Rapid spatio-temporal prediction of coastal urban floods based on deep learning approaches

Abstract

Accurate real-time prediction of flood inundation ranges and water depths is crucial for effective flood warning and disaster mitigation. However, traditional hydrodynamic flooding simulation methods based on physical mechanisms are very complex in modeling and time-consuming. Consequently, they often fall short in meeting the requirements of near real-time flood warnings. In this study, we explore the potential applications of three deep learning neural network architectures, namely Convolutional Neural Networks (CNN), Convolutional Neural Network with Batch Normalization (CNNBN) and Multilayer Perceptron (MLP), for the spatio-temporal explicit prediction and mapping of urban floods in the coastal city of Linhai in eastern China. The results demonstrate that these three deep learning models exhibit the capability to sensitively capture the dynamic changes in the process of floods evolution and simulate the depth of floodwater efficiently and accurately. Importantly, the deep learning models only take 14–24 s to generate the inundation depth map of the entire research domain, representing a significant improvement in prediction speed, exceeding that of traditional hydrodynamic models by >20 times. Due to the high computational efficiency, superior performance and simple modeling process, the deep learning models are expected to become a powerful tool for near real time flood early warning.