Classification of Seismic Events in the Mainland of China Based on Spectrograms and Model Interpretability

Abstract Accurate identification of seismic event types is crucial for seismic monitoring, early warning, and disaster prevention. Traditional classification methods relying on manual features, while deep learning approaches improve automation, still face challenges in practical application due to limited interpretability. This study collected 99,600 three‐component waveform records from 2,870 events including natural earthquakes, explosions, and collapses from China Digital Seismograph Network (2013–2024). We propose ResWaveQuake, a multi‐branch ResNet‐based model for single‐component classification using logarithmic time‐frequency spectrograms, incorporating convolutional wavelet transform, coordinate attention, and agent attention mechanisms. ResWaveQuake achieves 96.52% classification accuracy on the test set, employing a network‐based voting mechanism. To enhance decision transparency, a progressive interpretability framework is employed, combining Grad‐CAM++, Integrated Gradients, and occlusion testing to examine feature attribution patterns across different event types, source depths, and epicentral distances. The analysis reveals that earthquakes focus on high‐frequency body waves, while explosions and collapses emphasize S‐wave coda, with stable event‐specific patterns across distances and depths. These findings indicate that the model captures seismic propagation characteristics that are generally consistent with physical observations, offering insights into the signal features underlying automated seismic event classification.