Spatio-temporal deep networks with feature disentangling for advancing earthquake monitoring

Earthquake monitoring is essential for assessing seismic hazards and involves interconnected tasks such as phase picking and location estimation. Existing single-parameter estimation methods suffer from error accumulation caused by task interdependencies and typically rely on empirical values. Multi-parameter estimation methods often depend on data from multiple stations, posing challenges in modeling and revealing the inter-station relationships. To address these challenges, this study proposes a novel neural network, SINE, designed to simultaneously estimate key parameters in earthquake monitoring, including P-phase arrival time, location, and magnitude. SINE develops a multi-task framework that incorporates Graph Neural Networks (GNNs) and Bidirectional Long Short-Term Memory networks (BI-LSTM) to extract spatio-temporal features, effectively mitigating error accumulation across the tasks. Unlike previous GNN-based models, SINE incorporates a feature disentanglement structure to automatically identify multiple potential relationships between seismic stations. Additionally, the CNN-based parsing unit is employed to regress multiple seismic parameters simultaneously. Evaluation on datasets from Southern California and Italy shows that SINE outperforms existing DL models and traditional seismological methods. Furthermore, SINE effectively reduces inter-task dependencies, enhancing robustness in earthquake monitoring.