FocMech-Flow: Automatic Determination of P-Wave First-Motion Polarity and Focal Mechanism Inversion and Application to the 2021 Yangbi Earthquake Sequence

Abstract

Featured Application

P-wave first-motion detection; focal mechanism inversion.

Abstract

P-wave first-motion polarity is important for the inversion of earthquake focal mechanism solutions. The focal mechanism solution can further contribute to our understanding of the source rupture process, the fault structure, and the regional stress field characteristics. By using the abundant focal mechanism solutions of small and moderate earthquakes, we can deepen our understanding of fault geometry and the seismogenic environment. In this paper, we propose an automatic workflow, FocMech-Flow (Focal Mechanism-Flow), for identifying P-wave first-motion polarity and focal mechanism inversion with deep learning and applied it to the 2021 Yangbi earthquake sequence. We use a deep learning model named DiTingMotion to detect the P-wave first-motion polarity of 2389 waveforms, resulting in 98.49% accuracy of polarity discrimination compared with human experts. The focal mechanisms of 112 earthquakes are obtained by using the CHNYTX program, which is 3.7 times more than that of the waveform inversion method, and the results are highly consistent. The analysis shows that the focal mechanisms of the foreshock sequence of the Yangbi earthquake are highly consistent and are all of the strike-slip type; the focal mechanisms of the aftershock sequence are complex, mainly the strike-slip type, but there are also reverse and normal fault types. This study shows that the deep learning method has high reliability in determining the P-wave first-motion polarity, and FocMech-Flow can obtain a large number of focal mechanism solutions from small and moderate earthquakes, having promising application in fine-scale stress inversion.

1. Introduction

The focal mechanism solution reflects the seismogenic fault structure and regional stress field information, which is essential for understanding the regional geological structure and seismicity analysis [1,2,3,4,5]. Compared with moderate and strong earthquakes, small earthquakes (M < 3.0) are more numerous, occur more frequently, and have a wider distribution. The focal mechanisms of small and moderate earthquakes contain a wealth of fault structure information, which is of great significance for the simulation of the earthquake rupture process, the analysis of fault zone characteristics and fault properties, and the inversion of stress fields [4,6,7,8].

The inversion methods of focal mechanism solutions can be divided into two categories: waveform fitting and P-wave first-motion polarity. For moderate and strong earthquakes (M > 3.5), the waveform inversion method is commonly used to determine the focal mechanism [9,10,11,12]. This method uses information carried in the waveform and can give a good focal mechanism solution even when there are few stations. However, most of the earthquakes recorded by the regional seismic network are small earthquakes below M3.0, with short propagation distances and high-frequency components, making their waveforms difficult to fit and their focal mechanism solutions difficult to obtain. On the other hand, because the rupture of small earthquakes is relatively simple, the results based on the P-wave first-motion polarity are generally consistent with the results based on waveform fitting. Therefore, for small earthquakes (M < 3.0), the first-motion polarity of P-wave inversion is often used to retrieve focal mechanism solutions [1,6,13,14,15,16]. Manually identifying the first-motion polarity of a P-wave is stable, reliable, and easy to operate, and it was the earliest method used to study the source mechanism of earthquakes [17]. With the continuous advancement of technology, many focal mechanism inversion methods have been developed, such as [18,19], combining P-wave first-motion polarity and the S/P wave amplitude ratio to determine focal mechanisms, which was called the HASH method. Moreover, [20] used a high-frequency full waveform combined with P-wave first-motion and the S/P wave amplitude ratio to invert the focal mechanism of small earthquakes in a sparse network, and the GPAT [21] method combines P-wave first-motion polarity and generalized maximum amplitude to invert the focal mechanism of earthquakes.

The identification of the P-wave first-motion polarity was performed manually in the past. In recent years, with the construction of the National Intensity Rapid Report and Earthquake Early Warning Project, the number of seismic stations in China has increased from ~1200 to more than 15,000. Manual processing can no longer meet the needs for the development of earthquake monitoring. In order to quickly identify the polarity of a P-wave, seismologists have developed a variety of methods. For example, [22,23] proposed an autoregressive model to automatically determine P-wave first-motion polarity, [24] used the Bayesian probability density function to identify the polarity of a P-wave, and [25] focused on mutual information and maximum order statistics in information theory to determine the P-wave first-motion polarity. However, when the signal-to-noise ratio of the waveform is low, these automatic algorithms based on rules are unable to achieve reliable first-motion polarity discrimination, and manual review is required. In addition, these algorithms only use limited discriminant parameters, and it is difficult to reflect the complete waveform characteristics, which restricts the accuracy of the algorithm.

With the rise of deep learning, data-driven algorithms have made breakthroughs in tasks such as object detection and classification, and have been widely used not only in image recognition [26,27] and speech recognition [28,29], but also developed rapidly in the field of seismology, such as earthquake detection [30,31], seismic phase picking [32,33,34,35], earthquake classification [36,37], and focal mechanism inversion [38,39]. The P-wave first-motion polarity is usually divided into three categories: upward, downward, and uncertain. Therefore, the discrimination of incipient polarity can be regarded as a classification problem, and the automatic identification of first-motion polarity can be realized by using a large number of known labels for supervised learning. Specifically, Refs. [40,41,42,43] used data from Southern California, Japan, and Taiwan, China, and used a convolutional neural network to try to realize the automatic determination of the first-motion polarity of P-waves, and they achieved good results.

In this paper, based on a deep-learning P-wave first-motion picking model named DiTingMotion [44,45], we detected the P-wave first motions of the 2021 Yangbi Ms6.4 earthquake sequence and compared them with the manual picking results and the rule-based automatic algorithm. The P-wave first-motion polarity discriminative effect of the model was evaluated using 2389 P-wave first-motion polarities detected by DiTingMotion; the focal mechanisms of 112 earthquakes were retrieved by the CHNYTX program [16]. The FocMech-Flow process for the automatic inversion of focal mechanisms using P-wave first-motion was established. Compared to the results of other methods, the feasibility of automatic inversion of the focal mechanism solution was analyzed.

