等距映射和局部切空间排列降维后,低维流形坐标能够保留原始高光谱影像中地物光谱信息,用于提取原始影像的潜在特征。然而这两种流形方法的理论差异导致其低维坐标继承光谱信息的能力不同,对比这两种流形坐标可凸显出原始影像内部的潜在特征。因此,本文基于等距映射和局部切空间排列非线性降维,提出两种流形坐标的差异图法来提取高光谱影像内部的潜在特征。首先,根据流形坐标的光谱解释确定两种坐标的每一维代表相同的光谱信息。其次,根据相同的光谱特征,归一化两种流形坐标并调整坐标轴方向,统一两种坐标到相同的坐标框架。最后,通过加权流形图相减得到坐标差异图,采用经典的图像处理方法提取潜在特征。采用两个实验并对比等距映射和局部切空间排列方法的降维结果来验证本文方法。结果表明,流形坐标差异图能够成功提取单一流形结果无法得到的潜在特征,如靠河岸的浅水区域和大场景沼泽地中的低分辨率道路。这为高光谱影像的潜在特征提取研究提供了一种新方法。 After the isometric mapping and local tangent space reduction, the low-dimensional manifold coordinates can preserve the spectral information of the original hyperspectral image and extract the latent features of the original image. However, the theoretical differences between these two manifold methods lead to their different ability to inherit the spectral information from the low-dimensional coordinates. Contrasting these two manifold coordinates can highlight the latent features inside the original image. Therefore, based on the isometric mapping and the non-linear dimensionality reduction of the local tangent space arrangement, this paper proposes two difference graph methods of manifold coordinates to extract the latent features inside the hyperspectral image. First, the spectral interpretation of the manifold coordinates determines that each of the two coordinates represents the same spectral information. Secondly, according to the same spectral features, normalize the two manifold coordinates and adjust the direction of the coordinate axis, and unify the two coordinates into the same coordinate frame. Finally, the coordinate difference map is subtracted from the weighted manifold graph, and the latent image is extracted by the classical image processing method. Two experiments were carried out to verify the proposed method by comparing the reduced-dimensional results of isometric mapping and local tangent space arrangement. The results show that the manifold difference map can successfully extract the latent features that can not be obtained by the single manifold results, such as the shallow water area near the river bank and the low-resolution road in the big scene swamp. This provides a new method for the study of potential feature extraction of hyperspectral images.