UNet reimagined: advanced building extraction through dual attention and global-local synthesis

This is the authors' abstract. We don't add key points for this paper.

Remote sensing image segmentation, particularly building extraction, presents unique challenges due to the dense and intricate structures inherent in these images. This study addresses the critical limitation of feature retention during the downsampling process, which hampers the performance of conventional U-shaped networks when applied to remote sensing tasks. Motivated by the need for more accurate building extraction, we propose the Dual Skip Attention UNet (DSAUNet), a model incorporating a Dual Skip Attention Mechanism (DSAM) to enhance intermediate feature retention and mitigate information loss. Additionally, the integration of a Global-Local Focus Module (GLFM) enables our model to balance global context awareness with local detail precision, effectively capturing complex structural details such as corners. Extensive experiments demonstrate the effectiveness of DSAUNet, achieving an IoU of 88.24% and an [Formula: see text] score of 93.75% on the Vaihingen dataset, an IoU of 82.13% and an [Formula: see text] score of 90.19% on the Beijing dataset, and an IoU of 82.08% and an [Formula: see text] score of 90.16% on the Austin dataset. These results highlight the superior performance of our method in precision building extraction and its potential for advancing the automated interpretation of urban landscapes in remote sensing imagery.