Dynamic Landslide Susceptibility Assessment Using Machine Learning Models

Abstract. Landslide susceptibility assessments have traditionally used static rainfall statistics that do not reflect the actual meteorological conditions when slopes fail. This study develops a machine learning framework that aligns high-resolution radar rainfall (XRAIN, 250 m / 10 min) and modeled soil moisture (XSWI) with documented landslide occurrence times as dynamic triggering factors. Applied to the Heavy Rain Event of July 2018 in Hiroshima Prefecture, the framework combines watershed-based spatial cross-validation, systematic comparison of four class imbalance strategies (no treatment, sample weighting, random under-sampling, and SMOTE) across eight algorithms (XGBoost, LightGBM, CatBoost, HGBoost, Random Forest, Balanced Random Forest, Easy Ensemble Classifier, and Logistic Regression), and spatially explicit SHAP interpretation. Two key findings emerged. First, soil moisture state — not instantaneous rainfall intensity — was the dominant triggering predictor: XSWI variables ranked 2nd and 3rd in importance after slope angle, operating as independent axes (r = 0.074). The no-treatment condition consistently outperformed all resampling strategies across algorithms. Second, spatial SHAP mapping revealed that predisposing factors produce time-invariant contribution patterns governed by terrain, while dynamic triggers produce event-specific patterns reflecting rainfall distribution; their spatial overlap identifies the highest-risk locations. Time-series susceptibility maps confirmed that the framework captures within-event risk evolution as rainfall progresses — a capability unattainable with static approaches. These results indicate that incorporating occurrence-time-aligned soil moisture dynamics substantially improves both the predictive and explanatory capacity of landslide susceptibility assessment.