Building Footprint Delineation from UAV Image Using Deep Learning

Manual digitization of building footprints from high-resolution images takes more time and human resources, and it is more difficult for future updates. In this paper, building footprints are automatically delineated using a deep learning algorithm, which is a conditional generative adversarial network (CGAN). First, the red, green, blue (RGB) Unmanned Aerial Vehicle (UAV) image of the Yangon Technological University (YTU) campus is manually digitized for building footprints. Second, both UAV image and digitized images are sliced into smaller images, and the sliced images are grouped into a training dataset and validating dataset. Third, the training dataset is used to train CGAN, and after training, a validating dataset is used to test the performance of the trained CGAN. During training CGAN, the learning rate and the number of epochs are adjusted to get minimum training loss. The validated output images are then transformed into vector images, and compared with manually digitized images for evaluation. Three evaluation parameters: completeness, correctness, F1 score are used for evaluation of output images. According to the evaluation results, CGAN can automatically extract building footprints up to 83% completeness, 85% correctness, and 84% F1 score