关键词: 定日地震, 发震构造, 深度学习检测, 前震序列, b值

Abstract:

Using seismic data from the Tibet Regional Seismic Network between January 2021 and January 2025, we relocated 7, 951 earthquakes employing the double-difference algorithm. To ensure relocation reliability, we selected phase data with minimal travel-time residuals, constraining earthquake pairs to a maximum separation of 30km and requiring at least eight common phase arrivals per pair. This yielded 4, 370 high-precision relocations, with average relative errors of 0.130km(longitude), 0.131km(latitude), and 0.199km(depth). The relocated mainshock location is(28.501°N, 87.477°E)with a focal depth of 9.3km. The aftershock sequence extends approximately 70km in a nearly north-south direction, with depths mainly concentrated between 3 and 15km.

We conducted a detailed analysis of the foreshock activity preceding the M_S6.8 Dingri earthquake. Due to sparse station coverage near the epicenter, traditional seismic monitoring methods were insufficient for detecting small events. To address this, we applied the deep learning-based PhaseNet model to continuous waveform data from the nearest station(ZHF, ~50km from the epicenter), in combination with a single-station amplitude-magnitude empirical relationship for magnitude estimation. This approach significantly improved catalog completeness. Within the 56-hour window prior to the mainshock, we identified 90 seismic events, of which 80(88.9%)were microearthquakes with magnitudes M_L<2.0. In contrast, the regional network recorded only 8 events in the same period. A reliable single-station magnitude calibration was established(log₁₀A=0.77M_L+1.36) using 293 aftershocks. For commonly detected events with M_L<2.0, the average magnitude difference between the single-station and regional network methods was just 0.09, confirming the accuracy of the single-station approach. Based on the enhanced catalog and using the maximum curvature method accounting for magnitude uncertainty, the completeness magnitude was determined to be M_L1.10. The b-value, estimated via the maximum likelihood method, was 0.58±0.07. Relocation results show that foreshocks were spatially clustered within the eventual aftershock zone, approximately 20km from the mainshock epicenter. Eight foreshocks occurred within the final hour before the mainshock, with the largest(M_L3.7)occurring approximately one hour prior.

These findings demonstrate that deep learning-based, single-station detection methods can substantially enhance earthquake monitoring in regions with sparse seismic networks. The spatial and temporal characteristics of the foreshock sequence offer critical insights into earthquake preparation processes. The single-station magnitude estimation method presented here provides a valuable reference for seismic monitoring in similarly data-limited regions.

Key words: Dingri earthquake, seismogenic structure, deep learning detection, foreshock sequence, b value