Automated Analysis of Reflection Mode Terahertz Hyperspectral Images

Abstract

A suite of algorithms and associated procedures, originally developed for mineral exploration applications, are adapted for application to terahertz hyperspectral images measured in reflection mode. Such data are often quite noisy due to the low reflectivity of many materials at terahertz frequencies. The algorithms and procedures are based on an extended linear mixture model consisting of two parts. The first part, called the “foreground”, models the distinguishing parts of the spectra of materials (including mixtures) of interest (especially their diagnostic absorption features). The second part, called the “background”, models parts of the spectra that are typically of lesser interest, such as variation in low frequencies and water vapor. The model and procedures are exemplified with a spectral library of six materials and are applied to three hyperspectral images, one consisting only of pure pellets, some of which are not in the library, and two of which contain both pure and mixed pellets of three of the materials in the library. The associated procedures include the following: estimating the number of materials in the mixture at each pixel; identifying pixels with materials that are well modeled by the background terms only; identifying pixels with materials not in the library; and identifying pixels containing metal. Finally, this article concludes with a discussion of some outstanding issues.