Estimation of the number of endmembers via the hubness phenomenon

Estimating the number of endmembers in a hyperspectral image is an important first task in any unmixing chain. Several existing algorithms typically underestimate this number due to the high correlations encountered in spectral data sets. We present a new method for estimating the number of endmembers which does not suffer from this problem, based upon comparing the indegree distribution of the data set with artificially generated distributions. This approach exploits the hubness phenomenon, the observation that indegree distributions show a high dependence on the intrinsic dimensionality of the data set. We demonstrate that this method outperforms several popular alternatives on artificial data sets, and that realistic values are obtained on real hyperspectral data sets.