Folded-PCA Based Hybrid Dimension Reduction for Effective Classification of Hyperspectral Image

Abstract

Dimension reduction from higher dimensional hyperspectral image (HSI) data cube has grown into a significant area of research for efficient classification of ground objects. The HSI data cube is a set of numerous highly correlated narrow spectral bands. For effective classification of hyperspectral image, dimension reduction strategies are performed using feature extraction and/or feature selection methods. Standard unsupervised feature extraction method Principal Component Analysis (PCA) has been used frequently for band reduction. But PCA suffers from limitation such as failure of extracting inherent structure of HSI data because of its global variance dependency. Folded-Principal Component Analysis (FPCA), an improvement of PCA, overcomes this problem by considering both the global and local structures of HSI with less computation and memory requirements. In this paper, a hybrid approach is proposed where FPCA is applied to produce new features from the original spectra bands. Then feature selection is performed on the extracted features using normalized Mutual Information (nMI) to select the relevant features. Finally, Kernel-Support Vector Machine (K-SVM) is applied to estimate the classification accuracy of the reduced data cube. The proposed method (FPCA-nMI) is assessed on a real mixed agricultural dataset and achieved the highest classification accuracy of 97.92%, outperforming the baseline approaches.