Improved medium-to-short-term earthquake predictions in China in 2022

Abstract

Every year, China should determine annual seismic hazard regions for earthquakes of magnitude 5.0 and above in Mainland China in the next year. Meanwhile, the short-term earthquake potential in each seismic hazard region is evaluated by monthly analyzing geophysical observations in the same neighborhood and identifying anomalies therein. In 2022, China improved the strategies for these medium-to-short-term earthquake predictions. The annual seismic hazard regions are produced with the optimal methods selected by rigorous tests whose predictions are integrated through the Bayesian formula. And in short-term earthquake potential evaluation, in addition to tracking crucial anomalies, China pays attention to the changes in the number of observed anomalies. The practice in 2022 shows that more than 80% of earthquakes of magnitude 5.0 and above in Mainland China occurred in annual seismic hazard regions. Besides, the largest (M_s 6.9 Menyuan, Qinghai) and the most disastrous (M_s 6.8 Luding, Sichuan) earthquakes were predicted in both annual and short terms. Some crucial anomalies were observed before these two earthquakes, whose evolutions correlate well with the earthquake time and location. The statistics of the near-field geophysical observations indicate that, prior to the mainshock, the number of anomalies increased significantly and migrated toward the epicenter.

Keywords:

• Earthquake prediction in China in 2022
• annual seismic hazard region
• Bayesian formula
• short-term earthquake potential
• strategies for decision-making

1. Introduction

Earthquake prediction is a controversial topic (Scholz et al. 1973; Silver and Wakita 1996; Geller et al. 1997; Kafka and Ebel 2007) and is also one of the most challenging problems in the world at present (Jordan et al. 2006; Jordan and Jones 2010; Jordan et al. 2011). China is the only country that has established a national institution for earthquake prediction (Zhuang and Jiang 2012). The monitoring systems, data accumulation, and prediction experiences are valuable for studies of this kind. The China Earthquake Administration (CEA), established in 1971, is responsible for earthquake prediction in China. After 50 years of effort, the CEA has formed a feasible strategy for earthquake prediction. Correspondingly, a series of observation networks (Figure 1) and analysis methods have been developed (Shi et al. 2001). The medium-term earthquake prediction is to detect areas where earthquakes of magnitude 5.0 or above may occur in the next year, that is, to determine annual seismic hazard regions. The short-term prediction is to evaluate the earthquake potential in each seismic hazard region within the following short time period, which is performed by tracking anomalies in the near-field observations monthly (Yu et al. 2022).

Figure 1. The distribution of seismometry, crustal deformation, electromagnetism, groundwater, and GNSS observation stations in Mainland China. The inset map at the right-side bottom corner is the South China Sea Islands, and the red circles represent the 22 May 2021 M_s7.4 Maduo, 8 Jan. 2022 M_s6.9 Menyuan, and 5 Sep. 2022 M_s6.8 Luding earthquakes, respectively.

The CEA has established diverse disciplines to monitor the observation data of Seismometry, Crustal deformation, Groundwater, Electromagnetism, Gravity, Global Navigation Satellite System (GNSS), and so on, and submit prediction opinions based on the observed anomalies (Yu et al. 2022). The anomalies commonly adopted for medium-to-short-term earthquake prediction are derived from the morphological analysis of raw data and the comprehensive calculation. The morphological analysis method is to detect the changes in the seismicity pattern (Mogi 1990; Rundle et al. 2003) or identify anomalous annual variation and significant increases or decreases in the raw data of each discipline (Koizumi et al. 2004; Cicerone et al. 2009). For example, Ye et al. (2015) studied the characteristics of water radon changes before the great Wenchuan earthquake in 2008. He et al. (2016) computed atmospheric pressure and earth tidal coefficients in water level associated with the Wenchuan earthquake. Jing et al. (2018) studied the thermal anomalies associated with the 1997 Manyi and 2001 Kokoxili earthquakes in Tibetan Plateau, China using microwave remote sensing data. Jing and Singh (2021) explored the index of ozone anomalies (IOA) prior to the 2008 Wenchuan, the 2013 Lushan and the 2017 Jiuzhaigou earthquakes. The comprehensive calculation method extracts anomalies by evaluating fault motion and its stress state (Shao et al. 2022), seismic occurrence rate (Bodri 2001), seismic repeatability (Kragh and Christie 2002), MMEP (Yu et al. 2013), b value (Gulia and Wiemer 2019), Geodetic strain rate (Wu et al. 2021), seismicity enhancement/quiet (Keilis-Borok 1982), and so on. With the development of observation techniques and prediction methods, the hit rate of annual seismic hazard regions is gradually improving, with about 50% of earthquakes of magnitude 5.0 and above detected in recent years (Yu et al. 2022).

