Landslide susceptibility study using Bayesian networks combined with landslide physical mechanisms: a case study in Hebei Province, China

This study addresses the limitations of current landslide susceptibility assessment methods, which rely on machine learning with strong predictive power but lack interpretability and fail to reveal physical interactions among influencing factors. We propose a Bayesian network evaluation model that integrates mechanisms of landslide occurrence to achieve highly interpretable susceptibility mapping. In the Hebei Province case study, the model was constructed using 15 key factors and comprehensively evaluated through 10-fold cross-validation. The receiver operating characteristic (ROC) curve area (AUC) reached 0.88, with Accuracy, Recall, and F1 score all at 0.83, and the MCC at 0.65. These metrics collectively demonstrate the reliable predictive performance and robustness of the model in the study region. Compared to traditional black-box machine learning, this model clearly quantifies and reveals the interactions and mechanisms underlying the factors' effects. Further Bayesian analysis indicates that high-susceptibility landslide zones are primarily distributed in low-altitude mountainous areas and plain transitional zones, spatially adjacent to human-engineered features such as roads and buildings. The study not only generates highly reliable landslide susceptibility maps but also identifies high-risk corridors and urban areas requiring prioritized prevention in Hebei Province, providing a direct scientific basis for territorial spatial planning, major infrastructure layout, and precise regional disaster prevention measures.