Construction of rapid early warning and comprehensive analysis models for urban waterlogging based on AutoML and comparison of the other three machine learning algorithms

Abstract

Urban waterlogging often causes urban disasters, and the rapid early warning and comprehensive analysis of the urban waterlogging can help disaster defenses. However, the warning of waterlogging through the monitoring data cannot give grid distribution and the forecast of hydrological models cannot ensure rapid early warning. To obtain a grid rapid early warning result for a region, like an urban area, a method needs to be proposed which can meet the above problems. In this research, AutoML (automatic machine learning based on genetic algorithm) was recommended to construct the rapid early warning and comprehensive analysis models for urban waterlogging by compared with the other three machine learning algorithms, CatBoost (Categorical Boosting), XGBoost (eXtreme Gradient Boosting), and BPDNN (Back Propagation Deep Learning Neural Network). In the models, the forecast and historical precipitation obtained from the Integrated Nowcasting through Comprehensive analysis system (INCA), the difference of elevation, and the urban waterlogging risk maps provided by Tianjin Meteorological Administration were employed as the input sources. The input precipitation duration was determined as 12 h based on the sensitivity analysis of the influence of various precipitation duration on waterlogging depths. Due to the non-digital (discrete dataset) features, the urban waterlogging risk maps were transformed to the weight of each corresponding risk level according to the area of each risk level and the number of samples falling in each risk level. The difference of elevation was characterized by the average elevations of various distances from the points of concern. The output waterlogging depths were compared with the waterlogging depths monitored in Tianjin, China, whose quality was controlled by eliminating the records of the waterlogging depths lasting for a long time after the end of rainfall. The comparison of the models constructed by different methods demonstrated that the AutoML performed better (NSE and R2 > 0.92, CC > 0.95, RMSE1.1–1.9 cm) than the other three models. The forecast waterlogging depths by AutoML was also coherent with the monitoring waterlogging depths (NSE and R2 ≥ 0.9, CC ≥ 0.95, RMSE 1.7–2.2 cm). For that local topography and waterlogging risk are considered, the AutoML models can be used in the area without the monitoring of water level, quickly predict waterlogging depths and give spatial grid results for rapidly early warning.