Application of semi-supervised dimensionality reduction for hyperspectral image classification

Dimensionality reduction techniques have become an important issue concerns of hyper-spectral image processing and application A semi-supervised dimensionality reduction (SSDR) for classification of hyper-spectral image is applied in this paper. This method employed both labeled and unlabeled data with pairwise-constraints to obtain a set of projective vectors such that intrinsic structures of image as well as the pairwise constraints can be preserved in the projective low-dimensional space. To evaluate the method, a case study of Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) image is implemented, and the experimental results validate the applicability and effective of the algorithm. Comparisons with principal component analysis (PCA) and Fisher discriminate analysis (FDA) are also conducted, and the result demonstrates that the SSDR can significantly improve classification accuracy.