Reliable probabilistic landslide susceptibility mapping using a calibrated stacked ensemble with bootstrap-based uncertainty

Landslides are among the most damaging natural hazards, causing recurrent loss of life, infrastructure disruption, and socio-economic impacts in mountainous regions. Accurate landslide susceptibility mapping (LSM) is essential for hazard mitigation, yet robust regional-scale prediction remains challenging in geomorphically complex terrain. This study presents an integrated machine learning workflow for high-resolution LSM in the Upper Svaneti region of northern Georgia, a tectonically active and landslide-prone sector of the central Greater Caucasus. The framework combines exploratory geomorphic structure analysis with a spatially robust stacked ensemble and probabilistic uncertainty quantification. The approach integrates CatBoost, Random Forest, Extra Trees, and a multilayer perceptron within an XGBoost meta-learner trained on out-of-fold probabilities. Hyperparameters were optimized via successive halving under spatial cross-validation, and class imbalance was mitigated through controlled down-sampling. The stacked model achieved strong discrimination (AUC = 0.979, AP = 0.955) and reliable calibration (Brier = 0.054), outperforming individual learners. A geomorphic consistency analysis showed a systematic increase in susceptibility with increasing slope angle. Bootstrap ensembles with isotonic calibration were used to quantify predictive uncertainty and assess the stability of probabilistic performance across resampled training datasets. The resulting susceptibility maps are statistically robust, geomorphically consistent, and probabilistically interpretable, providing a reproducible workflow for susceptibility mapping and risk-informed planning in mountainous terrain. Presents a hybrid ensemble framework integrating heterogeneous classifiers within a stacked meta-learning architecture for high-resolution landslide susceptibility mapping. Applies spatial cross-validation and isotonic calibration to obtain robust generalization and probabilistically reliable susceptibility estimates. Quantifies predictive uncertainty using a 1,000-member bootstrap ensemble, demonstrating stable discrimination and well-calibrated performance across resampling iterations.