Genetic algorithm-based dimensionality reduction method for classification of hyperspectral images

Abstract Although hyperspectral imaging provides rich information owing to its high spectral resolution, this high dimensionality causes significant computational costs and a decrease in classification accuracy. In this study, a genetic algorithm (GA)-based size reduction method is proposed as a solution to this problem. With the proposed method, unnecessary and repetitive bands in hyperspectral data were eliminated, and only the most significant bands were selected, and classification was performed with support vector machines. The proposed approach offers higher accuracy and lower processing time compared to principal component analysis and minimum noise fraction, which are traditional size reduction methods. In the Indian Pines dataset, the number of bands was reduced from 200 to 85, and in the KSC dataset, it was reduced from 176 to 78, ensuring classification accuracy of 91.95–93.44% and 95.27–95.77%, respectively. These results show that GAs are an effective size reduction method in hyperspectral image classification.