A High Accurate Automated First-Break Picking Method For Seismic Records From High-Density Acquisition in Areas With A Complex Surface
A High Accurate Automated First-Break Picking Method For Seismic Records From High-Density Acquisition in Areas With A Complex Surface
A High Accurate Automated First-Break Picking Method For Seismic Records From High-Density Acquisition in Areas With A Complex Surface
12923
ABSTRACT
As the application of high-density high-efficiency acquisition technology becomes
more and more wide, the areas with complex surface conditions gradually become
target exploration areas, and the first-break picking work of massive low signal-to-
noise ratio data is a big challenge. The traditional method spends a lot of manpower
and time to interactively pick first breaks, a large amount of interactive work af-
fects the accuracy and efficiency of picking. In order to overcome the shortcoming
that traditional methods have weak anti-noise to low signal-to-noise ratio primary
wave, this paper proposes a high accurate automated first-break picking method for
low signal-to-noise ratio primary wave from high-density acquisition in areas with
a complex surface. Firstly, this method determines first-break time window using
multi-azimuth spatial interpolation technology; then it uses the improved clustering
algorithm to initially pick first breaks and then perform multi-angle comprehensive
quality evaluation to first breaks according to the following sequence: ‘single trace
→ spread → single shot → multiple shots’ to identify the abnormal first breaks;
finally it determines the optimal path through the constructed evaluation function
and using the ant colony algorithm to correct abnormal first breaks. Multi-azimuth
time window spatial interpolation technology provides the base for accurately picking
first-break time; the clustering algorithm can effectively improve the picking accuracy
rate of low signal-to-noise ratio primary waves; the multi-angle comprehensive qual-
ity evaluation can accurately and effectively eliminate abnormal first breaks; the ant
colony algorithm can effectively improve the correction quality of low signal-to-noise
ratio abnormal first breaks. By example analysis and comparing with the commonly
used Akaike Information Criterion method, the automated first-break picking theory
and technology studied in this paper has high picking accuracy and the ability to sta-
bly process low signal-to-noise ratio seismic data, has a significant effect on seismic
records from high-density acquisition in areas with a complex surface and can meet
the requirements of accuracy and efficiency for massive data near-surface modelling
and statics calculation.
∗ E-mail: xuyinpo0521@163.com
1228
C 2019 European Association of Geoscientists & Engineers
A first-break picking method for seismic records 1229
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1230 Xu et al.
Figure 1 Shot gather data. Shot gather data are divided into four quadrants.
the statistics difference between the peak value and the inflex- is comparatively worse for the data from areas with a complex
ion point. This statistics method is verbose and is affected by surface; however, such method is quick and easy to perform.
factors such as the similarity among seismic traces. Gelchinsky The third type of method is the artificial intelligence
and Shtivelman (1983) proposed a correlation method which method that uses multiple dimensions’ information of seismic
assumes that the impulse shape of each trace does not change, primary wave, for example, fractal travel-time picking tech-
but the actual situation is not consistent with the assumption. nology and neural network technology. Boschetti, Dentith
Such method can inhibit noise to some extent, but it is affected and List (1996) proposed a first-break detection method based
by factors such as the similarity of seismic traces; so the effect on fractal dimension, which is based on the feature that the
Figure 2 Shot gather data are divided into four quadrants, and the seismic traces in each quadrant are sorted according to offset from small to
big.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1231
Figure 3 Designed time window. The length of the time window by default is 100 samples.
fractal dimension of seismic trace changes as the signal appears first-break picking technology based on artificial intelligence.
in the seismic trace to determine the first break of the seismic Chen (2018) proposed an unsupervised machine learning al-
trace; however, when the high-frequency interference in the gorithm to help recognize seismic waveforms in microseismic
background is strong, after the fractal dimension detects the or earthquake data. He used an unsupervised machine learn-
approximate position of the first-break time, the first-break ing algorithm to help to cluster the time samples into wave-
picked by the extreme value method in the local range will be form points and non-waveform points. Convolutional neural
interfered by impulse, which affects the accuracy of first-break networks (CNNs) perform much better than the traditional
picking. In order to further improve the picking accuracy by multilayer perceptron on the waveform classification of the
this method and enhance its effectiveness and practicability, noisy microseismic data. Duan et al. (2018) constructed an
Tosi et al. (1999) and Jiao et al. (2000) improved the fractal architecture of CNNs in machine learning, helping to identify
dimension method. Murat and Rudman (1992) applied the poor picks across multiple traces. Tsai et al. (2018) proposed
neural network learning algorithm to the first-break picking. a deep semi-supervised neural network to achieve automatic
Because the algorithm has slow convergence, is prone to local picking for the first break in seismic data. Comparing with su-
extremum and has difficulty to determine network structure, pervised neural networks, our proposed network can be used
and because the BP network is a static network, it cannot for both labelled and unlabelled data set to achieve higher ac-
learn additional samples to expand the network knowledge. curacy. Chen et al. (2019) proposed a novel anti-noise CNN
It is difficult to meet the requirements of first-break picking architecture for waveform classification and propose to com-
in area with a complex terrain. In order to improve the ap- bine k-means clustering (KC) with CNN classification to pick
plication of neural network technology in first-break picking, arrivals (CNN-KC). A group of synthetic and real high S/N
many improved techniques have been proposed (Fahlman data sets shows that the proposed methods are much more
and Lebiere 1990; Baluja and Fahlman 1994; Muller et al. robust than the state-of-the-art STA/LTA method in picking
1998; Shimshoni and Intrator 1998; Mercier et al. 2006). microseismic events; however, features of primary wave in
Recently, with the quick development of artificial intelli- low S/N data are not clear, and the first-break time is drown
gence technology, the deep learning model has shown excel- by noise. It is difficult to identify the first breaks; at the same
lent performances in solving a wide range of problems in sci- time, this technology needs a lot of proved samples for learn-
entific and industrial fields (Ferrari, Lombardi and Signoroni ing. The proved samples of low S/N first breaks are few, so
2017; Sze et al. 2017; Zhang et al. 2017, 2018; Mahmud target training lacks. This technology is being researched and
et al. 2018). Many scholars began to research the automated developed. There is no mature technology and software.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1232 Xu et al.
Figure 4 Single-trace seismic record (top); data in time window (middle); mean point classification (bottom).
