Semi-supervised Locality Preserving Discriminant Analysis for Hyperspectral Classification

In hyperspectral image classification, labeled samples are usually limited and unlabeled data are available in large quantities. This paper presents a novel semi-supervised feature reduction methods, Semi-supervised Locality Preserving Discriminant Analysis (SLPDA), in order to solve the classification problem of hyperspectral imagery. The distinguished feature of the algorithm is the combination of the two graph-based dimensionality reduction methods, supervised LSDA and unsupervised NPE, in a way without extra tuning parameters. The proposed method preserves local geometrical structure in the labeled and unlabeled data, and discriminates data of different classes in labeled samples as far as possible. Experimental results demonstrated that our method had better performance than other traditional dimensionality reduction methods in data sets Indian Pine and Kennedy Space Center.