Semisupervised Band Selection from Hyperspectral Images Using Levy Flight-Based Genetic Algorithm

Abstract

Getting the advantages of hyperspectral remote sensing images depends on overcoming the challenges posed by their large number of bands. One approach is selecting appropriate bands by the metaheuristic methods. Initial versions of these methods usually suffer from trapping into local optima, so they do not work optimally in selecting the best band subset. The number of training samples is also important in the band selection. With insufficient training samples, even improved metaheuristic methods do not yield the desired result. However, providing sufficient ground data as training samples is costly. In this letter, an improved Levy flight (LF)-based version of the genetic algorithm (GA) is developed and used to select bands in a semisupervised manner. The number of training samples in the proposed semisupervised method is increased based on spatial adjacency and spectral similarity simultaneously. Our results show that in cases where the initial version of the GA fails to select appropriate bands, our improved LF-based version introduces a band subset that yields the desirable classification result. Also, the classification accuracies have been improved considerably using our proposed semisupervised method. In some of our experiments using a small number of training samples, the accuracy improvement is near 17%. The proposed method has been effective in the case of sufficient training samples too. In this case, the accuracy improvement is near 11% in some experiments. © 2004-2012 IEEE.