Linear feature extraction for hyperspectral images using information theoretic learning

In this paper, we propose a new linear feature extraction scheme for hyperspectral images. A modified Maximum relevance, Min redundancy (MRMD) is used as a criterion for linear feature extraction. Parzen density estimator and instantaneous entropy estimation are used for estimating mutual information. Using Instantaneous entropy estimator mitigates nonstationary behavior of the hyperspectral data and reduces computational cost. Based on proposed estimator and MRMD, an algorithm for linear feature extraction in hyperspectral images is designed that is less offended by Hueghs phenomenon and has less computation cost for applying to hyperspectral images. An ascent gradient algorithm is used for optimizing proposed criterion with respect to parameters of a linear transform. Preliminary results achieve better classification comparing the traditional methods.