As a prevalent geohazard in China, landslides significantly endanger socio-economic development and public safety. Studying the distribution, properties, and mechanisms of landslides using artificial intelligence algorithms could help to minimize landslide damage, thereby improving risk planning and management in the regional area. Landslide management is only as effective as the landslide susceptibility maps (LSMs) developed to determine what areas are prone to such events. The objective is to create LSMs for Mudanjiang City using a GIS-based hybrid machine learning approach. The alternating decision tree (ADT) model was employed in the study area, along with several ensemble methods: bagging-alternating decision tree (BADT), random subspace-alternating decision tree (RSADT), and real AdaBoost-alternating decision tree (RABADT). The model training and validation process used historical landslide data and ten factors. Several statistical metrics were used to evaluate the predictive ability of the models, along with the area under the receiver operating characteristic curve (AUC). The results indicate that each model performed well, with AUC values exceeding 0.85. The three hybrid models showed varying degrees of improvement over the single model (ADT). The real AdaBoost–alternating decision tree model achieved the best performance on the validation dataset (AUC = 0.95). This study used the quantile method, equal interval method, geometric interval method, and natural break method to compare and verify the effects of different zoning methods on LSMs. The four classification techniques were applied together in all models to categorize into the five landslide susceptibility classes Further comparison based on landslide point density indicates that the equal interval method produces better results in the study area. Maps derived from this research contribute to risk-sensitive land-use and infrastructure planning by informing strategic decisions.
Zhang et al. (Fri,) studied this question.