Integrating Hydrodynamic Mechanisms and Deep Learning for Real-Time Flood Inundation Forecasting

Extreme climate events are becoming increasingly frequent, and flood disasters have been one of the most frequent and devastating forms of such events, threatening the lives and property of coastal residents. To reduce the potential costs to residential lives and property, fast and reasonable predictions and decisions should be made for quick emergency response based on timely flood routing analysis. This study proposes a hybrid model that aims to achieve real-time forecasting of time-varying flood routing and inundation maps by integrating hydrodynamic analysis and deep learning. A computational fluid dynamics (CFD) database of 125 simulated flood scenarios is established under varying flood frequency and runoff roughness of potential routing areas. Various deep learning networks, such as the long short-term memory (LSTM) network, convolutional neural network (CNN), and transpose convolutional neural network (TCNN), are used in this study to develop the proposed hybrid model for real-time and visual flood routing analysis. The results show that the model can quickly generate flood inundation maps with the input of real-time water level monitoring histories along the drainage basin, which provides valid support for emergency decision-making in various flood scenarios.