A novel weighted average ensemble method for landslide susceptibility mapping: A case study in Yuanyang, China

Landslide susceptibility mapping is critical for risk assessment, but existing ensemble methods like VotingClassifier suffer from three unresolved limitations: static weight allocation that ignores spatial variability, lack of quantifiable uncertainty measures, and poor integration of interpretability tools. This study introduces a novel weighted average ensemble method that dynamically adjusts weights for Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) through 5-fold spatial cross-validation, improving prediction robustness across Yuanyang County’s 2240 km² of mountainous terrain (23°05′–23°15′N, 102°40′–102°50′E) with 817 validated landslides. The method tackles important issues by combining the best features of strong models while reducing the effects of related variables using composite indices (like a soil-lithology index based on a Pearson correlation of r = 0.81), backed by a thorough preprocessing process that includes Moran’s I-validated stratified sampling (I = 0.12), normalization that accounts for outliers (95th percentile), and spatial division with 500 m buffers. The novel ensemble achieved an accuracy of 84.32 % and an ROC AUC of 91.96 %, with sensitivity analysis via SHAP (SHapley Additive exPlanations) identifying rainfall (21 %), distance index (13 %), and elevation slope index (27 %) as dominant drivers, while uncertainty analysis revealed prediction intervals of ±0.62 width (95 % coverage). The resulting maps, validated through spatial consistency checks (AUC > 0.84), provide actionable tools for high-risk zones. This research improves landslide susceptibility mapping by developing a dynamic, uncertainty-based system that rectifies major weaknesses in static ensemble methods, thereby establishing a replicable standard for future investigations.