A Method of Water Body Extraction Based on Multiscale Feature and Global Context Information

Water body extraction is an essential mission in the field of semantic segmentation of remote sensing images. It plays a significant role in natural disaster prevention, water resources utilization, hydrological monitoring, and other territories. In practice, the background of the majority of water remote sensing images is complicated. Owing to insufficient semantic mining and rough water body boundary extraction, traditional segmentation methods may be unable to adequately distinguish water bodies. We put forward a multiscale feature and global context fusion network (MSGFNet). In addition, multiscale feature extraction and fusion (MSEF) module based on UperNet and global context enhancement block (GCE Block) are designed. The MSEF module is capable of handling complex scenes by dynamically capturing multiscale semantic information and fusing different layers of features. The GCE Block can help the network to infer the location, shape and contextual information of the water bodies. The GF-1 dataset and Sentinel-2 dataset are used for model training simultaneously. The experimental results indicate that the extraction accuracy of the MSGFNet proposed are superior than other methods on GF-1 dataset and Sentinel-2 dataset, with overall accuracy of 98.60% and 98.22%, respectively. Compared to UperNet, the overall accuracy increases by 1.28% and 0.85%, respectively. In conclusion, the learning method build upon multiscale features and global context information can effectively prohibit noise, heighten the extraction accuracy of water bodies under intricate background, as well as ameliorate the matter of inaccurate water edge segmentation.