As for low S/N first-break data, many scholars proposed and receiving by all receivers. In order to solve this prob-
the refraction interferometry to improve the S/N of primary lem, Bharadwaj et al. (2012) and Hanafy, Al-Hagan and Al-
wave. Dong, Sheng and Schuster (2006) proposed the basic Tawash (2011) proposed the supervirtual refraction interfer-
theory of refraction interferometry, that is, the interference ometry (SRI). This method is based on constructing virtual
method is used to extract the virtual refraction signal be- traces by cross-correlation and then performs convolution and
√
tween receivers, and its S/N can be increased by N times stacking for virtual traces and corresponding near traces to en-
(N is the quantity of shot points that produce virtual source hance refraction wave in raw data. The preparatory applica-
records). However, the ‘pseudo-event’ generated by the refrac- tion indicates that when picking first breaks in refraction wave
tion interferometry will reduce the wavelet resolution and the enhanced by SRI, the stability and accuracy of the picked first
stability of first-break picking. In order to solve the prob- breaks are greatly improved (Mallinson et al. 2011; Al-Hagan,
lem that the virtual refraction signal does not have accu- Hanafy and Schuster 2014), the accuracy of model created
rate excitation time, the supervirtual refraction interferom- by tomographic velocity is clearly improved (Mallinson et al.
etry is further developed, that is, obtain a virtual refraction 2011). In order to improve the stability and accuracy of the
record by cross-correlation and stacking, and then obtain a first-break picking in refraction wave, SRI has been improved
high S/N supervirtual refraction wave gather by convolution by scholars. The solved problems include the unevenness of
and stacking. However, this method can obtain the data by fold for far traces and near traces (An et al. 2017a) and low
shooting in virtual shot points and receiving by all receivers similarity of stacked traces (Lu Xuemei et al. 2018). Good
and cannot obtain the data by shooting in real shot points effect has been observed. However, the effect of SRI depends
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1233
Figure 5 Single-trace seismic record (top); data in time window (middle); mean points classification (bottom).
on the quantity of stacked shot points. The less the quantity thickness and velocity are relatively stable; however, because
of shot points involved in cross-correlation to generate virtual complex near-surface structures cause low S/N of primary
traces, the less is the stack time of virtual traces; the worse wave and serious cross-strata phenomenon, existing meth-
the ability of anti-noise, the lower is the accuracy; finally, the ods cannot effectively discriminate and correct abnormal first
worse is the effect of enhancement. breaks.
In recent years, in order to further improve the quality Based on the feature that the S/N of primary wave in
and quantity of automated first-break picking, Khan (2007) seismic data from high-density acquisition in areas with a
presented a completely automated two-step first-break pick- complex surface is low and on the base of fuzzy clustering al-
ing method. The first step picks the first breaks, while the gorithms researched by Zhu, Li and Zhang (2016) and Chen
second step acts as a quality control stage to validate the (2019), this paper studies the automated first-break picking
picked times and refine them. Sabbione et al. (2010) fitted method that combines the design of time window of primary
all picks to the best two straight lines per flank model. Then, wave, algorithms based on clustering, multi-angle compre-
picks with errors larger than 3σ are temporarily rejected and a hensive quality evaluation and ant colony algorithm. Firstly,
preliminary refraction model is obtained. First breaks are now multi-azimuth spatial interpolation technology is used to de-
repicked locally by the preliminary refraction model. This pro- termine the range of time window of primary wave, and then
cess usually leads to the correction of some picks. The above- the improved clustering algorithm is used to initially pick first
mentioned quality monitoring technology has better effect breaks. The adjacent traces’ waveform and energy are used
for the primary waves from the areas where the near-surface to correct the abnormal first break with big variation to right
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1234 Xu et al.
Figure 6 Initial picking result of the single shot based on the clustering algorithm.
position; then the quality of the picked first breaks is com- higher precision, but the workload is large, and therefore can-
prehensively evaluated through multi-angle according to the not meet the demands of real-time processing. At present, how
steps ‘single trace → spread → single shot → multiple shots’ to improve the accuracy of single-trace picking and how to ac-
to identify the abnormal first breaks; finally, the ant colony curately remove and correct abnormal first breaks are the key
algorithm is used to update the pheromone by the constructed and difficult issues for automated first-break picking. There-
evaluation function and determine the optimal path according fore, based on the principle of present automated first-break
to the change of the pheromone trajectory on different paths, picking methods, and after considering the random noise that
and then by combining reliable first breaks. The abnormal appears before primary wave, a kind of automated first-break
first breaks can be corrected to right position, so the accuracy picking method for massive low S/N data from areas with a
and quantity of first-break picking for massive low S/N data complex surface is studied, and the specific implementation
can be improved. steps are as follows:
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1235
Figure 7 The vector diagram of the relationship between shot points and receivers.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1236 Xu et al.
wave before arriving is affected by random noise. The change the control points of primary wave time window of each area,
of the noise before the arrival of primary wave is disorganized, the multi-azimuth spatial interpolation technology is used to
which cannot be represented by a definite mathematical for- determine the time window range of primary wave of each
mula, but the noise generally pertains to a random process of shot and each trace in the work area.
ergodic property. According to the definition of generalized The selection of the time window length is related to ter-
ergodic property (Wang et al. 2009), the time mean value is rain, data S/N etc. If the terrain undulates greatly, use eleva-
equal to the mean value of the total set, that is to say, the mean tion statics corrections to correct, select seed shots in different
value of the total set can actually be replaced by the average areas of geometry as much as possible. The length of the time
value of a sample function on the entire time axis. Based on window shall be relatively big and shall include primary wave.
the idea of clustering algorithm and by combining with the If the undulated terrain is small, select comparatively few seed
characteristics of the noise before the arrival of primary wave, shots; the length of the time window shall be relatively small.
the clustering algorithm can be used to pick single trace to im- As for low S/N data, the length of the time window shall be
prove the accuracy of single trace picking. The detailed steps small. Generally, the length of the time window is 100 sam-
are as follows: ples; the shorter the length of the time window, the smaller is
1.1 Centred at shot point, the shot line is the vertical axis, and the range for searching primary wave, and the higher is the
the receiver line is the horizontal axis. The seismic shot gather precision of picking. As for high S/N data, the length of the
data are divided into N areas in the rectangular coordinate time window shall be relatively big; generally, the length of
system according to equal angle, and the seismic traces in the time window is 200 samples.
each area are sorted according to offset from small to big. Figure 1 shows shot gather data, which are divided into
The time window length is given and then the primary wave four quadrants. Seismic traces in each quadrant are sorted
time window in each area is designed. Select the seed shots according to offset from small to big (Fig. 2). Seed shots in
at different positions in the work area and design primary different areas of geometry are selected to design primary
wave time window for the seismic trace area with the seed wave time windows (green curve in Fig. 3), and primary
shots with the same method mentioned above. According to wave are included in the range of the time window. Using
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1237
spatial interpolation technology, the range of primary wave the quantity of iterations. Let the initial quantity of iterations
time window for each shot and each trace is determined and a = 0.
finally applied to shot gather data (green curve in Fig. 6). 1.5 The samples in the data set {R̄(t)|t ∫[t1 , t2 ]} to be classified
1.2 Starting from the top of each primary wave time window, in the primary wave time window WS will be divided into one
the upper data with the length equal to the length of the of two clustering clusters according to the principle of mini-
primary wave time window are taken as noise data, marked mum variation of expected variance, that is, if the following
as WN, which is shown in the upper single trace seismic record formula is right:
between line 2 and line 1 in Fig. 4. The range between line 2
(a) (a)
and line 3 is the primary wave time window, marked as WS, E R̄ (t) L = min E R̄ (t) L (t ∈ [t1 , t2 ] , L = 1, 2) (1)
and the lengths of the two time windows are equal.
