Machine Learning-Based Precursor Detection Using Seismic Multi-Parameter Data

Abstract

The application of certain mathematical–statistical methods can quantitatively identify and extract the abnormal characteristics from the observation data, and the comprehensive analysis of seismic multi-parameters can study and judge the risk of the tectonic regions better than a single parameter. In this study, the machine learning-based detection of seismic multi-parameters using the sliding extreme value relevancy method, based on the earthquake-corresponding relevancy spectrum, was calculated in the tectonic regions in the western Chinese mainland, and the R-value evaluation was completed. Multi-parameter data included the b value, M value (missing earthquakes), ƞ value (the relationship between seismic magnitude and frequency), D value (seismic hazard), Mf value (intensity factor), N value (earthquake frequency), and Rm value (modulation parameter). The temporal results showed that the high-value anomalies appeared before most target earthquakes during the training period. Moreover, some target earthquakes also occurred during the advantageous extrapolation period with high-value anomalies. The spatial results showed that some months before the target earthquakes, there was indeed a significant abnormal enhancement area that appeared near the epicenter, and the anomaly gradually disappeared after the earthquakes. This study demonstrated that machine learning techniques for detecting earthquake anomalies using seismic multi-parameter data were feasible.

1. Introduction

Various seismological parameters may exhibit different degrees of precursor anomaly characteristics before major earthquakes and reveal certain laws of earthquake development from different perspectives. There were many studies on the b value within seismological parameters [1,2,3]. The b value is a critical parameter in seismic hazard studies [4], and a high b value indicates a larger proportion of small earthquakes, and vice versa. In addition, earthquake frequency is also a commonly used seismological parameter internationally [5,6]. What would be the effect of integrating these seismic parameters for analysis? There has been relatively little comprehensive study of seismic parameters, which could solve the limitations of single parameter and method approaches [7]. The comprehensive analysis of multiple methods based on the quantitative description of a single anomaly is of great significance in earthquake research. Based on this idea, the study was conducted on the quantitative identification and extraction of anomalies from observation data using mathematical–statistical methods. For example, Wang Haitao [8] transformed the time series of original data into a probability time series based on the corresponding relevancy spectrum in different investigation durations, and used the multi-point cumulative moving average method to obtain the time curve of average probability; then, the earthquake precursor anomalies were identified, providing quantitative single factor data for the comprehensive method. Bo [9] used the multi-point group slope method and composite information flow method to convert the deformation data into the ‘standardized information curve’. Wang [10] quantitatively identified and extracted anomalies from various seismic parameter data in the western section of the southern Tianshan of Xinjiang by using the multi-parameter sliding extreme-value relevancy based on the earthquake-corresponding relevancy spectrum. Lu also conducted a long-term tracking study on the central and southern sections of the Tanlu fault, utilizing the earthquake-corresponding relevancy spectrum of seismic multi-parameters, and found that the earthquake-corresponding efficiency calculated by combining multiple advantageous parameters was indeed better than that of a single parameter [11].

Some studies also found that although some calculated factors and models were different, the regional scales of the predicted earthquake were consistent [12,13,14,15]. Perhaps more earthquake factors could more easily predict the spatial–temporal and intensity information of earthquakes. Therefore, it is possible to try combining multiple different factors within a unified and reasonable physical framework for the extraction of seismic anomalies. For example, establishing a physically reasonable framework that combines the pattern informatics (PI) prediction method [16,17] with the load/unload response ratio (LURR), state vector (SV), and accelerating moment release (AMR) method [18] could be considered. Alternatively, the LURR method could be combined with other methods for seismic hazard assessment [19]. Scholars have also attempted to use the LURR method to study the electromechanical coupling process before large earthquakes, using geoelectric data and Benioff strain data from small earthquakes as input [20], and the results were both well. These studies fully demonstrated the advantages of using multiple different factors and models for comprehensive calculation. Recently, there have been many studies on the application of machine learning for earthquake precursor mining and analysis. Possible precursors from the surface to the ionosphere using machine learning techniques were analyzed [21]. And earthquake precursors could be detected by a novel machine learning-based technique with GPS data [22].

