A new sparsity-aware feature selection method for hyperspectral image clustering

In this paper a new feature selection method suitable for hyperspectral image clustering is presented. The proposed spectral band selection method selects bands that exhibit significant discrimination ability, based on the optimization of a sparsity promoting cost function. This allows clustering algorithms to export results of the same quality compared to cases where all spectral bands are used, while, in some cases, it allows the unravelling of some less-obvious patterns. Experimental results on real hyperspectral data sets highlight the enhanced performance of the proposed technique.