Matched affine joint subspace detection in remote hyperspectral reconnaissance

Abstract

The GLR (generalized likelihood ratio) test has been invoked for several decades as a prescription for generating target detection algorithms, when limited prior knowledge makes a theoretically ideal test inapplicable. Many popular HSI (hyperspectral imaging) detection algorithms rely ultimately on a GLR justification. However, experience with real-time remotely deployed detection systems indicates that certain heuristic modifications to the classic algorithm suite consistently produce better performance. A new target detection test, based on a Bayesian likelihood ratio (BLR) principle, has been used to explain these results and to define a broader class of detection algorithms. The more general approach facilitates the incorporation of prior beliefs, such as that gleaned from experience in measurement programs. A BLR test has been used to generate a new family of HSI algorithms, called matched affine joint subspace detection (MAJSD). Several examples from this class are described, and their utility is validated by detection comparisons.