A High-Resolution Aftershock Catalog for the 2014 Ms 6.5 Ludian (China) Earthquake Using Deep Learning Methods

Abstract

A high-resolution catalog for the 2014 Ms 6.5 Ludian aftershocks was constructed based on the deep learning phase-picking model (CERP) and seismic-phase association technology (PALM). A specific training strategy, which combines the advantages of the conventional short–long window average energy ratio algorithm (STA/LTA) and AI algorithms, is employed to retrain the CERP model. The P- and S-wave phases were accurately detected and picked on continuous seismic waveforms by the retained AI model. Hypoinverse and HypoDD were utilized for the precise location of 3286 events. Compared to the previous results, our new catalog exhibits superior performances in terms of location accuracy and the number of aftershock events, thereby enabling a more detailed depiction of the deep-seated tectonic features. According to the distribution of aftershocks, it can be inferred that (1) the seismogenic fault of the Ludian earthquake is the NW-trending Baogunao–Xiaohe Fault, (2) the Ludian aftershocks interconnected with the discontinuous NW-trending Baogunao–Xiaohe Fault, and they also intersected with the Zhaotong–Ludian Fault. (3) This suggests that the NE-trending Zhaotong–Ludian Fault may have been intersected by the NW-trending Baogunao–Xiaohe Fault, indicating that the Baogunao–Xiaohe Fault is likely a relatively young Neogene fault.

1. Introduction

The earthquake catalog is a crucial foundation for seismic research. A high-resolution regional catalog includes a large number of highly precisely located earthquakes. The greater the number of precisely located earthquake events, the more it helps to clearly reveal the physical structure of the underground media. This is crucial for various seismological research fields, such as seismic tomography, fault zone structures, stress states of the Earth’s interior, and seismic hazard assessments [1,2,3]. Generally, the process of constructing a seismic catalog involves multiple steps, including earthquake detection, seismic phase picking, phase association, and earthquake localization. Since the inception of seismology, seismologists have been dedicated to detecting more events from noise-rich continuous seismic waveforms, more precisely picking seismic phase arrivals, and more precisely locating earthquakes to construct high-resolution catalogs [4,5,6,7,8,9,10,11,12,13]. However, manual phase-picking methods are not only inefficient but also susceptible to picking errors. Moreover, traditional rule-based automated algorithms face challenges in balancing efficiency, accuracy, and completeness. For instance, a class of automated algorithms based on energy characteristic functions struggles to correctly identify seismic signals characterized by impulsive noise [13]. On the other hand, algorithms based on waveform similarity principles heavily rely on the diversity of prior template events, with low computational efficiency, making them less suitable for scenarios involving large datasets [11,12]. The key challenge encountered when using traditional automated algorithms to address seismic phase recognition issues lies in the inherent difficulty of mathematically describing these problems [14]. Deep learning technology allows computational models composed of multiple processing layers to learn mathematical representations with multiple abstraction levels. Its advantage lies in employing a data-driven learning approach to find solutions to problems [15]. This characteristic is particularly suitable for addressing phase-picking issues.

