Abstract:

Rapid and accurate earthquake magnitude estimation is crucial for post-earthquake emergency response and earthquake disaster prevention and mitigation. High-rate global navigation satellite system (HR-GNSS) observations can rapidly capture near-field coseismic deformation and provide important data for magnitude estimation of large earthquakes. Existing methods, however, still face magnitude saturation, low signal-to-noise ratios, uneven station distributions, and insufficient use of source-station geometry in deep-learning models. To address these problems, this study proposes an epicentral-distance geometry-assisted station-wise attention convolutional neural network (EDGSA-CNN) model. The model uses a convolutional neural network (CNN) to extract three-component displacement waveform features from multiple stations. It employs a self-attention mechanism to model feature correlations among stations and a multilayer perceptron (MLP) to nonlinearly encode epicentral distances. The waveform features and geometric auxiliary information are then fused to estimate the moment magnitude of large earthquakes. Experiments using synthetic data from the Chilean subduction zone show that EDGSA-CNN generally yields lower magnitude-estimation errors than the waveform-only CNN baseline and peak ground displacement (PGD) empirical model, while maintaining relatively low errors for high-magnitude samples. Epicentral-distance perturbation tests indicate that the model is robust to early epicentral-location errors to some extent. Cross-regional synthetic tests and tests using observations from real large earthquakes further indicate that the model exhibits a certain degree of applicability and stability across different data sources and observational conditions. This study can provide potential technical support for accurate magnitude estimation of large earthquakes and earthquake early warning applications.

Key words: high-rate GNSS, deep learning, attention mechanism, epicentral distance, magnitude estimation