Assessing landslide susceptibility using improved machine learning methods and considering spatial heterogeneity for the Three Gorges Reservoir Area, China

When conducting susceptibility evaluation for study areas of special significance, especially those with spatial heterogeneity of landslide development, it is easy to ignore the potential errors caused by spatial asymmetry of geographic factors and differences in landslide development when evaluating the whole area. This study proposed an evaluation method that breaks down the Three Gorges Reservoir Area (TGRA) into smaller regions and assesses the susceptibility of landslides to each sub-region in order to assess and resolve the effect of spatial heterogeneity within the entire reservoir area of the TGRA. This method uses a combination of certainty factors (CF) and machine learning models to identify the key factors of high susceptibility index. Three machine learning models—the support vector machine (SVM), the logistic regression (LR), and the gradient boosted descent tree (GBDT)—were improved in this study. These enhanced models incorporate CF, resulting in the creation of CF-LR, CF-SVM, and CF-GBDT models. The results of the zonal evaluation are superior to those of the direct overall assessment, according to the examination of receiver operating characteristic (ROC) curves, and CF-GBDT outperforms the other five models in terms of determining the susceptibility of the entire TGRA. The occurrence of regional heterogeneity in the TGRA is confirmed by the CF-GBDT model, which also takes into account the importance of landslide influence factors between Region I and Region II. By analyzing the impact of zonal evaluation on each district and county in the TGRA, the significance of zoning in the study of landslide susceptibility within large watersheds is emphasized, providing a new perspective for regional landslide susceptibility assessment.