Transformer for the Building Segmentation of Urban Remote Sensing

The automatic extraction of urban buildings based on remote sensing images is important for urban dynamic monitoring, planning, and management. The deep learning has significantly helped improve the accuracy of building extraction. Most remote sensing image segmentation methods are based on convolution neural networks, which comprise encoding and decoding structures. However, the convolution operation cannot learn the remote spatial correlation. Herein we propose the Shift Window Attention of building SWAB-net based on the transformer model to solve the semantic segmentation of building objects. Moreover, the shift window strategy was adopted to determine buildings using urban satellite images with 4 m resolution to extract the features of sequence images efficiently and accurately. We evaluated the proposed network on SpaceNet 7, and the results of comprehensive analysis showed that the network is conducive for efficient remote sensing image research.