In recent years, the technology used to construct seismic catalogs based on deep learning has experienced rapid development. This includes not only deep learning-based techniques for earthquake detection [16,17,18], seismic phase-picking [18,19], phase association [20], and automatic earthquake location workflows [21,22], but also the gradual improvement of AI earthquake benchmark datasets [23,24]. Perol et al. [16] first trained a deep convolutional neural network (CNN) with eight hidden layers and used it to detect roughly located earthquakes in the Oklahoma area. Mousavi et al. [17] first designed a hybrid model integrating a CNN with a recurrent neural network (RNN). This model utilized a CNN to classify earthquake events and noise, while an RNN was employed for the phase-picking of P- and S-waves within the time windows of the seismic events. For regression-based solutions, Zhu and Beroza [19] designed a U-shaped deep neural network with four downsampling and four upsampling processes. This network (PhaseNet) treated the phase-picking of seismic waves as a probability distribution of P, S, and noise at each point within a time series window, achieving pixel-level identification in seismic waveform data for the first time. They trained their PhaseNet with over 700,000 earthquake events from California, enabling the high-precision picking of P and S wave arrivals across different instrument types. Mousavi et al. [18] first introduced the attention mechanism into the problem of phase detection in seismology and presented a global deep learning model (EQtransformer, Stanford, CA, USA) for simultaneous earthquake detection and phase-picking. These four AI models represent the four types of neural network structures commonly used in seismology. Compared to traditional methods, they all showed improvements in detection accuracy, completeness, and computational efficiency. However, they shared a common issue: significant variation in generalizability across different tectonic regions. In the field of AI earthquake datasets, Mousavi et al. [23] published the Stanford Earthquake Dataset (STEAD), the first publicly released benchmark dataset for training AI seismic models, which contains approximately 3 million global seismic and noise records. However, most seismic data in STEAD originated from Europe and the USA, with only a small portion originating from mainland China. Later, Zhao et al. [24] launched the DiTing dataset, which primarily features seismic benchmark AI data from mainland China, including about 780,000 earthquake events. Additionally, phase association technology was a crucial aspect following phase detection that impacted earthquake localization. Zhu et al. [20] developed a Bayesian-based hybrid model for this purpose, enhancing the accuracy and efficiency of seismic phase association. Furthermore, to enhance the efficiency of the earthquake location workflow, Zhang et al. [21] and Zhou et al. [22] developed comprehensive workflow frameworks called LOC-FLOW and PLAM (Phase picking, Association, Location and Matched Filter), respectively. Both frameworks offered flexible interfaces to overcome format barriers between multiple processing steps and employed traditional methods for earthquake association. The main difference between these frameworks was that LOC-FLOW used a grid search for association and localization, which had lower computational efficiency, while PALM achieved association based on temporal and spatial differences between different stations, resulting in improved computational efficiency.

Despite significant progress over the past years, data-driven deep learning technology still encounters big challenges when confronted with complex regional structures and observation conditions in the real world. Jiang et al. highlighted that, when the EQTransformer and PhaseNet were simultaneously utilized for detecting the aftershock sequences of the 2021 Yangbi earthquake and the 2021 Maduo earthquake, the catalogs produced by the two AI models exhibited significant differences. This inconsistency between the two catalogs indicated that the regional tectonic structures exerted a significant influence on the generalization of AI models.

In this study, we designed a novel training strategy by retraining the AI model on a small sample set to improve its generalization in specific regional tectonic structural scenarios. This retraining strategy combined the advantages of AI algorithms and the traditional short-long window algorithm (STA/LTA) [13]. We utilized the PALM method to obtain seismic/noise sample sets specific to the local region, which were then used to train a CNN model. Simultaneously, a manually picked P and S phases dataset was employed to train an RNN model. The CNN model was employed for detecting seismic events, while the RNN model was dedicated to picking seismic phases. Finally, the retrained AI model was applied to the 2014 Ludian aftershocks, resulting in a high-resolution catalog. Our results not only offer essential foundational data for the study of small seismic activities in the Ludian region but also provide a viable solution to the current challenges of generalization in AI phase detection.

2. Tectonic Background

The 2014 Ms 6.5 Ludian earthquake occurred along the southeastern margin of the Qinghai–Tibet Plateau (Figure 1). This region, experiencing the influence of compressional force between the Qinghai–Tibet Plateau and the South China Block, is known for its complex and diverse geological structures, marked by active tectonic features. Historically, this area has been the location of intense seismic activity in mainland China, with a recorded history of 17 earthquakes greater than Ms 6.0, the most notable being the Ms 7.1 Yongshan–Daguan earthquake in 1974 [25]. The 2014 Ludian earthquake was closely associated with several active faults, including the Xiaojiang Fault, the Lianfeng Fault, and the Zhaotong–Ludian Fault. This seismogenic fault zone is a part of the boundary separating the Sichuan–Yunnan block and the South China block, and also the transition zone from the actively deforming sub-block of the Dalianshan to the relatively stable South China block [26]. In addition, the region encompasses other important fault zones such as the Anninghe Fault, the Zemuhe Fault, the Daliangshan Fault, and the Mabian–Yanjin Fault. Many moderate earthquakes have occurred along these faults in the past decades, including the 2003 Ludian Ms 5.0 and Ms 5.2 earthquakes, the 2004 Ludian Ms 5.6 earthquake, the 2006 Yanjin Ms 5.1 earthquake, and the 2012 Yiliang Ms 5.7 and Ms 6.5 earthquakes. These seismic activities underscore the active characteristics of the region’s tectonics. He et al. [27] and Xu et al. [28] suggest that the Daliangshan and Mabian fault zones, previously disjointed, represent emerging neo-seismic tectonic belts. Currently, these zones are undergoing a phase of structural connectivity, with the precise location of this recent linkage yet to be determined. Situated at the southern end of the Daliangshan Fault’s tail, the Ludian region provides a critical vantage point for understanding the structural connectivity of this fault and for assessing related seismic hazards. The 2014 Ludian earthquake thus offers valuable insights into the dynamics of tectonic activities in this seismically active region.

