Hyperspectral Image Segmentation, Feature Reduction and Clustering using k-means

In this study, a novel clustering methodology is proposed, which utilizes k-means sequentially for performing feature reduction, segmentation, and clustering on hyperspectral imagery, to extract information. The proposed methodology is a multi-stage framework. Initially, k-means is utilized to perform feature reduction. In the next stage, k-means is again deployed to perform hyperspectral image segmentation, using new feature set obtained from the first stage. Finally, k-means clustering is carried out on segmented hyperspectral image by making use of reduced feature set. To assess the performance of the proposed methodology, experiments are conducted over three sets of hyperspectral images. Purity and Normalized Mutual Information (NMI) score are used for evaluation. The experimental results prove that the proposed methodology has an edge over the other compared clustering methodologies. Results show incorporation of feature reduction and image segmentation techniques leads to significant improvement in accuracies.