Distinguishing mainshocks from foreshocks using spatiotemporal labeling and interpretable ensemble learning

Abstract

Rapid identification of whether a seismic event is a mainshock or a foreshock is crucial for minimizing public panic and enabling timely disaster response in earthquake-prone regions. However, this task remains difficult due to the physical similarity of seismic waveforms and the stochastic nature of earthquake sequences. This work presents a data-driven framework for mainshock classification based on an 88 -year USGS catalog, incorporating strict chronological separation and a bidirectional spatiotemporal labeling algorithm. Ensemble classifiers including XGBoost, LightGBM, CatBoost, and AdaBoost - were optimized using Bayesian methods. Physics-informed features such as local seismicity rate, cumulative energy release, and magnitude differential capture the evolving stress state preceding each event. Among the evaluated models, CatBoost demonstrated the strongest sensitivity to mainshocks, achieving an AUC of 0.965 and recall of 0.97 for safety-critical detection. SHAP interpretability further confirms that magnitude contrast relative to recent local seismic history is the dominant predictor of mainshock behavior. The proposed framework establishes a scalable foundation for real-time seismic role classification and supports the development of operational early-warning and disaster-response mechanisms.