Reduction of Hyperspectral image based on OSP and a Filter based on Bhattacharyya Distance

Abstract

Abstract— This article proposes a new method to reduce the dimensionality of the hyperspectral image of Pavia. It became obvious to reduce the hyperspectral image, before their classification. This reduction is done by several strategies and approaches according to the literature. The high dimensionality of the hyperspectral image was and remains a challenge to overcome. Since it contains labelled pixels that belong to the target area and others considered as intruders. For this reason, the proposed method aims to extract the classified pixels by the application of the orthogonal projection of the ground truth on the bands then a selection is made adopting a filter based on the minimization of the distance Bhattacharyya inter classes (one against one: band class against ground truth class). Two other distances Jeffries Matusita and Kullback Leibler were applied in the same level of the algorithm in order to validate the appropriate distance, also to confirm the reliability of the process of the new method. The results of the proposed method obtained by SVM-RBF and also KNN demonstrate a remarkable improvement in classification accuracy. The proposed procedure was able to reach over 94% for only 18 bands; hence a simple Bhattacharyya filter got just 91.45%. Keywords—Bhattacharyya, Classification, Extraction, Jefferies Matusita, Kullback Leibler, KNN, Selection, OSP, RBF-SVM. Reduction of dimensionality