面向高光谱图像分类的空谱半监督局部判别分析

发布时间：2018-05-08 23:07

本文选题：遥感 + 高光谱图像分类　；参考：《光学学报》2017年07期

【摘要】：针对传统的基于特征提取的高光谱图像分类算法大多只考虑光谱信息而忽略空间信息的问题,提出了一种基于空谱半监督局部判别分析(S3 ELD)和空谱最近邻(SSNN)分类器的高光谱图像分类算法。该算法结合高光谱图像的空间一致性,在利用标记样本的判别信息保持数据集可分性的基础上,定义空间近邻像元散度矩阵来保存像元的空间近邻结构,提出基于空谱距离的相似性度量并将其应用于局部流形结构的发现和SSNN的构建。S3 ELD算法不仅能揭示数据集的局部几何关系,而且增强了光谱域同类像元和空间域近邻像元在低维嵌入空间的聚集性。结合SSNN进行分类,进一步提升了分类精度。利用PaviaU和Salinas数据集进行的实验结果表明,S3 ELD算法的总体分类精度分别达到了92.51%和96.29%;与现有几种算法相比,该算法能更有效地提取出判别特征信息,并达到更高的分类精度。
[Abstract]:The traditional hyperspectral image classification algorithms based on feature extraction only consider spectral information and ignore spatial information. A hyperspectral image classification algorithm based on space-spectrum semi-supervised partial discriminant analysis (S3ELD) and space-spectrum nearest neighbor (SSNNN) classifier is proposed. Based on the spatial consistency of hyperspectral images and using the discriminant information of labeled samples to keep the separability of the dataset, the spatial nearest neighbor pixel divergence matrix is defined to preserve the spatial nearest neighbor structure of the pixel. The similarity measure based on space-spectrum distance and its application to the discovery of local manifold structure and the construction of S3 ELD algorithm by SSNN can not only reveal the local geometric relationship of the data set, but also apply the similarity measure based on space-spectrum distance. It also enhances the clustering of similar pixels in spectral domain and adjacent pixels in spatial domain in low dimensional embedded space. Combined with SSNN classification, the classification accuracy is further improved. The experimental results using PaviaU and Salinas datasets show that the total classification accuracy of S3 ELD algorithm reaches 92.51% and 96.29% respectively, and compared with the existing algorithms, the algorithm can extract the discriminant feature information more effectively and achieve higher classification accuracy.
【作者单位】：火箭军工程大学信息工程系;
【基金】：国家自然科学基金青年科学基金(61401471) 中国博士后科学基金(2014 M562636)
【分类号】：TP751

【相似文献】