 This paper investigates deep neural networks for accurate detection of building rooftops in aerial ortho-images using manually, labeled data in the Kitchener Waterloo area. The three methods UNET, FCN, and Dplab3 Plus were compared with training, validation, and testing sets. Dplab3 Plus achieved 63.8% IOU, 77.8% MIOU, 74% precision, and 78% F1 score. Improving performance with focal loss reduced training loss and increased convergence rate. Dplab3 Plus reached 93.6% average pixel accuracy with 65.4% IOU, 79% MIOU, 77.6% precision, and 79.1% F1 score. Ablation study showed that decreasing data volume negatively impacted performance with IOU, MIOU, precision, and F1 score mostly decreased. This article was authored by UACAI, HomeJahi, Guyan, and others. We are article.tv, links in the description below.