In the prediction of 2022, the CEA attempted to apply probability synthesis techniques to produce annual seismic hazard regions. The prediction effectiveness of the candidate methods is tested rigorously and retains the ones with the best performance. Taking the prediction effectiveness of each of the methods as the weight, the seismic hazard regions are derived from their Bayesian probability (Wang et al. 2019). Results show that more than 80% of earthquakes of magnitude 5.0 and above in Mainland China in 2022 occurred in annual seismic hazard regions, including the 8 January M_s 6.9 Menyuan, Qinghai, and 5 September M_s 6.8 Luding, Sichuan earthquakes. Moreover, the CEA has also improved the strategies for short-term decision-making and achieved outstanding results. Before the Menyuan earthquake, the Qinghai earthquake agency of CEA submitted prediction reports consecutively to the people’s Government of Qinghai province on December 30 and 31, 2021, while before the Luding earthquake, the CEA held a series of earthquake tendency consultations, pointing out the urgency of an earthquake of magnitude 6-7 in the Sichuan-Yunnan region, especially after the 1 June M_s 6.1 Lushan, and 10 June M_s 6.0 Barkam, Sichuan earthquakes (Xue et al. 2022).

In this paper, we would like to display the latest achievements made by the CEA to show the feasibility of earthquake prediction. The proposed model is based on the routine earthquake prediction work of the CEA, but there are important innovations, including the probability prediction method of annual seismic hazard regions and the determination method of earthquake risk on the short-term scale. These are valuable for earthquake prediction studies in areas with dense seismic monitoring networks worldwide. The content is composed of 5 sections. In addition to the introduction section, the “Annual earthquake prediction” section shows the method for determination of annual seismic hazard regions and analysis of the corresponding results; the “Short-term earthquake potential evaluation” section presents the techniques for short-term earthquake tendency assessment and the anomalies before the Menyuan and Luding earthquakes, and in the “Discussion” and “Conclusions” we would like to show some discussions and conclusions.

2. Annual earthquake prediction

2.1. Methods

The annual earthquake prediction in China aims to determine seismic hazard regions for earthquakes of magnitude 5.0 and above in the next year. In 2022, significant improvement was made by the CEA in the determination methodology of annual seismic hazard regions, promoting the application of the probability synthesis technique. The main procedures can be divided into two steps:

The first step is to identify areas with the high-probabilities of earthquakes of magnitude 5.0 and above in the next year, including two parts:

(1) To prepare candidate methods to determine annual seismic hazard regions. Each discipline applies the morphological analysis methods to its associated data and identifies the anomalous annual variation, trend turnings, and marked increases (or drops) from the crustal deformation, seismometry, GNSS, groundwater, gravity, and electromagnetism observations, respectively. And the areas covered by the anomalies of each discipline are delineated as the predicted areas. The crustal deformation observations include the borehole strain, borehole inclination, horizontal pendulum, vertical pendulum, cavity strain, and so on, electromagnetism consists of the georesistivity, geoelectric field, geomagnetism, and so on, and groundwater involves the water level, water temperature, radon, ion concentration, and so on. In addition, many comprehensive calculation methods, such as the b-value (Gulia and Wiemer 2019), LURR (Yin et al. 2000), GPS strain rates (Wu et al. 2011), and so on, are also applied to test observation data and produce the seismic hazard regions.

Note that these works are usually carried out every year, and the anomalies that occurred during the year are adopted to predict earthquakes in the next year. The predictions made by each discipline or comprehensive calculation method are regarded as the candidate methods for determining annual seismic hazard areas.

(2) To select sub-method with the optimal prediction performance from candidate methods and integrate their predictions with the Bayesian formula. The prediction effectiveness of all the candidate methods is tested using earthquakes over the last five years. The ones with the best performance for predicting the M_s ≥ 5.0 earthquakes in the next year are selected as the sub-method for producing the seismic hazard regions. We use the R-score for the evaluation (Shi et al. 2000; Yu et al. 2022). The R-score is expressed as: (1) $$R=\frac{N_H}{N_T}-\frac{S_P}{S_T},$$(1) where, S_P and S_T denote respectively areas of the seismic hazard regions and whole Chinese Mainland. N_H and N_T are numbers of hit and all target earthquakes during the prediction period, respectively. The temporal window length for the annual prediction is one year.

This equation is similar to one of the “loss functions” associated with the Molchan error diagram (Molchan 1991), Specifically: (2) $$L=1-\tau -\nu ,$$(2) where, τ and ν represent the rate of missed and false alarms, respectively.

Generally, the range of the R-score is between −1 and 1, and R > 0 is an effective prediction. The higher the R-score, the better the method.

