Dynamic evaluation of landslide susceptibility in a large-scale region based on time-series InSAR and multi-temporal cataloguing: a case study in Heifangtai, Gansu province

Landslide disasters are common in the mountainous regions of western China, making accurate classification of landslide hazard risk levels essential for effective geological disaster management. Data-driven models have achieved notable advancements in landslide susceptibility evaluation. However, they still encounter challenges such as a lack of dynamic feature data and an over-reliance on sample quality, which limits their effectiveness for large-scales. To address these issues, this paper introduces an integrated evaluation approach that combines time-series InSAR deformation data with data-driven models. This method initially incorporated time-series InSAR deformation monitoring data as dynamic factors, which were filtered through multivariate covariance analysis to construct a dynamic landslide susceptibility evaluation system. Additionally, only landslide samples were adopted to conduct large-scale landslide susceptibility analysis. Various machine learning algorithms, based on landslide and non-landslide samples, were also applied for model comparison. Building upon this foundation, the evaluation of landslide susceptibility was performed by integrating InSAR deformation data with multi-temporal cataloging. The Heifangtai region in Gansu Province was chosen as a case study. The results indicate that the maximum entropy (Maxent) model achieved the highest accuracy for large-scale susceptibility assessments. Incorporating time-series InSAR deformation into the dynamic landslide susceptibility model improved accuracy by about 0.36%, compared to models without this data. Additionally, the proportion of landslides identified in high and very high susceptibility zones increased by 4.29%. By using eight years of landslide catalog data as positive samples, the presented model achieved an accuracy of 99.26%, demonstrating that long-term, high-quality positive samples improve the precision and reliability of regional predictions. This study advances large-scale landslide risk assessment by integrating a dynamic evaluation system that accounts for data across multiple time periods.