A Comparative Study of Deep Learning Approaches to Rooftop Detection in Aerial Images

This paper investigates the deep neural networks for rapid and accurate detection of building rooftops in aerial orthoimages. The networks were trained using the manually labeled rooftop vector data digitized on aerial orthoimagery covering the Kitchener-Waterloo area. The performance of the three deep learning methods, U-Net, Fully Convolutional Network (FCN), and Deeplabv3+ were compared by training, validation, and testing sets in the dataset. Our results demonstrated that DeepLabv3+ achieved 63.8% in Intersection over Union (IoU), 77.8% in mean IoU (mIoU), 74% in precision, and 78% in F-1-score. After improving the performance with focal loss, training loss was greatly cut down and the convergence rate experienced a significant growth. Meanwhile, rooftop detection also achieved higher performance, as Deeplabv3+ reached 93.6% in average pixel accuracy, with 65.4% in IoU, 79.0% in mIoU, 77.6% in precision, and 79.1% in F-1-score. Lastly, in order to evaluate the effects of data volume, by changing data volume from 100% to 75% and 50% in ablation study, it shows that when data volume decreased, the performance of extraction also got worse, with IoU, mIoU, precision, and F-1-score also mostly decreased.