A multi-objective optimization framework for mudflow susceptibility mapping in the Yanshan Mountains: Integrating nondominated sorting genetic algorithm-II, random forest, and gradient boosting decision trees

Recurrent mudflow disasters in the Yanshan Mountains pose a significant impediment to regional sustainable development. To address the limitations of linear assumptions and the imbalance in single-objective optimization in traditional susceptibility assessments, this study develops a novel multi-objective framework integrating information value, random forest (RF), gradient boosting decision tree, and the nondominated sorting genetic algorithm II (NSGA-II). The framework synchronizes hyperparameter tuning and feature selection via NSGA-II, achieving synergistic improvements in precision (0.9248), recall (0.7972), and area under the curve [AUC (0.9323)]. Results indicate that the NSGA-II-optimized RF model outperforms other configurations, achieving superior performance (AUC = 0.9323, recall = 0.7972) and a 10.7% improvement in recall over the baseline RF model. Spatial mapping identifies very high susceptibility zones (12.7% of the study area) concentrated in the southern foothills, where steep slopes (6°–14°), intense rainfall (565–768 mm), and anthropogenic disturbances (e.g., mining, road construction) show strong spatial coupling. Model validation demonstrates a strong alignment with historical data, with 82.3% of recorded mudflow events localized within high- to very high-risk zones. Rainfall, slope, and lithology emerge as dominant controlling factors, while watershed area shows limited explanatory capacity. Future studies should adopt watershed-scale modeling to improve spatial heterogeneity analysis. This framework advances methodological innovations for multi-objective dynamic early warning systems, providing actionable insights for disaster mitigation in analogous mountainous regions.