African water body segmentation with cross-layer information separability based feature decoupling transformer
Due to climatic and geographical factors, Africa suffers from severe water scarcity. Deep neural networks (DNNs) based remote sensing water segmentation models are helpful for the observation and effective use of water resources. However, African water segmentation task faces three major challenges:...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225003887 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Due to climatic and geographical factors, Africa suffers from severe water scarcity. Deep neural networks (DNNs) based remote sensing water segmentation models are helpful for the observation and effective use of water resources. However, African water segmentation task faces three major challenges: the lack of data, existence of many small water bodies and high similarity between water and background. To solve the above-mentioned problems, we first establish AWS16K, a water segmentation dataset covering the whole African area, and generate masks by a manual and automatic hybrid method. Second, we study the water and background information in the high-level features of DNNs, and demonstrate the cross-layer separability of these two types of information. Third, we design an asymmetric cross-layer input-dependent Feature Decoupling Transformer (FDTran), to extract water features from information mixed high-level features, improving water segmentation performance. Finally, comprehensive experiments are conducted on AWS16K and two public image segmentation datasets, LandCover.ai and CrackVision12k, and our FDTran achieves state-of-the-art (SOTA) results compared to current semantic segmentation methods. Codes and datasets will be made publicly available at https://github.com/cv516Buaa/BinghaoLiu/tree/main/FDTran. |
|---|---|
| ISSN: | 1569-8432 |