Integrated InSAR and Machine Learning Reveal Soil Type as Primary Control on Rainfall-Triggered Landslide Susceptibility in Meizhou, China

The spatial controls on rainfall-triggered landslides remain elusive due to monitoring challenges in mountainous regions with frequent cloud cover. Here we fuse three complementary interferometric techniques—Small BAseline Subset (SBAS), Enhanced Small BAseline Subset (E-SBAS), and storm-pair Differential Interferometric Synthetic Aperture Radar (D-InSAR)—with Sentinel-2 imagery and seven machine learning classifiers to analyze the June 2024 landslide outbreak in mountainous Meizhou, Guangdong. Time-series interferometry captures centimeter-scale precursor motion, yet radar decorrelation in vegetated areas limits detection, underscoring the need for multisensor integration. After ingesting the full remote-sensing stack, the gradient boosting decision tree reveals soil types—especially the clay-rich red soils that mantle lower catchments—as the dominant control: within these zones, the model captures 69% of new failures inside just 18% of the landscape (AUC = 0.85), whereas slope angle and aspect rank second-order. Support vector machine performs optimally for historical records, while gradient boosting decision tree excels under extreme rainfall, reflecting temporal shifts in factor importance. By coupling near-real-time InSAR with soil-aware learning frameworks, our approach offers a practical route toward adaptive early warning and targeted mitigation across the red-soil belts of southern China.