Progressive neighbors pursuit for radar images classification

Abstract Finding appropriate class-separating metric and labeling rules is crucial in the construction of image classifiers. In this paper a Divergence-Chebyshev Neighbors Pursuit (DCNP) algorithm is proposed for rapid Polarimetric Synthetic Aperture Radar (PolSAR) image classification. First, an information-theoretic divergence is defined to measure the similarity of polarimetric features between pixels. Then a divergence-Chebyshev distance is defined to reveal the affinity of pixels in both the polarization and spatial domains. Moreover, inspired by human’s learning characteristic that the knowledge is learned little by little, the DCNP algorithm is designed to progressively determine the labels of unknown pixels. Some experiments are conducted on several real PolSAR image datasets and the results show that our method can achieve accurate classification with a small number of labeled data, and outperforms its counterparts in terms of several guidelines.