Li et al. (2025) PhysWRNet: A physics-guided deep learning framework for flood inundation mapping with SAR and hydrodynamic simulations

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-11-24
• Authors: Li Li, Yu Zhao, Shengwu Qin, Dianqi Pan, Jiquan Zhang, Aolin Li, Qinhong Hu
• DOI: 10.1016/j.jhydrol.2025.134662

Research Groups

• State Key Laboratory of Deep Earth Exploration and Imaging, College of Construction Engineering, Jilin University, Changchun, China
• College of Jilin Emergency Management, Changchun Institute of Technology, Changchun, China
• School of Environment, Northeast Normal University, Changchun, China
• Observation and Research Station of Geological Hazards and Geological Environment in Changbai Mountain Volcano, Ministry of Natural Resources, Changchun, China

Short Summary

This paper introduces PhysWRNet, a physics-guided deep learning framework that integrates Sentinel-1 SAR data and HEC-RAS flood probability maps to significantly improve flood inundation mapping accuracy and reduce errors compared to conventional deep learning methods. The framework achieves an overall accuracy of 90.2% and enhances boundary accuracy by 68.3%.

Objective

• To develop an efficient and reliable physics-guided deep learning framework for accurate flood inundation mapping, addressing limitations of existing methods in complex alluvial floodplains and SAR imagery interpretation.

Study Configuration

• Spatial Scale: Alluvial floodplains; applicable for large-scale flood mapping.
• Temporal Scale: Event-based; designed for timely emergency response and rapid detection.

Methodology and Data

• Models used: PhysWRNet (proposed framework), WVResU-Net (component of PhysWRNet), HEC-RAS (for generating flood-probability maps).
• Data sources: Multi-polarized Sentinel-1 (VV/VH) synthetic aperture radar (SAR) data, HEC-RAS flood-probability maps (as physics-guided prior), standard-resolution pixel-level labels (for training).

Main Results

• Improved flood boundary accuracy by 68.3% compared to conventional deep learning methods.
• Reduced overall mapping error by 48.6% compared to conventional deep learning methods.
• Effectively suppressed false alarms in non-inundated areas.
• Achieved an overall accuracy of 90.2% under full supervision with standard-resolution, pixel-level labels and a small labeled training set.

Contributions

• Introduces PhysWRNet, a novel physics-guided deep learning framework for accurate flood inundation mapping.
• Integrates eight scattering features derived from multi-polarized Sentinel-1 data with HEC-RAS flood-probability maps as a physics-guided prior.
• Designs a joint loss function for WVResU-Net to simultaneously optimize semantic segmentation accuracy and physical consistency.
• Significantly enhances flood boundary delineation accuracy and reduces overall mapping errors, particularly in challenging alluvial floodplains.
• Provides an efficient and reliable tool to support rapid and science-based decision-making in flood emergency management.

Funding

Not specified in the provided text.

Citation

@article{Li2025PhysWRNet,
  author = {Li, Li and Zhao, Yu and Qin, Shengwu and Pan, Dianqi and Zhang, Jiquan and Li, Aolin and Hu, Qinhong},
  title = {PhysWRNet: A physics-guided deep learning framework for flood inundation mapping with SAR and hydrodynamic simulations},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134662},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134662}
}