Building detection from high-resolution aerial images plays a vital role in various applications such as urban planning, disaster management, and 3D building modelling. Recent advances in deep learning have significantly improved the performance of building segmentation systems by leveraging the spatial and contextual information present in high-resolution images. In this study, we explore two semantic segmentation models for automatic building detection, from the family of U-Net architectures, trained on annotated image datasets that have been produced from the available aerial image, its corresponding ground truth and data augmentation techniques. The first is a classic U-Net model trained from scratch, while the second is a U-Net++ architecture that uses a pretrained 34-layer ResNet as its backbone. Both quantitative and qualitative examination of the results on two sample building blocks from the same aerial image showed excellent performance and significant improvement in segmentation accuracy compared to a baseline shallow architecture trained with traditional pixel-wise classification schemes. The evaluation metrics used to assess model performance are precision, recall, F1-score, and Intersection over Union. Visual examination of segmentation results on a building block from a different aerial image, for which no ground truth was available, shows that besides some segmentation inaccuracies primarily in shadow regions or regions with color similarities with the background, the models can produce acceptable segmentation results. By incorporating image pre-processing and post-processing techniques, it is expected to further improve the accuracy of the produced building footprints.