The strongest earthquakes in China occur in the western part of the Chinese mainland. This paper focuses on analyzing the characteristics of comprehensive anomalies before and after target earthquakes in the study area through the calculation of the comprehensive probability of seismic multi-parameters in several tectonic areas in the western Chinese mainland. Moreover, the evaluations of the R-value for these tectonic areas were calculated, and the effectiveness of corresponding earthquakes during the advantageous prediction time for the regions with anomalies was tested.

2. Materials and Methods

2.1. Materials

The study areas mainly encompassed 12 tectonic regions in the western Chinese mainland, including the western section of southern Tianshan, the middle of Tianshan, the northwest of Yunnan, the east of the border between Sichuan and Yunnan, the western portions of southern Xizang, Yutian, Xinjiang, and so on (Figure 1). The blue lines in Figure 1 represents the faults. Meanwhile, a completeness analysis and collation of earthquakes since 1991 were carried out. The completeness of earthquake catalogs directly affects the understanding of seismic activity patterns. For a certain tectonic zone, if the earthquake catalog is basically complete in a certain period, the annual frequency of earthquakes in a certain magnitude segment should basically be similar [23]. The seismic data in this study primarily relied on the local magnitude (M_L) and adopted the ≥M_L3.0 earthquake catalog from January 1991 to December 2021 provided by the China Earthquake Networks Center, with good completeness.

The appropriate earthquake catalogs for each region in the training database were selected, and the multi-parameter sliding extreme-value relevancy of the training database was calculated, combined with the target earthquakes. Then, the study extrapolated the probability and advantageous time range of target earthquakes in each study area and completed the calculation of the R-value evaluation for each study area.

2.2. Method

Assuming the observation sequence was ${{\displaystyle x}}_1$, ${{\displaystyle x}}_2$, …, ${{\displaystyle x}}_n$, ${{\displaystyle x}}_{\max }$ was the maximum value in the sequence, ${{\displaystyle x}}_{\min }$ was the minimum value. The mean value, $\overline{x}$, and standard deviation, $\sigma$, of the sequence ${{\displaystyle x}}_i$ (i = 1,2,…, n) could be calculated as follows:

$$\overline{x}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{{\displaystyle x}}_i}$$

(1)

$$\sigma =\sqrt{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2/(n+1)}$$

(2)

Based on the calculation of the mean value and standard deviations, the rules for the value range distribution interval are defined as follows:

$$D{x}_0\in (\overline{x}-k\sigma ,\overline{x}+k\sigma )$$

$$D{x}_1\in [\overline{x}+k\sigma ,\overline{x}+2k\sigma )$$

$$D{x}_m\in [\overline{x}+mk\sigma ,x_{\max }]$$

(3)

$$D{x}_{-1}\in (\overline{x}-2k\sigma ,\overline{x}-k\sigma ]$$

$$D{x}_{-m}\in [x_{\min },\overline{x}-mk\sigma ]$$

$D{x}_0$, $D{x}_1$, …, $D{x}_{-m}$ in the formula represent the observed values in different value ranges. Formula (3) calculates the frequency of the observed value sequence, ${{\displaystyle x}}_i$, distributed across different intervals, thereby constructing a range spectrum curve to define the distribution interval of the value ranges. Taking the range spectrum curve of the seismic parameter, b, in the western section of southern Tianshan as an example (Figure 2), 0 on the horizontal axis corresponds to the value range of $D{x}_0$, which was near the mean. Value 4 corresponds to the value range $D{x}_4$, −4 corresponds to the value range $D{x}_{-4}$, and so on. The vertical axis represents the frequency of data in the corresponding range.

