Find in Library

Detecting and localizing buildings is of primary importance in urban planning tasks. Automating the building extraction process, however, has become attractive given the dominance of Convolutional Neural Networks (CNNs) in image classification tasks. In this work, we explore the effectiveness of the CNN-based architecture U-Net and its variations, namely, the Residual U-Net, the Attention U-Net, and the Attention Residual U-Net, in automatic building extraction. We showcase their robustness in feature extraction and information processing using exclusively RGB images, as they are a low-cost alternative to multi-spectral and LiDAR ones, selected from the SpaceNet 1 dataset. The experimental results show that U-Net achieves a <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>91.9</mn><mo>%</mo></mrow></semantics></math></inline-formula> accuracy, whereas introducing residual blocks, attention gates, or a combination of both improves the accuracy of the vanilla U-Net to <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>93.6</mn><mo>%</mo></mrow></semantics></math></inline-formula>, <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>94.0</mn><mo>%</mo></mrow></semantics></math></inline-formula>, and <inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mn>93.7</mn><mo>%</mo></mrow></semantics></math></inline-formula>, respectively. Finally, the comparison between U-Net architectures and typical deep learning approaches from the literature highlights their increased performance in accurate building localization around corners and edges.