GU-Net3+: A Global-Local Feature Fusion Algorithm for Building Extraction in Remote Sensing Images

In remote sensing image building extraction, image regions with similar textures or colors often cause false positives and false negatives in building-detection. Global features can help the model better recognize the overall structure of large buildings and provide contextual background information when segmenting small buildings to avoid mis-segmentation. Local features focus on extracting details and edge information, enabling a precise description of the target’s local shape and texture details. In this study, we propose a building detection method that integrates global and local features. First, frequency domain transformation and a Convolutional Block Attention Module (CBAM) were applied to preprocess the remote sensing images, enhancing building details while suppressing irrelevant noise interference. Next, an auxiliary encoder was designed to dynamically adjust the orientation and scale of the receptive field, improving the precise localization of building edges and contours. Finally, a Global Cross-Attention module is developed to aggregate global features from different spatial locations, which are then fused with local features from skip connections as input to the decoder, enhancing information exchange across multi-scale and hierarchical features to improve building extraction accuracy. The experimental results demonstrate that the proposed algorithm achieves mIoU values of 93.8% and 81.05%, and F1-scores of 96.93% and 89.10% on the WHU and Massachusetts datasets, respectively, effectively addressing the false detection issues caused by similar textures or colors in the building extraction process.