Surface water extraction from remote sensing images of arid regions in Africa using a deep learning approach combining multiscale information

Surface water is of critical importance to the ecosystem, agricultural production and livelihoods of people in the arid regions of Africa. The surface water is characterized by significant differences in spectral features, complex morphological changes, and more fine streams, which reflects the climate and water resource conditions in Africa. To address the spectral and spatial variability of surface water of large areas, this paper proposes a new surface water extraction model. The model utilizes a VGG network to extract local information, while an MVT module is introduced to capture global contextual information in remote sensing images. Spatial attention and channel attention mechanisms are used to effectively fuse global and local information to more accurately recognize and extract continuous spatial relationships of water bodies. In this paper, Sentinel-2 multispectral remote sensing images are used to finely extract surface water in Egypt during the dry season using the proposed model. Experimental results show that the model not only improves the accuracy of fine water extraction, but also performs well in waters with complex backgrounds and boundaries.