Semi-Supervised Classification considering space and spectrum constraint for Remote Sensing Imagery

Abstract

Semi-Supervised Classification (SSC), which makes use of both labeled and unlabeled data to determine classification borders in feature space, has great advantages in extracting classification information from mass data. In this paper, a novel SSC method is proposed based on Remote Sensing Image character, which utilizes both spacial and spectrum feature relationship between labeled and unlabeled data. Spectral relationship of all label and data are determined by a super-ellipsoid of Gaussian Mixture Model (GMM) in spectral space, while their spacial relationships are constrained in image patches. In order to ensure all pixels in one image patch are belonging to one category, image is segmented by edge based Mean Shift, a novel image segmentation method, and divided into sub-patches when classification if desired. Spectrum and space features are combined by a Wins Get All (WGA) algorithm, in which category of each image patch is voted by all of its subordinate pixels. These pixels are classified firstly by Bayesian Inference based on GMM Probability Density Function (PDF). Then the entire SSC is completed in a loop process, in each iteration step, pixels are classified by WGA algorithm and utilized as label data in the next iteration. Experiments shows this method could get a high accuracy result with small samples.