Explainable AI uncovers key landslide drivers in northeastern Bangladesh

Landslides constitute a persistent geohazard in the hilly landscapes of northeastern Bangladesh, driven by intense monsoonal precipitation and escalating anthropogenic disturbances. Despite growing applications of machine learning (ML) in susceptibility mapping, frameworks that integrate predictive performance with model interpretability remain insufficiently developed. This study addresses this gap by implementing a hybrid ML-explainable artificial intelligence (XAI) framework to evaluate landslide susceptibility using four algorithms: Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost). Model training and validation were conducted through 10-fold cross-validation on a dataset comprising 140 landslide and non-landslide samples, incorporating topographic, geological, land surface, vegetation, anthropogenic, soil, hydrological, and environmental conditioning factors. All models demonstrated strong predictive capability, with SVM achieving the highest performance (Accuracy = 0.93; AUC = 0.97), followed closely by RF and GBM (Accuracy = 0.92), while GBM exhibited the highest recall (0.98), indicating superior sensitivity in detecting landslide occurrences. XGBoost also maintained competitive performance (Accuracy = 0.90). Spatial susceptibility mapping reveals a pronounced concentration of high- to very-high-risk zones in the central and southwestern regions, with model-dependent extensions toward the eastern margins. To move beyond “black-box” predictions, SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) were employed to decode model behavior, identifying elevation, slope, and stream density as dominant controlling factors. Elevation consistently emerged as the most influential predictor across all models, exhibiting a distinct nonlinear threshold effect around 25 m. These findings demonstrate that while ML models may yield comparable predictive accuracy, they differ substantially in their internal decision logic and spatial generalization patterns. By integrating robust predictive modeling with transparent interpretability analysis, this study advances a comprehensive framework for landslide susceptibility assessment, offering actionable insights for targeted land-use planning and risk mitigation in hazard-prone regions of northeastern Bangladesh.