Supervised unconstrained and constrained least squares unmixing in hyperspectral imagery

Spectral mixture analysis of hyperspectral imagery allows for detection, classification and quantification of targets present in the imaged scene. It can be difficult to characterize the performance of spectral unmixing techniques at detection, classification and quantification of field data at the subpixel level due to limited ground truth, especially for mixed pixels. Fortunately, the SpecTIR Hyperspectral Airborne Experiment (SHARE) 2012 contains a set of targets specifically designed to test spectral unmixing algorithms. In this paper we explore the performance of an unconstrained and three constrained least squares methods for supervised spectral unmixing. Each of the three methods provides an estimate of the abundance of known targets which can be used for detection, classification and quantification. A detailed evaluation of these spectral unmixing techniques on the SHARE 2012 hyperspectral data is used to demonstrate the performance of each method at supervised target detection, classification and quantification.