Band selection based hyperspectral unmixing

Hyperspectral unmixing is the procedure by which the measured spectrum of a mixed pixel is decomposed into a collection of constituent spectra, or end members, and their mixing proportions. However, due to the hundreds of spectral bands contained in the hyperspectral imagery, the large amount of data not only increase the computational loads, but also are unfavorable for the fast hyperspectral unmixing. Hence, dimensionality reduction, which selects the relevant range of wavelengths in the spectrum, is a necessary preprocessing step for hyperspectral unmixing. In this paper, in order to preserve the crucial and critical information, band selection techniques are firstly used to choose the appropriate bands from the original data, and then the unmixing methods are applied. Two recently proposed algorithms, affinity propagation and constrained nonnegative matrix factorization, are respectively adopted for the above two procedures. Experimental results show that the performance of the band selection based hyperspectral unmixing strategy is comparable to that without band selection.