2. FocMech-Flow Workflow

The workflow of FocMech-Flow for focal mechanism inversion is shown in Figure 1. The main steps consist of four parts:

• Seismic Event Screening. Seismic events satisfying focal mechanism inversion are selected from the observation report. Seismic events recorded by at least 10 stations and with a maximum gap angle less than 70° are selected.

• Data Preprocessing. According to the P-wave arrival time in the observation report picked by analysts, we take the P-wave arrival time as the center, cut a 4-s waveform window, and carry out de-mean, de-trend, and amplitude normalization processing.

• Determination of the P-Wave First-Motion Polarity. The preprocessed vertical component waveform is centered on the P-wave arrival time; we cut a 1.28-s waveform window, detected by DiTingMotion, and the polarity of the P-wave is given.

• Focal Mechanism Inversion. According to the earthquake location, station location, and regional velocity model, we calculate the takeoff angle, combined with the first-motion polarity, and invert the focal mechanism with the CHNYTX program.

The first-motion polarity of the P-wave is detected automatically by DiTingMotion. DiTingMotion was trained by [45,46] using the DiTing dataset [44] and the first-motion dataset from the Southern California Seismic Network [40]. DiTingMotion is a deep learning model based on a convolution neural network. The model consists of five convolution modules. The first two convolution modules contain two one-dimensional convolution layers and one maximum pooling layer, respectively. The last three convolution modules contain three one-dimensional convolution layers and one maximum pooling layer, respectively. In the last three convolution modules, each module outputs a set of polarity predictions and then determines the P-wave polarity as “Up” or “Down” by comparing the probability threshold. See Ref. [45] for details about the network structure, training process, and test results of DiTingMotion.

The CHNYTX program [16] is used for focal mechanism inversion, which is based on the grid research method [6,14]. By improving the calculation of the weight factor of the data points and the dynamic clustering of solution areas, the reliability of the focal mechanism is evaluated according to the number of clustering centers, the lowest weighted contradiction ratio, and the minimum root mean square (RMS) of the rotation angle.

3. Yangbi Earthquake Sequence

At 21:48 on 21 May 2021 (Beijing time), a Ms 6.4 earthquake occurred in Yangbi, Yunnan (Figure 2). The epicenter was located at 25.63° N, 99.92° E, with a focal depth of 8 km. Around 27 min before the mainshock, an M_L 5.4 foreshock occurred. The foreshock activity of this earthquake was very obvious, and more than 4600 aftershocks occurred two weeks after the mainshock [47,48,49,50,51]. There were 88 seismic stations around the mainshock, with good azimuthal coverage. The phase arrivals and P-wave first-motion polarity were picked manually by experienced analysts, which provided good observation data to study the focal mechanisms of this sequence and compare different methods for the automatic determination of the P-wave first-motion polarity [52,53,54,55,56,57].

The observation report and earthquake waveforms were provided by the Yunnan Earthquake Agency. We selected events with at least ten first-motion polarities [42]. There were 112 events with 2389 P-wave first-motion polarities that met the requirements. The magnitude range of 112 events was 1.6 ≤ M_L ≤ 5.7, of which 50 earthquakes were below M_L3.0.

4. The Detection Results of P-Wave First-Motion Polarity

The experts from the Yunnan Earthquake Agency processed 2389 P-wave first-motion polarities of 112 earthquakes, of which 1360 polarities were upward and 1029 polarities were downward. To ensure the reliability of the first-motion polarities, we conducted a review and corrected 13 polarities with obvious errors. The detection results of DiTingMotion showed that 2353 polarities were consistent and 36 polarities were inconsistent, with an accuracy rate of 98.49% (Table 1). Among them, there were 1376 polarities detected as upward, with an accuracy rate of 99.19%; there were 1013 polarities recognized as downward, with an accuracy rate of 97.38%. After visual inspection, the 36 waveforms that were identified incorrectly were mainly due to the low signal-to-noise ratio and difficulty in distinguishing first-motion polarity, which led to the inconsistency between DiTingMotion and manual picks (Supplementary Figure S1).

In order to compare the advantages and disadvantages of different methods for identifying the first-motion polarity of the P-wave, we selected the PhasePApy [23] and the Network for Polarity Classification (NPC) [59] model, and compared the detection results with DiTingMotion (Table 1). PhasePApy uses a rule-based automatic algorithm to distinguish P-wave first-motion polarities and divides them into three categories: upward, downward, and uncertain. The NPC model is developed on the basis of a convolutional neural network and attention mechanism, and outputs upward and downward results. The NPC model used in this study was trained with data from Southern California and Oklahoma, Japan, Italy, and Taiwan, Chinese Mainland.

It can be seen from Table 1 that the two models trained by deep learning have accuracy higher than 90% in the classification of the P-wave polarity. DiTingMotion used 641,025 P-wave first-motion polarity data from the Chinese Mainland in the past 10 years for training. It had the highest accuracy and was more suitable for detecting earthquakes in China. The NPC model used only part of the DiTing dataset, and the detection accuracy was slightly lower than that of DiTingMotion. However, PhasePApy discriminated the polarity according to the threshold given manually, and its ability to distinguish the characteristics of different waveforms was weak, which led to misdetection in nearly half of all cases. It can be seen that the deep learning models have obvious advantages over traditional rule-based methods in the detection of the P-wave first-motion polarity.

5. Robustness of P-Wave First-Motion Polarity

In Ref. [15], the authors calibrated the polarity error rate for the handmade first-motion polarities with similar events and found that ∼10% of the impulsive polarities and ∼20% of the emergent polarities were wrong. In order to test the robustness of the DiTingMotion, we flipped the amplitudes of 2389 waveforms and detected their polarities with DiTingMotion again. The test showed that for 2389 reversed waveforms, 56 polarities were misidentified. The detection accuracy was 97.66%, which was similar to that before the waveform flip. This test indicated that DiTingMotion has good robustness in determining the first-motion polarity.

