基于多分类器组合的高光谱图像波段选择方法

Multi-classifier combination-based hyperspectral band selection

由于高光谱数据具有波段多,数据量大等特点,对其进行降维处理成为高光谱遥感研究的一个重要问题。提出一种基于多分类器组合的高光谱波段选择方法,该方法通过遗传算法良好的寻优能力获得若干组较优初始波段子集,在此基础上使用这些波段子集训练若干个基分类器,进而利用改进的基于相同错误差异性度量的分类器选择方法选出部分较优分类器,实现波段选择的目的;最终通过局部精度分析的动态分类器选择实现多分类器组合决策。在公共测试数据集上的实验结果表明:与以往直接选择最优波段子集方法相比,提出的算法能够选择更多具有鉴别能力的波段,明显提高了分类正确率。

Due to the multi-waveband and massive data characteristics of hyperspectral data,dimension reduction is becoming a distinct problem in regards to hyperspectral remote sensing research. A hyperspectral band selection algorithm has been proposed based on a multi-classifier combination. This algorithm obtains several groups of sub-optimal initial waveband subsets through a genetic algorithm,which has better optimizing ability and on this basis trains several base classifiers with these waveband subsets,and further selects several member classifiers from the initial classifier pool by using an improved classifier selection method that is based on same-fault measures,realizing the purpose of the wave band selection. And finally,the multi-classifier combination decision is made through dynamic classifier selection based on the analysis of local classification accuracy（ DCS-LA）. The experimental results regarding the Indian Pine benchmark data set show that this new method can select these bands with more discriminative information and obviously improve the accuracy of the classification process.