Investigating of Geometrical and Asymptotical Properties of Hyperspectral Data for Discriminant Analysis

Hyperspectral images provide abundant information about objects. The high dimensionality of such images arise various problems such as curse of dimensionality and large hypothesis space. There are two methods to overcome the high dimensionality problem which are band selection and feature extraction. In this paper we present a feature extraction method based on an angular criterion; this method is defined so that minimizes angle between mean vector and samples with in each class and maximizes the angle between mean classes and simultaneously satisfies fisher criterion. It explores other aspects of pattern in feature space and tries to discriminate classes with respect to geometric parameters. We have employed the angular and the fisher criteria for feature extraction also the spectral angle mapper (SAM) and minimum distance (MD) classifiers are used for image classification. The results demonstrate that this method can improve the discrimination of objects in feature space and improve the classification accuracy of SAM classifier.