Dynamic Earth Observation for Landslide Susceptibility Mapping Using Machine Learning and InSAR-Derived Deformation

Abstract

Landslide susceptibility assessment is essential for hazard mitigation and sustainable urban planning, yet many existing approaches rely primarily on static environmental predictors and often neglect active slope deformation. This limitation is particularly relevant in rapidly urbanizing areas, where human activity may destabilize slopes and reactivate dormant landslides. This study develops a process-informed susceptibility framework that integrates LiCSBAS-derived SBAS-InSAR deformation into a MaxEnt model for the urban and peri-urban residential areas of Cluj-Napoca, Romania. Two comparative models were implemented: (i) a baseline model using conventional conditioning factors and (ii) an enhanced model incorporating Sentinel-1 LOS velocity derived from SBAS-InSAR time-series analysis (2020–2023). The study quantitatively evaluates the predictive value of deformation-informed susceptibility modelling under single-orbit InSAR conditions, independently validates velocity data using EGMS observations, and constitutes the first integration of InSAR-derived ground deformation into landslide susceptibility assessment for Cluj-Napoca. Results show moderate-to-strong agreement between SBAS and EGMS deformation data (Pearson correlation ≈ 0.7). Incorporating LOS velocity improved model performance (AUC from 0.809 to 0.833; p < 0.001) and increased the spatial correspondence between mapped landslides and high- and very high-susceptibility zones. The integrated framework enabled the identification of localized active instability zones and provided a practical basis for hazard-informed urban planning and land management.