Urban waterlogging prediction and risk analysis based on rainfall time series features: A case study of Shenzhen

Abstract

In recent years, the frequency of extreme weather has increased, and urban waterlogging caused by sudden rainfall has occurred from time to time. With the development of urbanization, a large amount of land has been developed and the proportion of impervious area has increased, intensifying the risk of urban waterlogging. How to use the available meteorological data for accurate prediction and early warning of waterlogging hazards has become a key issue in the field of disaster prevention and risk assessment. In this paper, based on historical meteorological data, we combine domain knowledge and model parameters to experimentally extract rainfall time series related features for future waterlogging depth prediction. A novel waterlogging depth prediction model that applies only rainfall data as input is proposed by machine learning algorithms. By analyzing a large amount of historical flooding monitoring data, a “rainfall-waterlogging amplification factor” based on the geographical features of monitoring stations is constructed to quantify the mapping relationship between rainfall and waterlogging depths at different locations. After the model is trained and corrected by the measured data, the prediction error for short-time rainfall basically reaches within 2 cm. This method improves prediction performance by a factor of 2.5–3 over featureless time series methods. It effectively overcomes the limitations of small coverage of monitoring stations and insufficient historical waterlogging data, and can achieve more accurate short-term waterlogging prediction. At the same time, it can provide reference suggestions for the government to conduct waterlogging risk analysis and add new sensor stations by counting the amplification factor of other locations.