TA-SUNet++: Twin Attention-based Swin-UNet++ model for building damage segmentation using remote sensing images

Accurate building damage segmentation from post-disaster remote sensing images remains challenging due to illumination variations, complex structural patterns and background interference. The conventional image processing approaches have been designed to detect the damages that utilised handcrafted features and struggle to capture complex structural variations. The advancements in deep learning have improved the capability of automated damage detection by enabling models to learn hierarchical spatial representations. In this work, Twin Attention‑based Swin-UNet++ (TA-SUNet++) model is proposed for accurate building damage detection from remote sensing images. The proposed Twin Attention-based Swin-UNet++ architecture is referred to as TA-SUNet++, where ‘TA’ denotes the proposed Twin Attention module and ‘SUNet++’ represents the Swin-UNet++ backbone. The enhanced images are processed through the proposed TA-SUNet++ architecture that considers the hierarchical feature extraction capability of the Swin Transformer with the dense skip connections of UNet++ and the TA mechanism to enhance both spatial and channel-wise feature representation. The proposed TA-SUNet++ achieved accuracy values of 98.9% for binary building-damage segmentation on xView2 and 98.6% for building-footprint segmentation on SpaceNet 2 Paris Buildings Datasets. Thus, the proposed approach provided improved segmentation accuracy in complex disaster scenarios by capturing local structural features and global contextual information from remote sensing imagery.