Small Target Building Extraction Using Dual-Branch Enhanced U-Net With Multi-Scale Attention Fusion

This study addresses critical challenges in small target building extraction from high-resolution remote sensing imagery, particularly instance separation difficulties and boundary precision limitations in dense urban environments where adjacent building structures often appear as connected regions in conventional segmentation outputs. An Enhanced U-Net architecture with dual-branch processing and multi-scale attention fusion was developed. The architecture integrates a ResNet-34 encoder backbone with a novel boundary-aware loss function and specialized dual-branch outputs for semantic segmentation and instance separation. Experimental validation across Massachusetts, WHU, and INRIA datasets demonstrates consistent performance improvements, achieving IoU scores of 71.8%, 77.9%, and 67.3% respectively, representing improvements of 4.7%, 4.7%, and 5.7% over baseline U-Net architectures (67.1%, 73.2%, and 61.6%). The proposed method also achieves F1-scores of 71.2%, 77.4%, and 66.8% with accuracy rates of 91.7%, 94.2%, and 89.4% across the three datasets. Systematic ablation studies confirm progressive contributions of attention mechanisms (67.1% to 70.1% IoU) and loss function components (achieving final 71.8% IoU), while cross-dataset generalization analysis reveals realistic transferability with performance degradation ranging from 0.6 to 14.1 percentage points across different geographical contexts. The dual-branch design philosophy successfully decouples semantic understanding from geometric boundary detection, enabling superior optimization for both building classification and instance separation tasks while maintaining computational efficiency essential for practical deployment scenarios, with enhanced extraction capabilities supporting automated urban monitoring, disaster response mapping, and smart city development initiatives that provide sophisticated spatial analysis capabilities essential for evidence-based urban planning and infrastructure management applications.