Objectives Landslide is one of the most frequent and destructive geological hazards in the world, posing significant threats to human lives, property, and natural environments. Accurate landslide susceptibility modeling is critical for disaster early warning and risk management. However, a major challenge in this process is the selection of non-landslide samples, which is often arbitrary and overlooks the fact that these samples might originate from potentially unstable areas, leading to reduced model accuracy.

Methods To address this, a novel negative sample selection strategy based on interferometric synthetic aperture radar (InSAR)-derived ground deformation data is proposed. Taking Hanyuan County, Sichuan Province as the study area, annual average ground deformation rates are generated using the small baseline subset InSAR technique. Negative samples are extracted from the regions which exhibit extremely low deformation rates, representing areas unlikely to experience landslides.

Results Results demonstrate that the InSAR-based sampling method significantly improves predictive performance in both extreme gradient boosting (XGBoost) and random forest (RF) models. The InSAR-RF model achieves the highest accuracy, with the area under the curve of 0.847 and the overall accuracy of 89.8%, representing 5.5% and 5.6% improvement, respectively, over traditional buffer-based sampling approaches. Additionally, to overcome the lack of dynamic features in conventional landslide susceptibility assessment, InSAR data are integrated with machine learning models, and an evaluation matrix is constructed to refine susceptibility mapping. The decision mechanisms within the models are further analyzed using shapley additive explanations to enhance interpretability.

Conclusions The InSAR-based negative sampling strategy not only improves model accuracy, but also strengthens the scientific reliability of landslide susceptibility assessments by integrating real-world deformation data. It offers a new approach for landslide hazard early warning and risk management.