Landslide susceptibility assessment along a highway corridor using a confidence-tiered inventory and an InSAR-enhanced， physics-informed convolutional neural network

Landslide susceptibility mapping along transportation corridors is often limited by two data-quality issues: inventories combine evidence of unequal reliability， and conventional data-driven models can overpredict high-susceptibility areas. We developed an InSAR-enhanced， physics-informed convolutional neural network (PI-CNN + InSAR) for a 71.6-km section of the Fengxian-Hezuo freeway on the northeastern Qinghai-Tibet Plateau， where loess-mudstone contact landslides predominate. The inventory was stratified into field-verified landslides as the reference standard， visually interpreted polygons as lower-confidence auxiliary samples， and InSAR-active zones as down-weighted soft positives. A factor-only CNN assigned mean susceptibility values of 0.42， 0.23， and 0.12 to verified， auxiliary， and background pixels， respectively， supporting the confidence-tiered design. Five-fold cross-validation against the verified inventory increased AUC from 0.909 ± 0.006 for an equally weighted CNN to 0.975 ± 0.001 for the proposed model. InSAR provided the largest gain， while confidence weighting further reduced overprediction; physical guidance mainly improved map conservatism. The very-high-susceptibility class occupied only 3.90% of the corridor but contained 66.4% of the verified landslide area. Site-scale limit-equilibrium analyses were consistent with near-critical stability and saturation-induced failure， supporting the physical consistency of the framework.