A novel method to estimate the number of endmembers in hyperspectral images based on the virtual dimensionality concept

Abstract

This paper presents a modification of the Harsanyi-Farrand-Chang (HFC) method to automatically estimate the number of endmembers of hyperspectral images when assuming a linear mixing model. The proposed unsupervised algorithm dynamically determines the probability of false alarm required as an input of the HFC method, with independence of the amount of noise, spatial dimensions, and/or the real number of endmembers of the hyperspectral image under processing. The Automatic HFC (A-HFC) method has been tested using synthetic and real hyperspectral images, demonstrating in all the cases an equal or superior performance than HFC in terms of precision when computing the number of endmembers. Additionally, for the real image Cuprite, the number of endmembers coincides with the number of endmembers obtained with the HFC method and it is similar to the one given by the hyperspectral subspace identification method (HySime), which definitely confirms the accuracy of the proposed method.