An invariant display strategy for hyperspectral imagery

Abstract

Remotely sensed data produced by hyperspectral imagers contains hundreds of contiguous narrow spectral bands at each spatial pixel. The substantial dimensionality and unique character of hyperspectral imagery requires display techniques that differ from traditional image analysis tools. This study investigated the appropriate methodologies for displaying hyperspectral images based on the physical principles of human color vision and a generalized set of linear transformations. Principal components (PC) analysis is a powerful tool for reducing the dimensionality of a data set, and PC-based strategies were explored in creating a broadly applicable image display strategy. It is shown that the invariant display strategy and generalized eigenvectors developed within this study offer a first look capability for a wide variety of spectral scenes. PC transformations utilizing this generalized set of eigenvectors allow for real time' initial classification. Detailed investigation of the relationship between the PC eigenvectors and dissimilar image content shows that this strategy is robust enough to provide an accurate initial scene classification.