Abstract
This study presents an efficient method for the automatic detection of primary (P) seismic waves, based on time-frequency analysis and spectrogram classification using the YOLO11x-cls deep learning model. Unlike traditional detection algorithms that rely on handcrafted feature extraction and domain-specific heuristics, the proposed approach directly processes visualized seismic signals generated via the continuous Morlet wavelet transform. The main output of the method is the precise identification of P-wave arrival times, which are critical for earthquake early warning systems.To train the model, a balanced dataset comprising 1700 labeled spectrogram images was prepared using seismic data from a single station, TLG, located near the city of Almaty. To evaluate the generalization capability of the model, testing was conducted on datasets acquired from three independent seismic stations situated within approximately 333 km of Almaty (two in Kyrgyzstan and one in Kazakhstan), as well as on recordings of global high-magnitude earthquakes (M ≥ 7.0) that occurred worldwide during the period from January 1, 2023 to December 31, 2024. The experimental results demonstrated the high effectiveness of the proposed method. On the independent seismic stations TARG and JNKS (Kyrgyzstan), as well as SHLS (Kazakhstan), the model achieved F1-scores ranging from 91% to 95%. In the combined evaluation across all three stations, the model yielded a total of 564 true positives, 40 false positives, and 35 false negatives. The aggregated performance metrics were as follows: precision – 93.4%, recall – 94.2%, and F1-score – 93.8%, confirming the model’s high accuracy and robustness in detecting P-waves on datasets that differ from the training distribution. Additionally, on recordings of global high-magnitude earthquakes, the model achieved an overall accuracy of 93.5%, further highlighting its reliability and applicability under conditions of worldwide seismic activity.