Mean shift based clustering of neutrosophic domain for unsupervised constructions detection

Automation has been a hot issue in constructions extraction, but there has not yet been a universally accepted algorithm. Commonly, constructions are extracted by user-defined thresholds, and they have to be adjusted with the variation of images and types of constructions. To overcome the shortages, an unsupervised algorithm to extract constructions is proposed in this paper. It adopts mean shift clustering in neutrosophic set domain to segment images, which makes it possible to detect constructions with a stable threshold. The algorithm is compared with three welcomed and recently developed supervised techniques by six study images with two sorts of resolutions. Experiments show that among the four algorithms, the method proposed in this paper performs best in constructions detection. It not only maintains the original shape of buildings, but also generates extracted constructions as a neat whole. Furthermore, the new method has stronger robustness when faced with images with different resolutions and imaging qualities. As tests show that the new algorithm can reach a kappa coefficient of 0.7704 and an accuracy of 89.8054%, which are relatively high in constructions extraction, it can be a robust unsupervised technique to extract constructions.