1.3 Perform a Hilbert transform on the original signal X(t)
in two time windows of WN and WS to obtain a virtual and R̄(t) ∈ L can be determined. Here, E( R̄(t))(a)
L is the ex-
seismic trace X̃(t), envelope of the original signal is E(t) = pectation after R̄(t) joins the corresponding cluster.
2 1.6 The expectation and variance of the two clusters after
X2 (t) + X̃ (t), R(t) = (X(t) + E(t))2 + (X(t) − E(t))2 is
reclassification are calculated. The first clustering centre is
obtained by the envelope. Use the sliding time window method
obtained by calculating the new data that come from the data
to calculate the mean value of each sample in two time win-
m of noise time window WN and the data selected from primary
dows: R̄(t) = 2m+1 1
k=−m R(t + k), m is the length of the
wave time window WS.
sliding window.
1.4 Divide the mean values of each sample in the two time (a+1)
windows of WN and WS into two clustering clusters 1 and (a+1) 1 T2
E R̄ (t) 1 = R̄(t) dt. (2)
2 , respectively. Calculate the expectation of the mean value T2(a+1) − T1(a+1) (a+1)
T1
R̄(t) of each point in the two time windows and take the
value of expectation as the two initial clustering centres: The second clustering centre is obtained by calculating
E( R̄(t))1(a) and E( R̄(t))2(a) , where the superscript a represents the remaining data of primary wave time window WS after
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1238 Xu et al.
some data of the primary wave time window WS are selected As for high S/N data, it is used as the discrimination that
to the first clustering centre: the variance (or expectation) of the mean value of samples
in cluster 1 is greater than(1 + ε0 ) times variance (or expec-
(a+1)
(a+1) 1 T4
tation) of the mean value of samples in the original noise
E R̄ (t) 2 = R̄(t) dt. (3)
T4(a+1) − T3(a+1) (a+1)
T3 window WN. At the primary wave boundary, the variance
and expectation of the mean value of samples in cluster 1
By using formulae (2) and (3), the variance of two clus- will suddenly change. As for the low S/N data, it is used as
tering centres can be obtained, respectively: the discrimination that the expectation of the mean value of
samples in cluster 2 is less than (1 + ε1 ) times expectation
(a+1) (a+1) 2
D R̄ (t) L = E R̄ (t) − E R̄ (t) L (L = 1, 2), (4) of the mean value of samples in cluster 1 , at the primary
wave boundary, the expectation of clustering 2 will gradu-
where T1(a+1) and T2(a+1) are the initial and end samples of the ally decrease, but the expectation of cluster 1 will gradually
data in cluster 1 after reclassification after a + 1 iterations, increase. As the two clustering centres are becoming stable, the
T3(a+1) and T4(a+1) are the initial and end samples of the data in variance of the two clusters does not change clearly. When the
cluster 2 after reclassification after a + 1 iterations. above conditions are met, the boundary position of the first
1.7 The variance (or expectation) of the mean value of sam- break is determined. Both ε0 and ε1 are more than zero. As for
ples in cluster 1 is greater than (1 + ε0 ) times variance ε0 , the higher the S/N, the bigger is the value of ε0 ; otherwise,
(or expectation) of the mean value of samples in original the smaller the value of ε0 is. As for ε1 , the lower the S/N is,
noise time window WN, or the expectation of the mean the smaller the value of ε1 is, and the more the quantity of
value of samples in cluster 2 is less than (1 + ε1 ) times iterations is.
expectation of the mean value of samples in cluster 1, The top part in Fig. 4 is a single trace seismic record. The
that is, if D[ R̄(t)]1(a+1) > (1 + ε0 )D[ R̄(t)](0) (a+1)
> area between lines 2 and 3 is the primary wave time window,
1 (or E( R̄(t))1
(0)
(1 + ε0 )E( R̄(t))1 ) or E( R̄(t))2(a+1) (a+1)
< (1 + ε1 )E( R̄(t))1 , iter- the area between lines 2 and 1 is the noise time window, and
ating ends; otherwise, go to step (1.5). the lengths of the two time windows are equal. The middle
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1239
part between lines 1 and 3 in Fig. 4 is seismic data; the bottom 2.4, and the S/N in the local scope of primary wave in Fig. 5
part between lines 1 and 3 in Fig. 4 is the mean value R̄(t) of is 1.2. The boundary point of the two clusters in Fig. 5 is at
samples after the transformation of seismic data. According the position of 1756 ms, and the peak point of primary wave
to the clustering analysis algorithm, the points between lines is shown in the middle of Fig. 5.
1 and 3 are divided into two categories, and the points in Figure 6 shows the initial picking result of the single
cluster 1 are noise (red), the points in cluster 2 are signal shot based on the clustering algorithm. The green line is
(yellow), the boundary point of the two clusters in the figure the primary wave time window. The average S/N of primary
is at the position of 1428 ms, which is the boundary of primary wave in the local scope of the shot is 2.8, but the average
wave. Search downward from this position within the range S/N of primary wave in the local scope of 635–667 traces
of a wavelet length, the position corresponding to maximum is 0.76, only a few primary waves are picked correctly, if
value is the peak point of the first break, as shown in the the abnormal first breaks are not removed, the near-surface
middle of Fig. 4. As for these seismic data, ε0 is 0.8 and ε1 is modelling and statics calculation accuracy will be seriously
0.3. The S/N in the local scope of primary wave in Fig. 4 is affected.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1240 Xu et al.
