64 / 2025-05-07 16:23:33

定性和定量评估滑坡敏感性模型鲁棒性：采用ZOO攻击和定量指标

Landslide susceptibility model,Robustness,ZOO attack,Random Forest,XGBoost

The main objective of the present study was to provide two novel methodological approaches for assessing the robustness of the landslide susceptibility map (LSM) model qualitatively and quantitatively, respectively. In this study, the Wulong District of Chongqing, China, was selected as the study area, and Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) models were used to construct the LSM for this region. Firstly, a total of 594 landslide locations were examined with 18 landslide-affected variables, of which 70% was used for training and 30% for testing. Secondly, Secondly, we utilized Zeroth-Order Optimization (ZOO) attack to generate adversarial samples and qualitatively assessed the robustness of RF and XGBoost by comparing the accuracy changes between LSM models constructed with original samples and those constructed with adversarial samples. Subsequently, we quantitatively assessed the robustness of RF and XGBoost models using the metrics "average bound" and "verified error". Finally, SHapley Additive exPlanations (SHAP) was employed to reveal the relationship between landslide-affected variables and landslides. The results show that the accuracy of the RF model (AUC = 0.9000) is higher than that of the XGBoost model (AUC = 0.8828), but the robustness of the XGBoost model (average bound = 0.5028, verified error = 0.4971) is higher than that of the RF model (average bound = 0.2321, verified error = 0.5371). Prediction ability and robustness are two key indicators to evaluate a model. High accuracy does not necessarily indicate high robustness. This is the first time to compare the robustness of RF and xgboost models. Our research provides a valuable reference for the selection and comparison of LSM models.