Taking the performance (i.e. R-score) of each method as the weight, the occurrence probability of earthquake (P) at any place can be calculated by: (3) $$P=1-\prod _{i=0}^N(1-c{R}_i),$$(3) where, N is the number of prediction methods used, R_i is the prediction performance of the i-th method, and c is defined as: (4) $$c=\frac{{1-(1-\overline{P_0})}^{1/N}}{\overline{R}},$$(4) where, $\overline{P_0}$ is the average probability, and $\overline{R}$ is the average R-score which can be defined by Equation (1)(1) $$R=\frac{N_H}{N_T}-\frac{S_P}{S_T},$$(1) .

The next step is to delineate the regions where earthquakes tend to occur. This is done by setting a threshold in the probability plot. Usually, the proportion of areas of the seismic hazard regions and whole Chinese Mainland should be considered. In addition, the faults that are at relative risk for strong earthquakes in the coming ten years in Mainland China (which have been provided by Shao et al. (2022) should also be included in the calculation.

2.2. Annual seismic hazard regions in 2022

In 2022, dozens of candidate methods were tested to produced annual seismic hazard regions in China. These methods have been studied previously (Bodri 2001; Kragh and Christie 2002; Yu et al. 2013; Gulia and Wiemer 2019; Wu et al. 2021). The predicted areas yielded by each of the methods are displayed in Figure 2. All of these methods were tested by using the R-score (Equation (1)(1) $$R=\frac{N_H}{N_T}-\frac{S_P}{S_T},$$(1) ) method to evaluate their predictive performance for earthquakes in the same year. For example, the predicted areas by the b value method in 2022 is shown in Figure 3. Also shown in the Figure are the earthquakes of magnitude 5.0 and above in 2022, which are adopted to evaluate prediction power of the b value method.

Figure 2. Predicted areas for the earthquakes of magnitude 5.0 and above in 2022 derived from all the candidate methods. The black lines indicate the national and provincial boundaries. The seismic hazard areas were displayed by the contours with different colors.

Figure 3. Seismic hazard regions made by the b value method and the earthquakes of magnitude 5.0 and above in Mainland China in 2022.

An earthquake falling within the predicted area is a correct prediction, and vice versa. Thus, the R-score in 2022 can be evaluated using Equation (1)(1) $$R=\frac{N_H}{N_T}-\frac{S_P}{S_T},$$(1) . We then calculated the average R-score for each method over the last 5 years, and selected six disciplines and six comprehensive calculation methods as the sub-method to produce annual seismic hazard regions. The predictive performance of these sub-method exceeds the average R-score of all the candidate methods shown in Figure 2.

We then applied the Bayesian formula (Equations 3(3) $$P=1-\prod _{i=0}^N(1-c{R}_i),$$(3) and 4(4) $$c=\frac{{1-(1-\overline{P_0})}^{1/N}}{\overline{R}},$$(4) ) for calculation of the probability distribution. The weights of each of the sub-method, which are the same as the R-scores obtained from the predictive performance tests over the last 5 years, are also displayed in Figure 4. The higher the R-score, the greater the weight of the sub-method. Figure 5 shows the annual seismic hazard regions for the M_s ≥ 5.0 earthquakes in Mainland China in 2022. The probability contours that account for about 20% of areas of the whole Chinese Mainland is indicated by the dashed lines.

Figure 4. The selected Sub-method for producing probability seismic hazard regions in Mainland China in 2022. The average R-score of each Sub-method over the last five years is represented by a histogram, in which the error bars show the fluctuation range of the R-scores in each year.

Figure 5. Annual seismic hazard regions in China in 2022 and the corresponding earthquakes of magnitude 5.0 and above in Mainland China. The dashed lines delineated the probability contours in which the proportion of areas of the seismic hazard regions and whole Chinese Mainland is 20%. earthquakes in 2022 but outside the annual seismic hazard regions are marked with dotted boxes.

It should be pointed out that our approach is just applicable in Mainland China because this approach is a combination of a series of sub-method for earthquake prediction. Some of the sub-method have been commonly adopted and validated by scientists worldwide, such as the PI and LURR methods (which have been adopted by the MMEP method in Figures 2 and 4). These type of methods are mainly based on the earthquake catalogs (Yin et al. 1995, 2001; Rundle et al. 2001, 2003). Nevertheless, there still are many sub-method that involve observation data of crustal deformation, groundwater, electromagnetism, and so on. These data are either not observed in other countries systematically or are less observed.

