An Improved Class-wise Principal Component Analysis Based Feature Extraction Framework for Hyperspectral Image Classification

Abstract

It has been a challenging task for researchers to improve the classification accuracy of hyperspectral images in recent years. High dimensionality or the high number of spectral bands with respect to a comparatively less amount of training sample causes higher misclassification rate. The problem is known as 'Hughes Phenomenon'. To overcome this problem many feature selection as well as feature extraction methods are applied to reduce the dimensionality of hyperspectral image. In this paper, we introduce a new framework of class-wise Principal Component Analysis (PCA) operation for the dimensionality reduction of hyperspectral images. Our proposed approach comprises four major steps. The first step involves separating the pixels according to different existing classes in the image. Secondly, we considered the problem of feature extraction with PCA. Thirdly, we took the same number of features, extracted with PCA, from each class considering their maximum average variance. In the final step, we used the SVM classifier to find out the classification results. Experiment results show that our proposed approach turns out to be better than the primary classification results. The primary classification accuracy shown by the SVM classifier is 89.02% and the classification accuracy of our proposed method is 95.44%.