Reducing the dimensionality of hyperspectral data using diffusion maps

We examine the analysis of hyperspectral data produced by the Hy-perspectral Core Imager of AngloGold Ashanti. The dimension of the data is reduced using diffusion maps and the data is then clustered using three divisive clustering strategies. Divisive k-means, PDDP and the NCut algorithm are used. It is shown that the clusterings produced are reasonably accurate compared to a reference clustering, but superior with respect to an internal quality evaluation. Moreover, using a divisive algorithm makes it possible to keep track of inter-cluster similarities. It is also shown that by embedding sample spectra in a dataset it is possible to identify particular minerals within the cluster.