Wrapper Based Principal Component Selection For Hyperspectral Image Classification

This paper proposes a Wrapper based Principal Component Selection (WPCS) method for improved classification accuracy of hyperspectral images. First, the hyperspectral image is transformed into feature space using the widely used Feature Extraction (FE) method such as the Principal Component Analysis (PCA). Next, the proposed WPCS method with an improved objective criterion has been applied on the PCA image to obtain the best heuristic set of Principal Components (PCs) with respect to the selected classifier. The generality of the method has been proven by using the widely used conventional Support Vector Machine (SVM) and a deep learning based 1-Dimentional Convolutional Neutral Network (1D-CNN) as wrappers. This paper also presents the comparison between the best heuristic PCs of SVM and 1D-CNN to draw interesting findings for improving classification accuracy. Simulations of the proposed method have been carried out on the real and challenging hyperspectral datasets such as Pavia University, Indian Pines, and Salinas, and the results are compared with representative methods such as SVM wrapper, K-Nearest Neighbour (KNN) wrapper and the state-of-the-art Distance Density based Convolutional Neural Network (DDCNN) method to substantiate the effectiveness of the proposed method.