An improved semi-supervised local discriminant analysis for feature extraction of hyperspectral image

Abstract

We propose an improved semi-supervised local discriminant analysis (ISELD) for feature extraction of hyperspectral image in this paper. The proposed ISELD method aims to find a projection which can preserve local neighborhood information and maximize the class discrimination of the data. Compared to the previous SELD, the proposed ISELD better models the correlation of labeled and unlabeled samples. Experimental results on an ROSIS urban hyperspectral image are encouraging. Compared to some recent feature extraction methods, our approach has more than 2% improvements as the training sample size changes.