Feature extraction based on multi-objective GA in hyperspectral images

Feature selection and feature extraction are very important in hyperspectral applications due to huge data amount of hyperspectral images. It is very necessary to reasonably combine feature selection with feature extraction because each of them has advantages and disadvantages. In process of combining them, one key problem has to be solved, i.e. how many bands should be selected and which bands should be utilized. In this paper, a multi-objective genetic algorithm (MOGA) is proposed to solve this problem. This method combines genetic algorithm (GA) with principal component analysis (PCA). It firstly divides original data into several sub-intervals of band numbers, and then the best sub-interval is obtained by rough search. Refined search is then conducted in this sub-interval and a subset is selected for feature extraction. MOGA uses GA for search, and separability measures and training classification accuracies are adopted as objective functions for different purposes. The numerical experiments are conducted on AVIRIS hyperspectral data with 224 bands. The experimental results prove the effectiveness and practicability of the MOGA. MOGA has better performance than those existing algorithms.