AR2Net: An Attention-Residual-Recurrent Network for Building Footprint Extraction from High-Resolution Aerial Imagery

Accurate building footprint extraction from high-resolution aerial imagery is essential for urban planning, cadastral mapping, disaster management, and digital twin development. Existing deep learning approaches often improve individual architectural components in isolation or rely on increasingly complex designs, at the expense of efficiency and generalization. This study proposes AR2Net, an encoder–decoder architecture integrating residual learning, recurrent convolutional refinement, and channel–spatial attention within a unified framework to improve optimization stability, spatial feature refinement, and building-relevant feature selectivity. A two-stage training strategy is introduced on a large-scale international dataset with OpenStreetMap-derived annotations before fine-tuning on a high-quality regional dataset from Iranian cities. AR2Net is evaluated through ablation experiments and comparisons with five widely used segmentation models (U-Net, U-Net++, DeepLabV3+, HRNetV2, and SegFormer), achieving the best overall performance (IoU = 88.35%, F1-score = 93.31%). A geometric accuracy assessment against manually digitized ground truth across 486,000 buildings yielded a mean positional error of 0.92 m. The extracted footprints were further used to generate LoD1 CityGML building models, illustrating a practical pathway toward 3D cadastral map production. These results demonstrate that AR2Net provides an accurate, geometrically reliable, and scalable solution for automated building footprint extraction.