An Interpretable Uncertainty-Aware Framework for Landslide Susceptibility Mapping Based on Weak Supervision and Probabilistic Inference

Landslide susceptibility mapping infers future slope-failure potential from observed landslides. This inference is uncertain because locations without recorded failures cannot be confirmed as stable or used as reliable negatives, and incomplete inventories, geomorphic complexity, spatial representation, sampling, and model dependence further increase uncertainty. To address these challenges, this study developed an interpretable, uncertainty-aware framework integrating joint SHAP–PFI feature assessment, weakly supervised negative-sample construction, and ensemble-based probabilistic inference. The framework yielded relative probabilistic susceptibility estimates, together with predictive uncertainty (Upred), characterizing predictive ambiguity, and ensemble-based model uncertainty (Umodel), characterizing bootstrap variability. The framework was applied in northern Noto, Japan, using a legacy inventory and an independent 2024 event-based landslide inventory at the pixel and slope-unit scales and across seven model families. Joint SHAP–PFI analysis combined contribution magnitude and predictive dependence to support scale-specific factor selection. Ablation analyses showed that weak supervision improved independent-event generalization and reduced uncertainty. For LightGBM, independent-positive Recall increased from 0.76 to 0.83, Upred decreased from 0.18 to 0.065, and Umodel from 0.019 to 0.0085. Ensemble learning quantified spatial uncertainty while maintaining independent-positive Recall comparable to both single-model controls. The framework was effective across tree-based, regression-based, and neural-network architectures, with LightGBM showing a stable cross-scale balance: independent-positive Recall was 0.83 and 0.77, and Upred was 0.065 and 0.10 at the pixel and slope-unit scales, respectively. Spatial analysis further showed that uncertainty highlights areas where susceptibility estimates require caution. Overall, the framework quantifies and reduces uncertainty while improving independent-event generalization, thereby supporting more reliable landslide susceptibility assessment.