Joint subspace detection of hyperspectral targets

Abstract

Joint subspace detection (JSD) arises from a Bayesian formulation of the binary detection problem, as contrasted with the "fixed but unknown parameter" approach that generates the generalized likelihood ratio (GLR) test. The Bayesian philosophy allows the incorporation of prior knowledge gleaned from empirical experience into the design of a detection algorithm. The knowledge appears in the form of probability distributions for parameters considered deterministic in the GLR method. An example of this principle, called complementary subspace detection, has been applied to hyperspectral data and, with appropriate subspace selection, is shown to outperform the traditional detection techniques over a wide range of assumed prior knowledge of target distribution.