Citation

Das, Atanu and Rajin, S. M. Abrar and Goh, Michael Kah Ong and Biswas, Shuvodip and Billah, Nabibun and Khan, Riasat (2025) Dual-Attention ResUNet With Masked Focal-Tversky Loss for Robust SAR-Based Flood Mapping. IEEE Access, 13. pp. 201460-201477. ISSN 2169-3536

Text 11267426.pdf - Published Version Restricted to Repository staff only Download (3MB)

Abstract

Floods are among the deadliest natural disasters, threatening lives and infrastructure worldwide, and require rapid and accurate mapping for effective emergency response. Synthetic Aperture Radar (SAR) technique is highly valuable because it operates independently of weather conditions and sunlight availability. However, flood segmentation in SAR imagery remains challenging due to speckle noise, diverse terrain characteristics, and uncertain labels in publicly available datasets. To address these challenges, we propose a dual-attention ResUNet that integrates channel recalibration and spatial attention to capture complex floodwater patterns more effectively. Unlike conventional U-Net architectures that rely solely on encoder–decoder skip connections for feature extraction, the proposed dual-attention ResUNet incorporates both Squeeze-and-Excitation channel attention and spatial attention mechanisms at multiple encoder–decoder scales, thereby enhancing feature discrimination and spatial context understanding in noisy SAR data. For datasets such as SEN1FLOODS11, where labels often contain uncertain pixels, we design a masked Focal–Tversky loss guided by pixel-level validity masks. This approach enables the model to exclude unreliable labels while addressing severe class imbalance. Additionally, conventional VV and VH polarizations are complemented with engineered spectral indices—such as normalized difference, root mean square intensity, and polarimetric ratio—to better distinguish flooded regions from surrounding land areas. Experimental results on the SEN1FLOODS11 and S1GFloods datasets demonstrate that our model achieves 97.50% accuracy, 92.65% IoU, and 96.18% F1 score on S1GFloods. Ablation studies show that dual attention provides an 11.85% IoU improvement over the baseline ResUNet, while the masked Focal–Tversky loss yields a 15.16% gain compared to BCE–Dice. Further hyperparameter tuning of its α, β, and γ parameters improves boundary sensitivity, and robustness tests confirm stable performance under varying speckle noise conditions. Grad-CAM++–based interpretability analysis verifies that the model focuses on meaningful SAR backscatter patterns, enhancing its transparency and reliability. The fully annotated SEN1FLOODS11 dataset and implementation codes can be found at: https://github.com/AbrarRajin/Dual-Attention-ResUNetwith-Masked-Focal-Tversky-Loss-for- Robust-SAR-Based-Flood-Mapping

Item Type:	Article
Uncontrolled Keywords:	Flood mapping
Subjects:	T Technology > TC Hydraulic engineering. Ocean engineering > TC530-537 River protective works. Regulation. Flood control
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	22 Dec 2025 03:38
Last Modified:	26 Dec 2025 03:55
URII:	http://shdl.mmu.edu.my/id/eprint/15093

Downloads

Edit (login required)