Building footprint extraction from aerial imagery through semantic segmentation techniques

Abstract The extraction of building footprints from aerial images is crucial for creating detailed topographic maps. Leveraging advancements in artificial intelligence, particularly deep learning, presents a promising avenue for enhancing the automation of this extraction process. This study aims to establish a methodology that employs deep learning semantic segmentation technology to detect and reconstruct building footprints from digital aerial images. The methodology encompasses several key steps: generating true-ortho images from multi-view aerial images, constructing dataset for building regions, employing deep learning semantic segmentation, and regularizing the detected building footprints. The test area is located in Hsinchu, Taiwan. The experiment utilizes 25cm aerial true-ortho images for building region detection through the TransUNet semantic segmentation method. An additional data covering approximately 7km 2 was utilized as an independent check region. The precision and recall metrics for building detection achieved 90% and 91%, respectively. Post-boundary regularization, the extraction of building footprints from the identified building regions was successfully executed. This proposed scheme shows the efficacy of automatically extracting building footprints by integrating aerial orthoimagery with deep learning technology, significantly benefiting the creation of precise topographic maps.