An attention-based multiscale few-shot network for water body extraction from GF-2 satellite imagery

The core to water resources mapping is water body extraction, a pivotal technique employed to distinguish areas of water bodies from land and other regions. Presently, water body extraction methods relying on remote sensing images predominantly utilize supervised learning approaches. However, these methods are limited to fixed areas and demand a substantial amount of manually annotation data for the accurate identification of water bodies. This paper introduces a novel perspective by treating water body extraction as a few-shot semantic segmentation task, aiming for generalization across diverse geographic regions. This paper proposes a hierarchical atrous spatial pyramid pooling (ASPP) attention module to enhance the model’s extraction capabilities in multi-scale water bodies to highlight the characteristics of water bodies at different scales. Quantitative evaluation on test dataset shows that our method achieves average precision of 94.01%, average recall of 90.81%, average IoU of 90.14%, and average Fscore of 91.31%. Comparative analyses also prove that our method can extract water bodies across geographical areas, providing new ideas for water resources mapping.