一种结合时序InSAR技术的可解释性滑坡易发性制图框架:以中国贵州省为例OA

To address the limitations of traditional landslide susceptibility assessments,namely the underutiliza-tion of dynamic surface deformation information and insufficient model interpretability,this study proposes an interpretable landslide susceptibility evaluation framework that integrates Stacking-InSAR technology with machine learning,using Guizhou Province,China as the study area.Annual mean surface deformation veloci-ties were derived from Sentinel-1A imagery and combined with fourteen static conditioning factors encompass-ing topography,land cover,climate,and human activities for modeling.Comparative experiments across mul-tiple machine learning models revealed that without InSAR-derived deformation data,the AUC values ranged from 0.896 to 0.921.Upon integrating deformation information,the overall model performance improved by 3.5%-5.6%,with the InSAR-TabPFN model achieving the optimal performance(AUC=0.951).Sus-ceptibility mapping results indicated that the areal proportion of high and relatively high susceptibility zones decreased from 22.1%to 13.11%,significantly reducing the misclassification rate of stable areas.SHAP inter-pretability analysis identified distance to rivers,surface deformation velocity,land use type,slope gradient,and distance to roads as the dominant controlling factors of landslide occurrence,with a notable interaction effect observed between surface deformation and distance to rivers.The findings demonstrate that the TabPFN model incorporating InSAR dynamic deformation data offers significant advantages in predictive accuracy.Fur-thermore,the integration of SHAP-based interpretability analysis enables the quantitative expression of indi-vidual factor contributions and their interaction mechanisms,effectively bridging the gap between predictive accuracy and model interpretability that has long constrained conventional landslide susceptibility assessments.