Non Zero Mean Adaptive Cosine Estimator and Application to Hyperspectral Imaging

Abstract

The Adaptive Cosine Estimator (ACE) has become a popular detection scheme in many applications. Similarly to the majority of detection schemes, it assumes a zero mean noise. In some domains, such as hyperspectral imaging, this assumption no longer holds and this algorithm has to be adapted. In this paper we revisit the use of ACE in a non zero mean context.We consider the case where the data under test and the training samples differ from one scaling factor on the mean and one scaling factor on the covariance matrix. We derive two-step generalized likelihood ratio tests for both the additive model and the replacement model and show that the new detectors differ in the way the mean value is removed. A real data experiment shows that they outperform the standard version.