Determination of optimum classifier and feature subset in hyperspectral images based on ant colony system

Recent advances in hyperspectral remote sensing technology allow the discrimination of complex land-cover classes that have similar spectral reflectance. The reliability and robustness of Support Vector Machines (SVM) in high dimensional space makes it an efficient tool for the classification of hyperspectral images. However, two optimization issues have a strong effect on the SVM performance: optimum parameter determination and optimum feature subset selection. There is an intricate relationship between these two issues, and we propose in this paper to determine the SVM parameters and the feature subset simultaneously by Ant Colony System (ACS). We compare our approach with three powerful meta-heuristic optimization algorithms, namely, Simulated Annealing (SA), Genetic Algorithm (GA), and Bees Algorithm (BA). Experimental results with two AVIRIS datasets clearly demonstrate the superiority of the ACS algorithm in that it improves the classification accuracy and decreases the size of the selected feature subset. In both datasets the classification accuracy improves by about 7 percent and 95 redundant bands are eliminated.