A method for identifying landslide potential based on GraphSAGE and SBAS-InSAR

Landslides are a prevalent and highly destructive geological hazard; identifying potential landslides is crucial for risk governance. This study takes Badong County as the research area and constructs a landslide identification system. Landslide susceptibility was evaluated based on XGBoost, LightGBM, 2D-CNN and GraphSAGE models, and the 34 ALOS-2 images were processed by using SBAS-insAR technology. Dynamic deformation and the optimal landslide susceptibility map (LSM) were fused to further complete landslide identification. A dual-evaluation system comparing conventional random hold-out with spatial block cross-validation (SBCV) revealed a 48% average performance drop under spatial isolation (Macro-F1 0.16–0.32 for SBCV vs. 0.50–0.91 for random hold-out). This proves that the traditional random hold-out method severely overestimates model generalizability in unseen regions. Under SBCV conditions, GraphSAGE demonstrated the highest spatial generalizability. Finally, through cross-validation with optical imagery, this integrated approach successfully identified 285 potential landslides, providing a rigorous and robust reference for regional risk governance.