2. Comprehensive quality evaluation to first break based on and the mean value of the positive (negative) waveform area
multiple angles of the wavelet with peak (trough) in the upper closest to the
After finishing the initial picking by using clustering al- primary wave are calculated. The corresponding mean values
gorithms, due to the impact of serious ambient interference, are A0 and A1 . In general, if the ratio A1 /A0 is greater than
some first-break pickings are incorrect, as shown in 635–667 the given threshold (1 + δ0 ) (δ0 > 0), the first break is an ab-
traces in Fig. 6, if these abnormal first breaks are put together normal first break. The parameter δ0 is set according to the
with the normal first breaks for statistics analysis, the correct- S/N of data and source type. The range of δ0 is from 0 to 1. If
ness of geological information result may be affected, if the the data are low S/N data from vibrator, δ0 shall be less than
abnormal first breaks can be eliminated, the result can be more 0.2; otherwise, δ0 shall be more than 0.2. If the data are low
in line with the objective underground situation; however, if S/N data from dynamite, δ0 shall be less than 0.5; otherwise,
some suspicious first breaks within the tolerance of the error δ0 shall be more than 0.5.
are thrown away, some important underground information This method is mainly used to remove the seismic trace
may be ignored. Therefore, at present the important ques- with strong noise interference.
tion is how to establish the criteria to identify abnormal first
breaks, and then accurately determine whether the suspicious 2.2 Quality evaluation to first breaks of adjacent traces of
first break is abnormal and should be eliminated to obtain single spread based on the permutation and combination
reliable first breaks. Based on the above, this paper proposes method
to judge the reliability of the picked first break by multiple For any spread, generally, the arrival time of pri-
angles to determine abnormal first breaks according to the fol- mary wave in the common-source point record continuously
lowing procedure: ‘single trace →single spread→single shot changes with the offset, so the first break of the adjacent
→multiple shots’. The specific methods are as follows: traces will not greatly jump, which provides the basic crite-
ria to detect abnormal first breaks. However, for areas with
2.1 Quality evaluation to the first break of single trace based a complex surface, the thickness and velocity of the low-
on the waveform area method velocity layer/subweathering layer change largely, the first
Take the picked first break as a reference point. If breaks of adjacent traces jump to some extent and do not
the peak (trough) is picked, the mean value of first breaks linearly change with offset, so the conventional sliding linear
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1241
n
fitting method cannot be used to evaluate abnormal first where j is one trace number of n traces selected from the C2n
breaks. Based on the above, the permutation and combination combination.
method is used to identify the abnormal first break of single In the formula, ti is the first-break time of the trace to
spread. be discriminated, t j is the first-break time of jth trace of n
Select the reference shot and calculate the mean value ω̄ adjacent traces selected from the C2n n
combination, β is a given
of the first-break time differences of the adjacent traces of the constant, the value of β is between 0 and 2, the lower the S/N
reference shot. In formula (5), m is the total quantity of traces of data, the smaller is the value of β. For general primary wave
of reference shots. seismic data, the value of β by default is 1. L is the quantity
m
of traces of shot to be evaluated.
ti − ti−1
ω̄ = (5) For the first-break time of any trace, calculate according
i=2
m− 1
n
to the formula (6). If a combination of C2n is met, the
Take the above ω̄ as the standard for any trace of any first-break time of the trace is reliable. Repeat the above
shot to be picked up. Use the permutation and combination calculations for common-source point gather to determine
method to calculate the relationship of first-break time be- the credibility of the first-break time of each trace.
tween the trace and adjacent traces. Generally select an odd This method is mainly used to remove the abnormal first
trace, such as 2n + 1. The trace to be discriminated is re- break with big jump when comparing with the first break of
garded as the selected trace, and the first-break time of the adjacent traces in areas with a complex surface.
trace is permuted and combined with the first-break times
n 2.3 Quality evaluation of first break of one shot based on the
of the remaining traces. There are C2n possible combinations
between the trace and the adjacent traces, and calculate the normal distribution method
relationship between mean value ω̄ and each combination, as In the single spread, the method (2.2) is used to deter-
in equation (6): mine the abnormal first break, and the remaining abnormal
first breaks have a certain continuity with the first break
ti − t j < β ω̄, (1 ≤ i ≤ L) , (6)
of adjacent traces. For such abnormal primary wave, the
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1242 Xu et al.
first-break time of adjacent spreads of single shot can be used otherwise, the value of α shall be small; the value of α by
for discrimination. default is 5.
According to the automated first-break picking method This method is mainly used to remove the abnormal first
in 1.1 based on the clustering algorithm, divide shot gather break that has good continuity in a time range but has a bigger
data into N quadrants with equal angle according the relief time difference if compared with the first break in the same
of terrain. If the surface of terrain is complex, the divided position in the adjacent spread.
quadrants shall be more. The first-break time with the same
offset in one quadrant changes little and meets normal dis- 2.4 Quality evaluation to first break of adjacent shots based
tribution, and the first-break time with big jump is regarded on the vector method
as abnormal first-break time. The first breaks with the same For any spread of one shot, if the time difference between
offset are extracted from each receiver line of each quadrant, the abnormal first-break time in a time range of this spread
and the first-break time is t1 , t2 . . . ty , where y is the quantity of and the normal first-break time in the same position in adja-
receiver lines, and the average value t̄ and standard deviation cent spread is relatively small, the above-mentioned method
σ of the first breaks in the group are calculated. If a first break (2.3) will fail. In order to solve this problem, it is necessary to
ti of the group meets |t̄ − ti | > ασ , the first break is abnormal. use the primary wave of adjacent shots to discriminate.
In each area, the first break of each trace is classified and dis- The primary wave is formed by the mixing of two or more
criminated according to offset, and the abnormal first break kinds of different types of waves (such as shallow refracted
can be automatically eliminated. σ is the standard variance, α waves, deep refracted waves, direct arrival waves and turning
is a given constant and the value of α is between 1 and 10 and waves from different depths) in some areas, which leads to the
is set according to surface conditions. If the surface is complex confusion of first-break phases. As for the area with complex
and the terrain undulates greatly, the value of α shall be large; structures, the subsurface velocity changes greatly, and the
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1243
types of primary waves are also different, which makes it given threshold, then the first break of the trace is abnormal.
impossible to predict the first-break time of adjacent shots by By using this method, the discrimination is performed shot
using the single time–distance equation. by shot and trace by trace, the abnormal first breaks can be
According to the assumption of surface consistency, as effectively eliminated.
for the seismic records from high-efficiency acquisition, the This method can solve the problem existing in identify-
change of first breaks of shots with adjacent spatial positions ing the abnormal first break in the method (2.2.3), and this
has good similarity, the shot density and trace density are method can also effectively solve the problem existing in areas
relatively high, and the first-break time difference in the same with a complex surface that the conventional method is to use
receiving point of the adjacent shots in the local range is not the velocity difference between the fitting velocity of adjacent
much different. Based on this criterion, the first breaks of the reliable first break and the fitting velocity of abnormal first
adjacent shots are used to discriminate anomaly. As for any break to discriminate, so the accuracy to identify abnormal
shot, the spatial position vector relationship of shot points first break by using adjacent reliable first break is improved.
relative to each trace can be obtained according to coordi- According to the above-mentioned four methods, the ab-
nates of shot points and receiving points. The vector position normal first breaks can be effectively identified in initial pick-
relationship between shot points and receivers of mth trace ing by using clustering algorithms. After quality evaluation is
in a shot is marked as m(s, d, a, tx , ty ), where s is the shot performed to the picked first breaks in Fig. 6, the abnormal
point number, d is the offset, a is the azimuth, tx and ty are first breaks in 635–667 traces are eliminated. The result is
the coordinates of receivers. Figure 7 is the vector diagram of shown in Fig. 8.
