PSAGNet: A Water Body Extraction Method for High Resolution Remote Sensing Images

Abstract

Water body extraction is a significant researching fields in remote sensing image interpretation. However, due to the great differences of water bodies in shapes and sizes, how to improve the accuracy of water body extraction is a research issue in recent years. In this paper, a gated convolution neural network based on pyramid split attention (PSAGNet) is proposed for water body extraction. The network includes two main modules: pyramid split attention module and gated convolution module. PSAGNet can extract small water bodies accurately in complex remote sensing scenes, because the gated convolution module can extract the shape features of small water bodies from the encoder’s shallow feature map. The pyramid split attention module can extract and fuse the high-order features of multi-scale water bodies, and provide valuable feature encoding for improving water body extraction accuracy. In addition, in order to address the problems of fuzzy boundary and imbalanced water and background in the dataset, a novel loss function called FE loss is proposed to train network. FE loss significantly sharpens the segmentation boundary and improves the extraction accuracy. Our model achieves the best performance with the precision of 93.52%, recall of 94.66%, and Intersection over Union (IoU) of 86.13% on Gaofen Image Dataset (GID). In general, our method achieves the purpose of accurately extracting small water bodies from remote sensing images, and can be widely applied in water body extraction of high-resolution remote sensing images.