Application-Dependent Wavelength Selection For Hyperspectral Imaging Technologies

Hyperspectral imaging has proven to provide benefits in numerous application domains, including agriculture, biomedicine, remote sensing, and food quality management. Unlike standard color imagery composed of these broad wavelength bands, hyperspectral images are collected over numerous (possibly hundreds) of narrow wavelength bands, thereby offering vastly more information content than standard imagery. It is this higher information content which enables improved performance in complex classification and regression tasks. However, this successful technology is not without its disadvantages which include high cost, slow data capture, high data storage requirements, and computational complexity. This research seeks to overcome these disadvantages through the development of algorithms and methods to enable the benefits of hyperspectral imaging in inexpensive portable devices that collect spectral data at only a handful (i.e., 5-7) of wavelengths specifically selected for the application of interest.This dissertation focuses on two applications of practical interest: fish fillet species classification for the prevention of food fraud and tissue oxygenation estimation for wound monitoring. Genetic algorithm, self-organizing map, and simulated annealing approaches for wavelength selection are investigated for the first application, combined with common machine learning classifiers for species classification. The simulated annealing approach for wavelength selection is carried over to the wound monitoring application and is combined with the Extended Modified Lambert-Beer law, a tissue oxygenation method that has proven to be robust to differences in melanin concentrations. Analyses for this second application included spectral convolutions to represent data collection with the envisioned inexpensive portable devices. Results of this research showed that high species classification accuracy (\u3e 90%) and low tissue oxygenation error (\u3c 1%) is achievable with just 5-7 selected wavelengths. Furthermore, the proposed wavelength selection and estimation algorithms for the wound monitoring application were found to be robust to variations in the peak wavelength and relatively wide bandwidths of the types of LEDs that may be featured in the designs of such devices