Robust Principal Components based on Projection Pursuit for hyperspectral band reduction

Supervised classification techniques use labeled samples in order to train the classifier. In a hyperspectral image, usually the number of such samples is limited, and as the number of bands available increases, this limitation becomes more severe. Such consequences suggest the need for reducing the dimensionality via a preprocessing method. This reduction should enable the estimation of feature extraction parameters to be more accurate. In order to mitigate this problem a technique referred to as robust principal component analysis by Projection Pursuit (pcaPP) is used in this paper. Our experiments on real hyperspectral dataset (AVIRIS 1992 Indian pine image), proves that the pcaPP technique is not only robust but offers more classification accuracy than classical PCA on the SVM classifier.