Sparse feature extraction for hyperspectral image classification

Due to the high dimensionality and redundant spectral information in a hyperspectral image (HSI), principal component analysis (PCA) and linear discriminant analysis (LDA) are commonly-used for its feature extraction. By converting PCA and LDA to regression problems and imposing l₁-norm constraint on the regression coefficients, sparse principal component analysis (SPCA) and sparse discriminant analysis (SDA) have been developed for improved feature extraction. Furthermore, recently sparse tensor discriminant analysis (STDA), reserving useful structural information and obtaining multiple interrelated is also proposed. Their performance in HSI classification is investigated in this paper. Experiment results demonstrate the effectiveness of these sparse feature extraction methods, especially for STDA, which outperforms the traditional linear counterparts without maintaining spatial relationships among pixels, such as PCA and LDA.