Effect of compression on detection in hyperspectral data

Hyperspectral sensors typically obtain tens to hundreds of spectral bands having bandwidths in the 1-10 nm range. Methods for exploiting hyperspectral data commonly use a number of dimension-reducing tools for compression or dimensionality reduction. The motivation for using these methods is to express the relevant information in the data in a smaller dimensional space to improve the computational complexity, reduce bandwidth, improve estimation error, or aid in visualization. These methods may be applied as precursors to detection processing, and include principal components projection and vector quantization. These methods effectively compress the dominant components of scene data. However, because of their bias towards high population materials in the scene, they are not always effective in detection of low concentration materials in hyperspectral data. We examine the effect of these common transforms on detection of man-made materials in collected hyperspectral data and compare to a newly developed method that seeks to preserve detectability of rare events in hyperspectral data.