FloodUnet: A Rapid Spatio‐Temporal Prediction Model for Flood Evolution Based on an Enhanced U‐Net

Abstract Flooding causes significant loss of life and economic damage and affects healthy development of society. Deep learning (DL) models demonstrate significant advantages in improving computational efficiency while maintaining accuracy. Existing research of predicting dynamic flood evolution still remains some gaps for predicting flooding maps from the initial time step, weak transferability for flood scenarios from unseen breaches, and potential enhancement of common neural network frameworks. This paper proposes a DL model called FloodUnet based on an improved U‐Net architecture to achieve rapid and accurate prediction of flood evolution. FloodUnet can predict a series of flooding depth maps and maintain high‐precision prediction. It achieves an average root mean square error of 0.2 m and an average Nash‐Sutcliffe Efficiency coefficient of 0.9 on testing sets of unseen breaches and inflows through a 4‐fold cross validation. It is three orders of magnitude faster than the hydrodynamic model with a 24‐hr lead time. It has obvious advantage in prediction accuracy compared to ordinary convolutional neural network and U‐Net. Residual module and channel attention mechanism can enhance feature representation for complex flood dynamics and ensures stability during multi‐step rolling prediction.