Uncertainty-Aware Bayesian Machine Learning for Landslide Susceptibility Mapping: in the Chattogram Metropolitan Hill System, Bangladesh

Landslide-prone hilly regions experiencing rapid urban expansion need susceptibility models that provide both robust predictive performance and transparent uncertainty estimates. This study develops an uncertainty-aware probabilistic framework for landslide susceptibility mapping in Bangladesh’s Chattogram metropolitan hill system, incorporating 14 conditioning factors: geomorphological, hydrological, environmental, and anthropogenic. Multicollinearity analysis, i.e., VIF and tolerance, verified the statistical stability of the predictors. We compared four classifiers, Bayesian Logistic Regression, Gaussian Process Classifier, L2-regularized Logistic Regression, and Random Forest, within a standardized preprocessing pipeline. Models were evaluated using cross-validation, holdout testing, and spatial block validation, with metrics for discrimination, robustness, and calibration. Random Forest had the strongest discrimination (ROC-AUC = 0.866; MCC = 0.611) on the holdout set, while the Gaussian Process Classifier showed better probability calibration (ECE = 0.084) with competitive discrimination (ROC-AUC = 0.821) among Bayesian models. Susceptibility and uncertainty were combined in a 3×3 scheme to visualize hazard and confidence. Mapping showed that 78.71% of landslides were in the Very High class, covering only 9.47% of the area (SCAI = 0.12). These results show that uncertainty-aware Bayesian modeling with bivariate spatial integration enhances interpretability and provides transparent hazard assessments in data-limited urban hill areas.[