s,μ are generated by ap-
Leonard M. 2000. Comparison of manual and automatic onset time
picking. Bulletin of the Seismological Society of America 90, 1384–
plying dilations, shear transformations and translations to a
1390. well-localized window function ψ(x).
Lim W.Q. 2010. The discrete shearlet transform: a new directional The generating function ψ should satisfy appropriate ad-
transform and compactly supported shearlet frames. IEEE Trans- missibility conditions (Kutyniok and Labate 2009). For the
actions on Image Processing 19, 1166–1180. point ξ = (ξ1 , ξ2 ) ∈ R̂2 , ξ1 = 0, R̂ represents frequency do-
Maxwell S. C., Rutledge J., Jones R. and Fehler M. 2010. Petroleum
main; let ψ be given by
reservoir characterization using downhole microseismic monitor-
ing. Geophysics 75, 75A129–75A137.
ψ̂a,s,μ (ξ1 , ξ2 ) = a 3/4 e−2πiξ μ ψ̂1 (aξ1 ) ψ̂2 a 1/2 (ξ2 /ξ1 − s) , (A5)
Zhao X., Li Y., Zhuang G.H., Zhang C. and Han X. 2016.
2-D TFPF based on contourlet transform for seismic random noise
attenuation. Journal of Applied Geophysics 129, 158–166.
APPENDIX
SHEARLET TRANSFORM
(A1)
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96
96 Y. Cheng, Y. Li and C. Zhang
where ψ̂a,s,μ is the Fourier transform of ψa,s,μ . The frequency we can see that each element ψ̂a,s,μ is supported on origin-
support of each function ψ̂a,s,μ can be written as symmetric trapezoids, the orientation is along a straight line
with the slope equals to S. The support becomes increasingly
sup pψ̂a,s,μ ⊂ ξ1 , ξ2 : ξ1 ∈ −2 a, −1 2a ∪ 1 2a, 2 a ,
√ elongated as a tends to zero. Shearlets show excellent local-
ξ ξ − s ≤ a . (A6)
2 1
ization and compact support characteristics, which makes the
Figure A1 gives the frequency support of shearlets for dif- shearlet transform more effective in capturing the geometric
ferent values of a and s (Guo and Labate 2009), from which structure of multidimensional data.
C 2018 European Association of Geoscientists & Engineers, Geophysical Prospecting, 67, 85–96