基于概率物理信息的机器学习区域滑坡易发性评估

Landslide susceptibility assessment (LSA) is critical for predicting potential landslides at regional scales. Existing machine learning based data driven models are highly dependent on data quality and generally lack physical interpretability, which often leads to limited generalization ability and physical inconsistency under complex geological and hydrological conditions. In contrast, physically-based models in mechanical principles can characterize the relationship between rainfall infiltration and slope stability, but their application at the regional scale remains constrained by spatial heterogeneity of parameters and limited model applicability. To address these challenges, this study proposes a probabilistic physics-informed machine learning (PIML) framework that explicitly incorporates geotechnical domain knowledge into the model training process. A simplified transient infiltration model and the first-order reliability method (PRL-STIM) are utilized to compute the factor of safety and failure probability, respectively. These physics-based outputs are integrated into a composite loss function designed to ensure both physical and risk consistency, thereby guiding the training of a neural network. The framework is validated using a rainfall-induced shallow landslide event that occurred in Niangniangba, Gansu Province, China, in 2013. A high-quality dataset is compiled, and a spatial cross-validation strategy is employed to assess the model′s generalization ability and predictive uncertainty in previously unseen areas. Results from five-fold spatial cross-validation indicate that the PIML model enhances the average AUC by 13.6% and reduces the scientific inconsistency index (SI) by 88.6% compared to the baseline model. These improvements demonstrate the proposed model′s enhanced robustness, physical consistency, and interpretability for regional LSA.