DAS-Unet for building extraction from high-resolution remote sensing images

Building extraction from high-resolution remote sensing images is critical for urban planning and disaster management but remains challenging due to complex boundaries and dense distributions. This paper proposes DAS-Unet, an enhanced U-Net–based model designed to improve segmentation accuracy. The model incorporates a strengthened backbone for multi-scale and edge feature learning, an optimized pooling module with multi-directional pooling and dilated convolution to capture fine details and global context, and an A2 attention module to refine local features while enhancing computational efficiency. Experiments on the Massachusetts dataset and a self-built Kunshan City dataset demonstrate the superior performance of DAS-Unet over U-Net and other state-of-the-art models. DAS-Unet achieves F1-scores of 91.45% and 91.08% and mIoUs of 84.72% and 84.36% on the respective datasets, confirming its effectiveness and robustness in building extraction tasks.