An Interpretable Stacking Ensemble Framework for Defining Rainfall–Geoenvironmental Coupling Thresholds of Shallow Landslides

Abstract

Rainfall-induced landslides exhibit marked spatio-temporal heterogeneity and multi-factor coupling. Existing susceptibility methods struggle to balance predictive accuracy with interpretability in nonlinear settings. This study addresses this gap by developing a Stacking ensemble framework integrated with SHAP analysis, using the July 2019 extreme rainfall-triggered landslides in northern Fujian and eastern Jiangxi as a case study. The framework builds an indicator system from multi-source data (topographic, geological, hydrological, and rainfall factors), screens key factors via correlation analysis, and constructs a Stacking model with Random Forest, XGBoost, and LightGBM as base learners and Ridge Regression as the meta-learner. The Stacking model outperforms individual models on the test set (AUC = 0.8935, F1 = 0.800) and captures 90.9% of known landslide points within high and very high susceptibility zones. SHAP-based interpretability identifies rainfall as the dominant control, with soil thickness, slope gradient, and lithology as secondary factors. Each factor exhibits nonlinear marginal effects and condition-dependent behavior, with significant rainfall–terrain and rainfall–geotechnical interactions. The proposed framework not only improves prediction stability but also provides critical interpretability and decision-making support for landslide risk identification under extreme rainfall conditions.