The precision of the P-wave arrival time will affect the discrimination of first-motion polarity, especially when there is strong noise before the P-wave, which may cause the opposite sign of the first motion. To investigate whether the detection results were affected by the P-wave arrival time, we added a random shift to the P-wave arrival times of 2389 waveforms to simulate the picking error. The predominant P-wave period of the local earthquakes recorded by the regional network was generally 0.1–0.2 s. We applied a random error within ±0.1 s to the arrival time of the P-wave, and cut the waveforms with the shifted arrival times. Then, we detected the polarities with DiTingMotion again. Tests showed that when the arrival time of the P-wave was incorrectly marked, the detection accuracy of P-wave first-motion polarities decreased by 5.61–22.1% (Table 2).

To further analyze the reason for the drop in accuracy, we calculated the signal-to-noise ratio (SNR) of the waveform. The formula for calculating SNR is as follows:

(1)$\mathrm{SNR}=10\times {\log }_{10}{\left(\frac{\mathrm{signal}}{\mathrm{noise}}\right)}^2$

where signal is the maximum absolute value of the amplitude in the signal window, noise is the root mean square of the waveform amplitude in the noise window, and the unit of SNR is dB. The signal window is taken as 0.5 s after the arrival time of the P-wave, and the noise window is taken as 0.5–1.5 s before the arrival time of the P-wave.

We take the SNR of 30 dB as the boundary, divide the data into two categories—high SNR and low SNR—and calculate the detection accuracy at different SNRs (Table 2). When the SNR is high, the arrival deviation has less influence on the accuracy of polarity detection; however, when the picked arrival time is earlier, the false detection rate is higher. It can be seen that when using DiTingMotion to detect the P-wave first-motion polarities, the picking accuracy of the P-wave arrival time should be as accurate as possible. For the manually picked arrival times, one can use deep learning-based methods, such as Refs. [32,34], etc., to detect the event waveform again, and select the waveform that is manually and automatically marked at the same time (the residual less than 0.1 s) to detect its first-motion polarities. On the one hand, this can ensure the reliability of the first-motion polarities. On the other hand, it can also improve the waveform utilization. In some cases, only the arrival time is picked manually, but the initial movement symbol is not given.

6. Focal Mechanism Solutions of the Yangbi Earthquake Sequence

Using 2389 P-wave first-motion polarities of the above 112 earthquakes, the focal mechanism solutions of 112 earthquakes were inverted by the CHNYTX program [16]. Referring to the evaluation metrics in [16] (Table 3), the inversion results were divided into three classes, A, B, and C. There were 43 in Class A, 9 in Class B, and 60 in Class C. The detailed parameters of the focal mechanism solutions are shown in Supplementary Table S1.

The focal mechanism solutions of approximately 30 earthquakes with M_L ≥ 3.5 have been obtained by using the waveform inversion method [52,53,54,55,57]. In this study, we determined 112 focal mechanism solutions using the P-wave first-motion polarity, which is 3.7 times that of the waveform inversion method (Figure 3). Among the 112 earthquakes, there are 40 earthquakes with M_L ≥ 3.5 and 50 earthquakes with M_L < 3.0, and the minimum magnitude is M_L1.6. This shows that when the stations are well distributed, the focal mechanism of a large number of small earthquakes can be retrieved using the P-wave first-motion data.

To verify the reliability of the new inversion results, we selected 10 earthquakes with focal mechanism solutions from published articles that have been commonly studied (Figure 4). We measured the consistency of the focal mechanism solutions [60] using the Kagan angle (KA). We calculated KA with the FMcenter program [61]. When KA is less than 25°, it can be considered that the focal mechanism solutions are consistent [62]. Figure 4 is the comparison result. The standard deviation of the KA obtained in this paper is less than 24°, which is in good agreement with previous results. It shows that the focal mechanism solution obtained by FocMech-Flow inversion is reliable.

The leftmost column is the earthquake information; the rightmost column is the standard deviation of the Kagan angle.

Previous studies have shown that the Yangbi earthquake sequence occurred on a secondary fault parallel to the Weixi-Qiaohou fault [52,53,55,56]. This right-lateral strike-slip fault can be regarded as the northward extension of the Red River fault zone [63]. We used the FMC program [64] to classify the focal mechanisms (Figure 5). The results showed that there were 31 strike-slip events (SS), 35 strike-slip types with normal fault components (SS-N), 17 normal fault types with strike-slip components (N-SS), 15 strike-slip types with reverse components (SS-R), 9 normal fault types (N), 2 reverse fault types (R), and 3 reverse fault types with strike-slip components (R-SS).

We divided the magnitude–time (M-T) diagram of the Yangbi earthquake sequence from 18 to 28 May into two phases: foreshock and aftershock sequences. Figure 6 shows the focal mechanism solutions of this study for some large earthquakes (M_L ≥ 3.5). We found that the focal mechanisms of the foreshocks were consistent, and they were all strike-slip-type events. The earthquakes were distributed along the NW-SE direction, and the largest M_L5.4 foreshock occurred at the southeast end of the foreshock sequence. The focal mechanisms of the aftershocks were complex. The focal mechanism of the mainshock was the strike-slip type, and the focal mechanism of the largest aftershock M_L5.1 was the strike-slip type with a normal fault component. Earthquakes at the periphery of the source region had prominent normal fault components, and very few of them had normal fault events. The focal mechanisms with the strike-slip and normal-fault components accounted for more than 92% of the total inversion results in this study, which was consistent with the regional tectonic background (Figure 5 and Figure 6). In particular, the focal mechanism solutions for the foreshock sequence were not only highly consistent, but their nodal planes had an NW-SE orientation (Figure 6b), which is consistent with the seismogenic tectonics of this earthquake, as suggested by previous studies. Although the focal mechanisms of the aftershocks were relatively complex, the orientation of the nodal planes of the source mechanism solution for most aftershocks was also consistent with the seismogenic tectonics. This indicates that the focal mechanisms obtained in this study are highly credible.

