[Objective] Global landslide models often ignore spatial heterogeneity and feature redundancy in complex gorge reservoirs, causing local biases. Furthermore, static models lack timeliness, increasing false-negative risks. To enhance accuracy for a lower Jinsha River reservoir, we propose a novel assessment method integrating spatial heterogeneity partitioning, feature selection, and dynamic InSAR deformation for correction. [Methods] First, the AGNES (agglomerative nesting) clustering algorithm was used to divide the study area into homogeneous sub-regions, and Geodetector was applied to optimize regional hazard factors. Then, hazard assessment models were constructed using multi-grained cascade forest (gcForest) and random forest (RF) algorithms. Finally, SBAS-InSAR (small baseline subset interferometric synthetic aperture radar) was utilized to extract surface deformation information, correcting the initial assessment via a hazard correction matrix. [Results] The gcForest model, accounting for spatial heterogeneity and feature optimization, achieved the best predictive performance with an AUC of 0.954. After introducing InSAR data for correction, the area proportion of low-hazard zones decreased by 17.29%, while medium-, high-, and extremely high-hazard zones increased by 14.46%, 2.48%, and 0.35%, respectively. Case validations confirmed that the corrected zonation aligns well with macroscopic surface deformations. [Conclusion] Feature optimization based on spatial zonation effectively mitigates spatial heterogeneity. Moreover, integrating InSAR deformation data better identifies potentially unstable areas. This method enhances assessment accuracy in complex environments, providing a reliable reference for disaster prevention in alpine gorge reservoirs.