Feature Extraction-Selection Scheme for Hyperspectral Image Classification Using Fourier Transform and Jeffries-Matusita Distance

Hyperspectral Image Classification represents a challenge because of their high number of bands, where each band represents a random variable in the classification system. In the first place, the computational cost can be higher because of large data volume during processing. In addition, some information can be redundant or irrelevant; furthermore, it maybe not a discriminatory. Consequently, a classifier has a little biased information related to the classes resulting in lower accuracy rates. In this work, we describe a novel methodology in performing feature extraction in classification as well as in efficient feature selection based on coefficients obtained via Discrete Fourier Transform (DFT) for signals by linking the bands of the images and making a selection by Jeffries-Matusita distance criterion. To test the experimental accuracy of current proposal, we employ three hyperspectral images justifying its performance against other state-of-the-art methods using Principal Components Analysis (PCA) feature extraction algorithm in combination with the Jeffries-Matusita distance criterion for its components selection and employing a Support Vector Machine (SVM) for classification.