The range spectrum curve is similar to a normal distribution, indicating that the range selection is appropriate. k and m are both parameters of the range spectrum curve, where k could adjust the size of each range interval and m could adjust the number of range intervals. This study conducted extensive data calculations and found that the best results were achieved by uniformly using k = 0.35 and m = 6 to eliminate errors caused by calculating with different parameters of the value range.

Based on the analysis of the value range spectrum, the earthquake-corresponding relevancy spectra (ECRSs) in different value ranges could be solved. By using the ECRS, one can analyze the basic abnormal characteristics of the original observation value sequence and determine the abnormal reliability attributes of data in different value ranges. Firstly, the target earthquakes for retrospective testing in different study areas must be determined. Then, the data falling into the value range from $D{x}_{-m}$ to $D{x}_m$ (from low to high) are counted, point by point, according to the time series of observation data. At the same time, the method needs to count whether there are target earthquakes in different inspection periods t separately.

According to the above rules, the corresponding number of occurred earthquakes, $nD{x}_m$, as well as the total number, $ND{x}_m$, in the corresponding value range are determined through the count of the observation value sequence, point by point. Then, the ratio is calculated by $PD{x}_m=nD{x}_m/ND{x}_m$. The ECRSs in different value ranges of the observation value sequence could be obtained by counting the corresponding results of all value ranges.

The time series of different seismic parameters based on the ECRS is $x_{ij}$ (i = 1, 2, …, n; j = 1, 2, …, k; k represent the different seismic parameter numbers). The relevancy time sequence, $p_{ij}$, corresponding to different parameters, j, could be converted from the $PD{x}_m$ in different value ranges, point-to-point. Then, the sliding average relevancy sequence values, ${\overline{p}}_{ij}$, of different seismic parameters, according to the lengths of different inspection periods, t, are obtained by the multi-point cumulative average and point-by-point sliding calculation methods.

Setting t = 12 (month) represents the different investigation durations. The sliding average relevancy sequence, ${\overline{p}}_{ij}$, for different seismic parameters is as follows:

$${\overline{p}}_{ij}=(p_{ij}+p_{(i+1)j}+\dots +p_{(i+t-1)j})/t, i=1,2,\dots ,n-t+1$$

(4)

By analyzing the sliding average relevancy sequences, ${\overline{p}}_{ij}$, of different seismic parameters, the precursor anomaly characteristics of individual seismic parameters could be quantitatively identified and studied.

Based on the above analysis, we could obtain the multi-point sliding extreme-value relevancy sequence, $M_{ij}$, of the sliding average relevancy sequence, ${\overline{p}}_{ij}$, of different parameters, point by point, and finally, calculate the average value of the sliding extreme-value relevancy sequences, $M_j$, of different parameters to obtain the multi-parameter sliding extreme-value relevancy value, $p_c$. In order to observe the curve changes more clearly, 6 points are selected as the sliding window lengths in this paper, and the comprehensive results of sliding, point by point, are marked on the time coordinates of the last point.

Suppose that j = 1, 2, …, k (k represents the different seismic parameters), w = 6 (w represents the sliding window lengths of different points).

$$M_{ij}=Max\left\{{\overline{p}}_{ij},{\overline{p}}_{(i+1)j},\dots ,{\overline{p}}_{(i+w-1)j}\right\},i=1,2,\dots ,n-w+1$$

(5)

$$M_j=\left(M_{1j},M_{2j},\dots ,M_{(n-w+1)j}\right),j=1,2,\dots ,k$$

(6)

In the formula, $M_{ij}$ represents the multi-point sliding extreme-value relevancy for different parameters; $M_j$ represents the multi-point sliding extreme-value relevancy of a single parameter, j.