It was found that there were differences in the consistency of the focal mechanisms between foreshocks and aftershocks [65,66,67], which was proposed to describe the seismicity with parameter K (K$\in$[−1, 1]) using the consistency of the focal mechanisms. High K means high consistency of the focal mechanism. In addition, KA can also reflect the consistency of the focal mechanisms. Except for the largest foreshock (M_L5.4) whose KA was more than 30°, the KA values of other foreshocks were all less than 25°, indicating that the focal mechanisms were consistent (Figure 7c). On the contrary, the KA of the aftershocks has a wide distribution range and low consistency. The seismicity parameter K and the KA of the foreshock and aftershock changed before and after the mainshock. The focal mechanisms of the foreshock sequence were consistent, all of which were of the strike-slip type with high K values, which reflects that the stress in the source area accumulated continuously before the mainshock, resulting in the stress field tending to be uniform. This is consistent with the previous research that considered that the foreshocks were controlled by a unified stress field, and the focal mechanism solutions had high similarity [8,65,68,69]. After the mainshock, the number of aftershocks with different types of focal mechanisms increased, the K value decreased, and the KA increased, all of which reflected the adjustment process of the stress field in the source area [8,70,71,72].

7. Conclusions

Taking the Yangbi earthquake sequence in 2021 as an example, this paper discusses the feasibility of the automatic discrimination of the polarity of P-wave first motion and its application in focal mechanism inversion and establishes a workflow named FocMech-Flow for automatic focal mechanism inversion by using P-wave first motion. The main conclusions are as follows.

(1) Compared with manual and rule-based methods, deep learning-based methods have high accuracy in determining P-wave first-motion polarity. Even if there is a 0.1-s error in the P-wave arrival time, the recognition accuracy of the P-wave first-motion polarity is still over 80%.

(2) Focal mechanism solutions of 112 earthquakes in the Yangbi earthquake sequence were determined with FocMech-Flow, which were in good agreement with previous results. FocMech-Flow can reduce the lower limit of focal mechanism inversion to approximately M_L2.0, and can obtain a large number of focal mechanism solutions for small and moderate earthquakes (M_L ≤ 3.0), which provides extensive data to depict the fine fault geometry and stress field inversion.

(3) The Yangbi earthquake sequence is mainly strike-slip. A few earthquakes have normal fault components. The focal mechanisms with strike-slip and normal fault components account for more than 92% of events. The focal mechanisms of the Yangbi foreshock sequence are highly similar to that of the mainshock. The consistency of focal mechanism solutions decreases for the aftershock sequence, which reflects the stress field adjustment after the earthquake.

Author Contributions

Conceptualization, S.L. (Shuai Li), L.F. (Lihua Fang) and Y.Z.; Data curation, S.L. (Shuai Li), S.L. (Shirong Liao) and L.F. (Liping Fan); Formal analysis, S.L. (Shuai Li); Funding acquisition, L.F. (Lihua Fang); Investigation, S.L. (Shuai Li); Methodology, L.F. (Lihua Fang) and Z.X.; Project administration, L.F. (Lihua Fang); Resources, L.F. (Lihua Fang), Z.X., Y.Z., S.L. (Shirong Liao) and L.F. (Liping Fan); Software, S.L. (Shuai Li), Z.X. and L.F. (Liping Fan); Supervision, S.L. (Shuai Li), L.F. (Lihua Fang), Z.X., Y.Z. and S.L. (Shirong Liao); Validation, S.L. (Shuai Li); Writing—original draft, S.L. (Shuai Li); Writing—review and editing, L.F. (Lihua Fang), Z.X., Y.Z. and S.L. (Shirong Liao). All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data and code are available upon request.

Acknowledgments

We would like to thank the Yunnan Earthquake Agency for providing high-quality observation reports and waveform data, Qiushi Zhai for providing the NPC model, and Shuzhong Sheng, Xiangyun Guo, ZhaoMing, and Zhe Zhang for the suggestions in preparing the manuscript. This research was supported by the National Key Research and Development Program of China (2021YFC3000702) and the Special Fund of the Institute of Geophysics, China Earthquake Administration (DQJB22X08). The figures were plotted with GMT6 (https://www.generic-mapping-tools.org, accessed on 15 January 2022) and Matplotlib (https://matplotlib.org, accessed on 15 January 2022).

Conflicts of Interest

The authors declare no conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. FocMech-Flow flow chart for focal mechanism inversion. A, B and C represent three levels of reliability of the focal mechanism solution. A is the best, B is the second, and C is the worst. P and T are the pressure axis and tension axis, respectively.

Figure 2. Seismic stations, faults, and epicenters of the Yangbi earthquake sequence. (a) The tectonic setting of the Yangbi earthquake. Black lines indicate faults [58], blue triangles indicate seismic stations, and red dots represent the epicenters of the sequence. The blue line in the inset indicates the boundary of the active block, and the yellow five-pointed star indicates the mainshock. (b) The distribution of the Yangbi earthquake sequence (red dots), the epicenters of 112 earthquakes used for focal mechanism inversion (blue dots), and the focal mechanism solutions of large earthquakes.

Figure 3. Statistics of the focal mechanism solutions for 1 ≤ ML ≤ 6.4. (a–f) denotes the magnitude distribution of the focal mechanisms obtained in this study and Refs. [52,53,54,55,57].

Figure 4. Comparison of focal mechanism solutions between different studies.

Figure 5. Types of focal mechanism solutions of the Yangbi earthquake sequence.

Figure 6. M-T diagram, earthquake distribution, and focal mechanism solutions for foreshocks and aftershocks. (a) M-T diagram of the Yangbi earthquake sequence. The yellow star indicates the mainshock. (b) Focal mechanism solutions of the foreshock sequence. (c) Focal mechanisms of the mainshock and aftershocks. The color of the beach ball represents different types of focal mechanisms (see Figure 5).

