Research on the Problem of Spatial Heterogeneity in Row Data and Generalization Capability for Landslide Susceptibility Assessment using the Physics-constrained U-net Model

Landslide susceptibility research serves as the primary approach for analyzing the future development of landslides, and could provide the scientific reference in mountainous area development strategic decisions. The accuracy of landslide susceptibility assessment mainly depends on input data and evaluation methods. To address the issues of significant impacts from raw data defects and low spatial resolution in susceptibility assessment results within traditional deep learning evaluation models, this study establishes a physical constrained U-Net model (PCUM). This method selects ten landslide assessment factors, including slope gradient, profile curvature, slope aspect, landform, river distribution density, annual average precipitation, annual average temperature, fault distribution density, lithology, and seismic intensity. Through adjusting the weights of the model's loss function by imposing explicit physical constraints, the model's performance was ultimately enhanced. Compared with other deep learning models, the evaluation results of PCUM demonstrate the advantage of maintaining spatial resolution consistent with the original input data, better revealing the spatial characteristics of landslide distribution and higher accuracy. This study focuses on the southeastern region of Tibet, and selects five typical regions to investigate the impact of spatial heterogeneity in raw data on the PCUM. The results indicate that the physically constrained U-Net model can effectively handle spatial heterogeneity in raw data and exhibits good generalization capabilities. This study may provide an effective reference for the generalizability research of landslide susceptibility assessment models.