Cluster-based virtual band construction for hyperspectral classification

Abstract

Hyperspectral image classification is challenged by high spectral dimensionality, inter- band redundancy, and limited labeled training data. This thesis evaluates virtual band construction as a dimensionality reduction strategy in which spectrally similar bands are grouped into clusters and merged into representative features using weighted linear combinations. Four datasets - Botswana, Indian Pines, Pavia University, and a field- collected Montana burn plot - were reduced to five virtual bands using five clustering methods and four merging strategies: uniform, entropy-based, inverse redundancy, and PCA-based. Classification performance was assessed using Random Forest and k-Nearest Neighbor classifiers under an 80/20 train-test split. Merging-based representations matched or exceeded the accuracy of conventional band selection baselines across all datasets, with the most pronounced improvements on spectrally complex scenes such as Indian Pines, where UMRMR clustering with PCA-based merging achieved 0.7455 RF accuracy compared to 0.5684 for selection without merging. Cluster weight analysis showed that merging strategies distribute contributions across neighboring wavelengths rather than isolating single bands, preserving intra-cluster spectral structure while maintaining a direct connection to the original wavelength measurements.