Enhancing landslide susceptibility mapping incorporating landslide typology via stacking ensemble machine learning in Three Gorges Reservoir, China

Abstract

Different types of landslides exhibit distinct relationships with environmental conditioning factors. Therefore, in regions where multiple types of landslides coexist, it is required to separate landslide types for landslide susceptibility mapping (LSM). In this paper, a landslide-prone area located in Chongqing Province within the middle and upper reaches of the Three Gorges Reservoir area (TGRA), China, was selected as the study area. 733 landslides were classified into three types: reservoir-affected landslides, non-reservoir-affected landslides, and rockfalls. Four landslide inventory datasets and 15 landslide conditional factors were trained by three Machine Learning models (logistic regression, random forest, support vector machine), and a Deep Learning (DL) model. After comparing the models using receiver operating characteristics (ROC), the landslide susceptibility indexes of three types landslides were acquired by the best performing model. These indexes were then used as input to generate the final map based on the Stacking method. The results revealed that DL model showed the best performance in LSM without considering landslide types, achieving an area under the curve (AUC) of 0.854 for testing and 0.922 for training. Moreover, when we separated the landslide types for LSM, the AUC improved by 0.026 for testing and 0.044 for training. Thus, this paper demonstrates that considering different landslide types in LSM can significantly improve the quality of landslide susceptibility maps. These maps in turn, can be valuable tools for evaluating and mitigating landslide hazards.