Estimation of the Intrinsic Dimension of Hyperspectral Images: Comparison of Current Methods

Abstract

The intrinsic dimension (ID) of a hyperspectral image (HSI) is an important prior knowledge for unsupervised unmixing. Incorrect determination of this number may have adverse effects on the unmixing results. Several methods have been developed to determine the ID, including Harsanyi–Farrand–Chang (HFC), Hysime, and random matrix theory (RMT). Previous work has shown that real HSI images could contain a certain amount of spectrally correlated noise, and noise approximation as well as ID estimation would suffer from it. This paper compares the performance of ID estimation methods with respect to various noise approximation methods, types of data, and parameters such as noise levels and correlation, noise approximation methods, and number of endmembers. It shows that a significant improvement can be obtained with most ID estimation methods by adapting them to the case where spectral correlation in the noise is considered as well.