In this paper an approach for dimension reduction of the hyperspectral image using scale invariant feature transform (SIFT) is introduced. Due to high dimensionality of hyperspectral cubes, it is a very difficult task to select few informative bands from original hyperspectral remote sensing images. Band with maximum amount of non-redundant information are chosen using the dissimilarity matrix obtained from scale invariant feature transformed image. The performance of the dimension reduction technique is analyzed by implementing a post-processing technique named spectral un-mixing. Spectral unmixing is the process of extracting end members and generating their abundance maps. End members are extracted with these selected informative bands using N-FINDR and abundance maps are generated using fully constrained least square estimation. The simulation software used for implementation of algorithms is MATLAB. Qualitatively and quantitatively the proposed feature based approach has been analyzed with application to spectral unmixing by comparing with two well-known existing dimension reduction techniques namely principal Component Analysis and Linear Discriminant Analysis. Hyperspectral images finds application in astronomy, agriculture, geosciences and surveillance.