A novel dependency definition exploiting boundary samples in rough set theory for hyperspectral band selection

Dimensionality reduction is considered to be a primary task for effective classification of hyperspectral images. In this work, a novel feature (band) selection technique based on rough set theory is presented to reduce the dimensionality of hyperspectral images. Here, a new definition of dependency measure in rough set theory is proposed by not only considering the objects in the positive region but also some objects from the boundary region. The proposed dependency definition is completely parameter free and computationally very cheap. Our technique, first, defines a novel criterion by combining the relevance and significance measure computed using the proposed dependency definition. Then, a first order incremental search is adopted to select the most informative bands by maximizing the defined criterion. The proposed band selection technique shows better result compared to the existing rough set based band selection techniques on three real hyperspectral data sets. (C) 2020 Elsevier B.V. All rights reserved.