COOPERATIVE MODEL BASED ON NMF AND SPECTRAL LIBRARY FOR HYPERSPECTRAL UNMIXING

The significance of hyperspectral unmixing is to separate endmembers from mixed spectral signatures and calculate their fractional abundances. Hyperspectral unmixing can be generally divided into blind and semi-blind unmixing, in which the former extracts endmembers directly from the given data and estimates their abundances, while the latter selects end endmembers from a prior overcomplete spectral library and calculate their abundances. However, the blind unmixing results based on non-negative matrix factorization (NMF) are usually of low accuracy and unstable, while the semi-blind unmixing results based on sparse regression have high accuracy but depend on the strict requirement that the spectral library must be overcomplete. In order to overcome the shortcomings of the two classical methods, a new cooperative model based on NMF and spectral library (called CoNMFSL) is proposed to solve the unmixing problem when the given spectral library is incomplete. Through the fusion of these two methods, our model is designed to make full use of partial prior information of the incomplete spectral library, that is, some endmembers are selected from the spectral library, and the remaining endmembers and their abundances are estimated by blind decomposition. Extensive simulations and real-data experiments demonstrate CoNMFSL can not only get more accurate and stable results than blind unmixing, but also relax the application conditions of semi-blind unmixing.

Automated summary

The significance of hyperspectral unmixing lies in separating mixed spectral signatures into endmembers and calculating their fractional abundances. However, blind unmixing based on non-negative matrix factorization (NMF) often yields inaccurate and unstable results, while semi-blind unmixing based on sparse regression requires an overcomplete spectral library. To overcome these limitations, a cooperative model called CoNMFSL is proposed, which combines NMF and spectral library to utilize partial prior information from the incomplete spectral library. Extensive simulations and real-data experiments demonstrate that CoNMFSL achieves more accurate and stable results compared to blind unmixing, and relaxes the application conditions of semi-blind unmixing.