Machine learning-based identification of key factors and spatial heterogeneity analysis of urban flooding: a case study of the central urban area of Ordos.

With global climate change and accelerating urbanization, urban flood is becoming more frequent worldwide. Understanding the urban vulnerability is crucial for making decisions on urban flood control. This study uses urban flood susceptibility (UFS) as an indicator, and comprehensively applies three machine learning models, XGBoost, CatBoost and LightGBM, in the Kangyi area of Ordos City. Combined with the Shapley Additive explanations method, the driving mechanism and spatial heterogeneity of flood susceptibility was explored in the study area. The results show: (1) Model performance comparison: All three models have high accuracy, with XGBoost performing well in overall classification (OA = 0.96) and CatBoost performing well in distinguishing flood/non-flood samples (AUC = 0.85). (2) Multi-model adaptability assessment: The proposed "model-factor-space" framework highlights the sensitivity of XGBoost to urbanization indicators, the ability of CatBoost to capture naturalgeographical elements, and the efficiency of LightGBM in analyzing terrain thresholds. (3) Dynamic thresholds and synergies: Impervious surface density (ISD) is the most critical factor, and when ISD > 0.2, the risk of flooding will continue to increase by 60%. Comprehensive analysis with spatial heterogeneity shows that high-risk areas are mainly affected by ISD, road density (> 10,000 m/km2) and low altitude (< 1300m) in urban built-up areas, while low-to-medium risk areas are sensitive to vegetation coverage (> 7,000) and Dis2Water bodies (> 1,500m). (4) Hierarchical governance strategy: A three-level spatial governance strategy is proposed: in the core area, priority is given to ISD control (< 0.2) and pipe network upgrades; in the transitional area, slope interception and ecological restoration are combined; and in the potential risk area, a multi-scale monitoring and early warning system is established for multi-scale monitoring.