Improving Landslide Prediction Accuracy by Incorporating PS-InSAR into a Random Forest Model

Abstract

Thailand has experienced increasingly frequent landslides, posing significant risks to highway infrastructure. This study develops a corridor-scale landslide susceptibility mapping (LSM) framework for Thailand’s highway network using a Random Forest (RF) model trained with conventional topographic, hydrological, and environmental conditioning factors. Model performance is evaluated using spatial cross-validation to mitigate spatial autocorrelation, yielding a mean ROC-AUC of 0.9626 ± 0.0058, indicating strong predictive capability under geographically independent testing. Within the available Sentinel-1 coverage areas (Chiang Mai and Yala), persistent scatterer interferometric synthetic aperture radar (PS-InSAR) time-series deformation information is incorporated through feature-level integration to produce a deformation-informed dynamic landslide susceptibility map (LDSM). The deformation-informed model achieves a mean ROC-AUC of 0.9936 ± 0.0108 under the same spatial validation setting, demonstrating improved discrimination performance within the InSAR footprints. Feature importance analysis indicates that terrain controls such as slope and roughness remain dominant predictors, while PS-InSAR deformation provides complementary time-dependent information that helps improve discrimination in actively deforming or marginally classified zones. Although dynamic susceptibility assessment is spatially restricted to areas with InSAR coverage, the results demonstrate that integrating time-series deformation indicators with machine-learning-based susceptibility modeling can refine hazard delineation along infrastructure corridors. The proposed footprint-aware framework provides a scalable approach for landslide risk assessment and infrastructure planning in regions where time-series InSAR data are available.