Then, the multi-parameter sliding extreme-value relevancy value, $p_c$, is calculated by averaging the values of $M_j$ from different parameters.

$$p_c=\sum _{j=1}^k{M}_j/k$$

(7)

3. Results

3.1. Multi-Parameter Relevancy Spectrum Calculation

The most reasonable single parameter and its threshold settings, within the probability spectrum of earthquake occurrence and the value range spectrum in each study area, were analyzed through time scanning with different statistical window lengths and step sizes. After years of analysis and experience summaries, the optimal seven parameters were ultimately selected to participate in the calculation of each research area, namely, the b value, M value (missing earthquake), and η Value—these three parameters could describe the relationship between seismic magnitude and frequency; the D value describes seismic hazard; the Mf value denotes the intensity factor; the N value denotes the earthquake frequency, which describes the intensity changes of seismic activity; and the Rm value is the environmental modulation parameter.

A large portion of the computational analysis was conducted across different research areas. Using the three study areas of border region as examples, e.g., Qinghai and Xizang, the western section of southern Tianshan, and the central and western sections of the Qilian Mountains, different time scanning scales were applied for calculation: 3 months, 6 months, and 12 months. The results showed that the curve did not change much, but the anomaly and earthquake-corresponding rates of the three study areas were relatively high at the 12-month scale (

Table 1. Anomaly-corresponding rate and earthquake-corresponding rate of different time scanning scales.

Name		Anomaly-Corresponding Rate (%)			Earthquake-Corresponding Rate (%)
Time scanning scale (month)		3	6	12	3	6	12
Study region	Border regions of Qinghai and Xizang	50.00	66.67	75.00	71.43	100.00	100.00
	Western section of southern Tianshan	50.00	66.67	100.00	71.43	85.71	100.00
	Central and western sections of the Qilian Mountains	77.78	85.71	85.71	80.00	80.00	80.00

). Therefore, this study uniformly used a 12-month time scanning scale for all study areas, and the seismic parameters and thresholds used in all study areas were also the same, which could help to avoid errors caused by differences in the time scale and parameters.

Taking the abnormal changes in the border areas of Qinghai and Xizang, where the Maduo M7.4 earthquake occurred on 22 May 2021, as an example, and based on the previous summary and analysis that 12 months was a good sliding window length for calculation, the window lengths of the time calculations for seven parameters were all selected as 12 months, and the calculation step sizes were all 1 month. The 6-point sliding average and 3-point sliding extreme-value algorithms were used to calculate the multi-parameter sliding extreme-value relevancy of earthquake data from January 1991 to December 2021. According to the time distribution of target earthquakes in the study area, it is necessary to ensure that the training period had sufficient sample data and a certain number of target earthquakes to test the extrapolation effect. Therefore, the deadline for the training period here was selected as December 2020, and the target earthquakes in the study area were set as M_L ≥ 6.3 earthquakes. The results that began in January 2021 were extrapolation values after the machine learning, and the threshold of the probability curve was taken as 0.5 times the standard deviation. The threshold line was obtained through a large number of calculations in 12 study areas in the western Chinese mainland. It was found that when 0.5 times the standard deviation of the multi-parameter probability result was taken as the threshold line, the corresponding effect between the probability value anomaly and the target earthquake was the best. Therefore, in order to eliminate the error caused by using different threshold lines, all study areas uniformly used 0.5 times the standard deviation as the threshold line.

Figure 3 shows the study area of the border regions of Qinghai and Xizang. Figure 4 shows the MSER results of the study area. The results show that the earthquake-corresponding rate was good during the training period from January 1991 to December 2020. During the training period, a total of seven target earthquakes occurred, all with high-value anomalies exceeding the threshold line before the earthquakes. The extrapolation period began in January 2021. The probability curve had been in an abnormally high-value state and reached its peak in January 2021. On 19 March 2021, the Xizang Biru M_L6.4 earthquake occurred in the study area, proving that the extrapolation effect was good. According to the retrospective study of earthquakes in this region, there was also an anomaly peak before the Qinghai Yushu M_L7.4 earthquake on 14 April 2010; the anomaly peak occurred in September 2009 and the Xizang Nierong M_L6.6 earthquake occurred on 24 March 2010 (about one month prior). The Maduo M_L7.9 earthquake was similar to the Yushu earthquake. In the process of the curve turning and falling, the Xizang Biru M_L6.4 earthquake first occurred on 19 March, then the Qinghai Maduo M_L7.9 earthquake occurred on 22 May (about two months later).