2.3. Results and comparison

In the annual predictions shown in Figure 5, it is clear that most earthquakes of magnitude 5.0 and above in 2022 occurred in areas with a high-level seismic hazard. There were 27 M_s ≥ 5.0 earthquakes on the Chinese Mainland, of which 22 occurred in the annual seismic hazard areas, including the 8 January M_s 6.9 Menyuan, Qinghai, 1 June M_s 6.1 Lushan, Sichuan, and 5 September M_s 6.8 Luding, Sichuan earthquakes. Just a few earthquakes occurred outside the seismic hazard areas. They are the 24 June M_s 5.1 Shanshan, Xinjiang, 19 October M_s 5.5 Mangya, Qinghai, and the 23 January M_s 5.8, 26 March M_s 6.0, and 15 April M_s 5.4 Delingha Qinghai earthquakes. In this case, the hit rate is roughly 0.81. Adopting Equation (1)(1) $$R=\frac{N_H}{N_T}-\frac{S_P}{S_T},$$(1) , we can get R = 0.61 (the ratio between the alarmed and whole areas is 0.20). This value is significantly higher than the average R-score (∼0.4) derived from previous years produced using the traditional approach (Yu et al. 2022).

Similarly, we have made the forward-looking prediction for the earthquake of magnitude 5.0 and above in 2023 in Mainland China, and the results were shown in Figure 6. Among the 11 earthquakes of magnitude 5.0 and above in 2023 on Chinese Mainland, 10 events occurred in the annual seismic hazard regions, except for the M_s6.1 Shaya, Xinjiang earthquake on January 30, 2023 (occurred in the Tarim Basin). The corresponding R-score reaches 0.7. Note that the 24 June Beibu Gulf and 23 Oct. Nan’ao, Guangdong earthquakes were not included in the statistics as they occurred in the sea area. Although the seismic activity in Chinese Mainland in 2023 is relatively lower, the prediction efficiency is better than that in previous years.

Figure 6. Annual seismic hazard regions in China in 2023 and the corresponding earthquakes of magnitude 5.0 and above in Mainland China.

Moreover, the annual seismic hazard regions are more effective in predicting earthquakes in a specific year. In Figure 5, we also show the 1249 M_s ≥ 5.0 earthquakes that occurred in Mainland China during 1970–2021, in which 801 earthquakes are located in the seismic hazard regions. The hit rate is about 64%. Since the seismic hazard regions of this study are specifically produced for predicting earthquakes in 2022, significant performance could be observed for predicting earthquakes in the year (∼81%). However, for testing earthquakes during 1970–2021, because of the random selection of the objects, the performance is therefore decreased (∼64%).

As a retrospective study, we also applied this probability prediction model to produce annual seismic hazard regions in Chinese Mainland in 2019 (Figure 7). In 2019, there were a total of 20 earthquakes of magnitude 5.0 and above in Mainland China, of which 4 did not occur within the seismic hazard regions, with an R-score of approximately 0.6. As a comparison, Figure 7 also shows earthquakes of magnitude 5.0 and above that occurred during 2019 to 2021. We noticed that just only 57.9% of earthquakes occurring in the seismic hazard regions, indicating a significant decrease in predictive efficiency of the approach.

Figure 7. Annual seismic hazard regions in China in 2019 and the corresponding earthquakes of magnitude 5.0 and above in Mainland China.

3. Short-term earthquake potential evaluation

3.1. Methods

The short-term earthquake potential assessment mainly aims at the time of earthquake that should occur in the annual seismic hazard regions. Specifically,

1. To seek anomalies from observation data. All the observation data, including seismic activity, crustal deformation, groundwater, electromagnetism, and so on, are tracked monthly using the morphological analysis and comprehensive calculation methods. The anomalies in each observation curve are identified using four methods: (1) the rate changes, (2) exceeding thresholds, (3) suddenly rise or drop, and (4) anomalous annual variation. The criteria for deeming an observation as an anomaly has been presented by CEA, and the anomalies are determined in each curve if its associated prediction indexes (critical value) are reached. Generally, this critical value is determined by the background fluctuations of the observation curve, and when the observed value is greater than twice the average background fluctuation, it is considered as an anomaly. When an anomaly occurs, the CEA will verify it within 24 h, removing the precipitation, barometric pressure, artificial activities, and other interference factors.

2. To establish indexes of each anomaly. The CEA counts the discernible anomalous changes in each curve calculated above and evaluates the correlation between the amplitudes and the time, location, and magnitude of ensuing earthquakes. This can also be assessed by the R-score, which is rewritten as:

(5) $$R_A=\frac{N_H}{N_T}-\frac{t_p}{t_T},$$(5) where, N_H and N_T are the numbers of detected and all target earthquakes within certain spatial and temporal scopes, and t_p and t_T denote the times of predicted and total periods, respectively. R_A changes with the amplitude setting. When R_A reaches the maximum, the corresponding amplitude of the anomaly and the time, space, and magnitude scales of earthquakes are defined as the prediction indexes. An observation with R_A exceeds 0.4 can be considered as a crucial anomaly. The larger the R_A, the higher the reliability of the anomaly.

