A greedy modular eigenspace-based band selection approach for hyperspectral imagery

Abstract

The greedy modular eigenspaces (GME) has shown effective in hyperspectral feature extraction. The GME was developed by grouping highly correlated hyperspectral bands into a smaller subset of band modular regardless of the original order in terms of wavelengths. It utilizes the inherent separability of different classes in hyperspectral images to reduce dimensionality and further to generate a unique GME feature. This paper takes advantage of the GME to develop a GME-based band selection (GMEBS) for hyperspectral imagery. It selects a subset of non-correlated hyperspectral bands for hyperspectral images using the unique ability of the GME in class separability. The proposed GMEBS algorithm provides a fast procedure to select the most significant features and speeds up the distance decomposition compared to GME features. It also avoids the bias problems of transforming the information into linear combinations of bands as does the traditional principal components analysis (PCA). The proposed GMEBS approach selects each band by a simple logical operation, call GME feature scale uniformity transformation (GME/FSUT), to include different classes into the most common feature modular subset of bands. Interestingly, experimental results show that this simple GMEBS approach is very effective and can be used as an alternative to other band selection algorithms.