Potential and limitations of band selection and library pruning in sparse hyperspectral unmixing

Sparse regression using spectral libraries is nowadays a widely used technique for hyperspectral data unmixing. Assuming that the potential endmembers are collected in a large database of spectra, sparse unmixing finds the fractional abundances of a reduced set of constituent materials by solving convex optimization problems which target, at the same time, low data reconstruction errors. The large amount of data, jointly with other limitations related to the internal characteristics of the large spectral libraries (such as spectra similarity), affect the performance of sparse unmixing algorithms in terms of accuracy and running time. Recently, an efficient method based on a multi-measurement vector (MMV) approach was proposed to select, from a large library, suitable spectra for unmixing. Many research efforts have also been devoted to data dimensionality reduction techniques, of which band selection is very popular. In this work, we investigate the effects on the unmixing performance when the two techniques are applied simultaneously. Our experiments show that important improvements can be achieved, depending on several factors, such as: data noise, number of endmembers present in the dataset, number of spectra retained from the library, the order of operations, accuracy of data subspace estimation.