• 3. To design strategies for short-term decision-making. Besides the anomalies and their associated indexes acquisition, the CEA should draw up strategies for evaluating short-term earthquake potential in a region. This can be implemented through statistics of the common characteristics of the anomalies before large earthquakes in the same neighborhood. Here, every anomaly is regarded as a tracked object. The prediction strategy usually contains the criteria as: 1. The earthquake should occur in the seismic hazard regions; 2. Existence of anomalies with the higher reliability; 3. The number of anomalies increases significantly.

Taking the Menyuan and Luding earthquakes as examples, we show the strategies for the short-term prediction of earthquakes in detail.

3.2. The Menyuan and Luding earthquakes

On 8 January 2022, the M_s 6.9 Menyuan earthquake occurred in the border areas between Gansu and Qinghai Provinces (Figure 8a). This is the largest earthquake in Mainland China in 2022. The surface fracture zone of this earthquake is composed of two parts: the western segment of the Lenglongling fault with NWW trending and the eastern end of the Tolaishan fault with near EW trending, and the focal depth is 10 kilometers. The strikes of these two faults are respectively 291° and 86.9°, with extension lengths of 26 and 3.5 kilometers. The seismogenic fault is characterized by sinistral strike-slip movement, and the maximum co-seismic dislocation is 2.77 m.

Figure 8. The mainshock and aftershock locations of the Menyuan and Luding earthquakes. (a): the Menyuan, Qinghai earthquake sequence. (b): the Luding, Sichuan earthquake sequence. Dark lines indicate active faults. LLL-F: Lenglongling fault, TLS-F: Tolaishan fault, CM-EB-F: Changma-Ebo fault, XSH-F: Xianshuihe fault.

The M_s 6.8 Luding earthquake occurred in Sichuan province on 5 September (Figure 8b), which killed more than 100 people. This earthquake occurred in the Moxi section of the Xianshuihe fault, which is dominated by the sinistral strike-slip movement along NW trending. The fracture length is 30–40 km, with a maximum fracture displacement of 1.2 m. The aftershocks are distributed along the NW and SE directions, with an extended length of about 60 km, and the fault dip is nearly vertical to the southwest.

As shown in Figure 5, both Menyuan and Luding earthquakes occurred in the annual seismic hazard regions in 2022.

3.3. Example of pre-seismic anomaly

3.3.1. Before the Menyuan earthquake

The short-term anomalies recorded by the near-field stations played a crucial role in predicting the Menyuan earthquake. Since 2021, the 22 May 2021 M_s 7.4 Maduo, 8 January 2022 M_s 6.9 Menyuan, and 26 March 2022 M_s 6.0 Delingha earthquakes have successively occurred in Qinghai Province. It is clearly that significantly pre-seismic changes could be observed before these earthquakes (Figure 9).

Figure 9. Time series of some crucial anomalies before the Menyuan earthquake, with the time of mainshocks in each map indicated by vertical arrows. (a): EW component of Menyuan borehole strain, (b): NS component of Gaotai borehole tilt, (c): Pingan groundwater level, (d): Xining gaseous radon, (e): Zuoshu groundwater level, (f): Gonghe groundwater temperature. The pre-seismic anomalies in each map are marked in red, and the numbers on the marked regions are the amplitude of the changes.

The Menyuan borehole strain decreased about 43 days before the Menyuan mainshock, while the significant enhancements in Gaotai borehole tilt, anomalous annual variation in Pingan groundwater level, and drops in Xining gaseous radon were observed months before the Menyuan and Delingha earthquakes. In addition, the Zuoshu groundwater level and Gonghe groundwater temperature dropped markedly before the Maduo, Menyuan, and Delingha earthquakes. It is clear that the amplitude of these anomalies far exceeds twice the background fluctuation amplitude of each observation curve.

We also note that except for these pre-seismic changes, few changes can be regarded as an anomaly (above the threshold) in each observation curve, indicating significantly precursory changes for ensuing large earthquakes. From Equation (5)(5) $$R_A=\frac{N_H}{N_T}-\frac{t_p}{t_T},$$(5) , the R_A values of the Menyuan borehole strain, Gaotai borehole tilt, Pingan groundwater level, Xining gaseous radon, Zuoshu groundwater level and Gonghe groundwater temperature are respectively 0.83, 0.75, 0.79, 0.72, 0.5, and 0.58.

3.3.2. Before the Luding earthquake

The most crucial anomaly before the Luding earthquake is the seismicity pattern which shows that a seismic gap of magnitude 4.0 and above occurred in the Sichuan-Yunnan area from February 2019 to January 2022 (Figure 10a).

