A priori fully constrained least squares spectral unmixing based on sparsity

Abstract

Fully constrained least squares (FCLS) method has been widely applied in estimating land cover abundances within an image pixel through spectral unmixing. With FCLS, if given a large spectra library, researchers have to examine the appropriateness of all endmember signatures, and may mistakenly choose the ones that do not exist in the pixel. Such erroneously selected endmember signatures may lead to overestimation of the non-existing endmembers. In this article, A priori FCLS unmixing method based on sparse unmixing was been pro posed. This methodology contains three steps, including: 1) estimating the abundance of each mixed pixel using sparse unmixing with all the endmembers, 2) chosing the number, type, and signatures of endmembers in every mixed pixel and 3) estimating land cover abundances by FCLS only with chosen endmember signatures. Experimental results suggest that the developed a priori FCLS method is with apparently better performance when compared to those of traditional FCLS and sparse unmixing models. Especially, the improvements of the a priori FCLS are more significant with a larger spectral library and/or a better signal to noise ratio (SNR).