Linearly constrained minimum variance beamforming approach to target detection and classification for hyperspectral imagery

Abstract

Subspace projection and maximum likelihood have been used for hyperspectral image classification. One of their disadvantages is the requirement of a priori knowledge about material signatures resident in images which may be difficult to obtain in many practical applications. In order to resolve this problem, an approach, called Constrained Energy Minimization (CEM) was developed for target detection where the only required knowledge is the target signature of interest. It was derived based on the Minimum Variance Distortionless Response (MVDR) in array processing. However, CEM has been also shown to be very sensitive to the knowledge being used to describe the target signature. This paper develops a more general approach, called Linearly Constrained Minimum Variance (LCMV) beamforming which includes CEM as a special case. Unlike CEM which considers a single constraint, LCMV uses multiple signatures to extend the capability of CEM to various target classification. More specifically, LCMV can be implemented as an estimator for target classification while CEM can be used as a detector for target detection.