WaterPLNet: A noise-robust weakly supervised network for water extraction from SAR imagery using point labels

Existing methods for water extraction from Synthetic Aperture Radar (SAR) imagery rely on labor-intensive and costly pixel-level annotations, which demand expert knowledge of SAR backscatter characteristics, significantly hindering the large-scale application of water extraction models. Although Weakly Supervised Semantic Segmentation (WSSS) can mitigate this label-dependence limitation, the label sparsity inherent in existing methods limits the accuracy of water extraction, exacerbated by speckle noise and insufficient regional spatial semantic information. To address these challenges, a noise-robust weakly supervised Network for high-precision Water extraction (WaterPLNet) from SAR imagery is proposed that leverages only Point Labels. First, a Noise-Robust Network (NRNet) with an asymmetric dual-branch structure is proposed, which enhances feature robustness and mitigates noise-induced accuracy degradation by introducing Gaussian perturbations. Second, a Prior-Guided Region Expansion (PGRE) module is designed to identify potential water regions and mitigate false negatives arising from insufficient spatial semantic information. Finally, a Low-Backscatter Ambiguity Reduction (LBAR) module is proposed to leverage point labels and low-scattering priors, thereby reducing confusion between water and similar low-backscatter land covers. Experiments conducted on three datasets of SAR images collected from the Gaofen-3 and Sentinel-1 satellites covering typical spatial resolutions (1 m, 3 m and 10 m) demonstrate that WaterPLNet reduces reliance on labeled data while achieving state-of-the-art performance compared to other WSSS approaches using point labels. Specifically, WaterPLNet achieves an overall accuracy of over 93% across all datasets, with an mIoU that outperforms other methods by an average of 5.48%. The code is publicly available at: https://github.com/Kiwi-xd/WaterPLNet .