Dimensionality reduction strategies can be broadly categorized as band selection and feature extraction. Researchers and analysts from the remote sensing community give greater preference to band selection over feature extraction as the latter modifies the original reflectance values of hyperspectral data, making it difficult to understand the behavior of the materials in terms of their reflectance values. However, feature extraction strategies have their own advantages which cannot be ignored. Thus, a two-level, PCA-based band selection framework is proposed to unify the two dimensionality reduction strategies so that benefits of both the strategies can be derived. The proposed approach selects bands based on their relationship with a given set of principal components explained in terms of component loadings, thus keeping the original bands intact. Additionally, contrary to the popular notion that the complete information of all bands is adequately coalesced in the top principal components, middle principal components play a far stronger discriminative role when the competing classes are spectrally confusing to each other. Thus, for each level of classification, a different range of principal components is used to select the bands, on the basis of the level of spectral similarity expected between the classes at each level. Experimental results indicate that the proposed two-level band selection algorithm can select bands with varying levels of discriminative capabilities to effectively classify hyperspectral images consisting of classes spectrally very similar in nature.