BSG-WSL: BackScatter-guided weakly supervised learning for water mapping in SAR images
Extracting and analyzing water resources in Synthetic Aperture Radar (SAR) images is crucial for flood management and environmental resource planning due to the ability to monitor ground all-weather and all-time. However, extracting water entirely from high-resolution SAR images in diverse scenarios...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-02-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225000329 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Extracting and analyzing water resources in Synthetic Aperture Radar (SAR) images is crucial for flood management and environmental resource planning due to the ability to monitor ground all-weather and all-time. However, extracting water entirely from high-resolution SAR images in diverse scenarios is challenging due to variable water shapes, many low-intensity land covers similar to water, and scarce labels. In this article, a BackScatter-Guided Weakly Supervised Learning (BSG-WSL) framework based on image-level labels is proposed for water extraction with the requirement of high generalization and low labeling. In BSG-WSL, a BackScatter-Guided Network (BSGNet) is proposed, where the backscatter information of water is used to guide the feature extraction process, yielding precise Class Attention Maps (CAMs) of water. Then, a morphological pseudo-label optimization algorithm is designed to employ CAMs to generate high-quality pseudo-labels. Finally, a confidence cross-entropy loss is introduced to utilize pseudo-labels to train the extraction model and achieve precise water extraction in different scenarios. Experiments on three datasets of SAR images from the GF-3 and Sentinel-1B satellites verify that the proposed method achieves state-of-the-art performance compared to other weakly supervised methods based on image-level annotations. |
|---|---|
| ISSN: | 1569-8432 |