A Comprehensive Review on Effect of Band Selection on the Recital of Hyper-spectral Image Classification

In every research field algorithms have been realized by various authors. These algorithms like geo-spatial-based land cover and land use research are to reassess in day by day. The recital of land cover and land use (LCLU) nomenclature of hyper-spectral image chiefly depends on two principal concerns listed, namely (i) huge number of predictive pixels with hundreds of spectral bands as dimensionality and (ii) noisy and redundant bands that may mislead the classification accuracy. When compared with a number of spectral bands due to less training sample instances, they have sceptical collision on the accuracy of supervised classifiers which is called as the Hughes effect. This paper is to study the result of reduction in dimensionality by selecting relevant bands and eliminating irrelevant and redundant ones by varied feature selection techniques. Once the bands are selected, they will be supplied to unlike classifiers namely support vector machine (SVM), Bayes and decision tree classifiers to examine the effect on classification accuracy and also estimate the decency of fit. In this regard, we employ two hyper-spectral image data sets, namely Indian Pines and Botswana, which are used. With a minimal spectral bands subset, it can achieve maximum classifier accuracy with the help of support vector machine-REF, joint mutual information (JMMI) and high-dimensional model representation (HDMR), and the feature selection methods are proposed.