Urban area extraction from Polarimetric SAR imagery using only positive samples

In this paper, we present a study of extracting urban areas from Polarimetric Synthetic Aperture Radar (PolSAR) images using only positive samples. We solve this problem by learning a standard binary classifier (urban/non-urban) given an incomplete set of positive samples (urban) and a set of unlabeled samples (some of which are urban and some of which are non-urban) based on the work of Elkan and Noto. Our experiments on RADARSAT-2 fully PolSAR data show that learning with only positive samples can significantly reduces the manual work to select completed positive and negative samples that required by a traditional binary classifier, while providing satisfactory results. Meanwhile, multiple diverse features can be effectively combined for better extraction accuracy.