3. Data and Methods

3.1. Seismic Data

In this research, we gathered comprehensive seismic data to construct a high-resolution aftershock catalog of the 2014 Ludian earthquake. The data collection involved continuous waveform and manually picked phase data from the Sichuan Seismic Network, the Yunnan Seismic Network, and the Qiaojia–Ludian seismic station network. The Qiaojia–Ludian network consisted of 12 permanent stations and 12 campaign stations, which were deployed near the Xiaojiang–Zhaotong–Ludian Fault Zone, allowing for a robust coverage of the 2014 Ludian earthquake. The campaign stations used Güralp CMG-3T seismometers (Güralp Systems Ltd., Aldermaston, UK) and Reftek 130S data loggers (Trimble Navigation Ltd., Sunnyvale, CA, USA), with a frequency range of 120 s–50 Hz. The permanent stations used Güralp CMG-3ESPC seismometers (Güralp Systems Ltd.) and EDAS-24I data loggers (Beijing Gangzhen Technology Development Co., Ltd., Beijing, China), with a frequency range of 60 s–50 Hz, and the seismometers are mostly installed on bedrock. The initial step in data handling was a rigorous assessment of data quality from all seismic stations. This involved evaluating the continuity of waveform data and excluding stations with consistently poor data quality. Then, we meticulously selected 12 campaign stations and 12 regional permanent stations to locate Ludian aftershocks, considering data quality and geographical significance. The selection criteria aimed to achieve an optimal balance between station quality and location to enhance aftershock location effectiveness. The observation period for these 24 seismic stations spanned from 1 August 2014 to 31 December 2015, with data continuity exceeding 90%. To verify the consistency of instrument responses, especially in campaign stations, we conducted comparative analyses using instrument responses to surface waves from global earthquakes greater than Ms 7.0. This comparison, which involved examining recordings from different stations for the same global seismic events, focused particularly on the response to long-distance seismic surface waves. This procedure confirmed the consistency of instrument responses across stations, which was crucial for the accurate determination of earthquake magnitudes.

The construction of the seismic AI training dataset was a crucial step in the development of a reliable and effective deep learning model for seismic event detection and phase picking. Our dataset was structured into three distinct sets: seismic event samples (positive), noise samples (negative), and seismic phase-picking samples. The CNN model was tailored to learn from the event and noise samples, while the RNN model was specifically trained using the seismic phase-picking samples.