the spatial position relationship between shot point and re-
ceiver. Centred at the shot point to be discriminated, R is the 3. Abnormal first-break correction based on ant colony opti-
radius. The first-break times in the same position of all shots mization
in the local range are superimposed to obtain the mean value. After the above-mentioned first-break quality evaluation,
If the absolute value of the difference between the first-break the abnormal first breaks can be effectively eliminated. After
time of a trace of a shot and the mean value is greater than the removing the abnormal first breaks, in low S/N data, a small
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1244 Xu et al.
number of first breaks with high credibility still exist in the is as follows: In the selected first-break time window, seis-
remaining data. These remaining first breaks can be used to mic data are classified according to the positive waveform
correct the eliminated first breaks to enhance the quantity of and negative waveform. Ants are randomly placed at differ-
traces for picking, especially those with near offset, which is ent peaks (or troughs), and each ant searches traces in seismic
used to meet the requirements of near-surface modelling and data one by one and finds a reasonable peak (or trough) point
statics calculation. Based on the above, an abnormal primary in the local range according to the steepness s and the time
wave correction method based on ant colony optimization is window w in the next adjacent trace to form a foraging tra-
introduced. jectory with the peak (or trough) point as mark. At the same
Ant colony optimization is a probabilistic algorithm used time, the pheromone of the corresponding trace is updated.
to find optimal paths in graphs. The basic idea of the algo- Then, iterate continuously based on the set maximum quan-
rithm comes from the shortest path principle of ants foraging tity of iterations; the ant itself will select the next peak (or
in nature, when searching for food sources, ants can release trough) point according to pheromone. Of course, the thicker
special pheromones on the way they pass, so other ants in a the pheromone of the peak (or trough) of a wavelet, the greater
certain range can detect them. Ants in the same ant colony can is the probability of selection, and the pheromone itself has
sense pheromones and their strength. Later ants tend to move a certain volatilization effect, so multiple high pheromone
toward higher pheromone concentrations, and the pheromone concentration paths are formed after the pheromone is accu-
left by the movement will strengthen the original pheromone, mulated and volatilized; finally, based on the first break with
so the subsequent ants will have a higher probability to select high credibility, connect abnormal first breaks according to
this path; thus, a positive feedback mechanism forms. Then, the information concentration path, the first-break peak (or
subsequent ants have a larger probability to choose a shorter trough) is determined in a local range.
path, and finally the optimal path can effectively be found. In order to improve the accuracy of correcting abnormal
Based on the basic idea of ant colony optimization, the first breaks, seismic data in the first-break time window shall
ant colony algorithm is improved to adapt to correct abnor- be classified according to the positive waveform and negative
mal first break to improve the quantity of traces and the pre- waveform using the ant colony algorithm. If classified accord-
cision for automated first-break picking. The specific method ing to the positive (negative) waveform, the negative (positive)
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1245
Figure 18 Distribution chart of the picking result according to offset by using the AIC method (left) and the new method (right).
waveform should be zeroized. Figure 9 shows the result of where I0 represents the mean value of the amplitude of the
the classification according to the positive waveform. As for wavelet peak (trough) of the first break tracked near the
the classified data, the evaluation function is constructed ac- position of ant, P0 represents the mean value of samples
cording to the properties (amplitude value, steepness, wavelet included in the positive (negative) waveform of the tracked
waveform etc.) of the primary waves of the adjacent traces. first-break wavelet, I(w, s) represents the amplitude of a
Under normal circumstances, the amplitude value of the peak wavelet peak (or trough) point in the range of a given
(or trough) of the adjacent primary wave generally meets the steepness s and time window w of a trace adjacent to the ant’s
rule of continuous change, and the primary wave has the char- path, P(w, s) represents the quantity of samples included in
acteristics of smoothness and continuity; however, in complex the positive (negative) waveform of the wavelet in the range
areas, the surface terrain undulates greatly, the lateral velocity of the corresponding steepness s and time window w, and s
in near surface changes drastically, the lithologies are variable, represents the steepness of the tracked first break near I0 .
and base rock outcrops, which inevitably causes that primary The pheromone is updated according to the evaluation
waves change suddenly and have worse continuity and low function (7). The optimal path is finally determined accord-
S/N. The position of the primary wave of adjacent trace ing to the change of pheromone trajectory in different paths.
should be selected according to the amplitude value, steep- Based on the primary wave with high credibility, connect ab-
ness, wavelet waveform and other information of the peak normal first breaks according to pheromone concentration
(or trough) of the wave. Then, the position of the first-break paths. Then the abnormal primary wave can be corrected to
dot of adjacent trace can be determined. So the evaluation the right position.
function can be constructed, as shown in equation (7): Ants starting from different positions can generate a pri-
mary wave layer line with high pheromone concentration in
E (I0 , I (w, s) , P0 , P (w, s))
seismic data according to formula (7). After crawling sev-
I P eral times, the pheromone concentration in the primary wave
= log2 1 + 0 0 , (7)
I0 I (w, s) − I0 + P0 P (w, s) − P0 layer line is obviously higher than that in ambient, which
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1246 Xu et al.
Table 1 Comparison of picking accuracy rates obtained by different methods in three working areas
Decreased percentage of
Manual picking AIC method New method inaccurate picking
provides the base for extracting primary wave later. Usually, spatial interpolation technology, return the designed primary
the quantity of iterations for ant colony tracking is set to 40– wave time window of each trace to shot gather data. Then
60 times, and the pheromone matrix and the pheromone chart use Hilbert transform to obtain the expectation and variance
are obtained according to the ratio between the pheromone of each sample in noise time windows and first-break time
of each point and the maximum pheromone concentration of window, divide the samples within first-break time window
the seismic data in the first-break time window. Figure 10 into two clusters according to variance; iterate continuously
is a pheromone plot generated by ant tracking after 40 it- to determine the position of peak (trough) of primary wave.
