Application and Comparison of Multiple Machine Learning Models in Flood Susceptibility Assessment in the Beijing-Tianjin-Hebei Region of China

Haijun LiJiubo Dong^*Yaowen ZhangHongtao LiuYixin PangShuiqing ZhouBingbin Du

• Institute of Disaster Prevention, Sanhe, China

The confluence of extreme precipitation and rapid urbanization has led to a marked increase in flood risk across the Beijing-Tianjin-Hebei(BTH)region.To this end, conducting a thorough flood susceptibility assessment is of paramount importance to safeguard the region and ensure its sustainable development. Based on historical flood disaster records, 15 flood related influencing factors such as elevation, average annual rainfall, and the Normalized Difference Vegetation Index(NDVI) were selected as the initial variable set. A flood disaster susceptibility evaluation framework was established through multicollinearity analysis and feature selection based on the Information Gain Ratio(IGR). Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting(XGBoost),and Multilayer Perceptron (MLP) models were employed to conduct the susceptibility assessment. The predictive performance and susceptibility zoning outcomes of the models were systematically compared using the Area Under the Receiver Operating Characteristic Curve(AUC) and a set of statistical evaluation metrics, including accuracy, Kappa coefficient, and sensitivity. Research findings demonstrate that (1) elevation, distance from rivers, average 24-hour maximum rainfall, and slope constitute the primary controlling factors for flood occurrence in the BTH region;(2) Very high and High susceptibility zones are primarily concentrated in topographic transition zones, critical nodes of the river system, and key flood storage and detention areas. The high and relatively high susceptibility zones identified by the four models show a strong spatial consistency with the actual distribution of flood disasters, and exhibit minimal overfitting. (3) The AUC validation results of the four models are as follows: XGBoost (0.938) > RF (0.920) > MLP (0.867) > SVM (0.854). Among these models, XGBoost produced the smallest proportion of high-susceptibility zones, demonstrating a superior ability to accurately identify areas with the highest potential flood risk. This study provides a scientific foundation for flood risk management in the BTH region and holds significant practical value for improving regional flood control strategies and spatial planning.