Feature reduction of hyperspectral image for classification

Abstract

Informative feature extraction from hyperspectral image (HSI) is the primary and most challenging task in the hyperspectral data processing. The rich source of HSI information provides effective ground cover analysis which requires high computational cost and, using the original, classification accuracy suffers from the curse of dimensionality. Therefore, feature reduction has been applied through feature extraction and feature selection. The popularly used unsupervised feature extraction method, Minimum Noise Fraction (MNF), has been applied but the computational cost is high. This paper proposed a band grouping technique using Normalized Mutual Information (NMI) and applies MNF to each individual group called BgMNF. Feature selection can be done with NMI. The extracted feature can be classified using kernel Support Vector Machine (SVM) for performance analysis. Two real HSI is used in experimentation that demonstrates the proposed technique significantly improves the classification accuracy as well as computational cost as compared with the studied methods.