Real-time processing of an unsupervised constrained linear spectral unmixing algorithm

Most spectral unmixing techniques require complete a priori knowledge of spectral endmembers in an image scene to be implemented. This knowledge is usually obtained from a complete hyperspectral image scene in either a supervised or unsupervised fashion, which prohibits real-time implementation. In this paper, we present a unique constrained spectral unmixing technique, which develops this knowledge in real-time on a pixel-by-pixel basis. Experiments using HYDICE data were conducted to demonstrate this proposed real-time approach for target detection and classification in hyperspectral imagery. Small modifications to the real-time algorithm were made in the experimental section to alleviate practical problems associated with a true pixel-by-pixel implementation