Abstract: In order to extract the local characteristics of hyperspectral data, a dimensionality reduction method of hyperspectral imagery manifold learning based on spectral gradient angle was proposed from the nonlinear structure of hyperspectral imagery. Locality preserving projection (LPP) of localized manifold learning algorithm was performed to reduce the dimensionality of hyperspectral remote sensing data. In order to improve the distance metric, similarity measurement of spectral gradient angle, which can better characterize local features of hyperspectral images, was applied to LPP method. The real hyperspectral images were subjected to dimensionality reduction experiments.The results were better than the original LPP method and the LPP method using the spectral angle. The results show that the proposed method is superior to LPP method and LPP method using the spectral angle in the spectral normalized eigenvalues. Meanwhile, the proposed method can also obtain a good performance in information retainment and have better local information retention. Therefore, the manifold learning method with spectral gradient angle has a better performance in dimensionality reduction of hyperspectral images.