基于粒子群算法和序贯搜索的高光谱波段选择

Hyperspectral Band Selection Based on Particle Swarm Optimization and Sequential Search

波段选择是降低高光谱数据量,克服地物分类中Hughes现象的有效手段。子集生成方式和评价准则是选择算法的两要素。提出一种混合随机搜索与启发式搜索的子集生成方法。该方法在随机搜索中嵌入启发式搜索,对由离散粒子群优化算法每次迭代更新的种群利用序贯搜索进行局部微调,提高了随机搜索的精度。这种嵌入微调也保证了优化算法解的有效性。高光谱波段选择与分类实验比较了该方法与混合遗传算法、标准遗传算法和顺序前向浮动选择算法的性能,表明算法能选择出评价准则意义下更好的子集。

Band selection can cut down a large amount of hyperspectral data and alleviate the Hughes phenomenon in supervised classification of ground objects. The generation and evaluation of subsets are two key factors for selection algorithm. A hybrid scheme of random search and heuristic search is proposed to generate the band subset. The method embeds the sequential search into the evolution optimization for better performance of the fine tune in local search space. Thus, it behaves well in both global and local cases. Furthermore, the embedding scheme guarantees the validity of solutions for the optimization algorithms. The performance of the proposed method, the hybrid genetic algorithm （HGA）, the standard genetic algorithm （SGA） and the sequential forward floating selection （SFFS） are compared in the experiments on band selection and classification with the hyperspectral data sets. Results show that the proposed method can obtain the best subsets according to the evaluation criterion.