Zhou et al. (2025) MSFlood-Net: A physically informed deep learning model integrating multi-source data for flood inundation mapping

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-11-13
• Authors: Lv Zhou, Tsuhan Chen, Fei Yang, Yuanjin Pan, Ling Huang, Xiang Huang, Po‐Chien Lai
• DOI: 10.1016/j.envsoft.2025.106779

Research Groups

• College of Geomatics and Geoinformation, Guilin University of Technology, Guilin, China
• Technology Innovation Center for Geohazard Monitoring and Risk Early Warning, Ministry of Natural Resources, Beijing, China
• Geoscience and Survey Engineering College, China University of Mining and Technology-Beijing, Beijing, China
• School of Remote Sensing and Geomatics Engineering, Nanjing University of Information Science and Technology, Nanjing, China
• Guangxi Water & Power Design Institute Co., Ltd, Nanning, China

Short Summary

This study introduces MSFlood-Net, an enhanced U-Net-based deep learning model that integrates multi-source remote sensing and topographic data with physical information for accurate flood inundation mapping. The model achieves 97.187 % accuracy and a 96.756 % F1 score, demonstrating superior robustness in complex environments compared to baseline models.

Objective

• To develop an enhanced deep learning model (MSFlood-Net) that integrates multi-source remote sensing and topographic data with physical information to improve the accuracy and robustness of flood inundation mapping, particularly in complex and challenging environments.

Study Configuration

• Spatial Scale: Regional to local scale, focusing on diverse environments including rivers, reservoirs, and urban transition areas, addressing both large-scale and fine-scale inundation dynamics.
• Temporal Scale: Aims for timely and continuous flood monitoring, addressing challenges posed by short-lived and transient flood events.

Methodology and Data

• Models used: MSFlood-Net (an enhanced U-Net-based deep learning model incorporating multi-scale attention and dilated convolutions), U-Net (baseline), DeepLabV3 (baseline).
• Data sources: Synthetic Aperture Radar (SAR) imagery, optical imagery, Digital Elevation Model (DEM), Height Above Nearest Drainage (HAND). Training and evaluation dataset built upon the publicly available GF-FloodNet dataset.

Main Results

• MSFlood-Net achieved a flood inundation mapping accuracy of 97.187 % and an F1 score of 96.756 %.
• The model significantly outperformed U-Net and DeepLabV3 baseline models.
• It demonstrated strong robustness in delineating flood extent across rivers, reservoirs, and urban transition areas.
• Physically informed inputs effectively reduced false positives caused by shadows, clouds, and wet surfaces.

Contributions

• Introduction of MSFlood-Net, a novel physically informed deep learning model that integrates multi-source remote sensing (SAR, optical) and topographic data (DEM, HAND) for improved flood inundation mapping.
• Enhancement of feature representation in complex terrain through the incorporation of multi-scale attention and dilated convolutions within a U-Net architecture.
• Demonstrated superior performance (accuracy and F1 score) and robustness compared to existing deep learning baselines, particularly in mitigating common false positives.
• Provides a practical and scalable solution for flood monitoring, supporting integration into broader environmental modeling systems.

Funding

• Not specified in the provided text.

Citation

@article{Zhou2025MSFloodNet,
  author = {Zhou, Lv and Chen, Tsuhan and Yang, Fei and Pan, Yuanjin and Huang, Ling and Huang, Xiang and Lai, Po‐Chien},
  title = {MSFlood-Net: A physically informed deep learning model integrating multi-source data for flood inundation mapping},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106779},
  url = {https://doi.org/10.1016/j.envsoft.2025.106779}
}