Abstract: Although hyperspectral data has been widely utilized in recognition and classification of materials, it suffers from large data size and high correlation between the bands, which may decrease the classification accuracy and hinder its applications. In this paper, previous band selection methods are analyzed for the above problem, and a new band selection algorithm for hyperspectral data based on k-means clustering and supervised classifications is proposed. Firstly K-mean classification method is used to cluster bands of hyperspectral data into several sets. Then band selection based on genetic algorithm is developed by using the classification accuracy as the cost function criterion. Hyperspectral data of leaves is used for classification to testify the effectiveness of this band selection algorithm. As shown in the experimental results, the proposed method can achieve high performance for classification applications.

Key words: genetic algorithm, band selection, K-mean clustering, classification of hyperspectral data

摘要： 高光谱数据在物质分类识别领域得到了广泛应用，但存在数据量大、波段间相关性高等问题，严重影响分类精度及应用。针对以上问题分析了已有的波段选择方法，提出了基于波段聚类及监督分类的遗传算法，对高光谱数据进行波段选择：采用[K]均值聚类算法对波段数据进行聚类分析，构造波段子集合；利用分类器族分类精度构造适应度函数，采用遗传算法对波段子集合进行优化选择。最后用阔叶林高光谱数据对提出的算法进行对比实验，实验结果表明针对分类应用，提出的算法能够非常有效地选择高光谱谱段。

关键词: 遗传算法, 谱段选择, [K]均值聚类, 高光谱数据分类