Article Preview
Top1. Introduction
The well-known Hughes problem in HSI lowers the classification performance due to its large dimensionality and limited training samples (Subhashree et.al, 2019). The solution to the above said problem can be achieved, either using selection or reduction framework. Presently the Band Selection (BS) is focused and further exploration on similar approaches are presented. Due to the presence of too many, unwanted, redundant and nosy bands make the classification task more erroneous and complex in hyperspectral images. Band selection techniques handle this problem appropriately by selecting the more informative bands subset without disturbing the actual bands structure of the hyperspectral image. If the Ground Truth (GT) information is incorporation then, the BS is expressed as supervised (Ram et al.,2019a, He et al., 2010) or semi-supervised and if GT not incorporated then, it is termed as unsupervised (Swarnajyoti et al., 2015, Sen et al., 2012) techniques. On the other hand, both the spectral (Ram et al., 2019b) and/or spatial information are generally incorporated for BS or classification purposes.
Various pre-processing steps as; Low SNR band removal, enhancement, de- noising can be helpful for BS (R.N.Patro et al., 2019). Such a denoising framework incorporating the PC (Principal Component) is presented in (Divyesh et al., 2019). The heuristic approaches for BS (Ram et al., 2019c, Kalidindi et al., 2020, Rodrigo et al., 2013) are resource consuming process and persistence of results cannot be fully justified because of the randomness of the optimization approaches. Several supervised classifiers enabled clustering (Xianghai et al., 2016) are helpful, but the availability of GT is not always confirmed. The discussion of BS approaches is restricted to feature based clustering, ranking. Such an k-means based clustering (Qi et al., 2016, Martin et al., 1996, Aloke et al., 2015) and manifold ranking (Veera et al., 2017) is proposed, but the bad cluster may still affect the classifier results.