Figure 7. Temporal variation in KA angle and K value in the Yangbi earthquake sequence. (a) KA angle of the foreshock sequence. (b) KA angle of the aftershock sequence. (c) K value for the consistency of the foreshock and aftershock (light green indicates KA ≤ 25°), and the red dotted line separates the foreshock and aftershock. The red dotted lines in (a,b) indicate 25° KA angle.

Table 1

Comparison of polarity classification results obtained with different methods.

Method	Up	Down	Uncertain	TP	FP	TP/ALL
DiTingMotion (This Study)	1376	1013	0	2353	36	98.49%
PhasePApy [23]	768	509	1112	1275	2	53.37%
NPC [59]	1366	1023	0	2261	128	94.64%

Note: Uncertain indicates the number of polarities that cannot be determined; TP and FP represent the numbers of consistent and inconsistent polarities between automatic methods and manual discrimination; ALL represents 2389 P-wave first-motion polarities used for test; TP/ALL represents the accuracy.

Table 2

Detection accuracy of P-wave first-motion polarity before and after shifting the arrival time.

Random Error	0	±0.1 s	+0.1 s	−0.1 s
Accuracy (all)	98.49%	84.26%	76.39%	92.88%
Accuracy (SNR ≥ 30 dB)	99.60%	87.69%	82.85%	95.58%
Accuracy (SNR < 30 dB)	97.57%	81.12%	71.96%	92.22%

Table 3

Focal mechanism solution quality evaluation [16].

Classes	Evaluation Threshold
A	Ψmin ≤ 0.15, and NC = 1, and RMS ≤ 15°
B	Each parameter is between Class A and Class C
C	Ψmin ≥ = 0.30, or NC ≥ 2

Note: Ψ_min denotes the minimum weighted contradiction ratio; N_C denotes the number of clustering centers; RMS denotes the minimum root mean square of the rotation angle.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/app13042233/s1, Figure S1: Inconsistent discrimination between DiTingMotion and analysts. The red dotted line indicates P-wave arrival time. Light blue shading region indicates the DiTingMotion detection window (±0.64 s centered at the P-wave arrival time); Figure S2: Consistent discrimination between DiTingMotion and analysts. The red dashed line indicates the arrival time of P-wave. Light blue shading indicates the DiTingMotion detection window (±0.64 s centered at the P-wave arrival time); Table S1: The source parameters of 112 earthquakes in this paper.

References

1. Xu, Z.; Liu, Y.; Zhang, Y. On the characteristic of direction of the earthquake stress field around the Beijing aera. Acta Seismol. Sin.; 1979; 1, pp. 121-132.

2. Gephart, J.W.; Forsyth, D.W. An Improved Method for Determining the Regional Stress Tensor Using Earthquake Focal Mechanism Data: Application to the San Fernando Earthquake Sequence. J. Geophys. Res.; 1984; 89, pp. 9305-9320. [DOI: https://dx.doi.org/10.1029/JB089iB11p09305]

3. Wan, Y.; Sheng, S.; Huang, J.; Li, X.; Chen, X. The Grid Search Algorithm of Tectonic Stress Tensor Based on Focal Mechanism Data and Its Application in the Boundary Zone of China, Vietnam and Laos. J. Earth Sci.; 2016; 27, pp. 777-785. [DOI: https://dx.doi.org/10.1007/s12583-015-0649-1]

4. Sheng, S.; Chen, G.; Xu, X.; Wan, Y.; Li, H.; Gong, M.; Wang, X.; Zhang, X.; Zhang, X. Identification of seismogenic faults between focal nodal planes based on tectonic stress fields with applications to the Yingjiang area. Chin. J. Geophys.; 2022; 65, pp. 4273-4283. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg2022Q0001]

5. Jia, K.; Zhou, S. Inference of source parameters of historical major earthquakes from 1900 to 1970 in southwestern China and analysis of their uncertainties. Chin. J. Geophys.; 2012; 55, pp. 2948-2962. (In Chinese) [DOI: https://dx.doi.org/10.6038/j.issn.0001-5733.2012.09.014]

6. Xu, Z.; Yan, M.; Zhao, Z. Evaluation of the direction of tectonic stress in north China from recorded data of a large number of small earthquakes. Acta Seismol. Sin.; 1983; 5, pp. 268-279.

7. Luo, Y.; Ni, S.; Zeng, X.; Zheng, Y.; Chen, Q.; Chen, Y. A shallow aftershock sequence in the north-eastern end of the Wenchuan earthquake aftershock zone. Sci. China Earth Sci.; 2010; 40, pp. 677-687. [DOI: https://dx.doi.org/10.1007/s11430-010-4026-8]

8. Cheng, Y.; Ben-Zion, Y. Variations of Earthquake Properties Before, During, and After the 2019 M7.1 Ridgecrest, CA, Earthquake. Geophys. Res. Lett.; 2020; 47, e2020GL089650. [DOI: https://dx.doi.org/10.1029/2020GL089650]

9. Zhao, L.; Helmberger, D.V. Source Estimation from Broadband Regional Seismograms. Bull. Seismol. Soc. Am.; 1994; 84, pp. 91-104.

10. Zhu, L.; Helmberger, D.V. Advancement in Source Estimation Techniques Using Broadband Regional Seismograms. Bull. Seismol. Soc. Am.; 1996; 86, pp. 1634-1641. [DOI: https://dx.doi.org/10.1785/BSSA0860051634]

11. Zhu, L.; Zhou, X. Seismic Moment Tensor Inversion Using 3D Velocity Model and Its Application to the 2013 Lushan Earthquake Sequence. Phys. Chem. Earth Parts ABC; 2016; 95, pp. 10-18. [DOI: https://dx.doi.org/10.1016/j.pce.2016.01.002]

12. Wang, X.; Zhan, Z. Moving from 1-D to 3-D Velocity Model: Automated Waveform-Based Earthquake Moment Tensor Inversion in the Los Angeles Region. Geophys. J. Int.; 2020; 220, pp. 218-234. [DOI: https://dx.doi.org/10.1093/gji/ggz435]

13. Li, Q.; Zhang, Z.; Jin, Y.; Yu, X.; Li, Z. Focal mechanisms of tangshan earthquakes. Seismol. Geol.; 1980; 2, pp. 59-67.

