SEISDETNET: ARTIFICIAL NEURAL NETWORK FOR SEISMIC EVENT DETECTION. PART 2: MODEL ASSESSMENT

Based on a combination of convolutional and fully connected neural networks, we developed the SeisDetNet model to distinguish seismic events from seismic noise using waveform records. We also proposed an algorithm for detecting seismic events in continuous records of arbitrary duration. It processes waveforms in a sliding window, feeding successive 1-minute segments into SeisDetNet and producing a time series of seismic event probabilities. Extending this approach to analyze records from multiple seismic stations demonstrated promising results. Specifically, the model successfully detected all events in the KNET seismic catalog of the Research Station of the RAS for January−March 2024, including low-energy earthquakes (K ≤ 7). Furthermore, testing the model for detecting aftershocks of the strong Uchturpan earthquake with Mw=7 that occurred on January 22, 2024, at 18:09 UTC near the China-Kyrgyzstan border, showed results comparable to manual processing.