Comparative performance of fractal based and conventional methods for dimensionality reduction of hyperspectral data

•Dimensionality reduction applied on spectral domain of hyperspectral data.•Fractal dimension has been used as a feature in the reduced dimensional set for the first time.•Enhanced the location information along an SRC by introducing multiple fractal dimensions along a single SRC.•Nature of spectral response unaltered in the reduced domain, thus keeping the physical meaning of the curve unaltered.•While compared with conventional methods, statistically equivalent classification accuracy has been achieved at a much reduced computational burden.

Although hyperspectral images contain a wealth of information due to its fine spectral resolution, the information is often redundant. It is therefore expedient to reduce the dimensionality of the data without losing significant information content. The aim of this paper is to show that proposed fractal based dimensionality reduction applied on high dimensional hyperspectral data can be proved to be a better alternative compared to some other popular conventional methods when similar classification accuracy is desired at a reduced computational complexity. Amongst a number of methods of computing fractal dimension, three have been applied here. The experiments have been performed on two hyperspectral data sets acquired from AVIRIS sensor.