Effects of SBAS-InSAR Deformation Integration Methods and Machine Learning Model Selection on Landslide Susceptibility Mapping

Landslide susceptibility mapping (LSM) is effective in revealing the spatial distribution and probability of landslides. Its accuracy is crucial for minimizing casualties and property loss. However, current LSM methods primarily rely on static indicators, that do not consider the significant surface deformation during landslide events, resulting in inadequate reliability of the assessment results. This study aims to enhance the accuracy and reliability of landslide susceptibility assessment by integrating surface deformation data extracted using Small Baseline Subset-Interferometric Synthetic Aperture Radar (SBAS-InSAR) technology with static LSM results. We selected Jinping Autonomous County as the study area and performed a landslide susceptibility evaluation based on 11 static evaluation factors using three machine learning models: Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). Comparative analysis of the results indicated that RF and XGBoost exhibited higher accuracy in susceptibility evaluations. Joint training and the weighted sum method were used to integrate the assessment results of the RF and XGBoost models with surface deformation rates. Four dynamic landslide susceptibility models (joint training RF (JTRF), weighted sum RF (WSRF), joint training XGBoost (JTXGBoost), and weighted sum XGBoost (WSXGBoost)) were developed for LSM. The results demonstrate that integrating dynamic surface deformation data with static LSM outcomes generates more reasonable and accurate landslide susceptibility maps. This research provides valuable insights for optimizing the integration of surface deformation data with static LSM results and a scientific basis for governments to identify and manage high-risk areas, increasing the capacity for landslide disaster prevention and response.