基于近邻子空间划分的高光谱影像波段选择方法

A Hyperspectral Band Selection Method via Adjacent Subspace Partition

在降低高光谱遥感影像数据的冗余度方面,波段选择一直是一种有效的方法.近年来,提出了许多用于高光谱波段选择的聚类算法,但大多数算法只有在选择足够多的聚类中心时才能够表现出良好的性能.在选择少量波段时,往往效果很不理想,不能满足实际使用的目的.而且,随着聚类中心数量的增加,大多数波段选择算法的精度存在不同程度的下降趋势.针对当前基于聚类的波段选择方法存在对聚类中心数的强敏感性和选择的特征波段子集高相关性的问题,提出了一种基于近邻子空间划分的波段选择方法(SEASP).该方法主要包括近邻子空间划分和特征波段选取两个步骤.考虑到高光谱波段之间的有序性,SEASP首先计算出相邻波段之间的相关系数,得到相关系数向量.若两个波段之间的相关性在某个区间内最小,即相关系数的变化率在该区间内最大,说明这两个波段在很大概率上不属于同一组,为两个相邻分组之间的分割点.因此,在相关系数向量的基础上,计算出其对应的若干个极小值,通过极小值的选取来确定最终划分的子空间.最后以信息熵为度量标准从划分的子空间中选出特征波段子集.在3个公开数据集的实验结果表明,提出的SEASP算法与其他算法相比,不仅原理简单,而且在精度和计算效率方面,均表现出了更好的效果.

Band selection is regarded as an effective method for reducing the redundancy of hyperspectral remote sensing images.In recent years,many clustering algorithms have been proposed for the selection of hyperspectral bands,but most of them perform well only when enough clustering centers are selected.When selecting a small number of bands,the results of these algorithms are often not ideal and are unsuitable for practical usage.Furthermore,the accuracy of most band selection methods tends to decrease when the number of selected bands increases.To ad-dress the high correlation of selected feature band subsets and the sensitivity to the number of cluster centers in the current clustering-based band selection methods,this study proposes a simple yet effective hyperspectral band selec-tion method via adjacent subspace partition(SEASP).The proposed algorithm comprises two parts:the partition of subspace and selection of feature bands.By considering the order between adjacent hyperspectral bands,the SEASP calculates the correlation of the adjacent bands to first obtain the correlation coefficient matrix.If the degree of the correlation between two bands is the smallest within an interval,i.e.,the rate of change of the correlation coefficient is the largest within that interval,these two bands should not belong to the same group in a large probability,indicating a segmentation point of two adjacent groups.Thus,the corresponding minimum values via the correlation coefficient matrix are obtained,and the final subspace is then determined through the selection of the minimum values.Finally,the subset of feature bands is selected from the subspace based on the information entropy.Extensive experiments on three public datasets show that the proposed SEASP not only has a simple form in principle but also shows better results in terms of accuracy and computational efficiency than other state-of-the-art algorithms.