Dimensionality reduction methods address the challenges associated with high dimensional hyperspectral data by giving a low-dimensional representation preserving only the common, global spectral information. However, while classifying the image using reduced representation of the samples, subtle, discriminative information required to distinguish between similar classes may get disappeared. In our paper, we have demonstrated data transformed using how middle principle components emphasize the subtle differences between such classes. Based on this premise, a two-level feature extraction framework for classifying hyperspectral images consisting of similar and distinct classes is proposed. Experimental results indicate that the framework is able to address the issue of confusion in discriminating between similar classes efficiently.