China's first sub-meter building footprints derived by deep learning

The high spatial resolution building footprints are crucial for understanding urban development and its associated applications. However, up to now, the sub-meter-level building footprint data of China is still lacking. The challenges arise from two aspects: 1) the number of training samples is inadequate for large-scale building extraction. 2) the accuracy and efficiency of current models are insufficient to conduct large-scale building extraction. Therefore, we propose a framework for large-scale building extraction in this study, including semi-automated sample generation, building extraction model, model training, and post-processing. Specifically, the main technical contributions include: 1) BldgNet (Building Extraction Network) is proposed, including the Large Window Attention, Edge Attention, and Distribution Alignment Module with consideration of spatial contextual information, to address the challenge of the multi-scale building extraction, building boundary delineation, and class imbalance, respectively; 2) a semi-supervised training approach is proposed for large-scale building extraction, leveraging the incomplete information from OpenStreetMap (OSM) to enhance the diversity of building samples and the robustness of the model. Meanwhile, we created an open-source Global Building Dataset (GBD) comprising approximately 800,000 high-resolution (0.25 m) samples. This dataset incorporates diverse building styles worldwide, offering support for global building extraction. Based on the constructed sample set and the proposed deep net, we generated China's first sub-meter (0.5 m) building footprint dataset (CBF). Through testing on 750,000 buildings from 350 cities, the overall F1 score for CBF reached 83.71%. Finally, we validated that the proposed building extraction model can achieve satisfactory results compared to existing representative deep networks. GBD and CBF datasets can be publicly available and downloadable via https://zenodo.org/doi/10.5281/zenodo.10043351.