A heterogeneous ensemble landslide susceptibility assessment method based on InSAR and geographic similarity extended landslide inventory

Abstract

Landslide Susceptibility Mapping is a vital tool for assessing landslide risks and designing mitigation strategies. Traditional methods rely on field-based landslide inventories, which are time-consuming and costly, limiting their applicability in complex terrains and their ability to detect potential landslides in steep or subtle surface change areas. Based on multi-temporal Synthetic Aperture Radar Interferometry (InSAR) surface monitoring, this study innovatively integrates Geomorphic Environmental Similarity analysis (GES) to extract potential samples from areas with similar geographical characteristics to known landslides. Considering the complex landslide mechanisms in Fengdu County of the Three Gorges Reservoir Area, this study further develops a heterogeneous ensemble model (Stacking) that integrates RF, LightGBM, and XGBoost algorithms to effectively capture nonlinear relationships and evaluate local landslide susceptibility. Results show that the GES-SAR-Stacking model outperforms traditional methods, with AUC improved by 2.27 %, accuracy by 1.34 %, precision by 3.05 %, recall by 2.32 %, and F1-score by 2.12 %. This study proposes a method that combines InSAR-based sample selection with ensemble learning, providing a dynamic and accurate solution for early warning and risk assessment in complex terrains. This method demonstrates high landslide detection capability in r reservoir banks, foothill foot slopes, and regions with intensive human activities, showing great potential for geological hazard management.