摘要
针对因数据冗余及Hughes现象带来的高光谱数据分类精度降低问题,提出一种基于样本依赖排斥图的非负稀疏嵌入投影降维(NSEPSRG)算法.首先,利用非负稀疏表示方法,得到样本的非负稀疏重构权重矩阵.然后,利用样本的先验类别信息,构建样本依赖排斥图,有助于避免误分类和提高分类精度.最后,为保持每个样本间的稀疏结构关系和各样本的内在流形结构不变,根据非负稀疏重构权重矩阵和样本依赖排斥图的邻接矩阵,将样本嵌入投影到低维子空间,有助于从高维高光谱数据中提取信息量大的光谱波段,从而使得到的分类图像更清晰、平滑.AVIRIS高光谱数据上的实验结果表明,运用支持向量机对经过NSEPSRG降维处理后的高光谱数据进行分类,分类整体精度和Kappa系数分别达到了87.87%和0.856 6.
In order to solve the problem of low classification accuracy of hyperspectral data which is resulted from high redundancy and Hughes phenomenon, a non-negative sparse em- bedding projection based on sample-dependent repulsion graph （NSEPSRG） algorithm is pro- posed for dimensionality reduction of hyperspectral data. At first, a non-negative sparse recon- struction weight matrix is obtained by using a non-negative sparse representation method. Then, a sample-dependent repulsion graph is constructed by taking advantage of class label in- formation, which is helpful for avoiding misclassification and improving classification accuracy. At last, in order to keep the sparse structure between samples and the inner manifold structure in each sample unchanged, high-dimensional samples are projected to a low-dimensional sub- space according to the non-negative sparse reconstruction weight matrix and the adjacency ma- trix of the sample-dependent repulsion graph, which is conducive to extracting large-scale information from high-dimensional hyperspectral data and consequently rer and smooth classification image. Experimental results on AVIRIS hyperspectr that the overall c assificatio tively in the use of support ands with tains clea- data show n accuracy and Kappa coefficient reach 87.87~/00 and 0. 856 6 respec- vector machine to classify the hyperspectral data that has been dealtwith NSEPSRG.
出处
《中国矿业大学学报》
EI
CAS
CSCD
北大核心
2012年第6期1010-1017,共8页
Journal of China University of Mining & Technology
基金
国家自然科学基金项目(60974050
61072094)
教育部新世纪优秀人才支持计划(NCET-08-0836
NCET-10-0765)
高等学校博士学科点专项科研基金项目(20110095110016)
霍英东教育基金会青年教师基金项目(121066)
中央高校基本科研业务费专项资金项目(JC111231)
关键词
高光谱数据
降维
非负稀疏表示
样本依赖排斥图
hyperspectral data
dimensionality reduction
non-negative sparse representationsample-dependent repulsion graph
