Seismic event classification in north China based on machine learning

Automated classification of seismic events is critical for earthquake monitoring and explosion detection, particularly in tectonically active regions, such as North China, where the waveform features of earthquakes and explosions are highly similar. This study compared feature-based machine learning (ML) and image-based deep learning (DL) methods in event- and station-level classification frameworks. The dataset consisted of 1,847 events and more than 43,000 vertical-component waveforms with two input types, 40-dimensional feature vectors for ML and spectrogram images for DL. The results showed that the event-level models consistently outperformed the station-level models, achieving over 98% accuracy; the station-level models performed well above 94%. On the test set, the ML and DL models exhibited comparable performance; however, the ML models demonstrated better generalization and lower computational demands. In contrast, DL models required fewer manual interventions. The misclassification analysis revealed distinct error patterns across the model types, indicating potential complementarity. These findings highlight the importance of model choice based on the input type, data granularity, and generalization needs. Although DL models are well suited to automated processing, ML approaches provide more robust and efficient solutions for real-world deployment.