Abstract
This paper presents a method to extract rooftops from aerial images with only visible red, green, and blue bands of data. In particular, it does not require near-infrared data, lidar, or multiple viewpoints. The proposed method uses shadows in the image in order to detect buildings and to determine a set of constraints on which parts can or cannot be rooftops. We then use the grabcut algorithm to identify complete rooftop regions and a method to make corrections that simulate a user performing interactive image segmentation in order to improve the precision of our results. The precision, recall, and F-score of the proposed approach show significant improvement over two very recently published papers. On our test dataset, we observe an average F-score of 89% compared to scores of 68% and 33%.