Mapping and interpretability of aftershock hazards using hybrid machine learning algorithms

This study addresses gaps in aftershock prediction research by proposing an interpretable hybrid machine learning model that leverages multi-source data. The model overcomes challenges related to the selection of influencing factors, model types, prediction result visualization, and decision mechanism interpretability. It integrates mainshock factors, geological features, site characteristics, and terrain conditions using geospatial information system (GIS) technology. By employing the stacking algorithm to optimize and combine XGBoost and LightGBM models, the proposed model significantly improves the prediction performance. Visualization through aftershock hazard mapping offers a robust tool for aftershock warning. The Shapley additive explanations (SHAP) model is used to explain the decision-making process from both global and local perspectives. Results show that, compared to the optimized XGBoost-CMA_ES and LightGBM-CMA_ES hybrid models, the stacking model achieves area under the curve (AUC) increases of 7.71% and 5.72% on the test set, respectively, with a maximum prediction accuracy of 0.9344. The hazard zoning map identifies high-risk areas mainly around fault lines and near the epicenter. As hazard levels rise, the proportion and density of aftershocks in these areas increase. The SHAP model results highlight the distance to fault as the most critical factor. The study integrates local explanations with on-site investigations, effectively visualizing the contributions of different factors to aftershocks. This research provides new tools and methods for enhancing aftershock warning and response.