Spatiotemporal Graph Convolutional Networks for Earthquake Source Characterization.

Area of Science:

• Seismology
• Geophysics
• Artificial Intelligence

Background:

• Accurate earthquake location and magnitude are crucial in seismology.
• Deep learning shows promise in seismological tasks, but many methods use single-station data.
• Multiple seismic stations offer more comprehensive information for source characterization.

Purpose of the Study:

• To develop a Spatiotemporal Graph Neural Network (STGNN) for improved earthquake location and magnitude estimation.
• To leverage geographical and waveform data from multiple stations for enhanced earthquake analysis.
• To compare STGNN performance against existing deep learning models.

Main Methods:

• Developed a Spatiotemporal Graph Neural Network (STGNN) that uses multi-station seismic data.
• Constructed dynamic graphs using geographical and waveform information with adaptive message passing.
• Applied STGNN to earthquake data from the Southern California Seismic Network and Oklahoma.

Main Results:

• STGNN achieved more accurate earthquake epicenter locations compared to baseline models.
• Depth and magnitude prediction performance was comparable to baselines, indicating a general challenge for all tested models.
• The study demonstrated the effectiveness of GNNs in analyzing multi-station seismic data.

Conclusions:

• Spatiotemporal Graph Neural Networks show significant potential for improving automatic earthquake epicenter estimation.
• Utilizing data from multiple seismic stations within a GNN framework enhances earthquake source characterization.
• Further research is needed to improve depth and magnitude prediction accuracy in seismic event analysis.