This paper introduces the use of clustering algorithm using weighted features in cluster space classification of hyperspectral data. The best weights are obtained via an optimization process to seek the most compact clusters. This procedure is integrated in a cluster space classification, where the distribution of class of interest data is represented by the set of the clusters generated, instead of adopting the Gaussian distribution assumption. In essence, this is a combined supervised and unsupervised classification methodologies. Experiments were conducted using a HyMap data and the advantages offered by this nonparametric multi-signature classification scheme are demonstrated with improved classification accuracy.