City scale urban flooding risk assessment using multi-source data and machine learning approach

Abstract

With the frequent occurrence of extreme rainfall and the acceleration of urbanization, the issue of urban flooding worldwide has gained increasing prominence. City scale flooding risk assessment is critical for urban safety and renovation, yet faces challenges such as data complexity, accuracy and interpretability. In this study, a machine learning approach incorporated with multi-source big data was developed to perform a city-scale urban flooding assessment. The developed approach was demonstrated in China’s largest city of Shanghai. Five ensemble learning models including categorical boosting (CatBoost), extreme gradient boosting, random forest, light gradient boosting machine and adaptive boosting, were employed for establishing the relationship among a variety of geological, natural, social-economical factors and urban flooding events. It was found that all the ensemble learning models achieved prediction reliability of over 80% for the city-scale flooding events; specially, the CatBoost model had the relatively best performance, offering 95% prediction of the actual flooding events. With the CatBoost model, Shapley additive explanations, partial dependency plot and individual conditional expectation plot were further employed to probe the quantitative effects of a variety of factors on urban flooding. It was revealed that areas with higher road network density, slighter topography gradient, closer distance to rivers, higher gross domestic product and population density are more prone to urban flooding. Further, city-scale risk map was generated, showing downtown areas exhibits higher flooding risk than the suburban areas. Therefore, urban flooding prevention strategies were provided.