A Spy Positive and Unlabeled Learning classifier and its application in HR SAR image scene interpretation

In this paper, we present a Spy Positive and Unlabeled Learning (SPUL) classifier. It is a novel two-step strategy of implementing a positive-and-unlabeled-sample-based classifier. In the first step, by using spy detection, the unlabeled samples are divided into unreliable positive and reliable negative samples. In the second step, the classifier is built using labeled positive, unreliable positive, and reliable negative samples with different and suitable weights. The proposed SPUL classifier is incorporated into a One-Class-Extraction (OCE) framework for High Resolution (HR) Synthetic Aperture Radar (SAR) image scene interpretation. The performance of the SPUL classifier and the SPUL-based OCE framework is presented and analyzed on a TerraSAR-X HR SAR image.