Hyperspectral imaging produces high-dimensional data that challenge conventional classification due to redundancy and computational burden. In this study, we evaluate virtual band construction as a dimensionality reduction strategy, combining clustering and merging techniques to generate representative spectral features. Four datasets—Botswana, Indian Pines, Pavia, and a Montana burn plot—were analyzed using five clustering methods and four merging strategies, with performance assessed via Random Forest and k-Nearest Neighbor classifiers. Results indicate that merging a cluster of bands consistently maintains or improves classification accuracy compared to normal band selection methods when using these classifiers. This work demonstrates the potential of virtual bands as a viable alternative to traditional band selection, providing a framework for efficient hyperspectral image classification.
