Toward an optimal analysis of hyperspectral data

In hyperspectral data materials of practical interest usually exist in a number of states and are observed in a number of conditions of illumination. It is thus necessary to characterize them not with a single spectral response but with a family of responses. One of the most versatile means for representing such a family of responses quantitatively is to model each by a normal distribution, as this makes possible classification by assigning the class to a sample based on Bayes rule. This leads to competitive performance only under special circumstances. The purpose of this dissertation is to improve the quantitative definitions of classes by replacing the classical normal model with more flexible and powerful alternatives and thereby investigate the relationship between class definition precision and classification accuracy. One other factor that directly affects the precision of class definitions is the number of labeled samples available for training. Characterizing class data with a limited set of labeled samples may have severe consequences. In a typical setting a remote sensing analyst should either sacrifice from the classifier performance by confining himself to the already available labeled data set or commit more time and effort to acquire more labeled samples both of which comes at a price. The present study also addresses this problem by proposing a semi-supervised binary classifier which seeks to incorporate unlabeled data into the training in order to improve the quantitative definitions of classes and hence improve the classifier performance at no extra cost.