Identifying Seismogenic Anomalies in Satellite Microwave Emissions by Decoupling Environmental Confounders via Physics-Informed Deep Learning

Abstract

Identifying reliable microwave emission anomalies related to impending earthquakes (EQs) remains challenging due to their intricate coupling with surface environmental dynamics. While previous studies rely heavily on historical baselines to remove background emissions, they largely overlook the terrestrial confounding factors inherent within the seismic influence period, compromising the reliability of obtained anomalies. To bridge this gap, we present a novel microwave-physics-informed deep learning (DL) framework that captures the spatiotemporal dependencies of satellite observations across both historical baselines and concurrent influence periods. Distinct from purely data-driven approaches, we embed explicit microwave physical constraints—namely, dual-polarization consistency and gradient monotonicity—directly into the loss function. Validated across multiple EQ cases, our model effectively decouples overwhelming meteorological and surface interferences (e.g., soil moisture variations, precipitation) from the microwave observations, isolating residuals highly associated with targeted EQs. Comparative analysis demonstrates the framework’s superiority over traditional techniques, achieving an average 22% increase in contrast-to-noise ratio (CNR) and a 260% enhancement in average anomaly intensity. Ablation studies further confirm that microwave physical constraints significantly enhance reconstruction fidelity, improving the correlation coefficient by 3% and reducing root-mean-square error (RMSE) by 36%. Crucially, validation against ancillary datasets reveals that our framework can account for nonseismic drivers. Consequently, the derived residuals are largely immune to these confounders and inherently seismogenic, substantially reducing the need for post-hoc exclusion analyses. This work provides a physically interpretable solution for identifying reliable seismic precursors using multisource satellite remote sensing, bearing profound implications for the future development of an operational, precursor-based EQ forecasting framework.