The spatial scanning calculation of the border regions of Qinghai and Xizang adopted the same seven dominant seismic parameters, which were used in the time scanning. M_L ≥ 3.0 earthquakes from January 1991 to December 2020 were selected as the training data. The probability spectrum calculation analysis was carried out with a spatial window length of 2° × 2° and a step size of 0.2° × 0.2°, with a duration of 1 year and a step size of 1 month. January 2021 was the starting time for extrapolation testing. The Xizang Biru M_L6.4 earthquake on 19 March 2021 and the Qinghai Maduo M_L7.9 earthquake on 22 May 2021 both occurred in the study area. From the results in Figure 5, it could be seen that there were almost no high-value anomalies in the study area from January to May 2019. In June 2019, sporadic high-value anomalies began to appear near the Xizang Biru region. Afterward, the abnormal area and amplitude gradually expanded and strengthened, reaching their maximum value in December 2020. The Biru M_L6.4 earthquake occurred in the abnormal area in March 2021. At the same time, abnormal high points began to appear between Yushu and Maduo in Qinghai from March 2021 (inside the red dashed box in Figure 5). Although the abnormal points were small, they showed a gradually increasing trend. The Maduo M_L7.9 earthquake occurred to the northeast of the abnormal points in May 2021. Afterward, the abnormal values in both regions gradually weakened, and the high-value abnormal points near the Maduo earthquake completely disappeared in December 2021.

It was found that there were similar highly abnormal changes in 2010. From January 2008 to 2009, the anomaly gradually increased, and almost reached its maximum value in December 2009. In March 2010, the Xizang Nierong M_L6.6 earthquake occurred near the anomaly concentration area. At the same time, the abnormal high-value points also gradually increased near the Yushu region in December 2009 (inside the red dashed box in Figure 6). In April 2010, the Qinghai Yushu M_L7.4 earthquake occurred near the high-value points. Afterward, the abnormal areas gradually weakened, and almost all of them disappeared in December 2010 (Figure 6).

3.2. Risk Assessment of the Multi-Parameter Relevancy Spectrum in the Western Chinese Mainland

The R-value evaluation is a widely recognized method used for evaluating the effectiveness of earthquake prediction [24,25]. The R-value evaluation is calculated by the difference between the earthquake prediction accuracy rate and the time occupancy rate of prediction. The specific formula is as follows:

$$R=A-O=\frac{N_a}{N_t}-\frac{S_a}{S_t}$$

(8)

$A$ represents the earthquake prediction accuracy rate, $O$ represents the time occupancy rate of the predicted area, $N_a$ represents the number of predicted earthquakes successfully, $N_t$ represents the total number of earthquakes that should be predicted, $S_a$ represents the time occupied by prediction, and $S_t$ represents the total time spent on prediction research.

Assuming that earthquakes occur independently of each other, the number of earthquakes that occur within a certain period should follow the discretization state of the Poisson distribution. The probability of a single earthquake occurring during this period is represented by the time occupancy rate, $O$, of the predicted area, and the corresponding significance level should be satisfied as follows [26]:

$$\begin{array}{l}\alpha ={\displaystyle \sum _{i=N_a}^{N_t}\left[\left(N_{ti}{O}^i{(1-O)}^{N_t-i}\right)\right]}\\ N_{ti}=\frac{N_t!}{i!(N_t-i)!}\end{array}$$

(9)

In the formula, $\alpha$ represents the significance level, taken as 5%. Based on the known $A$ value, the value $O$ is calculated, and the corresponding $R$ score for the significance level $\alpha$ is obtained and represented by $R$. When $R$ > $R_0$, the predicted result is considered to have high statistical significance [27,28].

