Landslide susceptibility assessment (LSA) plays a critical role in disaster prevention and mitigation. However, many conventional approaches fail to adequately capture (i) spatial interactions among slope units, (ii) temporal dynamics induced by rainfall, and (iii) fundamental geotechnical mechanisms. This study presents a physics-informed deep learning framework that simultaneously models spatiotemporal dependencies across slope units while enforcing constraints derived from soil mechanics. The study area is represented as a dynamic graph, where each slope unit is modeled as a node connected to its neighboring units. A hybrid neural architecture is designed to extract local features from antecedent rainfall sequences and to propagate global contextual information across the terrain. A key innovation lies in the explicit incorporation of the infinite-slope stability model into the training objective, thereby ensuring that model predictions adhere to limit-equilibrium principles. Extensive comparative evaluations against established baseline methods demonstrate that the proposed framework achieves superior accuracy in identifying high-susceptibility slopes and delivers more reliable predictions of slope stability states. In addition to improved performance, the framework enhances model interpretability by enabling the inference of physically interpretable parameters, such as effective soil strength. As such, the proposed approach provides a robust, physics-aware solution for dynamic landslide hazard assessment, with direct application in rainfall-induced landslides early warning.
 

landslide susceptibility assessment,physics-informed deep learning,spatiotemporal modeling,graph neural networks,rainfall-induced landslides warning