۱٬۳۸۵ / ۵٬۰۰۰
Today, hyperspectral images are considered a powerful and efficient tool in remote sensing due to the wealth of spectral information and provide the possibility of distinguishing between similar complications. Considering the stability of support vector m
More
۱٬۳۸۵ / ۵٬۰۰۰
Today, hyperspectral images are considered a powerful and efficient tool in remote sensing due to the wealth of spectral information and provide the possibility of distinguishing between similar complications. Considering the stability of support vector machines in spaces with high dimensions, they are considered a suitable option in the classification of hyperspectral images. Nevertheless, the performance of these classifiers is influenced by their input parameters and feature space. In order to use support vector machines with the highest efficiency, the optimal values of the parameters and also the optimal subset of the input features should be determined. In this research, the ability of the genetic algorithm as a meta-heuristic optimization technique has been used in determining the optimal values of support vector machine parameters and also selecting the subset of optimal features in the classification of hyperspectral images. The practical results of applying the above method on the hyperspectral data of AVIRIS sensor show that the input features and parameters each have a great effect on the performance of support vector machines, but the best performance of the classifier is obtained by solving them simultaneously. In the simultaneous solution of parameter determination and feature selection, for Gaussian kernel and polynomial, 5% and 15% increase in accuracy was achieved by removing more than half of the image bands. Also, the gradual cooling simulation optimization algorithm was implemented in order to compare with the genetic algorithm, and the results indicate the superiority of the genetic algorithm, especially with the large and complicated search space in the simultaneous solution approach of parameter determination and feature selection.
Manuscript profile
