Duangkhwan et al. (2025) Enhancing flood forecasting with deep learning: A scalable alternative to traditional hydrodynamic models

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-12-19
• Authors: Weeraphat Duangkhwan, Chaiwat Ekkawatpanit, Chanchai Petpongpan, Duangrudee Kositgittiwong, So Kazama, Yusuke Hiraga, Chai Jaturapitakkul
• DOI: 10.1016/j.envsoft.2025.106841

Research Groups

• Civil Engineering Department, Faculty of Engineering, King Mongkut's University of Technology Thonburi, Bangkok, Thailand
• Department of Civil and Environmental Engineering, Tohoku University, Sendai, Japan

Short Summary

This study proposes an integrated deep learning framework, combining LSTM and CNN, to emulate computationally intensive HEC-RAS 1D/2D models for flood forecasting. The framework significantly reduces computational demands while maintaining high accuracy in predicting river water levels and flood inundation maps.

Objective

• To develop an integrated deep learning framework to emulate traditional HEC-RAS 1D/2D hydrodynamic models, thereby significantly reducing computational demands for real-time flood forecasting.

Study Configuration

• Spatial Scale: Pattani river basin
• Temporal Scale: Not explicitly defined for the study period, but the framework aims for real-time flood forecasting.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), HEC-RAS 1D/2D (for emulation target and training data generation).
• Data sources: Observed hydrological data, flood inundation maps generated from HEC-RAS 1D/2D simulations.

Main Results

• The LSTM model achieved good accuracy in predicting river water levels.
• The CNN model effectively translated water level predictions into flood depth maps, demonstrating close agreement with HEC-RAS outputs.
• The integrated AI-based framework significantly accelerates flood simulations while maintaining high accuracy compared to traditional hydrodynamic models.

Contributions

• Proposes an integrated deep learning framework (LSTM for water level, CNN for inundation) that effectively emulates complex HEC-RAS 1D/2D models.
• Significantly reduces computational demands for flood forecasting, offering a scalable alternative to traditional hydrodynamic models.
• Ensures physical consistency in the CNN by learning the relationship between river water levels and flood inundation by mimicking overflow results of 1D/2D models.
• Provides a promising tool for real-time flood prediction and large-scale flood risk assessment.

Funding

• No funding information was provided in the article text.

Citation

@article{Duangkhwan2025Enhancing,
  author = {Duangkhwan, Weeraphat and Ekkawatpanit, Chaiwat and Petpongpan, Chanchai and Kositgittiwong, Duangrudee and Kazama, So and Hiraga, Yusuke and Jaturapitakkul, Chai},
  title = {Enhancing flood forecasting with deep learning: A scalable alternative to traditional hydrodynamic models},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106841},
  url = {https://doi.org/10.1016/j.envsoft.2025.106841}
}