DMAM-UNET:An Improved Unet Semantic Segmentation for Water Body Extraction from Remotely Sensed Image

Water body extraction and mapping is an effective way for water resource management. With the development of remote sensing technology and the acquisition of remote sensing data, lots of methods were proposed for water body extraction. However, these existing methods were mainly used for large and continuous water and cannot get good accuracy when applied on broken and discontinuous water bodies. To address this problem, a double multiplication attention module UNet (DMAM-UNet) modified from the UNet network is proposed in this paper. DMAM-UNet mainly improves from UNet on network structure, loss function and data processing. In this study, four different types of broken and discontinuous water body extractions from Landsat were employed to test the performance of the DMAM-UNet method. The results of the four experiments showed that the proposed DMAM-UNet obtains better extraction accuracy than the two other related algorithms.