14. Xu, X.; Xu, Z.; Zhang, D. A probabilistic grid test method of determining earthquake focal mechanisms using P-wave onset polarity data. Seismol. Geomagn. Obs. Res.; 1995; 16, pp. 34-42.

15. Hardebeck, J.L.; Shearer, P.M. A New Method for Determining First-Motion Focal Mechanisms. Bull. Seismol. Soc. Am.; 2002; 92, pp. 2264-2276. [DOI: https://dx.doi.org/10.1785/0120010200]

16. Yu, C.; Tao, K.; Cui, X.; Hu, X.; Ning, J. P-wave first-motion focal mechanism solutions and their quality evaluation. Chin. J. Geophys.; 2009; 52, pp. 1402-1411. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.0001-5733.2009.05.030]

17. Chen, Y.; Liu, R. The use of the first P motion data for earthquake focal mechanism determination: A tutorial(1). Seismol. Geomagn. Obs. Res.; 2021; 42, pp. 1-14. [DOI: https://dx.doi.org/10.3969/j.issn.1003-3246.2021.03.001]

18. Hardebeck, J.L.; Shearer, P.M. Using S/P Amplitude Ratios to Constrain the Focal Mechanisms of Small Earthquakes. Bull. Seismol. Soc. Am.; 2003; 93, pp. 2434-2444. [DOI: https://dx.doi.org/10.1785/0120020236]

19. Hardebeck, J.L. HASH: A FORTRAN Program for Computing Earthquake First-Motion Focal Mechanisms. 2008; Available online: https://www.geo.uib.no/seismo/SOFTWARE/SEISAN/OLDER_VERSIONS/SEISAN_10.4.1/alpha/PCWIN/INF/hash.pdf (accessed on 15 January 2021).

20. Li, J.; Zhang, H.; Kuleli, H.S.; Toksoz, M.N. Focal Mechanism Determination Using High-Frequency Waveform Matching and Its Application to Small Magnitude Induced Earthquakes: Focal Mechanism by Waveform Matching. Geophys. J. Int.; 2011; 184, pp. 1261-1274. [DOI: https://dx.doi.org/10.1111/j.1365-246X.2010.04903.x]

21. Yan, C.; Xu, L. An inversion technique for the mechanisms of local and regional earthquakes: Generalized polarity and amplitude technique (I)—Principle and numerical tests. Chin. J. Geophys.; 2014; 57, pp. 2555-2572. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg20140816]

22. Nakamura, M. Automatic Determination of Focal Mechanism Solutions Using Initial Motion Polarities of P and S Waves. Phys. Earth Planet. Inter.; 2004; 146, pp. 531-549. [DOI: https://dx.doi.org/10.1016/j.pepi.2004.05.009]

23. Chen, C.; Holland, A.A. PhasePApy: A Robust Pure Python Package for Automatic Identification of Seismic Phases. Seismol. Res. Lett.; 2016; 87, pp. 1384-1396. [DOI: https://dx.doi.org/10.1785/0220160019]

24. Pugh, D.J.; White, R.S.; Christie, P.A.F. Automatic Bayesian Polarity Determination. Geophys. J. Int.; 2016; 206, pp. 275-291. [DOI: https://dx.doi.org/10.1093/gji/ggw146]

25. Pei, W.; Zhou, S. Automatic P-wave polarity determination and focal mechanism inversion based on maximum order statistics and its application in Xiaojiang Fault Zone, Yunnan. Chin. J. Geophys.; 2022; 65, pp. 992-1005. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg2022P0313]

26. Farabet, C.; Couprie, C.; Najman, L.; LeCun, Y. Learning Hierarchical Features for Scene Labeling. IEEE Trans. Pattern Anal. Mach. Intell.; 2013; 35, pp. 1915-1929. [DOI: https://dx.doi.org/10.1109/TPAMI.2012.231] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23787344]

27. Tompson, J.; Jain, A.; LeCun, Y.; Bregler, C. Joint training ofa convolutional network and a graphical model for human poseestimation. Adv. Neural Inf. Process. Syst.; 2014; 2, pp. 1799-1807.

28. Mikolov, T.; Deoras, A.; Povey, D. Strategies for training large scale neural network language models. Proceedings of the 2011 IEEE Workshop on Automatic Speech Recognition & Understanding; Waikoloa, HI, USA, 11–15 December 2012; pp. 11-15.

29. Sainath, T.N.; Mohamed, A.; Kingsbury, B.; Ramabhadran, B. Deep Convolutional neural networks for LVCSR. Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing; Vancouver, BC, Canada, 26–31 May 2013; pp. 8614-8618.

30. Perol, T.; Gharbi, M.; Denolle, M. Convolutional neural network for Earthquake Detection and Location. Sci. Adv.; 2018; 4, e1700578. [DOI: https://dx.doi.org/10.1126/sciadv.1700578]

31. Meier, M.; Ross, Z.E.; Ramachandran, A.; Balakrishna, A.; Nair, S.; Kundzicz, P.; Li, Z.; Andrews, J.; Hauksson, E.; Yue, Y. Reliable Real-Time Seismic Signal/Noise Discrimination with Machine Learning. J. Geophys. Res. Solid Earth; 2019; 124, pp. 788-800. [DOI: https://dx.doi.org/10.1029/2018JB016661]

32. Zhu, W.; Beroza, G.C. PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method. Geophys. J. Int.; 2018; 216, pp. 261-273. [DOI: https://dx.doi.org/10.1093/gji/ggy423]

33. Zhou, Y.; Yue, H.; Kong, Q.; Zhou, S. Hybrid Event Detection and Phase-Picking Algorithm Using Convolutional and Recurrent Neural Networks. Seismol. Res. Lett.; 2019; 90, pp. 1079-1087. [DOI: https://dx.doi.org/10.1785/0220180319]

34. Mousavi, S.M.; Ellsworth, W.L.; Zhu, W.; Chuang, L.Y.; Beroza, G.C. Earthquake Transformer—An Attentive Deep-Learning Model for Simultaneous Earthquake Detection and Phase Picking. Nat. Commun.; 2020; 11, 3952. [DOI: https://dx.doi.org/10.1038/s41467-020-17591-w] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32770023]

35. Zhu, J.; Li, Z.; Fang, L. USTC-Pickers: A Unified Set of Seismic Phase Pickers Transfer Learned for China. Earthq. Sci.; 2022; 36, pp. 1-11.

