A fusion-based approach to improve hyperspectral images’ classification using metaheuristic band selection

Abstract

Although the high number of bands in hyperspectral remote sensing images increases their usefulness, it also causes some processing difficulty. In supervised classification, one problem is decreasing classification accuracy due to the insufficient training samples against the bands. A way to deal with this problem is the selection of appropriate bands by the metaheuristic methods. Because of the stochastic search, the selected bands differ in any implementation of a metaheuristic method. So, the results obtained from the classification of these different band subsets will also have some differences. In this study, a fusion-based approach has been proposed to improve the classification of hyperspectral remote sensing images by multiple implementations of a metaheuristic method for band selection. We have tested the proposed method using ten metaheuristic methods with different objective functions on four well-known datasets. The results show the proposed fusion-based approach successfully improves the classification accuracy in all experiments. The accuracy improvement varies depending on the metaheuristic method, the objective function, and the dataset and ranges from 0.4% to 15.7%. The proposed method improves the classification of complex datasets more and affects weaker objective functions considerably. The results also show the proposed method brings the accuracy of different metaheuristic methods close to each other and reduces the sensitivity of selecting the proper ones. Thus, an automated classification system can be obtained using a parameter-less method.