Feature reduction and classification of high dimensional hyperspectral data

The problems of feature reduction and classification in high-dimensional hyperspectral (HS) data are addressed. Both feature selection and feature extraction algorithms are developed for dimensionality reduction. The new high-dimensional branch and bound feature selection algorithm uses the Kallbuck-Leibler distance to select a subset of 30 features out of the original high-dimensional features and uses the modified Branch and Bound algorithm to select an optimal subset from the 30 features. For feature extraction, the new high-dimensional generalization discriminant features select projection vectors that maximize a new criterion function that provides both class generalization and discrimination. These algorithms are shown to be of use on HS data for two product inspection problems and for two target detection problems. These applications include agricultural product inspection of almonds and corn kernels in HS data and detection of military vehicles and land mines in multispectral (MS) imagery. The results obtained with these new algorithms are compared to other well-known algorithms. For agricultural product inspection, it is always preferable to favor one class over the other, i.e., it is important to select a preferable operating point rather than the one with highest overall classification rate. The new confidence-clustering radial basis function (RBF) neural network adjusts local RBF parameters to generate preferable operating points.