erations. In Fig. 10, colour represents pheromone concentra- Secondly, perform quality evaluation to first breaks, mainly
tion: red represents the highest pheromone concentration, and including use waveform area method in single trace to calcu-
blue represent the smallest pheromone concentration. We can late the ratio between the area of first-break wavelet and the
see that after ant tracking, the high concentration pheromone area of adjacent wavelet (upper), identify abnormal first break
curve coincides well with the first breaks in seismic data. The according to the given threshold; use combination method in
ant colony algorithm can be used to determine the optimal single spread to calculate the time difference between the first
path. Then, by combining with the reliable primary waves, breaks of adjacent traces of current shot, set the threshold pa-
the removed abnormal first breaks of the 635–667 traces are rameter according to the S/N of primary wave data, calculate
well connected by the pheromone curve. The picking result the permutation and combination relationship between the
is shown in Fig. 11, where the abnormal first breaks are cor- first-break time of current trace and that of the adjacent trace
rected to the right position. to identify abnormal primary wave trace by trace; use normal
In summary, the new automated first-break picking distribution method in the same shot to classify shot gather
method mainly includes four steps: design primary wave time data according to the divided quadrants. If the terrain changes
window, the first-break picking based on clustering algo- greatly, the divided quadrants shall increase. The first-break
rithm, the quality evaluation to first breaks and the abnormal time with the same offset in one quadrant changes little and
first-break correction technology, all of which are automati- meets normal distribution. Give the threshold according to
cally calculated. The specific operation procedure is shown in the S/N of primary wave data and terrain undulation and
Fig. 12. identify abnormal first breaks segment by segment. Use the
We describe the work process as follows. Firstly, divide vector method in adjacent shots, according to the assump-
the shot gather data to be picked into N quadrants; the seismic tion of surface consistency. If the seismic records are from
trace gather data in each quadrant are sorted according to off- high-density acquisition, the change of first breaks of shots
set from small to big. Give a time window length, design time with adjacent spatial positions has good similarity. Stack the
window in each quadrant, use the same method to select quad- first-break times in the same position of adjacent shots in
rants with seed shots and design time window, determine the local range to calculate their mean value. If the difference be-
time window range of primary waves of each quadrants using tween the first-break time of a trace of a shot and the mean
Table 2 Comparison of picking efficiency between the AIC method and the new method
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1247
value is more than the given threshold, this first break of noises not only contaminate effective reflection information
the trace is regarded as abnormal first break. Finally, use the but also blur the primary waves of single shot, which leads to
ant colony algorithm to construct evaluation function to up- the S/N of primary waves from vibroseis, which is generally
date pheromone, determine the optimal path according to the low. Therefore, the difficulty of automated first-break picking
change of pheromone trajectory in different paths, and based increases. The first-break picking by general methods is not
on the first break with high credibility, correct abnormal first ideal for low S/N data from vibroseis, which causes that it is
breaks in local range to improve the precision of correcting difficult to calculate statics.
abnormal first break. Figure 13 shows the seismic data from vibroseis in an
area. Each shot has 15,360 traces, and the receiving interval
is 25 m. The overall S/N of primary waves is relatively low, and
EXAMPLES
the average S/N in the local range is 1.272. The first breaks of
At present, seismic records from vibroseis are obtained by far offset and near offset are clear. However, the first breaks
cross-correlation between sweep signal and vibration signal. in the middle traces are complicated due to complex near-
Theoretically, the phase of seismic record of single shot from surface structure, and the cross-strata phenomenon is serious.
vibroseis should be zero, but due to earth filter function and The existing automated first-break picking methods cannot
seismic instrument response, the phase of seismic record of effectively identify. In practice, the AIC method has certain
single shot from vibroseis is hybrid phase. The noises in seis- anti-noise performance; therefore, we compare the commonly
mic records of single shot obtained by cross-correlation be- used AIC method with the new picking method. We can pick
tween sweep signal and vibration signal mainly include side first breaks using the AIC method to calculate the first-break
lobe noise, linear interference and random interference. These time of single trace and using the sliding linear fitting method
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1248 Xu et al.
Figure 20 Stack section whose first breaks are picked by the AIC method and processed by the tomographic statics method.
to identify the abnormal first break. Also, using the spline A comprehensive quality evaluation to the picked first
interpolation technique to correct abnormal first break, the breaks by using clustering algorithms is performed in multiple
final picking effect is shown in Fig. 14. It can be seen that the angles according to the sequence of single trace, single spread,
abnormal first break with middle offset cannot be effectively single shot, and multiple shots to identify the abnormal first
removed. The first breaks picked on the left side are not in breaks. The result is shown in Fig. 16. All the abnormal first
the same stratum, cross-strata phenomenon is serious and the breaks can be effectively eliminated.
error is large. In fact, as for the data from the areas with com- We can correct the eliminated first break by the ant
plex near-surface structure, the discrimination method such colony algorithm and by combining with the reliable first
as sliding linear fitting can only identify the remote abnormal break. Firstly, seismic data should be classified according to
first break, and the first-break time corrected by the spline in- the negative waveform, the positive waveform is zeroized,
terpolation technology has certain deviation from the actual and then the pheromone of the foraging trajectory is updated
position. according to the constructed evaluation function. Then deter-
The improved clustering method is used for single-trace mine the optimal path. Finally, by combining with the reliable
picking. At the same time, the first breaks and waveforms of first breaks, the abnormal first breaks are corrected to right
adjacent traces are used to correct the abnormal first break positions. It can be seen that by using the ant colony algorithm,
with large variation. The initial picking effect is shown in the abnormal first breaks in 14004–14034, 14136–14141,
Fig. 15. Compared with Fig. 14, the single-trace picking 14234–14243, 14299–14322, and 14333–14347 traces are
accuracy is improved, but the low S/N primary waves in corrected to right positions (Fig. 17).
14004–14034, 14136–14141, 14234–14243, 14299–14322 In order to compare the picking accuracy between the
and 14333–14347 traces cannot be picked effectively. new method and the conventional AIC method, the picking
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1249
Figure 21 Stack section whose first breaks are picked by the new method and processed by the tomographic statics method.
inaccuracy percentage C% is introduced for evaluation. We In order to verify the accuracy rate of an automated first-
assume that the picking accuracy rate of the conventional break picking method proposed in this paper, actual data from
AIC method is A% and the picking accuracy rate of the new three different work areas are selected for testing, according
method is B%, then the following formula can be obtained: to the respective features of the seismic data from three work
C% = ( 100−A
B−A
)/100. The more the picking inaccuracy per- areas. The automated picking parameters of the conventional
centage C% decreases, meaning that there are more accu- AIC method and the new method are properly adjusted. The
rately picked traces by the new method compared with AIC principle is that it is better to leave a deficiency uncovered than
method. to have it covered without. Discretion is adopted to ensure
The first breaks in the seismic data are not clear due the quality and quantity of the picked first break. At the same
to correlation noise, impulse mutation, and bad traces etc. time, manual picking is performed to obtain accurate reference
The picking results of the first breaks of the whole work first-break time. If the error between automated picking and
area are statistically analysed. The accuracy rate of the AIC reference first-break time is within a sampling interval, the
method is 67.5%, and the accuracy rate of the new method is picked first break is right. The picking results obtained by
79.4%. The quantity of inaccurately picked traces decreases three methods in three working areas are shown in Table 1.