In this paper, the R-values of the multi-parameter calculation results of 12 tectonic regions in the western Chinese mainland were evaluated, and finally, 7 tectonic regions passed the test (

Table 2. Characteristics of multi-parameter relevancy spectrum anomalies in the western Chinese mainland.

No.	Region	Three Elements of Prediction			R-Value
		Time/Day	Intensity/ML	Location
1	Central Tianshan	290	≥6.0	Central Tianshan	R = 0.55 R0 = 0.38
2	Western section of southern Tianshan	380	≥6.3	Western section of southern Tianshan	R = 0.44 R0 = 0.29
3	Border regions of Qinghai and Xizang	260	≥6.3	Border area of Qinghai and Xizang	R = 0.68 R0 = 0.35
4	Yutian, Xinjiang	190	≥5.5	Yutian, Xinjiang	R = 0.58 R0 = 0.29
5	Nima in Xizang	100	≥5.5	Nyima in Xizang	R = 0.32 R0 = 0.28
6	The central and western section of the Qilian Mountains	170	≥5.5	The central and western section of the Qilian Mountains	R = 0.54 R0 = 0.35
7	Eastern of Southern Xizang	370	≥6.0	Eastern of Southern Xizang	R = 0.51 R0 = 0.36

), namely, central Tianshan, the western section of southern Tianshan, the border regions of Qinghai and Xizang, Yutian in Xinjiang, Nima in Xizang, the central and western sections of the Qilian Mountains, and the eastern section of southern Xizang. The criterion for determining anomalies was that the relevancy value exceeded the threshold line, and an abnormal peak appeared. The time among the three elements was considered the advantageous extrapolation time when calculating the optimal R-value by different times (Table 2). The extrapolation time period started with the abnormal peak time. Taking central Tianshan in Table 2 as an example, if its relevancy value exceeded the threshold line and reached an abnormal peak in October 2019, its advantageous extrapolation time was 290 days. Therefore, the validity period should be until July 2020. The intensity of the predicted earthquake was based on the minimum magnitude of target earthquakes used for simulation training in the training period. The predicted location refers to the area studied and calculated.

3.3. Extrapolation Inspection

The research data used in this paper were all up to December 2021. Among the seven areas that passed the R-value evaluation test, in addition to the border regions of Qinghai and Xizang, there were also areas with high anomalies, such as the western section of southern Tianshan and the central and western sections of the Qilian Mountains. The advantageous extrapolation times for the western section of southern Tianshan and the central and western sections of the Qilian Mountains were 380 days and 170 days after the multi-parameter relevancy value exceeded the threshold line and experienced abnormal peaks, respectively (Table 2).

Figure 7 shows the study area of the western section of southern Tianshan. Figure 8 shows that the target earthquakes in the western section of southern Tian Shan had good earthquake-corresponding rates during the training period (January 1991 to October 2016). There were seven target earthquakes during the training period, and all high-value anomalies exceeding the threshold line appeared before the earthquakes. Taking November 2016 as the starting time for extrapolation, the relevancy curve was still at a high value. The Aktao M_L7.0 earthquake in Xinjiang occurred in the study area on 25 November 2016, which showed that the extrapolation effect was good. However, there was no high-value anomaly before the Jiashi M_L6.7 earthquake in Xinjiang on 19 January 2020, which was considered a missed earthquake. Afterward, the relevancy curve continued to rise, and it exceeded the threshold line, reaching an abnormal peak in February 2021, indicating the possibility of ≥M_L6.7 earthquakes within the study area. According to Table 2, the advantageous time for the multi-parameter relevancy spectrum in this region was 380 days. Assuming the peak time of the anomaly in February 2021 as the starting time, the validity period would be until 16 February 2022. On 13 February 2022, the Tajikistan M_S6.1 earthquake occurred near the anomaly area, corresponding to this high-value anomaly.