For the generation of phase-picking samples, 55 regional permanent stations from the Sichuan Seismic Network were selected to create sample datasets for training the RNN model. These seismic records, encompassing continuous waveforms and P- and S-wave phase data, were collected from 1 January 2008 to 31 December 2012. The accuracy of the P- and S-wave phase arrivals was ascertained using the theoretical travel time versus epicentral distance relationship. We extracted earthquake events from 24 h continuous waveforms based on the arrival times of P- and S-wave phases. These events were then represented as 30 s seismic waveform segments, with the P- and S-wave arrival times annotated on each segment. To ensure the quality of the data, the signal-to-noise ratio (SNR) for all samples was calculated, and samples with exceptionally high or low SNR were excluded. Ultimately, this meticulous process yielded a dataset comprising 121,507 event samples annotated with manual phase labels, providing a rich and reliable foundation for the subsequent deep learning training and aftershock cataloging. For the generation of positive (seismic event) and negative (noise) samples, we employed a multifaceted approach integrating the STA/LTA, Kurtosis [29], and seismic association algorithms embedded within the PALM framework. This approach was applied to waveform data from the Ludian region to extract both event and noise samples efficiently. The Kurtosis algorithm was an automatic S-wave onset-picker, which used kurtosis-derived characteristic functions and eigenvalue decompositions on three-component seismic data. The Kurtosis algorithm outperformed the STA/LTA (short-term average/long-term average) algorithm in terms of accuracy and the number of picks, especially for S-wave. The detection of P and S arrival pairs for seismic events was accomplished using both the STA/LTA picker and the Kurtosis picker. Once all P and S arrival pairs were detected, the seismic association algorithm was employed to associate phase arrivals across all stations, thus forming distinct seismic events. These detected events constituted our event sample dataset. Noise samples were randomly extracted from 24 h continuous waveform recordings. A 30 s window was deemed a noise sample if it lacked any P or S phase arrivals. This method ensures the inclusion of diverse noise characteristics in the training dataset. In the end, we obtained 15,833 event samples. To augment the event samples, we added noise and randomly shifted time windows, resulting in a fivefold increase. A total of 75,327 noise samples were obtained through random selection. Since noise data were abundant and diverse, we did not perform data augmentation on the noise data to maintain its original complexity and diversity. After assembling all sample sets, we conducted a thorough analysis of various attributes within the dataset. This included assessing the distribution of signal-to-noise ratios, epicentral distances, azimuths, and the relationship between travel time and epicentral distance (see Figure 2). Such an analysis is essential to understand the dataset’s diversity and to ensure that the deep learning models are trained on a representative and comprehensive sample of seismic data. For a visual understanding of these distributions, refer to Figure 2 in the paper. This figure illustrates the range and variability of the aforementioned attributes within the dataset, offering insights into the data’s complexity and the challenges it poses for AI-based seismic analysis. This comprehensive preparation and analysis of the seismic AI training dataset lay the groundwork for the effective training of deep learning models, enabling them to accurately discern between seismic events and noise and to proficiently identify seismic phases, thereby contributing to the advancement of seismic research and aftershock analysis.

3.2. Methods

The comprehensive AI detection workflow is depicted in Figure 3. Unlike the procedural framework presented in Zhou et al. [30], the seismic phase data used in training the RNN model were manually picked by experts. In contrast, the sample dataset in Zhou et al. [30] was obtained through detection by the PAL-picker. The AI model in this study integrates a hybrid CNN and RNN structure, an approach previously developed by Zhou et al. [14]. Within this combined CNN and RNN model, the CNN deep neural network is composed of eight convolutional layers, Rectified Linear Unit (ReLU) non-linear activation functions, Max Pooling layers, and fully connected layers. The forward propagation procedure is defined by a loss function based on the L1 norm, while the backward propagation utilizes the Adam optimizer. The RNN features two bidirectional Gated Recurrent Unit (GRU) layers, which process data both forward and backward in time. In the RNN, each layer’s current state is influenced by both the input at the present time step and the hidden state from the preceding time step.

Phase association incorporates clustering analysis in both time and space domains [22]. Temporal association is achieved by searching for clusters of earthquake occurrence times. Spatial association is accomplished through grid searching for the hypocenter position with the minimum travel time residual. The phase association procedure ensures the detection of the same seismic signal by a minimum of four stations, thereby reducing the likelihood of misidentified signals. The magnitude estimation is determined by calculating the body wave magnitude using the S-wave amplitude. The earthquake localization utilizes Hypoinverse [7] for absolute location and HypoDD [10,31] for relative relocation.

3.3. AI Model Training

To improve the performance of the AI model under specific tectonic and observation conditions, we rapidly constructed a sample dataset for the local region and retrained the AI model. The methodology for curating the AI training dataset has been delineated in the preceding section. Before feeding samples into the AI model, data augmentation techniques were employed on the original dataset to enhance its robustness. This included temporal adjustments of the sampling window and the infusion of varying degrees of white noise. Specifically, the P-wave arrival served as the temporal anchor around which five random shifts within a 15 s interval were executed, each accompanied by the introduction of white noise ranging from 0 to 40%.

