Dau-Net: A Novel Water Areas Segmentation Structure For Remote Sensing Image

Water information extraction is always an important aspect of remote sensing image analysis. However, in the actual water images of remote sensing, the backgrounds of water areas are mostly complex buildings and vegetation, which interferes with water detection. In addition, traditional water detection methods were not able to accurately identify small tributaries and edge information. In order to improve to the accuracy of water segmentation, a dense skip connections network with multi-scale features fusion and attention mechanism is proposed for water areas segmentation. The proposed method obtains water feature information by using U-shaped Network (U-Net) as the backbone network and performs dense skip connections between nested convolution modules to reduce the semantic gap between the feature maps of the codec sub-networks and reduce the gradient vanishing problem of network training. In the proposed model, deep feature information is fused with shallow feature information through multi-scale attention modules. The shallow features and large-scale attention modules are used to locate the main body of the water area, while the deep features and small-scale attention modules are used to fine segment the edge of the water area. Combined with the above features and attention modules, waters can be extracted from the backgrounds. Finally, the segmentation accuracy of edge regions is further improved by Conditional Random Field (CRF). The experimental results from China HJ-1A (HJ-1B) satellite imageries and National Aeronautics and Space Administration (NASA) Land Remote-Sensing Satellite (System, Landsat-8) imageries show that the proposed method outperforms the existing methods.