Mixed-pixel classification for hyperspectral images based on multichannel singular spectrum analysis

In this paper, a spectral unmixing technique based on multichannel singular spectrum analysis (MSSA) is applied to derive quantitative information about general land-cover types whose spectra can be determined from the image. The proposed approach can tolerate white noise in the linear model; moreover, we also provide an automatic mechanism to eliminate the undesired singular values as much as possible to get better results. Several experiments for hyperspectral images were conducted to validate the spectral unmixing procedure. Comparisons with the least square orthogonal subspace projection approach were also given.