by 36.6%. Figure 18 show the distribution chart of the pick- An analysis of three picking results and error statistics shows
ing result according to offset by using the AIC method and that if compared with the AIC method, the new method can
the new method. The density and quantity of the first breaks effectively reduce the inaccurately picked traces by more than
picked by the new method are obviously higher than those by 20%, especially the quantity of picked low S/N primary waves
the AIC method, and all the picking accuracy rates in near with near offset and the picking accuracy can be improved. If
offset, middle offset and far offset are improved. compared with the accuracy rate by manual picking, although
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1250 Xu et al.
the picking accuracy rate by the new method is slightly worse, two green boxes (CMP numbers 300–480 and 600–750) is
the gap continuously decreases as the S/N of primary wave enhanced, and compared with Figs 19 and 20, the continuity
increases. of seismic events in the area with intrusive igneous rock in
In order to improve the efficiency of automated pick- the red box becomes better, especially the continuity of seis-
ing, according to the feature that the new picking method mic events in shallow layers is clearly enhanced. The results
uses data, only the data in the first-break window are read, show that the new method has strong anti-noise ability and
and unnecessary data I/O are reduced. Different time window can effectively pick low S/N primary waves. It can also be seen
lengths for the same seismic data are designed to test the ef- from the processing results that in the area where the S/N of
ficiency of automated picking, and the results are shown in primary wave in seismic data is high, the quantity of traces for
Table 2. As the length of first-break time window increases, first-break picking has little impact on imaging quality, but in
the quantity of picked traces per minute for both methods the area where the S/N of primary wave is low, the quality and
gradually reduces. Compared with the AIC method, the new quantity of picked first breaks have a large impact on imaging
method is slightly inefficient, but it can meet the demands for quality.
real-time automated picking in data processing.
In order to further demonstrate the scientificality and ra-
CONCLUSIONS
tionality of the new method proposed in this paper, a detailed
analysis of the vibroseis data from Dataset3 work area is car- This paper proposes automated first-break picking technology
ried out. Overall, the S/N of primary wave is relatively low, for massive low S/N data. The technology includes contents
and the primary waves of the traces with near offset are drown such as design primary wave time window range, the algo-
by interference waves. Due to the energy attenuation and the rithm based on clustering, multi-angle comprehensive quality
complex near-surface structure, the first breaks in far offset evaluation, and abnormal primary wave correction based on
are rather messy, and the cross-strata phenomenon is serious, ant colony algorithm etc. and has the following features:
which makes it very difficult to automatically pick the first 1 Use multi-azimuth time window spatial interpolation tech-
breaks. In order to compare the first-break picking effects of nology to conveniently and quickly determine primary wave
two methods, firstly, the stack section is obtained by pro- time window range, which provides the base for accurately
cessing with the elevation statics method (Fig. 19). Overall, picking first-break time.
except for the seismic events in the red box that are discontin- 2 As for low S/N first breaks, although the noise change with
uous, the continuity of the events in other positions is better. time before the arrival of primary wave is disorderly and can-
The main reason is that the area in the red box is an intru- not be represented by a definite mathematical formula, the
sive igneous rock. Automatically pick first breaks by using the noise generally belongs to random process of ergodic prop-
existing methods and the new method, and then process the erty, based on this feature, a clustering algorithm is proposed,
picking results by the tomographic statics method. The results and on the base of using energy feature to describe raw seismic
are shown in Figs 20 and 21, respectively. Figure 20 shows data, this algorithm selects the variance (expectation) of av-
a stack section whose first breaks are automatically picked erage value of samples as the feature function of the data for
by the AIC method. Compared with the section from the el- analysing. This algorithm can effectively separate the signal
evation statics method, the continuity of the events from the and noise and improve the accuracy of picking first break in
shallow strata to the deep strata in the two green boxes (CMP single trace.
numbers 300–480 and 600–750) becomes worse or discon- 3 As for the discrimination to abnormal primary waves, the
tinuous. This occurs mainly because the lateral strata change principle of ‘single trace → single spread → single shot →
significantly. First breaks picked by the existing method show multiple shots’ is proposed to judge first breaks by the multi-
a cross-strata phenomenon. The quantity of traces for first- angle comprehensive method such as the energy of primary
break picking is small, and there exist a large number of ab- wave, waveform, relationship between adjacent traces, near-
normal first breaks. Figure 21 shows the stack section whose surface velocity and the spatial position relationship between
first breaks are automatically picked by the new method. The shot points and receiver, which can effectively identify abnor-
abnormal first breaks are eliminated by quality evaluation, mal first breaks.
and the quantity of traces for first-break picking is increased 4 Ant colony algorithm is used. Update the pheromone by the
by using the ant colony algorithm. The continuity of the seis- constructed evaluation function and determine the optimal
mic events from the shallow strata to the deep strata in the path according to the change of the pheromone trajectory on
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
A first-break picking method for seismic records 1251
different paths, and then by combining reliable primary wave, Coppens F. 1985. First arrivals picking on common-offset trace col-
ensure that the picked first breaks belong to the same hori- lections for automatic estimation of static corrections. Geophysical
Prospecting 33, 1212–1231.
zon, so the quality of the abnormal primary wave correction
Dong S., Sheng J. and Schuster G.T. 2006. Theory and practice of
is improved. This algorithm can recognize the abnormal first
refraction interferometry: 76th SEG Annual Meeting, Expanded
break with small cross-strata, which reduces human interven- Abstracts, 3021–3025.
tion in the first-break picking process and obviously improves Duan X.D., Zhang J., Liu Z.Y., Liu S., Chen Z.B. and Li W.P. 2018.
the efficiency of first-break picking of seismic data. Integrating seismic first-break picking methods with a machine
learning approach. SEG Technical Program Expanded Abstracts,
2186–2190.
ACKNOWLEDGEMENTS Earle P. and Shearer P. 1994. Characterization of global seismograms
using an automatic picking algorithm. Bulletin of the Seismological
The study is supported by the National Science and Society of America 84, 366–376.
Technology Major Project (2017ZX05018-003) of China, Fahlman S.E. and Lebiere C. 1990. The cascade—correlation learning
jointly funded by National Key S&T Special Projects architecture. In: Advances in Neural Information Processing Sys-
tems, Vol. 2 (ed D.S. Touretzky), pp. 524–532. Morgan Kaufmann,
(2016ZX05024001003), Open Fund for SINOPEC Key Lab-
San Mateo.
oratory of Geophysics (WTYJY-WX2017-01-01), Innova-
Ferrari A., Lombardi S. and Signoroni A. 2017. Bacterial colony
tion Team of Youth Science and Technology in Southwest counting with convolutional neural networks in digital microbi-
Petroleum University (2017CXTD08). ology imaging. Pattern Recognition 61, 629–640.
