A Hybrid Approach for Supervised Spectral Band Selection in Hyperspectral Images Classification

Abstract

Recently, hyperspectral imagery has been very active research field in manyapplications of remote sensing. Unfortunately, the large number of bandsreduces the classification accuracy and computational complexity whichcauses the Hugh phenomenon. In this paper, a new hybrid approach for bandselection based is proposed. This approach combines the advantage of filterand wrapper method. The proposed approach is composed of two phases: thefirst phase consists to reduce the number of bands by merging the highlycorrelated bands, and, the second phase uses a wrapper approach based on SinCosine Algorithm to select the optimal band subset that provides a highclassification accuracy. In addition, a new binary version of Sin CosineAlgorithm is proposed to adapt it to the band selection problem. Theperformance evaluation of the proposed approach is tested on three publiclyavailable benchmark hyperspectral images. The analysis of the resultsdemonstrates the efficiency and performance of the proposed approach.