Sgformer: A Local and Global Features Coupling Network for Semantic Segmentation of Land Cover

With the introduction of Earth observation satellites, the classification technology through high-definition remote sensing images appeared. After decades of evolution, the land cover classification method in high-definition satellite maps has been gradually improved. Recently, high-definition remote sensing maps have been applied to land cover classification. Nowadays, classification methods using high-definition maps have these following problems. First, the traditional land cover classification methods cannot process the rich details in high-definition maps. Second, there are different acquisition conditions in the maps of different regions, which leads to distortion, deformation, and illumination blur of remote sensing images. Third, the existing methods are unable to provide a good generalization performance. To address these issues, a dual-branch parallel network structure is proposed, called Sgformer, to improve the performance of the transformer in the context of high-definition remote sensing maps. The network enhances perceptual learning with convolution operators that extract local features and a self-attention module that captures global representations. Local information and global representations with semantic divergence are fused through a feature coupling module. At last, a decoder is designed to maximize the preservation of local features and global representations and to better recover high-definition feature maps. The results of semantic segmentation experiments show that the methodology in this study has higher accuracy than the other methodologies.