On the use of the Hotelling's T2 statistic for the hierarchical clustering of hyperspectral data

In this work we propose a hierarchical clustering methodology for hyperspectral data based on the Hotelling's T
²
statistic. For each hypespectral sample data, the statistical sample mean is calculated using a window-based neighborhood. Then, the pairwise similarities between any two hyperspectral samples are computed based on the Hotelling's T
²
statistic. This statistic assumes a Gaussian distribution of the data while hyperspectral data have been observed to be long tailed distributed. In order to improve the statistic robustness we use the Fixed Point estimates, and compare them to the classical sample mean estimator. The similarities are then used to hierarchically cluster the hyperspectral data. We give some preliminary qualitative results of the proposed approach over the Indian Pines hyperspectral scene. Results show that the use of the Fixed Point estimator does not significantly affect the clustering results. Further work will be focused on the use of the robust Hotelling statistic.