Refined landslide inventory and susceptibility of Weining County, China, inferred from machine learning and Sentinel‐1 InSAR analysis

Abstract Landslides are widely distributed mountainous geological hazards that threaten economic development and people's daily lives. Interferometric synthetic aperture radar (InSAR) with comprehensive coverage and high‐precision ground displacement monitoring abilities are frequently utilized for regional‐scale active slope detection. Moreover, InSAR measurements that characterize ground dynamics are integrated with conventional topographic, hydrological, and geological landslide conditioning factors (LCFs) for landslide susceptibility mapping (LSM). Weining County in southwest China, with complex geological conditions, steep terrain, and frequent tectonic activities, is prone to catastrophic landslide failures. In this study, we refined the landslide inventory of Weining County using one ascending and one descending Sentinel‐1 dataset acquired during 2015–2021 through a small baseline subset InSAR (SBAS InSAR) analysis. We then combine the LOS measurements from both datasets using multidimensional SBAS to obtain time series two‐dimensional (2D) displacements to characterize the kinematics of active slopes. Hot spot and cluster analysis (HCA) was carried out on 2D displacement rate maps to highlight clustered deformed areas and suppress noisy signals that occurred on single pixels. Two hundred fifty‐eight landslides (including 71 active identified in this study) are used to construct 76,412 positive samples for LSM. In our study, the HCA maps, instead of the 2D displacement maps, are integrated with conventional LCFs to form an LCF_HCA set to feed support vector machine (SVM), Random Forest (RF), extreme Gradient Boosting (XGBoost) and Light Gradient‐Boosting Machine (LightGBM) models. A conventional LCF (LCF_CON) set and an integrated 2D displacement maps (LCF_2D) set have also been adapted for comparison. The performance of the tree‐based ensemble methods distinctly outperforms the SVM model. In the meantime, models' performances using the LCF_HCA set are superior to that of the other 2 LCF sets from all evaluation metrics. The ranks of HCA maps increased compared with 2D displacement maps from feature importance analysis, which might lead to the better performance of models using the LCF_HCA set. With the continuous accumulation of SAR images, ground dynamic characteristics from InSAR can offer us opportunities to understand landslide kinematics and enhance LSM.