Dynamic landslide susceptibility assessment in Zhouqu County, China, based on SBAS-InSAR and a Stacking ensemble learning model

Abstract

Landslide susceptibility assessment is an important basis for regional landslide risk management. Zhouqu County is located in a landslide-prone region along the eastern margin of the Qinghai–Tibetan Plateau; however, existing susceptibility studies in this area still suffer from limitations in negative sample selection, dynamic information integration, and model interpretability. This study proposes a dynamic landslide susceptibility assessment framework based on a Stacking ensemble learning model that incorporates SBAS-InSAR–derived deformation information. Using 74 descending Sentinel-1 SAR images, line-of-sight (LOS) deformation rates ranging from −78.6 to 96.5 mm/yr were retrieved. Negative samples were selected from extremely low-velocity deformation zones and further constrained by spatial exclusion from documented landslides to reduce label noise, and a deformation-driven dynamic correction matrix was introduced to revise the static susceptibility results. The results show that the Stacking model achieves an AUC of 0.986 and an overall accuracy of 95.0%, outperforming the individual base learners. Dynamic correction reduced the very low susceptibility zone from 48.4% to 30.4% and the very high susceptibility zone from 10.5% to 1.7%, effectively identifying actively deforming areas overlooked by the static model while reasonably downgrading stabilized historical landslides. The multi-temporal comparison between 2023 and 2024 further demonstrates that InSAR deformation information can significantly enhance the temporal sensitivity of susceptibility maps. Moreover, feature importance rankings exhibit notable variations across different assessment periods, highlighting the necessity of periodic susceptibility reassessment in tectonically active regions.