This study applies an integrated Geographic Information System (GIS) machine learning framework for rockfall susceptibility mapping in Huinan County, Jilin Province, China. Twelve conditioning factors were compiled, and 51,876 slope units were delineated using the curvature-watershed method. A balanced labeled dataset (75 rockfall units and 75 non-rockfall units) was constructed, where non-rockfall units were randomly sampled from a negative candidate pool outside a 1 km exclusion buffer to mapped rockfall units. A Multi-Layer Perceptron (MLP) optimized by genetic algorithm (GA) and trained with the ADAM optimizer was implemented. To reduce the instability of a single small test split, model performance was evaluated using repeated stratified random splitting (70/30, 100 repeats), and mean ± standard deviation metrics were reported. The proposed MLP achieved ROC-AUC = 0.862 ± 0.064 and PR-AUC = 0.870 ± 0.059, with Accuracy = 0.788 ± 0.067 and F1-score = 0.783 ± 0.074. The very high susceptibility class occupies 14.30% of the area yet contains 76.92% of the mapped rockfall points, indicating strong spatial concentration of high-risk locations. SHAP interpretation and dependence analysis identify relief, distances to roads, aspect and distances to river as dominant controls.
Du et al. (Wed,) studied this question.