Abstract:

The principal component analysis (PCA) method combined with JPEG2000 is widely used in hyperspectral image compression. However, the covariance matrix of the PCA only represents the second order statistics. In many applications of hyperspectral image analysis, only preserving the information of the second order statistics is not sufficient. Taking the anomalous pixels for example, more subtle information needs to be captured by using higherorder statistics. To solve this problem, we propose a hyperspectral image compression algorithm using mixed PCA/ICA in conjunction with JPEG2000 standard. Firstly, the PCA is performed for the original hyperspectral image to find the eigenvector matrix WPCA  corresponding to the first m largest eigenvalues. Then, the ICA is employed for the remaining eigenvectors to find n  eigenvector matrix WICA. Finally, the mixed projection matrix, original hyperspectral image and it's mean vectors are embedded into the JPEG2000 bitstream for compression. At different bit rates, the performance of the mixed PCA/ICA+JPEG2000 is evaluated by spatial correlation coefficient (ρ), signal noise ratio (SNR) and spectral angel map (SAM). Experimental results reveal that the proposed algorithm is not only better in spectral correlation reduction, but also can improve spectral fidelity and protect anomaly pixel information.

Key words: hyperspectral image compression;principal component analysis (PCA);independent component analysis (ICA);JPEG2000