Clustering of Hyperspectral Images using Entropy based Multiple Features (Bands) Set Selection

This study suggests a novel segmentation-based clustering algorithm that applies entropy based local feature selection to choose the top bands for each cluster. A framework with numerous steps constitutes the proposed methodology. k-means is initially used as an image segmentation technique. Then, Shannon Entropy is utilized to determine the significant clusters from the generated segments. Finally, cluster map is obtained after merging insignificant clusters into significant clusters. Further, multiple feature set obtained through entropy are also utilized while performing clustering. The performance of the proposed methodology is examined using three sets of hyperspectral images. Adjusted Mutual Information and Overall Accuracy are used as evaluation criteria. The results of the study demonstrate that the proposed methodology performs better than the other segmentation methodologies that were evaluated. The results indicate that accuracy is much enhanced by selecting multiple feature set while performing clustering.