| Perhaps the most common way to distinguish materials is by color. For example, this is typically how one determines, from some distance, whether a material on the ground is grass (green), soil (brown), or asphalt (black). To accomplish this, most digital cameras (along with the human eye) produce images that are comprised of three different colors, called spectral bands: red, green, and blue. The combination of these bands enables material discrimination. Working in the same way, but on a much larger scale, hyperspectral imaging sensors produce images that are comprised of hundreds of spectral bands. This combination of bands enables more accurate and sensitive discrimination of the materials in a scene.;One of the most common ways to make hyperspectral images useful is to perform spectral unmixing. This process can determine what types of materials are in the image as well as where those materials are located within the image. When this is done without access to some sort of reference library of material spectra (i.e. material colors), the processing is called blind or unsupervised spectral unmixing. One of many methods for performing blind spectral unmixing is independent component analysis (ICA). ICA is an unmixing approach that produces outputs, called independent components, that are statistically independent from one another. One problem associated with ICA in the context of the spectral unmixing problem is scale ambiguity. The problem arises because multiplication by a constant value does not affect the independence of two random variables. Scale ambiguity hinders interpretation of spectral unmixing results by preventing comparison of different materials (since they may be scaled differently).;In this dissertation, ICA is examined as a spectral unmixing approach. Various processing steps that are common in many ICA algorithms are assessed to determine their impact on spectral unmxing results. Synthetically-generated, but physically-realistic data are used to allow the assessment to be quantitative rather than qualitative only. Additionally, two algorithms, class-based abundance rescaling (CBAR) and extended class-based abundance rescaling (CBAR-X), are introduced to enable accurate rescaling of independent components. Experimental results demonstrate the improved rescaling accuracy provided by the CBAR and CBAR-X algorithms, as well as the general viability of ICA as a spectral unmixing approach. |
