Extraction of small water body information based on Res2Net-Unet

Abstract

As the extraction of small water bodies in remote sensing images has problems such as water line interruption and pretzel phenomenon, in order to be able to improve the extraction accuracy of small water bodies, this paper proposes a small water body extraction method based on Res2Net- Unet. The method uses the encoder and decoder structure of the UNet model. Firstly, the ResNet-50 network of the Res2Net module is used as an encoder, thus exploiting the feature information at multiple scales in the image. Secondly, a hybrid domain attention mechanism is incorporated into the decoder structure to fully mine the spatial and channel features in the image. Finally, a jump connection is added between the encoder and decoder to better fuse the features extracted by the encoder and decoder. Experiments on the Chinese Gaofen-1(GF-1) image datasets from two study areas show that the method in this paper is feasible for more complete and more accurate extraction of small water bodies compared with common deep learning models.