Synergistic Use of Multi-Temporal InSAR and Machine Learning for Dynamic Landslide Susceptibility Mapping

Landslides pose a significant hazard to life and property, particularly in mountainous regions. This study develops a comprehensive landslide susceptibility model by integrating Earth observation, in-situ data, and advanced MTInSAR techniques. The study compares random forest (RF) and Xtreme Gradient Boosting (XGBoost) algorithms for landslide susceptibility mapping. Employing the XGBoost algorithm, the model achieved results with an AUC ranging from 0.93 to 0.97 and an overall accuracy of 0.89. The model categorized areas into four landslide susceptibility levels ranging from no susceptibility to high susceptibility. Importantly, the model demonstrated explainable performance across all susceptibility classes. By incorporating MTInSAR data, the model accurately captures dynamic ground deformation, enhancing landslide susceptibility mapping. The resulting susceptibility map highlights areas along the Pacific coast as being particularly vulnerable. This research underscores the importance of considering temporal changes in environmental conditions for landslide risk assessment.