Automated Building Segmentation Using Half-UNet with Multi-Scale Residual Attention and Self-Calibrated Pixel Attention

Building detection is essential for updating geographic information and urban planning but remains challenging in Geoinformatics, particularly when buildings blend with their surroundings. Traditional automated methods often struggle with accuracy due to similar colors, textures, and shapes between buildings and their environments. This study introduces an improved Half-UNet model enhanced with Multiscale Residual Attention (MRA) and Self-Calibrated Pixel Attention (SC-PA) to refine feature extraction at multiple scales and enhance pixel-level representation. These additions improve the model's ability to differentiate objects from backgrounds while maintaining computational efficiency. The model's lightweight architecture reduces resource requirements, making it more cost-effective for real-world applications. Trained and evaluated on the Massachusetts Buildings Dataset, the proposed model outperforms existing methods, including DeepLabv3+ (Xception), RefineNet (ResNet50), U-Net, Half-UNet, SegNet, and Bayesian-SegNet. It achieves an F1-score of 83.23% and an IoU of 71.35% with only 0.44 million parameters. These results highlight its effectiveness and efficiency for automated building segmentation tasks.