Gelchinsky B. and Shtirelman V. 1983. Automatic picking of first
arrival and parameterization of travel-time curves. Geophysical
ORCID Prospecting 31, 915–928.
Hanafy S.M., Al-Hagan O. and Al-Tawash F. 2011. Super-virtual
Yinpo Xu https://orcid.org/0000-0003-2633-9427 refraction interferometry: field data example over a colluvial wedge.
81th SEG Annual Meeting, Expanded Abstracts, 3814–3818.
Hatherly P. 1982. A computer method for determining seismic first
REFERENCES arrival times. Geophysics 47, 1431–1436.
Al-Hagan O., Hanafy S.M. and Schuster G.T. 2014. Iterative super- Jiao L. and Moon W.M. 2000. Detection of seismic refraction signals
virtual refraction interferometry. Geophysics 79, Q21–Q30. using a variance fractal dimension technique. Geophysics 65, 286–
Allen R. 1978. Automatic earthquake recognition and timing from 292.
single traces. Bulletin of the Seismological Society of America 68, Khan K.A. 2007. Artificial intelligence based automated first break
1521–1532. picking and quality control. SEG Annual Meeting, 1103–1107.
An S.P., Hu T.Y., Liu Y.M., Peng G. and Liang X. 2017a. Automatic Lee M., Byun J., Kim D., Choi J. and Kim M. 2017. Improved modi-
first-arrival picking based on extended super-virtual interferometry fied energy ratio method using a multi-window approach for accu-
with quality control procedure. Exploration Geophysics 48, 124- rate arrival picking. Journal of Applied Geophysics 139, 117–130.
130. Lv X.M., An S.P., Hu T.Y. and Cui Y. 2018. Similarity weighted
Baer M. and Kradolfer U. 1987. An automatic phase picker for lo- super-virtual interferometry to enhance first breaks. Acta Scien-
cal and teleseismic events. Bulletin of the Seismological Society of tiarum Naturalium Universitatis Pekinensis (in Chinese) 54, 87–93.
America 77, 1437–1445. Mahmud M., Kaiser M.S., Hussain A. and Vassanelli S. 2018. Ap-
Baluja S. and Fahlman S.E. 1994. Reducing network depth in the plications of deep learning and reinforcement learning to bio-
cascade-correlation learning architecture. School of Computer Sci- logical data. IEEE Transactions on Neural Networks 29, 2063–
ence, Carnegie Mellon University, 1–11. 2079.
Bharadwaj P., Schuster G., Mallinson I. and Dai W. 2012. Theory of Mallinson I., Bharadwaj P., Schuster G. and Jakubowicz H. 2011.
supervirtual refraction interferometry. Geophysical Journal Inter- Enhanced refractor imaging by supervirtual interferometry. The
national 188, 263–273. Leading Edge 30, 546–550.
Boschetti F., Dentith M. and List R. 1996. Afractal-based algorithm Marsden D. 1993a. Static correction a review. Part I. The Leading
for detecting first arrivals on seismic traces. Geophysics 61, 1095– Edge 12, 43–49.
1102. Marsden D. 1993b. Static correction a review. Part II. The Leading
Chen Y.K., Zhang G.Y., Bai M., Zu S.H., Guan Z. and Zhang M. Edge 12, 115–120.
2019. Automatic waveform classification and arrival picking based Marsden D. 1993c. Static correction a review. Part III. The Leading
on convolutional neural network. Earth and Space Science 6, 1244– Edge 12, 210–216.
1261. Mercier D., Gaillard P., Aupetit M., Maillard C., Quach R. and
Chen Y.K. 2018. Automatic microseismic event picking via unsu- Muller J.-D. 2006. How to help seismic analysts to verify the French
pervised machine learning. Geophysical Journal International 212, seismic bulletin? Engineering Applications of Artificial Intelligence
88–102. 19, 797–806.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252
1252 Xu et al.
Muller S., Garda P., Muller J.-D., Crusem R. and Cansi Y. 1998. Tosi P., Barba S., Rubeis V.D. and Luccio F.D. 1999. Seismic signal
A neuro-fuzzy coding for processing incomplete data: application detection by fractal dimension analysis. Bulletin of the Seismologi-
to the classification of seismic events. Neural Processing Letters 8, cal Society of America 89, 970–977.
83–91. Tsai K.C., Hu W.Y., Wu X.Q., Chen J.F. and Han Z. 2018. First-
Murat M. and Rudman A. 1992. Automated first arrival picking: a break automatic picking with deep semi-supervised learning neu-
neural network approach. Geophysical Prospecting 40, 587–604. ral network. SEG Technical Program Expanded Abstracts, 2181–
Peraldi R. and Clement A. 1972. Digital processing of refraction data: 2185.
study of first arrivals. Geophysical Prospecting 20, 529–548. Wang Y.D. and Wang J. 2009. Random signal analysis foundation.
Sabbione J. and Velis D. 2010. Automatic first-breaks picking: new Beijing, Publishing House of Electronics Industry.
strategies and algorithms. Geophysics 75, 67–74. Wong J., Han L., Stewart R.R., Bentley L.R. and Bancroft J.C. 2009.
Sabbione J. and Velis D. 2013. A robust method for microseismic Geophysical well logs from a shallow test well and automatic time-
event detection based on automatic phase pickers. Journal of Ap- picking on full-waveform sonic logs. CSEG Recorder 34, 20–29.
plied Geophysics 99, 42–50. Zhang N. and Ding S.F. 2017. Unsupervised and semi-supervised
Shimshoni Y. and Intrator N. 1998. Classification of seismic signals extreme learning machine with wavelet kernel for high dimensional
by integrating ensembles of neural networks. IEEE Transactions data. Memetic Computing 9, 129–139.
on Signal Processing 46, 1194–1201. Zhang N., Ding S.F., Zhang J. and Xue Y. 2018. An overview on
Stevenson R. 1976. Microearthquakes at Flathead Lake, Montana: restricted boltzmann machines. Neurocomputing 275, 1186–1199.
A study using automatic earthquake processing. Bulletin of the Zhang N., Ding S.F., Zhang J. and Xue Y. 2017. Research on point-
Seismological Society of America 66, 61–79. wise gated deep networks. Applied Soft Computing 52, 1210–1221.
Sze V., Chen Y.H., Yang T.J. and Emer J. 2017. Efficient processing Zhu D., Li Y. and Zhang C. 2016. Automatic time picking for micro-
of deep neural networks: a tutorial and survey. Proceedings of the seismic data based on a fuzzy C-means clustering algorithm. IEEE
IEEE 105, 2295–2329. Geoscience and Remote Sensing 13, 1900–1904.
C 2019 European Association of Geoscientists & Engineers, Geophysical Prospecting, 68, 1228–1252