Figure 10. The seismic gap of magnitude 4.0 and above before the luding earthquake and its evolution. (a): Seismic gap before the Luding earthquake, (b): seismic gap before the Luhuo earthquake, (c): TIme intervals between the M_s ≥ 4.0 earthquakes within the seismic gap and the large earthquakes in or around the seismic gap following the seismic quiescence of 500 days. (d): seismic quiescence area after the 30 September 1972 M_s 5.7 Kangding swarm, (e): magnitude-time plot of the M_L ≥ 4.0 earthquakes before the Luhuo earthquake in the seismic quiescence area of (d), (f): seismic quiescence area after the 10 June 2022 M_s 6.0 Barkam swarm, (g): magnitude-time plot of the M_L ≥ 4.0 earthquakes before the Luding earthquake in the seismic quiescence area of (f).

3.3.2.1. On the medium-to-long-term scale

The statistical results show a similar seismic gap in this region from June 1970 to April 1972. And the seismic gap was followed by the 6 February 1973 M_s 7.6 Luhuo, Sichuan earthquake (Figure 10b). Since 1970, six seismic quiescence of 500 days have been detected in this area, five of which were followed by the M_s ≥ 6.7 earthquakes within a time frame of several years (Figure 10c). From Equation (5)(5) $$R_A=\frac{N_H}{N_T}-\frac{t_p}{t_T},$$(5) , the R_A of this anomaly should be 0.42.

3.3.2.2. On the medium-term scale

On April 8, 1972, the M_s 5.2 Kangding, Sichuan earthquake broke the seismic quiescence in the seismic gap of 1970–1972, after that a series of earthquakes of magnitude 5.0–6.0 occurred at the edge of the seismic gap, including the 30 September 1972 M_s 5.7 Kangding swarm (Figure 10b). Similarly, after the seismic gap of 2019–2022 was broken by the 2 January 2022 Ninglang, Yunnan M_s 5.5 earthquake, the 1 June 2022 M_s 6.1 Lushan, Sichuan earthquake, and the 10 June 2022 M_s 6.0 Barkam, Sichuan swarm successively occurred on its edge (Figure 10a). The time intervals between the earthquake that broke the seismic quiescence (Kangding or Ninglang) and the ensuing mainshock (Luhuo or Luding) are about 10 and 9 months, respectively.

3.3.2.3. On the short-term scale

After the activity of earthquakes of magnitude 5–6 at the edge of the seismic gap, a seismic quiescence area occurred again in this region. The seismic quiescence in 1972 was after the Kangding swarm (Figure 10d), which lasted 4.0 months until the Luhuo mainshock (Figure 10e). The scope of the seismic quiescence area outperforms the seismic gap that appeared earlier. Similarly, the seismic quiescence in 2022 was between the Barkam swarm and Luding mainshock (Figure 10f) and lasted 2.8 months (Figure 10g). The durations of seismic quiescence before the mainshock are just several months in each case.

3.3. The number of pre-seismic anomalies

In addition to the crucial anomalies tracking, the CEA emphasizes changes in the number of observed anomalies to make short-term decisions. Figure 11a shows the number of short-term anomalies per month observed by the stations in Gansu-Qinghai region from January 2021 to May 2022. A short-term anomaly is defined as the changes recorded at an observation station within a time frame of six months before the mainshock, such as the significant enhancements in Figure 9b, anomalous annual variation in Figure 9c, and drops in Figure 9d. The observation stations are mainly divided into three disciplines, that is, the crustal deformation, groundwater, and electromagnetism. Note that the data presented in this paper are recorded before the earthquake rather than the retrospective tests. Figure 11b presents spatial distribution of the stations that detected anomalies at the peak value in December 2021 in Figure 11a and c is the statistics of epicentral distance of the stations in Figure 11b and occurrence time of the observed anomalies.

Figure 11. The temporal and spatial evolutions of the short-term anomalies before the Menyuan earthquake. The green, blue, and dark yellow lines represent respectively the number of anomalies recorded by the crustal deformation, groundwater, and electromagnetism stations, while the purple line denotes the number of anomalies observed by all these stations. (a): numbers of anomalies observed per month by the stations in Gansu and Qinghai regions from January 2021 to May 2022. The spatial distribution of stations that reported short-term anomalous changes at time of t_p are shown in (b). (c): variation of epicentral distance of the stations in (b) with occurrence time of the corresponding anomalies and its linear fitting.

