Classification and anomaly detection algorithms for weak hyperspectral signal processing

Abstract

In applications involving weak light signal like hyperspectral or time distributed signals obtained in applications involving laser induced fluorescence spectral detection, fluorescence lifetime imaging, Raman Spectroscopy or hyperspectral imaging in low light environment, the photons arrive at such a rate that they can be counted or have to be intensified to obtain a usable signal. Detection and classification algorithms need to be designed and evaluated for weak hyperspectral signal processing. A new algorithm, Adaptive Shot Noise (ASN) based on the assumption that a signal respects the Poisson multivariate distribution has been developed using the method of the maximum likelihood. This algorithm demonstrates the capability to be used for detection and classification. Using Monte Carlo simulations its performances are compared with the Adaptive Coherence Estimator (ACE) classification and with an Integrated Signal Algorithm (ISA) and ACE for detection. This new algorithm provides a small increase in performance compared to ACE in very weak signal conditions for classification and in some conditions better performance over both ACE and ISA in detection. The algorithm behavior like ACE shows sensitivity to assumption on the spectral characteristics of the source for the detection, which is not the case for ISA.