Multiscale spectral-spatial hyperspectral classification using Multilinear PCA and contourlet transform

Multilinear principal component analysis (MPCA) is explored for hyperspectral imagery classification in this study. MPCA is first applied to the original image data, in 3 ^rd -order tensor form, to reduce the dimensionality. Then, a multiscale transform technique (i.e., contourlet transform, etc.) is applied to each of the principal components (PCs) generated from MPCA. The resulting transform coefficients can be used as spatial features. These multiscale spatial features are further combined with spectral features (e.g., PCs) to yield spectral-spatial features for hyperspectral classification. The experimental result shows that improvement is obvious when the reduced dimensionality is small.