The hyperparameters, notably the learning rate and batch size, are pivotal in the AI training process, influencing model convergence, the risk of overfitting, and computational efficiency. An inordinately high learning rate can precipitate non-convergence, while an excessively low learning rate may unduly protract the convergence timeline. Similarly, a minuscule batch size could prove inadequate in counterbalancing the stochastic influence on gradient estimation, whereas an excessively large batch size could lead to protracted iteration durations. For the training executed in this study, we utilized hardware equipped with a GeForce RTX 3090 GPU, boasting 24 GB of memory (Nvidia Corporation, Santa Clara, CA, USA). A learning rate of 0.001 was established for both CNN and RNN training, with a batch size of 512. The training dataset was apportioned into training, validation, and test sets following a 7:2:1 ratio.

Detection accuracy was quantified as the proportion of correct predictions derived from the training dataset during the CNN training epoch, while validation accuracy was ascertained using the validation dataset. An observed increment in detection accuracy concomitant with a decrement in validation accuracy typically denotes the phenomenon of overfitting within the AI training regime. The 30 s sampling window was dissected into multiple time steps with a granularity of 0.5 s. The accuracy of these time steps, along with the validation rate, was defined as the likelihood of accurately predicting the P- and S-wave phases within these temporal increments during the RNN training phase. The picking uncertainty is characterized as the temporal precision in identifying the P- and S-wave arrivals within the training/validation datasets during the RNN training phase. The trajectories of detection/validation accuracy, picking uncertainty, and time step accuracy are graphically represented in Figure 4.

3.4. Earthquake Detection, Phase Picking, Association and Location

Utilizing the retrained CNN and RNN models on continuous waveform data, we conducted earthquake detection over a 30 s window with a 15 s sliding step, applying a 1–20 Hz bandpass filter to 24 h three-component waveforms. P- and S-wave phases were concurrently picked within this framework, and surface wave amplitudes were quantified within an amplitude window extending from 1 s pre P-wave to 5 s post S-wave. After obtaining the P- and S-wave phases for all stations, clustering of phase arrival times is achieved using a threshold of 2.0 s for grid search travel time residual, and a requirement of at least 4 stations simultaneously recording the same seismic signal. This process culminated in initial detections comprising 3624 seismic events and 25,125 P- and S-wave phase arrivals. A juxtaposition of the AI-determined arrival times with those manually picked revealed an average temporal discrepancy for P-waves of 0.02 s (standard deviation of 0.32 s), and for S-waves, an average discrepancy of 0.11 s (standard deviation of 0.44 s).

For absolute earthquake localization, the Hypoinverse software (Version 1.40) was harnessed. We adopted an average strategy for the initial velocity model, utilizing an averaged three-layer model of Fang et al. (Figure 5) [25]. Throughout the iterative inversion process, station weights were modulated based on the root mean square of the travel time residuals and epicentral distance. A residual cutoff threshold was established, affording full weight to stations with residuals under 0.3 s, nullifying weights for residuals exceeding three times the cutoff residual (0.9 s), and implementing weighted interpolation for intermediate values according to a cosine function curve. A distance cutoff was set at 40 km, with a cutoff range spanning 40–120 km.

Subsequent to the absolute localization, relative localization was performed employing the HypoDD algorithm (Version 1.3). The double-difference method incorporates an initial relative location derived from travel time measurements, further refined by cross-correlation to correct temporal disparities, thereby augmenting the precision of the relative locations. The parameters included a maximum station-event distance of 150 km, an event-pair distance constraint of 6 km, and a minimum of 8 phases per event pair. After two cycles and eight iterations, the inversion parameters, inclusive of travel time residuals and horizontal and vertical discrepancies, were stabilized. During the cross-correlation for travel time difference calculations (cc), the maximum distance between events pairs was set to 4 km, and the maximum epicentral distance for stations was 120 km. Template windows are defined as a P-wave before 0.5 s and after 3.5 s, and S-wave before 0.3 s and after 4.5 s, using a 2–15 Hz bandpass filter. The velocity model is the same as the one used by Fang et al. (Figure 5). Finally, we obtained high-precision location results for 3286 events.

