Building extraction with SGAU-Net: a deep learning approach for high-resolution remote sensing imagery

Building extraction from high-resolution remote sensing imagery plays a crucial role in urban planning, disaster response, and land management due to its ability to provide accurate and up-to-date spatial information. To tackle the challenges of complex background interference and blurred building boundaries in remote sensing segmentation, this study presents SGAU-Net, an enhanced UNet-based model incorporating attention mechanisms. The encoder is replaced with VGG16 to boost feature extraction robustness. Furthermore, a Spatial-Guided Attention Gate is applied to the first three skip connections to emphasize salient spatial regions, while an Efficient Local Attention module is embedded in the fourth to capture advanced spatial–channel dependencies. Experiments on the INRIA dataset demonstrate that SGAU-Net achieves 85.29% IoU, 92.21% Precision, and 96.69% Accuracy, confirming the effectiveness of the proposed attention modules and skip connection enhancements.