Reference station-based transfer learning for earthquake anomaly extraction from borehole strain data related to two earthquakes in China

Earthquake precursor extraction from geophysical observation data remained challenging due to diverse signal characteristics and complex earthquake processes. This study proposed STMN-EQA (spatiotemporal multi-scale network for earthquake anomaly extraction), integrating segmented variational mode decomposition (SVMD), TimesNet temporal modeling, and graph neural networks (GNN) for multi-station spatial analysis. We trained the model exclusively on high-quality data from reference stations in seismically inactive regions to learn normal strain patterns and then directly transferred the pre-trained model to analyze borehole strain data for two earthquakes in China. Our temporal analysis of anomalies revealed consistent sigmoidal anomaly accumulation process 1–3 months before both events and spatial analysis of anomalies revealed that inflection points and anomaly counts at monitoring stations are correlated with epicentral distance, which may be related to earthquake processes. The framework achieved superior performance compared to conventional anomaly detection methods. Furthermore, the reference station-based transfer learning strategy combined with multi-domain analysis integrating temporal, spatial, and frequency information offered a novel solution for earthquake anomaly extraction while advancing our understanding of multi-scale earthquake preparation mechanisms.