The Frequency–Magnitude Distribution (FMD) between the AI catalog of this study and those compiled by Fang et al. and the China Earthquake Network Center (CENC) was contrasted (Figure 6). This comparative analysis indicated that the AI-generated catalog exhibits superior detection capabilities relative to the catalogs by Fang et al. [25] and CENC.

4. Results

4.1. Aftershocks Space Distribution, Temporal Evolution and Focal Mechanism

The analysis of the aftershock sequence following the 2014 Ludian Ms 6.5 earthquake elucidates a distinct L-shaped conjugate distribution, suggesting a compound fracture orientation primarily along the east–west (E–W) and northwest–southeast (NW–SE) axes. The E–W oriented section spans approximately 23 km in length and 5 km in breadth, while the NW–SE extension measures about 18 km in length and similarly 5 km across. In terms of depth, the aftershocks predominantly clustered within a stratum extending from 5 to 15 km beneath the surface. The intersection of the conjugate faults marks the zone of maximal hypocentral depth, where the concentration of aftershocks is notably dense. Moving laterally from this central intersection, there is a discernible gradation towards more superficial seismic events (Figure 7).

The aftershocks also exhibited distinct spatiotemporal variation characteristics. Initially, within the first 2 to 5 h post-main shock, the aftershocks predominantly aligned in a northwest–southeast (NW–SE) orientation, forming a strip-like distribution. This pattern underwent a notable shift approximately 5 h later, with the emergence of a northeast-southwest (NE–SW) directional trend in the aftershock distribution. This NE–SW orientation became increasingly pronounced over the subsequent 24 h. Remarkably, after five days, the aftershock sequence evolved to display an asymmetric conjugate distribution, further illustrating the dynamic nature of the seismic event’s aftershock activity.

The focal mechanism solutions for the two nodal planes, as published by the Global Centroid Moment Tensor (GCMT) [32,33], were consistent with the dominant orientations of the L-shaped conjugate distribution of aftershocks. On the other hand, this spatiotemporal patterning of the Ludian aftershock sequence concurred with the observations presented in the works of Fang et al. [25] and Wang et al. [34]. The consistency of these findings across independent studies lends credence to the interpretation of the tectonic behavior in the aftermath of the Ludian earthquake and reinforces the understanding of seismic dynamics in conjugate fault systems.

4.2. Seismic Rate Evolution

The seismic rate, defined as the frequency of earthquakes occurring per hour, peaked within the initial day following the main shock, and thereafter exhibited a general diminishing trend. This decay in seismic rate over time is a typical characteristic of aftershock sequence, as postulated by established seismic laws, such as Omori’s law for aftershock temporal decay. The observed magnitude-time relationship and seismic rate for the aftershocks also presented conspicuous diurnal patterns. These patterns were discerned to be linked to the diurnal variations in the ambient noise level, which in turn affected the earthquake detection capability of the monitoring system. During periods of heightened daily noise—such as human activity during daytime hours—the ability to detect smaller seismic events is often compromised. Conversely, the relative quietude of nighttime generally corresponds to higher detection rates of smaller aftershocks.

Furthermore, the seismic detection capability is also temporally affected by the ‘tail waves’ of larger seismic events. These trailing seismic waves generate a transient increase in the noise floor, which can substantially diminish the detection efficiency for smaller magnitude earthquakes in the period following significant aftershocks. The impact of these larger events on detection capability is visually represented in Figure 8, which likely includes a time series plot showing the variation in seismic rate alongside the occurrence of larger aftershocks. Such fluctuations in detection capability necessitate careful consideration when analyzing seismicity rates and the corresponding magnitude distributions. These variations underscore the importance of incorporating noise level assessments and potential detection biases into the seismic analysis to ensure the accurate interpretation of seismicity patterns and the underlying physical processes driving the aftershock sequence.

