Identification of Landslide Risks in the Subtropical Hilly Regions of Southern China Using Integrated Multi-Source Synthetic Aperture Radar Interferometry and Machine Learning

Abstract

The subtropical hilly regions of southern China are characterized by dense vegetation and highly concealed landslides, making it difficult for traditional, single-source remote sensing methods to meet disaster prevention needs. The core scientific contribution of this study is the development of a hierarchical, progressive hazard identification framework that bridges the gap between InSAR deformation detection and landslide risk identification. This study focuses on Mayang County, Hunan Province, China, and combines time-series InSAR data from C-band Sentinel-1 and L-band ALOS-2 with a random forest (RF) algorithm to construct an early-stage identification model for landslide risks. All SAR data were processed under controlled baseline conditions (perpendicular baseline <150 m; polarization: VV for Sentinel-1, HH for ALOS-2). By screening highly reliable deformation points through dual-source cross-validation and integrating nine evaluation factors including slope, we established a two-layer coupled identification model combining InSAR deformation and susceptibility indices at the slope unit scale. The results showed that the dual-source InSAR approach achieved an identification accuracy of 71% (precision 68%, recall 65%, F1-score 0.66, Cohen’s κ 0.62), significantly outperforming single-source methods (62% for Sentinel-1 alone and 58% for ALOS-2 alone); the AUC was 0.815 under spatial block cross-validation, with an out-of-bag error of 16.8%. The dual-source InSAR approach identified a total of 59 potential hazard sites, 83.1% of which were located in medium- to high-risk zones. Following field surveys and LiDAR verification, 35 of these were confirmed as active landslide sites, demonstrating identification accuracy significantly superior to that of a single data source. The multi-source coupling framework proposed in this study effectively overcomes the decoherence issues associated with single-source SAR data in subtropical vegetated areas, providing reliable technical support for the early identification of landslides in humid hilly regions of southern China.