基于K-L散度与光谱可分性距离的波段选择算法  

Band selection algorithm for hyperspectral imagery based on K-L divergence and spectral divisibility distance

作　　者：杨佳[1] 华文深[1] 刘恂[1] 马左红 

机构地区：[1]军械工程学院 [2]解放军73101部队

出　　处：《应用光学》2014年第1期71-75,共5页Journal of Applied Optics

摘　　要：波段选择是高光谱降维的常用手段,文中从波段选择应遵循的3个原则出发设计了一种基于信息散度与光谱可分性距离的波段选择算法。将高光谱数据中每个波段的光谱分量看作一个一维向量,使用K-L散度表示其相互之间的信息量,选出信息量大且相似性最小的波段组合;根据每个波段中不同地物光谱可分性距离的计算,得到可分性较大的波段组合;将两组波段组合取交集,即得到最优组合波段。为了验证算法的有效性,将选出的最佳3个波段进行伪彩色合成,对其进行光谱角制图分类,分类精度达到92.2%,Kappa系数为0.88.Band selection is a common approach to reduce data dimensionality of hyperspectral imagery. A new band selection algorithm for hyperspectral imagery based on Kullback-Liebler （K-L） divergence and spectral divisibility was proposed. Firstly by considering every waveband spectral component as the one-dimensional vector, the K-L divergence was used to quantify the information amount among them to select the bands combination with maximum information a mount and minimum similarity. Then the spectral divisibility distance between two ground ob jects was calculated to remove redundant bands from the selected bands combination and make the final selected bands combination contain optimal divisibility. Hyperspectral imagery was tested for classification to testify the effectiveness of this band selection algorithm. As shown in the experimental results, the classification accuracy percentage is 92.2% and Kappa coeffi cient is 0.88, when using spectral angle mapping on the pesudo color synthesis image of 3 se lected hands.

关 键 词：高光谱成像 波段选择 K-L散度 光谱可分性 光谱角制图 

分 类 号：TN911.73[电子电信—通信与信息系统]