Extraction of Built-up Areas From Fully Polarimetric SAR Imagery Via PU Learning

In this paper, we propose a PU learning (i.e., learning from positive and unlabeled data, which trains a binary classifier using only PU examples) based method for extracting the built-up areas (BAs) from fully polarimetric synthetic aperture radar (PolSAR) imagery. The key feature is that there are no labeled negative training data, thus the traditional classification techniques are not applicable. To solve this problem, we use a two-step strategy-based PU learning. In the first step, an improved algorithm yields reliable negative samples from an unlabeled set. In the second step, we apply a support vector machine iteratively to these negative samples, existing positive samples and the remaining unlabeled samples. Finally, we select a classifier after convergence. To make the method suitable for BA extraction from PolSAR imagery, an extended scattering mechanism-based statistical feature using the adaptive model decomposition is introduced as the feature descriptor. Experimental results for RADARSAT-2 PolSAR data sets demonstrate the effectiveness of our method, which achieves satisfactory accuracy with less manual labeling.