Principal components-based display strategy for spectral imagery

Abstract

A new pseudocolor mapping strategy for use with spectral imagery is presented. This strategy is based on a principal components analysis of spectral data, and capitalizes on the similarities between three-color human vision and high-dimensional hyperspectral datasets. The mapping is closely related to three-dimensional versions of scatter plots that are commonly used in remote sensing to visualize the data cloud. The transformation results in final images where the color assigned to each pixel is solely determined by the position within the data cloud. Materials with similar spectral characteristics are presented in similar hues, and basic classification and clustering decisions can be made by the observer. Final images tend to have large regions of desaturated pixels that make the image more readily interpretable. The data cloud is shown here to be conical in nature, and materials with common spectral signatures radiate from the origin of the cone, which is not (in general) at the origin of the spectral data. A supervised method for locating the origin of the cone based on identification of clusters in the data is presented, and the effects of proper origin orientation are illustrated.