Also shown in Figure 11 are the time and location of the Maduo, Menyuan, and Delingha earthquakes. Although the DLH Groundwater (marked by a dashed circle) anomalously changed before the Menyuan earthquake, it is more closely related to the M_s 6.0 Delingha earthquake on 26 March 2022. The distance to this mainshock, which is just 120 Km, is far less than that to the Menyuan epicenter (∼345 Km). The variation of amplitudes of observed anomalies in each curve in Figure 9 correlates well with the location of earthquakes (Figure 11). Since the Menyuan station is very close to the epicenter of the Menyuan earthquake (about 37 Km), pre-seismic changes were observed clearly before this Mainshock. Tectonically, the Gaotai, Pingan, Xining, and Zuoshu stations are all located in the Qilian block. This is the same as the Menyuan and Delingha earthquakes. Thus, significant anomalies were observed before these two earthquakes and were not observed before the Maduo earthquake that occurred on the Bayan Har block. Markedly pre-seismic change can be found if the observation station and the detection earthquake are on the same active tectonic block and hard to be found if they are on separate blocks. More interestingly, the Gonghe groundwater temperature displayed anomalous changes before all the Maduo, Menyuan, and Delingha earthquakes with less sensitivity, possibly because the station is located on the Qaidam block, adjacent to the blocks of these earthquakes.

Similarly, Figure 12 shows evolution of the number of anomalies observed by the stations in Sichuan province from January 2021 to October 2022 (Figure 12a), the distribution of stations that recorded short-term anomalies at the peak value in August 2022 (Figure 12b), and the relationship between the epicentral distance of anomalous stations and occurrence time of anomalies (Figure 12c).

Figure 12. The temporal and spatial evolutions of the short-term anomalies before the Luding earthquake. All the descriptions of (a), (b) and (c) are the same as the Figure 11. (a) except for the observation time period (from January 2021 to October 2022) and location of the mainshock (in Sichuan province).

Comparing the results shown in Figures 11 and 12, we found that:

1. Significant increases in the number of anomalies, within a short time frame, before earthquakes of magnitude 6.0 and above can clearly be observed. The time series in Figures 11a and 12a are at a low level most of the time and reach the peak value a few months before the mainshocks. The anomaly numbers peaked about 1, 2, and 1 months before the Maduo, Menyuan, and Delingha earthquakes and 3 and 1 months before the Lushan and Luding earthquakes.

2. The number of increased anomalies correlates well with the magnitude of subsequent earthquakes. The increased number before the Maduo, Menyuan, and Delingha earthquakes were 12, 7, and 4, respectively, while the seismic magnitude of the corresponding mainshocks were 7.4, 6.9, and 6.0. Similar increases in anomalies (6 and 12) can also be observed before the Lushan (M_s6.1) and Luding (M_s6.8) earthquakes. The results may support the notion that the more the increased anomalies, the greater the ensuing earthquake.

3. The pre-seismic anomalies tended to converge to the epicenter. Before the Menyuan and Luding earthquakes (Figures 11b and 12b), the anomaly stations appeared to be mainly distributed about 300-400 kilometers centered around the epicenters. However, this range is gradually decreasing over time. When approaches the earthquake, the epicentral distance of stations can be reduced to several kilometers (Figures 11c and 12c).

4. Discussions

In addition to the seismic activity, the monitoring capability of earthquake has significant impacts on the predictive performance of current model. For example, the 22 May 2021 M_s7.4 Maduo earthquake did not occur within the predicted seismic hazard regions (Figure 7). This earthquake is located in an area with weak monitoring capabilities (Figure 1). There are few monitoring stations near the epicenter. The situation of the 30 Jan. 2023 M_s6.1 Shaya, Xinjiang earthquake is also the same (see Figures 1 and 6). On the contrary, the 8 Jan. M_s6.9 Menyuan and 5 Sep. M_s6.8 Luding earthquakes occurred in the densely monitored areas (Figure 1), and captured by the annual seismic hazard regions (Figure 5). Therefore, the lack of sufficient earthquake-related monitoring data, which cause the differences between the annual seismic hazard regions in different years, limited the application of current model. Fortunately, the CEA has covered earthquake-related monitoring networks in the seismic active areas in Mainland China.

The annual seismic hazard regions in 2022 manifested good performance for earthquakes of magnitude 5.0 and above on Chinese Mainland, with R-score greater than 0.6. On the other hand, however, the uncertainties (nearly 40%) still present in the prediction, including the false positives and false negatives. In our practice, the alarmed area is fixed, that is, 20% of the area of Chinese Mainland. Therefore, we focus on the improvement of the hit rate of earthquakes to reduce false negatives. Among the undetected earthquakes in Figure 5, only one earthquake is far from the annual seismic hazard regions, and most of them occur nearby. To enhance the hit rate, we have been selecting high-quality prediction methods and integrating them through the Bayesian formula to obtain annual seismic hazard regions. The sub-method used for calculating probability distribution not only take into account the regional stress level, active tectonics, seismicity patterns, and anomalies in geophysical observations but also considers the movement of active tectonics and their associated stress state. More importantly, the prediction performance of each sub-method has been tested strictly by the R-score method, and provide quantified weights. The R-scores of annual predictions in 2019, 2022, and 2023 are 0.6, 0.61, and 0.7, respectively, showing a gradual increase. Some earthquakes that occurred in the areas with weaker seismic activity, such as the 18 Dec. 2023 M_s6.2 Jishishan, Gansu, and 6 August 2023 M_s5.5 Pingyuan, Shandong earthquakes (Figure 6), falling in the predicted annual seismic hazard regions.

