Deep Learning for Building Footprint Extraction Using UAV-Based Orthoimages

Building footprints is of great importance for various applications, especially with the advent of highly detailed remote sensing imagery that facilitates the identification of structures such as buildings. Recently, deep learning techniques transformed this field by overtaking traditional methods based on manually created features. At the same time, unmanned aerial vehicles (UAVs) have proven invaluable tools for capturing highly detailed aerial images of the Earth's surface for urban feature extraction. In this study, we used a DJI Mavic 2 equipped with a Hasselblad camera to capture 474 aerial images of the urban area on the west bank of the Euphrates River in Iraq. These images underwent several processing steps, including aerial triangulation, dense image matching, and point cloud generation to create digital surface models (DSMs) and orthophotos. The primary objectives of this study are to utilize UAV-based orthophotos for extracting building footprints, as orthophotos are free from tilt and relief displacement, ensuring that the rooftops align accurately with the footprints. This study aims to achieve a quick and highly detailed method for extracting building footprints using UAVs. Initially, we used standard deep-learning models developed to extract building footprints. These models performed poorly because these models were trained in different environments. To address this limitation, we iteratively trained the Mask R-CNN with orthophotos at three resolutions: 1.5 cm/pixel, 10 cm/pixel, and 20 cm/pixel. We then tested the newly trained Mask R-CNN models, which delivered promising results. Finally, we evaluated and confirmed the effectiveness and suitability of our proposed method for extracting buildings from UAV orthophotos.