Optimizing reservoir landslide susceptibility mapping with physics-enhanced data-driven models

Landslides in reservoir areas pose substantial risks to hydropower facilities, surrounding infrastructure, and the safety of local populations. Landslide Susceptibility Mapping (LSM) evaluates the likelihood of landslide occurrences, aiding in the mitigation and prevention of these risks. Data-driven LSM faces reliability constraints due to its inherent uncertainty and limited interpretability. This study constructs a physics-enhanced data-driven model to innovatively map landslide susceptibility in wide reservoir areas, considering the physical effects of impoundment-stage reservoir water-level rise. Surface deformation data acquired through InSAR technology are merged with geomorphological features to create a comprehensive inventory of active landslides in the Lianghekou Reservoir area. Subsequently, results from physics-based models are incorporated as factors into the data-driven model, merging the predictive strengths of data-driven models with insights from physics-based analyses. This integration not only enhances the accuracy of the LSM model but also improves its interpretability. Additionally, SHAP (SHapley Additive exPlanations) clarifies how various conditioning factors and enhancement strategies shape the model’s performance. It also reveals the key drivers of landslide susceptibility during reservoir impoundment. The results indicate that the physics-based model makes a notable contribution, playing a crucial role in model classification decisions. This study provides new insights into integrating data-driven and physics-based approaches within LSM, aiding in the accurate localization and prevention of landslide hazards.