China is attempting to join the Collaboratory for the Study of Earthquake Predictability (CSEP). With the support of the national key research and development project of China, we are establishing the CSEP test center in China. CSEP is the most outstanding earthquake prediction research program internationally. They have developed hundreds of statistical prediction models that can be used for earthquake prediction on multiple time scales (Schorlemmer et al. 2007). In addition, the “Operational earthquake prediction” developed after the 2009 M_s6.3 L'Aquila, Italy earthquake (Jordan et al. 2011) further explored earthquake hazard probability assessment and introduced the concept of “probability gain” (Marzocchi and Lombardi 2009; Zhuang 2011; Marzocchi et al. 2014; Herrmann et al. 2016). Through international cooperation in data sharing and method development, we should improve the predictive efficiency of annual seismic hazard regions.

Our probability prediction model has been deployed by the CEA and played an important role in the earthquake prediction work in China, currently. The prediction results for earthquakes of different magnitude ranges based on these probabilistic seismic hazard regions will be submitted to Chinese government to improve disaster preparedness and response approaches, monthly (short-term prediction) and annually (medium-term prediction). However, our model still has problems in the practice, for example, earthquakes often do not occur in areas with the highest probability of risk (Figures 5–7). In order to ensure consistency and improve the reliability across different seismic regions, we suggest strengthening the standardized management of monitoring data to ensure its continuity and stability. Meanwhile, we should make great efforts to develop high-quality prediction methods, and carry out standardized evaluation (e.g. the R-score method) of the predictive effectiveness for each of the methods. By clarifying the magnitude and scale of observed anomalous changes and their relationship with ensuing large earthquakes, we can establish criteria for pre-seismic anomaly determination, and thereby improve the accuracy of predictions.

Finally, we need to point out that earthquake prediction is a difficult task because the priory information should be provided before a future earthquake. This is even more serious for short-term earthquake potential evaluation due to the less precursory available. The sharply increased anomalies before a large earthquake imply the accelerated release of energy in the seismogenic area during the establishment of criticality, which can be explained using the rock constitutive relationship and its dynamic change (Wawersik and Brace 1971; Yin et al. 2000). With the increase of tectonic stress, the development of cracks lead to a decrease in the strength of the source rock, allowing us to observe the accelerated energy release before earthquakes (Bowman et al. 1998). During this process, new strain release points should be created, which may lead to the emergence of anomalies (Ma et al. 2012; Chen et al. 2021). With the evolution of stress field in the source media, strain release points may eventually concentrate near the epicenter. Thus, as the time of the mainshock approaches, the distance between the stations that detected anomalies and the epicenter decreased (Figures 11c and 12c). This results further suggest the correlation between the observed anomalies and mainshocks.

5. Conclusions

We show that China has made certainty progress in both annual and short-term earthquake predictions in 2022, suggesting the possibility of earthquake prediction using empirical methods. The improvements on current techniques can be summarized as: (1) To apply probability synthesis technique for annual seismic hazard areas determination, and (2) To establish prediction indexes and design appropriate strategies for short-term decision-making. This augments the quantification and standardization of earthquake prediction, and its associated studies are of great significance for disaster preparedness and response approaches. There is no denying that earthquake prediction is still one of the most difficult problems, but China, like seismologists worldwide, is making valuable contributions in this field. Their efforts indicate that earthquake prediction requires not only the innovation of methods but also objective and standard evaluation frameworks and reasonable prediction strategies.

Data and resources

The location of observation stations in Figure 1 was retrieved from China Earthquake Networks Center (CENC). The earthquake catalog used for producing Figures 2–12 was also from the CENC. The observation data before the Menyuan and Luding earthquakes shown in Figures 11 and 12 were retrieved from the regular prediction reports of CENC. All the data were archived and available in the Mendeley data repository (http://dx.doi.org/10.17632/scfxwdy795.1).

Author contributions

Huaizhong Yu., Rui Yan, and Jie Liu wrote the main manuscript text, Zhengyi Yuan prepared Figure 1, Shiguang Deng prepared Figures 2, 4 and 5, Mengyu Xie prepared Figure 3, Yawei Ma and Xiaotao Zhang prepared Figures 6 and 7, Gang Li, Yuchuan Ma and Zeping Li prepared Figures 8, 9 and 11, Yan Xue and Zeping Li prepared Figures 10 and 12. All authors reviewed this manuscript.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work has been supported by the Earthquake Joint Funds of NSFC (Grant No. U2039205) and the National Key Research and Development Project of China (Grant No. 2018YFE0109700).