Vision-optimized Image-adapted Projections for Visualization of Hyperspectral Imagery

Hyperspectral data visualizations are useful as a background layer to labeling information in the hyperspectral scene such as classification information, locations, or geographic features. Given a hyperspectral image H , where the ith-jth pixel Hij is a d-dimensional vector representing reflectance at d wavelengths, any dimensionality-reduction method an be used to reduce the d dimensions down to three dimensions, and then display the three dimensions as the R, G, and B channels of a standard display. In particular, principal components analysis is a standard method for hyperspectral visualization, dating back to 1973 [1]. However, problems with PCA and most image-adaptive visualizations to date include: