<p>This study proposes a framework that integrates a probabilistic physically based model with machine learning (ML) techniques for improved landslide susceptibility assessment. A physically based model and Monte Carlo simulation was employed to estimate probability of slope failure and identify stable areas for non-landslide sample extraction. Subsequently, several ML models, including random forest, extreme gradient boosting, and categorical boosting, were trained and combined using a stacking ensemble optimized by the grey wolf optimizer. Model interpretability was enhanced using shapley additive explanations (SHAP). The proposed framework was applied to Saka Town, Japan, where a severe rainfall-induced landslide event occurred in 2018. Results show that the physics-informed sampling strategy significantly improves model performance, achieving higher AUC values (0.859–0.868) compared to random sampling (0.837–0.848). Spatial analysis demonstrates that susceptibility maps generated using the proposed approach are more coherent and better aligned with observed landslide distributions. SHAP analysis further confirms that slope, elevation, and curvature are the most influential factors, and that the physics-informed approach yields more consistent and physically meaningful feature contributions at both global and local scales. Overall, the proposed hybrid framework provides a robust and reliable tool for landslide susceptibility mapping and hazard assessment.</p>