The Aktao M_L7.0 earthquake in Xinjiang on 25 November 2016 occurred during the extrapolation test period. The spatial evolution of the MSER anomaly in the western section of the southern Tianshan Mountains (Figure 9) showed that the anomaly gradually increased from June 2016, reaching its maximum intensity in September 2016, and the Aktao M_L7.0 earthquake occurred near the anomaly area in November. Afterward, the anomaly gradually weakened until it disappeared.

The training period in the central and western sections of the Qilian Mountains (Figure 10) was from January 1991 to December 2015, during which, a total of 10 target earthquakes occurred. Among them, eight target earthquakes had high-value anomalies exceeding the threshold line before the earthquakes, and two target earthquakes were missed (Figure 11). In January 2016, as the starting time for extrapolation, the relevancy curve continued to decline after reaching the abnormal peak in July 2018, and reached its lowest point in February 2019. During this period, there was no corresponding target earthquake. Afterward, the abnormal curve rose again and reached its peak in July 2019. Two months later, on 16 September 2019, the Gansu Zhangye M_L5.5 earthquake occurred. The relevancy curve had been continuously decreasing and remained below the threshold line after the Zhangye M_L5.5 earthquake in Gansu. The curve began to turn and rise in March 2021, and the Aksai M_L6.0 earthquake in Gansu occurred on 26 August 2021 during the curve ascent process, exceeding the threshold line. The curve reached its peak again in January 2022. Table 2 shows that the advantageous extrapolation time in the central and western sections of the Qilian Mountains was 170 days, so the validity period should be until 20 June 2022. From January 2022 to 20 June 2022, the study region experienced the Qinghai Menyuan M_L7.1 earthquake on 8 January 2022, the Gansu Sunan M_L5.5 earthquake on 17 March 2022, and the Qinghai Delingha M_L6.3 earthquake on 26 March 2022, respectively, proving that the extrapolation effect in this study area was very good.

The extrapolation test period for the central and western sections of the Qilian Mountains began in January 2016. The spatial evolution of the MSER anomaly before the 8 January 2022 Qinghai Menyuan M_L7.1 earthquake showed that the anomaly amplitude in the study area began to increase from August 2020, with the largest anomaly increase in November 2021 (Figure 12). The red pentagram in Figure 12 represents the epicenter position of the Qinghai Menyuan M_L7.1 earthquake in January 2022.

4. Discussion

The abnormal patterns of different parameters before the target earthquakes were different. There were differences in the starting times, peak values, and end times of the anomalies. The comprehensive analysis method of seismic parameters can combine single parameters related to physical processes at different stages of earthquake preparation, which could extract the comprehensive abnormal characteristics during the earthquake preparation process more accurately. Wang [10] conducted a comprehensive multi-parameter study on earthquakes from 1979 to 2008 in the western section of southern Tianshan, and the results showed that target earthquakes in the study area exhibited significant high-probability anomalies of multiple parameters 1–2 years prior to the earthquakes. This research conclusion was consistent with the advantageous extrapolation time of 380 days in the western section of southern Tianshan in this paper. In addition, the anomaly-corresponding rate is the ratio of the number of anomalies corresponding to the target earthquake to the total number of anomalies. The earthquake-corresponding rate is the ratio of the number of target earthquakes with anomalies before the earthquake to the number of target earthquakes. By statistically analyzing the anomaly-corresponding rates and earthquake-corresponding rates of single parameters and multiple parameters in different inspection periods and regions of Xinjiang (

Table 3. The anomaly-corresponding rates and earthquake-corresponding rates of single parameters and multiple parameters in different regions of Xinjiang [10].

Study Region	Single Parameter		Multi-Parameter
	Anomaly-Corresponding Rate (%)	Earthquake-Corresponding Rate (%)	Anomaly-Corresponding Rate (%)	Earthquake-Corresponding Rate (%)
West section of Tianshan	54.32	46.88	66.67	68.75
Baicheng-Kuche region	49.66	58.29	70.50	95.00
Korla region	55.64	63.69	70.59	83.33
Keping region	73.12	55.71	100.00	77.50

), it was found that the anomaly-corresponding rates and earthquake-corresponding rates of multiple parameters were indeed higher than the predictive efficiency of a single parameter.

