Urban Flood Depth Prediction and Visualization Based on the XGBoost-SHAP Model

With the warming of the global climate and the acceleration of urbanization, the intensity and frequency of urban floods pose increasingly significant threats to cities. The prediction of urban flood depth is an important non-engineering measure used to mitigate the hazards associated with urban flooding. In this study, a novel approach based on the XGBoost-SHAP is proposed to predict the urban flood depth, which compensates for the black-box shortcomings of the single XGBoost model and performs a global interpretation of the selected predictive factors. Meanwhile, Genetic Algorithms (GA) are employed to achieve the optimization of XGBoost model parameters. This approach is applied to a case study in Zhengzhou city, China, an area prone to frequent flooding due to its extensive built-up regions. The results suggest the Root Mean Square Error (RMSE) and Nash-Sutcliffe Efficiency (NSE) between observed and predicted values of urban flood depth are 0.065 and 0.885, which satisfy the demand. Besides, the results of relevance and significance analysis indicate that rainfall factors are more closely linked to flood than geographical factor. The results provide a reference for urban flood forecasting, factor analysis and the development of disaster reduction measures.