Chen et al. (2026) FloodUnet: A Rapid Spatio‐Temporal Prediction Model for Flood Evolution Based on an Enhanced U‐Net

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-01-30
• Authors: T. Chen, J. Tian, J. Sun, Z. Zhang, Hua Chai, B. Lin, X. Fu
• DOI: 10.1029/2025wr041427

Research Groups

Not specified in the abstract.

Short Summary

This paper proposes FloodUnet, an improved U-Net deep learning model, to rapidly and accurately predict dynamic flood evolution, demonstrating high precision and transferability to unseen flood scenarios while being significantly faster than hydrodynamic models.

Objective

• To propose a deep learning model (FloodUnet) based on an improved U-Net architecture for rapid and accurate prediction of flood evolution, addressing existing gaps in predicting flooding maps from the initial time step, weak transferability for unseen breaches, and potential enhancement of common neural network frameworks.

Study Configuration

• Spatial Scale: Not specified in the abstract.
• Temporal Scale: Prediction of flood evolution over a 24-hour lead time, generating a series of flooding depth maps.

Methodology and Data

• Models used: FloodUnet (an improved U-Net deep learning architecture incorporating residual modules and a channel attention mechanism).
• Data sources: Not specified in the abstract, but implies simulated or observed data for training and testing, specifically "testing sets of unseen breaches and inflows" used in a 4-fold cross-validation.

Main Results

• FloodUnet achieves an average root mean square error of 0.2 m for flooding depth prediction.
• It achieves an average Nash-Sutcliffe Efficiency coefficient of 0.9 on testing sets involving unseen breaches and inflows.
• The model is three orders of magnitude faster than a hydrodynamic model for a 24-hour lead time prediction.
• FloodUnet demonstrates superior prediction accuracy compared to ordinary convolutional neural networks and standard U-Net architectures.
• Residual modules and channel attention mechanisms enhance feature representation for complex flood dynamics and ensure stability during multi-step rolling prediction.

Contributions

• Proposes a novel deep learning model, FloodUnet, specifically designed for rapid and accurate dynamic flood evolution prediction.
• Addresses the challenge of predicting flooding maps from the initial time step.
• Improves transferability for flood scenarios originating from unseen breaches.
• Enhances common neural network frameworks (U-Net) through the integration of residual modules and channel attention mechanisms for better feature representation and prediction stability.
• Demonstrates significant computational efficiency gains (three orders of magnitude faster) compared to traditional hydrodynamic models while maintaining high accuracy.

Funding

Not specified in the abstract.

Citation

@article{Chen2026FloodUnet,
  author = {Chen, T. and Tian, J. and Sun, J. and Zhang, Z. and Chai, Hua and Lin, B. and Fu, X.},
  title = {FloodUnet: A Rapid Spatio‐Temporal Prediction Model for Flood Evolution Based on an Enhanced U‐Net},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041427},
  url = {https://doi.org/10.1029/2025wr041427}
}