Related studies had also found that—when combined with methods such as PI, LURR, SV, and AMR—the earthquake prediction performance was better than that of a single method [18]. At present, this study mainly applied different seismological parameters for comprehensive analysis through machine learning. This method could also comprehensively calculate other observation data of earthquake precursors, such as deformation, electromagnetic, and underground fluid anomalies. It was even possible to merge and summarize all seismic observation data to complete a comprehensive probability analysis. At the same time, by analyzing the spatial evolution of the MSER during the extrapolation test periods, it was found that the target earthquakes generally occurred after the maximum amplitude of anomaly enhancement was reached, and the epicenter was located in or near the anomaly concentration area. This indicated that the model obtained by using this method for machine learning of historical earthquakes had a good extrapolation effect, which can provide a reference for the seismic hazard of the study area.

In addition, the R-value in this manuscript is different from the linear correlation coefficient, R, which measures the relationship between two sets of data. The R-value evaluation is calculated by the difference between the earthquake prediction accuracy rate and the time occupancy rate of prediction, as outlined in Formula (8). The best prediction performance corresponds to paying the minimum cost (O→0) at the highest accuracy rate (A→1), that is, R→1. On the contrary, the worst prediction performance corresponds to paying the maximum cost (O→1) at the lowest accuracy rate (A→0), that is, R→−1. It is generally believed that when R > 0, the success rate of the forecasting is higher than the random probability, and this result has a certain predictive effectiveness [25]. R₀ represents the derived R-value based on the significance level of 5% (95% confidence level). Therefore, when $R$ > $R_0$, the predicted result is considered to have high statistical significance [27,28].

5. Conclusions

Twelve regions in the western Chinese mainland were studied using a comprehensive multi-parameter method, selecting sufficient earthquake events for machine learning to extrapolate and predict the risk of the study areas. The results showed that there were three structural regions with obvious anomalies during the extrapolation period, namely the border regions of Qinghai and Xizang, the western section of southern Tianshan, and the central and western sections of the Qilian Mountains. And the three structural regions all had corresponding target earthquakes within the advantageous extrapolation time. A total of seven advantageous seismological parameters were selected for training and calculation. In order to eliminate errors, all 12 study areas in this paper used the same seismic parameters, parameter thresholds, spatiotemporal scanning scales, and anomaly threshold lines for their calculation data.

In the calculated results, the extrapolation times of the border regions of Qinghai and Xizang started from January 2021, and the results showed that the relevancy curve of extrapolation was in a highly abnormal state. During the advantageous extrapolation time, the Xizang Biru M_L6.4 earthquake on 19 March 2021 and the Qinghai Maduo M_L7.9 earthquake on 22 May 2021 occurred, which were similar to the change states of the previous Xizang Nierong M_L6.4 earthquake on 24 March 2010 and the Qinghai Yushu M_L7.4 earthquake on 14 April 2010. Both spatial evolution processes were also very similar, with earthquakes occurring in areas with abnormally high values and their vicinity. This indicated that the seismic activity in this area had a certain regularity. If similar anomalies occur again in the future, it will be necessary to be vigilant about the risk of short-term continuous occurrences of earthquakes in this region.

The extrapolation effects in the western section of southern Tianshan and the central and western sections of the Qilian Mountains were both highly well through the machine learning-based detection of seismic multi-parameters. The anomaly-corresponding rate and earthquake-corresponding rate were all very high in the calculated results. The application of the comprehensive analysis of seismic multi-parameters in the western Chinese mainland could not only infer the urgency of the target earthquakes (in terms of time) but also distinguish the possible areas where the target earthquakes occurred from the spatial distribution of abnormal areas. In all, the results showed that the extrapolation effect using the machine learning method in the western Chinese mainland was very good.