Improvement of deep learning Method for water body segmentation of remote sensing images based on attention modules

Traditional remote sensing water identification methods lack texture and shape information extraction, and the previous algorithms lack the versatility of satellite images of different resolutions. Therefore, for the remote sensing image water segmentation task, this paper first collects GF-2 images to establish a remote sensing image water segmentation dataset. PSPNet, Deeplab v3+, and U-Net have achieved good training results on this dataset. Secondly, to further improve the accuracy of water body segmentation, an attention module is introduced in the feature fusion part of the U-Net model to improve the feature fusion efficiency. Among them, a channel-spatial attention module, CBAM, performs the best. The experimental results show that the U-Net model introduced with CBAM has various degrees of improvement in the six evaluation indicators of remote sensing water body segmentation. The IoU and MIoU of CBAM-vgg16-UNet reach 92.66% and 95.91%, respectively. Finally, the experimental results show that the method also performs well on GF-1, GF-6, Landsat8 and EO-1 datasets, verifying the generality of the network.