基于流形对齐的高光谱遥感图像降维和分类算法  

Manifold alignment for dimension reduction and classification of multitemporal hyperspectral image

作　　者：鲁锦涛[1] 马丽[1] 

机构地区：[1]中国地质大学(武汉)机械与电子信息学院,武汉430074

出　　处：《国土资源遥感》2017年第1期104-109,共6页Remote Sensing for Land & Resources

基　　金：国家自然科学基金项目"基于流形学习进行图结构设计的高光谱图像分类技术研究"(编号:61102104)资助

摘　　要：多时相的高光谱遥感图像数据处理中会出现地物光谱特征漂移的现象。为了提高源域数据已有知识对目标域数据分类的精度,采用了基于流形对齐的分类算法。先用一个标准的线性或非线性的降维方法将2个高光谱遥感数据集映射到低维(流形)空间中,再用Procrustes分析方法将其低维嵌入之间的平移、旋转和缩放因子剔除,得到数据集间的最优对齐,最后用最近邻算法进行分类。对多个不同时相高光谱遥感图像进行实验,并对比了已有的流形对齐算法,结果表明本算法具体较好的迁移能力和分类效果。For multitemporal hyperspectral images, the spectral characteristics of the same land cover object may vary significantly. Therefore, manifold alignment algorithm was employed to find a feature space in which data distributions of both images become the same. The method includes three steps. Firstly, a standard linear or nonlinear dimension reduction method is used to reduce the dimensionality of hyperspectral images. Secondly, the Procrustes analysis method is utilized to remove the translatio that the optimal alignment between the two data sets can be applied for classification. Experimental results using approach can obtain performances which are superior nal, rotational and scaling components from one set so achieved. Finally, the nearest neighbor algorithm is muhitemporal hyperion images demonstrate that the proposed to those of several popular manifold alignment methods.

关 键 词：PROCRUSTES 流形对齐 多时相 高光谱遥感图像 降维 分类 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]