面向高光谱图像分类的空谱判别分析

针对传统的基于特征提取的高光谱图像地物分类算法大多只考虑光谱信息而忽略空间信息的问题,提出了一种面向高光谱分类的半监督空谱全局与局部判别分析(S3 GLDA)算法。该算法首先利用少量标记样本保存数据集的线性可分性和全局判别信息,再依靠较多的无标记的空间局部近邻像元来揭示局部判别信息和非线性局部流形,使高光谱遥感图像的光谱域全局判别结构和空间域局部判别结构在低维特征空间同时得以保留,并在输出特征中自动融入了空间信息,构成了半监督的空谱判别分析。在Indian Pines和PaviaU数据集的实验表明,总体分类精度分别达到76.24%和82.96%。与现有几种算法比较,该算法有效提高了输出特征在低维空间的判别能力,更好地揭示了数据集的内在非线性多模本质,有效提升了高光谱图像数据集的地物分类精度。

Spatial-spectral discriminant analysis for hyperspectral image classification

The traditional hyperspectral image classification methods consider only spectral information while spatial information is ignored. To address this problem, a semi-supervised spatial-spectral global and local discriminant analysis (S³GLDA) algorithm for hyperspectral image classification was proposed. The method firstly made use of a few labeled samples to preserve the linear separability and global discriminant information of the data set, then the local discriminant information and nonlinear manifold was uncovered by the unlabeled spatial neighbors. The spectral-domain global discriminant structure and spatial-domain local discriminant structure were exploited simultaneously and the spatial information was incorporated into the output low-dimension features automatically, which constitute the semi-supervised spatial-spectral discriminant analysis. The overall classification accuracies reached 76.24% and 82.96% on the Indian Pines and PaviaU data sets, respectively. Compared with several existing methods, the proposed algorithm can effectively improve the discriminant ability of the output features in the low-dimension subspace, which can uncover the intrinsic nonlinear multi-modal structure of the data set and obtain higher ground objects classification accuracy.