Abstract:

In order to reduce the size and computational complexity of hyperspectral data as well as to avoid Hughes phenomenon in the following classification,a novel hyperspectral image dimensionality reduction method on the basis of fireworks algorithm ( FWA) is proposed.In this method,the weighted sum of compactness coefficient and separation coefficient is used as the criterion for band selection,and a search is constantly performed and updated until the algorithm is convergent.Thus,an optimal band combination is successfully obtained.Then,a comparison among the dimensionality reduction methods respectively based on FWA,genetic algorithm ( GA) and tabu search ( TS) algorithm is made with AVIRIS and ROSIS datasets.The results indicate that FWA-based method is of lower computational complexity and smaller time consumption because it helps select fewer band combinations; and that,when the obtained band combination is used for hyperspectral image classification,the proposed FWA-based method is the best because it improves classification accuracy and Kappa coefficient by 0. 06% ～ 4. 72% and 0. 00 ～ 0. 09,respectively.

Key words: image classification, hyperspectral image, dimensionality reduction, fireworks algorithm, intelligent optimization algorithm