36. Li, Z.; Meier, M.-A.; Hauksson, E.; Zhan, Z.; Andrews, J. Machine Learning Seismic Wave Discrimination: Application to Earthquake Early Warning. Geophys. Res. Lett.; 2018; 45, pp. 4773-4779. [DOI: https://dx.doi.org/10.1029/2018GL077870]

37. Miao, F.; Carpenter, N.S.; Wang, Z.; Holcomb, A.S.; Woolery, E.W. High-Accuracy Discrimination of Blasts and Earthquakes Using Neural Networks With Multiwindow Spectral Data. Seismol. Res. Lett.; 2020; 91, pp. 1646-1659. [DOI: https://dx.doi.org/10.1785/0220190084]

38. Kuang, W.; Yuan, C.; Zhang, J. Real-Time Determination of Earthquake Focal Mechanism via Deep Learning. Nat. Commun.; 2021; 12, 1432. [DOI: https://dx.doi.org/10.1038/s41467-021-21670-x]

39. Kuang, W.; Yuan, C.; Zhang, J.; Zhang, W. Relative Focal Mechanism Inversion and Its Application to Ridgecrest Sequence. Seismol. Res. Lett.; 2022; 13, pp. 305-317. [DOI: https://dx.doi.org/10.1785/0220210370]

40. Ross, Z.E.; Meier, M.; Hauksson, E. P Wave Arrival Picking and First-Motion Polarity Determination With Deep Learning. J. Geophys. Res. Solid Earth; 2018; 123, pp. 5120-5129. [DOI: https://dx.doi.org/10.1029/2017JB015251]

41. Hara, S.; Fukahata, Y.; Iio, Y. P-Wave First-Motion Polarity Determination of Waveform Data in Western Japan Using Deep Learning. Earth Planets Space; 2019; 71, 127. [DOI: https://dx.doi.org/10.1186/s40623-019-1111-x]

42. Uchide, T. Focal Mechanisms of Small Earthquakes beneath the Japanese Islands Based on First-Motion Polarities Picked Using Deep Learning. Geophys. J. Int.; 2020; 223, pp. 1658-1671. [DOI: https://dx.doi.org/10.1093/gji/ggaa401]

43. Zhai, Q.; Peng, Z.; Chuang, L.Y.; Wu, Y.; Hsu, Y.; Wdowinski, S. Investigating the Impacts of a Wet Typhoon on Microseismicity: A Case Study of the 2009 Typhoon Morakot in Taiwan Based on a Template Matching Catalog. J. Geophys. Res. Solid Earth; 2021; 126, e2021JB023026. [DOI: https://dx.doi.org/10.1029/2021JB023026]

44. Zhao, M.; Xiao, Z.; Chen, S.; Fang, L. DiTing: A Large-Scale Chinese Seismic Benchmark Dataset for Artificial Intelligence in Seismology. Earthq. Sci.; 2022; 35, pp. 1-11.

45. Zhao, M.; Xiao, Z.; Zhang, M.; Yang, Y.; Tang, L.; Chen, S. DiTingMotion: A Deep-Learning First- Motion-Polarity Classifier and Its Application to Focal Mechanism Inversion. 2022; in review

46. Xiao, Z. Deep-Learning for Data Processing in Seismology; Institute of Geology and Geophysics, Chinese Academy of Sciences (CAS): Beijing, China, 2022.

47. Zhou, Y.; Ghosh, A.; Fang, L.; Yue, H.; Zhou, S.; Su, Y. A High-Resolution Seismic Catalog for the 2021 M_S6.4/M_W6.1 Yangbi Earthquake Sequence, Yunnan, China: Application of AI Picker and Matched Filter. Earthq. Sci.; 2021; 34, pp. 390-398. [DOI: https://dx.doi.org/10.29382/eqs-2021-0031]

48. Zhou, Y.; Ren, C.; Ghosh, A.; Meng, H.; Fang, L.; Yue, H.; Zhou, S.; Su, Y. Seismological Characterization of the 2021 Yangbi Foreshock-Mainshock Sequence, Yunnan, China: More than a Triggered Cascade. J. Geophys. Res. Solid Earth; 2022; 127, e2022JB024534. [DOI: https://dx.doi.org/10.1029/2022JB024534]

49. Liu, M.; Li, H.; Li, L.; Zhang, M.; Wang, W. Multistage Nucleation of the 2021 Yangbi M _S 6.4 Earthquake, Yunnan, China and Its Foreshocks. J. Geophys. Res. Solid Earth.; 2022; 127, e2022JB024091. [DOI: https://dx.doi.org/10.1029/2022JB024091]

50. Yang, T.; Li, B.; Fang, L.; Su, Y.; Zhong, Y.; Yang, J.; Qin, M.; Xu, Y. Relocation of the Foreshocks and Aftershocks of the 2021 Ms 6.4 Yangbi Earthquake Sequence, Yunnan, China. J. Earth Sci.; 2022; 33, pp. 892-900. [DOI: https://dx.doi.org/10.1007/s12583-021-1527-7]

51. Zhu, G.; Yang, H.; Tan, Y.J.; Jin, M.; Li, X.; Yang, W. The Cascading Foreshock Sequence of the Ms 6.4 Yangbi Earthquake in Yunnan, China. Earth Planet. Sci. Lett.; 2022; 591, 117594. [DOI: https://dx.doi.org/10.1016/j.epsl.2022.117594]

