Yu et al. (2026) Flood inundation monitoring with multi-source satellite imagery based on deep learning and explainable frameworks

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-30
• Authors: Hongjie Yu, Yue‐Ping Xu, Yintao Huang, Yen‐Ming Chiang
• DOI: 10.1016/j.jhydrol.2026.135409

Research Groups

Institute of Water Science and Technology, Zhejiang University, Hangzhou 31058, China

Short Summary

This study introduces a U-Net deep learning model for flood inundation monitoring by integrating multi-source Sentinel-1 (SAR) and Sentinel-2 (Multispectral) satellite imagery, demonstrating enhanced accuracy and providing insights into model behavior using an explainable framework.

Objective

• To develop and evaluate a deep learning framework (U-Net) for flood inundation monitoring using integrated multi-source satellite imagery (Sentinel-1 and Sentinel-2), and to understand the model's decision-making process through an explainable framework (Grad-CAM).

Study Configuration

• Spatial Scale: Regional to global applicability, tested on real-world cases.
• Temporal Scale: Event-driven, for emergency response and continuous monitoring.

Methodology and Data

• Models used: U-Net (deep learning for segmentation), Grad-CAM (explainable framework).
• Data sources: Sentinel-1 (Synthetic Aperture Radar), Sentinel-2 (Multispectral), auxiliary information (e.g., Digital Elevation Model - DEM).

Main Results

• The U-Net model achieved strong segmentation performance on Sentinel-1 imagery, with an overall accuracy of 0.95 and an intersection over union (IoU) of 0.33. It produced a 1.41% false-water rate and a 34.62% false-dry rate.
• When Sentinel-1 and Sentinel-2 data were jointly used as inputs, overall accuracy increased to 0.97 and IoU remained at 0.33, while the false-positive rate decreased to 0.86% and the false-negative rate decreased to 27.85%, outperforming models trained on single data sources.
• The incorporation of auxiliary information, such as DEM, further enhanced the model’s performance.
• Analysis using Grad-CAM revealed the transparency of shallow layers in the deep learning model, emphasizing the necessity of deep learning architecture in multi-source satellite imagery applications.

Contributions

• Enhanced the utilization efficiency of multi-source satellite data for flood inundation monitoring.
• Improved the understanding of deep learning models used for flood inundation monitoring through the application of an explainable framework (Grad-CAM).
• Demonstrated superior performance of integrated SAR and Multispectral satellite data for flood mapping compared to single-source approaches.

Funding

Not specified in the provided text.

Citation

@article{Yu2026Flood,
  author = {Yu, Hongjie and Xu, Yue‐Ping and Huang, Yintao and Chiang, Yen‐Ming},
  title = {Flood inundation monitoring with multi-source satellite imagery based on deep learning and explainable frameworks},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135409},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135409}
}