The development of ground truth data and accuracy assessment of hyperspectral image classification and spectral unmixing

This work includes both theoretical and practical aspects of hyperspectral image classification and unmixing assessment. The theoretical aspect includes a study of end-to-end performance of hyperspectral classification and unmixing systems. Specifically, it compares widely used current state-of-the-art algorithms with those developed at the University of Puerto Rico at Mayaguez. These include algorithms for image enhancement, band subset selection, feature extraction, supervised and unsupervised classification, and constrained and unconstrained abundance estimation. The classification algorithms are compared in terms of percentage of correct classification. The unmixing algorithms are compared using a new procedure to evaluate unmixing performance is described in this work and tested using coregistered data acquired by various sensors at different spatial resolutions. Techniques for image complexity analysis currently available for automatic target recognizers are studied and adapted in an attempt to predict the performance of the classifiers for different image classes. Performance is generally specific to the image used. The practical aspect included both acquisition and manual classification of ground truth data. Most important among this work is a 1m ground truth map of the Enrique reef in La Parguera Puerto Rico. In addition, 158 calibrated spectral measurements along with their GPS location at submeter resolution were collected. Finally, a hyperspectral image from Hyperion was coregistered with the 1m ground truth to determine the abundaces for each of the pixels in the low spatial resolution image.