Unsupervised estimation of the number of endmembers in hyperspectral data

Estimate the number of endmembers in hyperspectral data is a significant step in the process of the spectral unmixing. Since most images contain noise that is not independent and identically distributed (i.i.d.) across bands, methods that assume i.i.d. noise are often avoided. This paper introduces a noise-whitening process into the state-of-the-art signal subspace estimation method. The noise covariance matrix of the image data is used to whiten the noise variances to unity, then the subset of eigenvalues that best represents the signal subspace is selected in the least squared error sense. The results derived from the classical methods and the improved method using simulated hyperspectral data and real hyperspectral data are presented and discussed.