On the use of spectral libraries to perform sparse unmixing of hyperspectral data

In recent years, the increasing availability of spectral libraries has opened a new path toward solving the hyperspectral unmixing problem in a semi-supervised fashion. The spectrally pure constituent materials (called endmembers) can be derived from a (potentially very large) spectral library and used for unmixing purposes. The advantage of this approach is that the results of the unmixing process do not depend on the availability of pure pixels in the original hyperspectral data nor on the ability of an endmember extraction algorithm to identify such endmembers. However, resulting from the fact that spectral libraries are usually very large, this approach generally results in a sparse solution. In this paper, we investigate the sensitivity of sparse unmixing techniques to certain characteristics of real and synthetic spectral libraries, including parameters such as mutual coherence and spectral similarity between the signatures contained in the library. Our main goal is to illustrate, via detailed experimental assessment, the potential of using spectral libraries to solve the spectral unmixing problem.