Effects of band selection on the hyperspectral classification

In the hyperspectral theory, data reduction techniques play an important role in the classification processing as hyperspectral imagery contains an immense amount of data posing many challenging problems such as data storage, computational efficiency, and the curse of dimensionality. Hyperspectral band selection technique is a well-known dimensionality reduction approach which retains the physical meaning of the data. It selects a set of bands from the input hyperspectral dataset which comprises the information needed for subsequent hyperspectral image spectroscopy. The majority of the existing state-of-the-art dimensionality reduction methods set criteria to the spectral information which is derived by the whole wavelength in order to define the optimum spectral subspace. These criteria are not associated with the particular classification task but with the data statistics, such as correlation and entropy values. However, each spectral signature of a particular material has spectral characteristics which contribute to distinguish it from other spectral signatures at specific sequential wavelengths. This paper focuses on investigating the effects of band selection on the classification by exploiting the information of sequential bands. More precisely, it is explored 1) whether classification can be optimized when a different set of initial bands is selected per category; 2) whether there is an optimum subset of sequential bands which lead to more accurate classification results. Experiments comprise application of a well-known classification method, the support vector machine (SVM), on real hyperspectral dataset using all the possible subsets of p sequential bands, where p is equal to the dimensionality of the signal subspace. Evaluation of the classification accuracy leads to remarkable conclusions.