Data-Driven Approach for Hyperspectral Band Selection

The primary objective of dimensionality reduction (DR) methods is to remove the information redundancy from the spectral dimension of hyperspectral imagery (HSI) to improve classification accuracy. In most of the DR methods, the required number of selected/extracted bands is given by the user. However, in reality, it is difficult to perceive the required number of bands before the analysis starts. A particular number of (selected or extracted) features cannot represent different datasets equally well, and the number of discriminating bands are data dependent. Hence, a data-driven approach is required to be established that will not only find the informative and discriminating bands but also provide their optimum number. This chapter focuses on this aspect of HSI band selection and proposes two data-driven band selection algorithms (unsupervised and supervised). The unsupervised data-driven band selection (BS) approach employs multi-featured analysis and signal-to-noise-ratio (SNR)-based band prioritisation for selecting discriminating bands. The signal quantisation process is used in the supervised data-driven approach for distinctly identifying each class signature pattern using a minimum number of bands, which are again refined by removing highly correlated bands. Both methods are compared with state-of-the-art band selection methods and found to be effective in selecting distinct bands from HSI.