Unsupervised Cluster-Wise Hyperspectral Band Selection using clustering Silhouette information

Due to its multiple bands, a hyperspectral image may provide fine details about a scene, but the high dimensionality may cause both overfitting of the classifier and the Hughes phenomenon.
Band selection can alleviate this problem.
So, this paper addresses this issue by improving a recently published unsupervised band selection algorithm. Specifically, we propose the use of Silhouette information in order to determine the number of clusters ---and iterations--- of the method. This number had been so far determined by the user.
Experiments show that the improved algorithm achieves similar results to those of the original method, with the advantage of being independent of user’s knowledge.