Evaluation of different metaheuristic optimization algorithms in feature selection and parameter determination in SVM classification

In various studies, classification method and input features are two main factors that have significant effects on the results. In high dimensional non-linear problems, SVM was suggested as the superior classification. In these cases, the kernel parameters are often determined by the trial-and-error approach, which leads to reduce reliability of the results and automation level. On the other hand, because of the similarity of pixel spectral behavior, the classification accuracy will be reduced using only spectral bands in complex urban areas. To overcome this limitation, using additional features (e.g., textural and elevational features) was suggested in many studies. However, due to high variety of textural features in terms of type and direction, the presence probability of dependent features will increase by using all the features, which results in classification inefficiency. In addition to relatively high automation, in this paper, metaheuristic optimization algorithms were used to optimize simultaneously SVM in feature selection (FS) and parameter determination (PD) process as a solution due to being independent of image type and scene. There are few comprehensive evaluations in this field in various studies. For this purpose, a comprehensive research of the most efficient optimization algorithms in SVM (i.e., ACOR, GA, ICA, and PSO) was carried out in different ways and by different input features. Moreover, the results were compared to random forest (RF) classification in terms of FS process and accuracy. The optimized SVMs were implemented on two different image scenes (i.e., simple suburban and complex urban areas) in order to show the robustness of the optimized methods in terms of image type and scene. The results were evaluated by five quantitative criteria and McNemar’s test. Also, the approximate time calculations, the number of optimized features, and parameters were presented for each image scene. In comparison with using only spectral bands, the results showed that the optimized SVMs improved the overall accuracy (OA) by 12% and kappa coefficient (KC) by 15% using independent features (Z score of 320 at 95% confidence interval). Moreover, the ICA in conjunction with SVM can provide more accurate results rather than other optimization algorithms. By applying optimized features, OA and KC were improved by 4.88 and 5.9% in the simple suburban scene and 21.82 and 40.21% in the complex urban scene in comparison with using all the input features, respectively. The higher improvement was in the complex image scene because FS is more important in these scenes.