A novel technique to detect a suboptimal threshold of neighborhood rough sets for hyperspectral band selection

Abstract

Neighborhood rough sets (NRS), an extension of rough sets, are widely used for feature selection. Although NRS have the advantage of dealing with the continuous data, success of the NRS-based feature selection techniques is strongly dependent on a predefined threshold value which determines the size of neighborhood granule. In this paper, we have proposed a novel technique to detect a suitable threshold of NRS for hyperspectral band selection. Our proposed technique analyzes the changes in boundary regions to select a suitable threshold value that keeps less uncertain boundary samples into positive region and more uncertain boundary samples into boundary region of the decision attribute. The effectiveness of the proposed technique is assessed by using different data sets.