面向目标探测的高光谱图像层次聚类波段选择  

Band selection based on hierarchical clustering for hyperspectral target detection

作　　者：何元磊[1] 刘代志[1] 易世华[1] 黄世奇[1] 

机构地区：[1]第二炮兵工程学院,西安710025

出　　处：《仪器仪表学报》2011年第4期825-830,共6页Chinese Journal of Scientific Instrument

摘　　要：针对高光谱图像波段间的相关性高、信息冗余大从而影响目标探测的问题,提出层次聚类波段选择(HC-BBS)。首先以ROC曲线线下面积(AUC)为指标确定最佳聚类个数,然后对原始波段凝聚聚类,再在聚类后的每类波段中选择最能代表该类的波段组成最终的波段子集,保证了目标探测算子获得最佳的探测效果。对AVIR IS获取的2幅真实高光谱图像进行了实验,结果表明,HC-BBS优于另外2种波段选择方法,其选出的波段分别占据全部波段的9%和3%,目标探测算子ACE和AMF的探测率较全波段分别提高了30%和15%。A band selection method based on hierarchical clustering（HC-BBS） is proposed to deal with the high correlation and redundant information of hyperspectral imagery that may affect target detection performance.The area under ROC curve（AUC） is taken as the criterion for determining the best number of clusters,the original bands are grouped with agglomerative hierarchical clustering and the most representative band is selected from each cluster to make the target detector achieve its best performance in final band selection result.Experiments were conducted on two real hyperspectral data sets acquired by AVIRIS and other two unsupervised band selection methods were used for comparison.As shown in the experimental results,less than 9% and 3% of the full bands are selected by HC-BBS method respectively,and the probabilities of detection of ACE and AMF are improved by 30% and 15%,respectively.

关 键 词：高光谱图像 波段选择 目标探测 层次聚类 ROC曲线线下面积 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]