Identification and susceptibility assessment of landslide disasters in the red bed formation along the Nanjian-Jingdong Expressway

• Published in: Ecological indicators 2025-01, Vol.170, p.113002, Article 113002
• Main Authors: Cao, Yifan, Zhao, Zhifang, Wen, Mingchun, Zhao, Xin, Zhou, Dingyi, Qin, Jingyi, Ouyang, Liu, Cao, Jingyao
• Format: Article
• Language: English
• Subjects: CatBoost; Landslide; SBAS-InSAR; Stacking; Susceptibility Assessment; XGBoost
• ISSN: 1470-160X
• DOI: 10.1016/j.ecolind.2024.113002

•Conducting landslide susceptibility assessments in the red bed strata region.•Using Sentinel-1 data for deformation monitoring from two orbits.•Combining optical imagery to identify landslide disaster points.•Employing four machine learning models for landslide susceptibility evaluation. The red bed strata region is characterized by distinct interbedded soft and hard water–rock properties and significant water sensitivity, resulting in the frequent occurrence of landslide disasters. Despite the widespread application of Interferometric Synthetic Aperture Radar (InSAR) technology in landslide identification, challenges such as low recognition rates and difficulties in objective assessment continue to persist. This study focuses on a section of the Nanjing Expressway in the western part of Yunnan Province as the research area and utilizes Small Baseline Subset InSAR (SBAS-InSAR) technology in conjunction with optical imagery to identify landslide disaster points. By analyzing nine evaluation indicators, this study assesses the susceptibility of landslide disasters in the research area by applying Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Stacking Ensemble Strategies (Stacking). Furthermore, Receiver Operating Characteristic (ROC) curves are used to evaluate the accuracy of the models and analyze the relative importance of each evaluation factor. The results indicate that: (1) the analysis of 245 datasets of ascending and descending orbits from 2017 to 2022 yielded deformation rates, with maximum positive and negative deformation rates of 37.02 mm/yr and −46.47 mm/yr, respectively. In combination with optical imagery data, a total of 521 landslide disaster points were identified. (2) In comparison to individual machine learning models, the Stacking demonstrated superior performance, with prediction capabilities and accuracy that surpassed other models. The Area Under the Curve (AUC) values increased by 2.55 %, 2.82 %, and 5.39 % compared to RF, XGBoost, and CatBoost, respectively. The findings from the stacking reveal that high-risk areas comprise 12.29 % of the total area of the research zone, predominantly located along the northern highway, where average annual rainfall and topographic relief are the primary driving factors for landslide occurrences.