5. Discussion

5.1. The 2014 Ludian Earthquake’s Seismogenic Fault and Its Tectonic Implications

Within the vicinity of the epicenter, the NE-oriented Zhaotong–Ludian fault is primarily recognized for its thrust faulting. However, the focal mechanism of strike–slip and the spatial-temporal pattern of aftershocks, which align predominantly along NW–SE and near E–W directions, imply that the Zhaotong–Ludian fault may not be the only seismogenic fault. The presence of the NW-oriented Baogunao–Xiaohe fault, noted for its left-lateral strike–slip movement, provides an alternative tectonic feature that corresponds more closely with the aftershock distribution and focal mechanism. Particularly within the first 5 h following the main shock, the predominance of aftershocks in the NW–SE direction, coupled with their depth characteristics consistent with a strike–slip fault, suggested that the Baogunao–Xiaohe Fault may have been the seismogenic fault for the Ludian earthquake. Synthesizing the aftershocks’ spatial-temporal pattern, the focal mechanism, and the intrinsic characteristics of the regional faults, we inferred that the seismogenic fault responsible for the Ludian earthquake is the NW-oriented Baogunao–Xiaohe fault.

The distribution of the NW-trending Ludian aftershocks, intersecting the NE-trending Zhaotong–Ludian fault, suggests that the former may be transecting the latter. This intersection could imply that the Zhaotong–Ludian fault, despite its longstanding geological presence, has been intersected and possibly offset by the younger Baogunao–Xiaohe fault, a tectonic feature that may have developed during the Cenozoic era [25]. Furthermore, geological structural maps indicate that the Baogunao–Xiaohe Fault is a minor and less distinct fault located north of the Zhaotong–Ludian Fault. On the south side of the Zhaotong–Ludian Fault lies another small fault, with both minor faults being separated by the Zhaotong–Ludian Fault. Based on these, we assumed that the 2014 Ludian earthquake may have interconnected the Baogunao–Xiaohe Fault across the north and south sides of the Zhaotong–Ludian Fault.

5.2. The Future Application of This Retraining Strategy

Previous researchers have attempted to train high-generalization AI models using large-scale global datasets to address phase-picking challenges [18,19,35]. However, these efforts have fallen short of achieving the desired results in complex real-world scenarios. This suggests that the approach of relying on massive global datasets to solve the generalization problem may not be feasible. In contrast, this study explores the use of small, easily obtainable datasets to retrain AI models for specific scenarios, aiming to improve model applicability in those specific contexts. We employed both traditional automated algorithms, STA/LTA and Kurtosis, for P and S phase detection, applying multi-station constraints to eliminate erroneous data. Data augmentation techniques were applied to expand the dataset for training precise AI models. Therefore, this re-training strategy is highly adaptable and flexible, theoretically applicable to different earthquake events and tectonic regions. Our successful results further demonstrated that this AI re-training strategy can generate a high-performance AI model suitable for specific scenarios. However, it relies on a well-distributed network. The well-distributed network influences the accuracy of the phase association process and further affects the generation of high-accuracy event samples.

6. Conclusions

A high-resolution catalog for the 2014 Ms 6.5 Ludian aftershocks is constructed based on an AI-picker model. During the AI model training process, we designed a specific training strategy that combines the advantages of the STA/LTA and AI algorithms. Our successful retraining and detection results indicate that this training strategy for building a sample set in a specific tectonic region to retrain the AI model can improve the generalization performance of the AI model in the specific region. Compared to the previous results from the Fang et al. [25] and China Earthquake Networks Center (CENC), our result exhibits superior performance in location accuracy and the number of aftershock events. According to the accurate distribution of aftershocks, we conclude that (1) the seismogenic fault of the Ludian earthquake is the NW-trending Baogunao–Xiaohe Fault; (2) the Ludian aftershocks interconnected with the discontinuous NW-trending Baogunao–Xiaohe Fault, and intersected with the Zhaotong–Ludian Fault; and (3) this suggests that the NE-trending Zhaotong–Ludian Fault may have been intersected by the NW-trending Baogunao–Xiaohe Fault, indicating that the Baogunao–Xiaohe Fault is likely a relatively young Neogene fault.