52. Li, X.; Jin, M.; Huang, Y.; Cha, W.; Wang, J.; Li, S. Staged Characteristics on the Consistency Parameters of the Focal Mechanisms of the 2021 Yangbi MS6.4 Earthquake Sequence in Yunnan. Earthq. Res. Adv.; 2021; 34, 100061. [DOI: https://dx.doi.org/10.1016/j.eqrea.2021.100061]

53. Wang, Y.; Hu, S.; He, X.; Guo, K.; Xie, M.; Deng, S.; Ma, Y. Relocation and focal mechanism solutions of the 21 May 2021 Ms6.4 Yunnan Yangbi earthquake sequence. Chin. J. Geophys.; 2021; 64, pp. 4510-4525. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg2021P0401]

54. Wang, Y.; Zhao, T.; Hu, J.; Liu, C. Relocation and Focal Mechanism Solutions of the 2021 Yangbi, Yunnan M_s6.4 earthquake sequence. Seismol. Geol.; 2021; 43, pp. 847-863.

55. Guo, X.; Yin, H.; Wang, Z.; Yang, H. Earthquake centroid, seismic moment tensor and dynamic environment analysis of the Ms6.4 earthquake sequence in Yangbi, Yunnan on May 21, 2021. Seismol. Geol.; 2021; 43, pp. 806-826.

56. Lei, X.; Wang, Z.; Ma, S.; He, C. A preliminary study on the characteristics and mechanism of the May 2021 MS6.4 Yangbi earthquake sequence, Yunnan, China. Acta Seismol. Sin.; 2021; 43, pp. 261-286. [DOI: https://dx.doi.org/10.11939/jass.20210100]

57. Zhao, B.; Gao, Y.; Ma, Y. Relocations, focal mechanisms and stress inversion of the May 21th 2021 Yangbi Ms6.4 earthquake sequence in Yunnan, China. Chin. J. Geophys; 2022; 65, pp. 1006-1020. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg2022P0497]

58. Wang, G.; Wu, Z.; Peng, G.; Liu, Z.; Luo, R.; Huang, X.; Chen, H. Seismogenic fault and it’s rupture characteristics of the 21 May, 2021 Yangbi Ms 6.4 earthquake: Analysis results from the relocation of the earthquake sequence. J. Geomech.; 2021; 27, pp. 662-678. [DOI: https://dx.doi.org/10.12090/j.issn.1006-6616.2021.27.04.055]

59. Zhai, Q. Understanding Microseismicity Behavior and Their Response to Earth Processes by Improving Earthquake Catalogs; Georgia Institute of Technology: Atlanta, GA, USA, 2022.

60. Kagan, Y.Y. 3-D Rotation of Double-Couple Earthquake Sources. Geophys. J. Int.; 1991; 106, pp. 709-716. [DOI: https://dx.doi.org/10.1111/j.1365-246X.1991.tb06343.x]

61. Wan, Y. Determination of Center of Several Focal Mechanisms of the Same Earthquake. Chin. J. Geophys.; 2019; 62, pp. 4718-4728. (In Chinese) [DOI: https://dx.doi.org/10.6038/cjg2019M0553]

62. Zheng, J.; Chen, Y. Regional deviatoric moment tensor inversion based on Langston’s decomposition and Hilbert transform constraints and its application. Acta Seismol. Sin.; 2012; 34, pp. 171-190.

63. Chang, Z.; Chang, H.; Zang, Y.; Dai, B. Recent active features of Weixi-Qiaohou fault and its relationship with the honghe fault. J. Geomech.; 2016; 22, pp. 517-530.

64. Álvarez-Gómez, J.A. FMC—Earthquake Focal Mechanisms Data Management, Cluster and Classification. SoftwareX; 2019; 9, pp. 299-307. [DOI: https://dx.doi.org/10.1016/j.softx.2019.03.008]

65. Chen, Y. Consistency of Focal Mechanism as a New Parameter in Describing Seismic Activity. Chin. J. Geophys.; 1978; 21, pp. 142-159. (In Chinese)

66. Wilding, J.D.; Ross, Z.E. Aftershock Moment Tensor Scattering. Geophys. Res. Lett.; 2022; 49, 14. [DOI: https://dx.doi.org/10.1029/2022GL098473]

67. Rodkin, M.V. Patterns of Seismicity Found in the Generalized Vicinity of a Strong Earthquake: Agreement with Common Scenarios of Instability Development. Extreme Events and Natural Hazards: The Complexity Perspective; Geophysical Monograph Series; Sharma, A.S.; Bunde, A.; Dimri, V.P.; Baker, D.N. American Geophysical Union: Washington, DC, USA, 2012; Volume 196, pp. 27-39. ISBN 978-0-87590-486-3

68. Wang, J.; Diao, G. Consistent CMT Solutions from Harvard University before the Great Earthquakes in Kurile Islands and its Significance for Earthquake Prediction. China. Acta Seismol. Sin.; 2005; 27, pp. 178-183.

69. Wan, Y. Study on Consistency of Focal Mechanism of Mainshock and That of Preshocks in Landers and Hector Mine Earthquake in United States. Earthq. Res. China; 2008; 24, pp. 216-225.

70. Ickrath, M.; Bohnhoff, M.; Bulut, F.; Dresen, G. Stress Rotation and Recovery in Conjunction with the 1999 Izmit Mw 7.4 Earthquake. Geophys. J. Int.; 2014; 196, pp. 951-956. [DOI: https://dx.doi.org/10.1093/gji/ggt409]

71. Sheng, S.; Meng, L. Stress Field Variation During the 2019 Ridgecrest Earthquake Sequence. Geophys. Res. Lett.; 2020; 47, e2020GL087722. [DOI: https://dx.doi.org/10.1029/2020GL087722]

72. Trugman, D.T.; Ross, Z.E.; Johnson, P.A. Imaging Stress and Faulting Complexity Through Earthquake Waveform Similarity. Geophys. Res. Lett.; 2020; 47, e2019GL085888. [DOI: https://dx.doi.org/10.1029/2019GL085888]