With the rapid development and wide application of optical remote sensing image technology, accurate classification of optical remote sensing images has farreaching research significance. The high-dimensional features extracted by traditional feature extraction methods are mixed with much redundant information, and the classification process can lead to over-fitting. The traditional linear dimension reduction algorithm cannot maintain the internal structure of the original data, and is easy to cause data distortion. We propose an optical remote sensing image classification algorithm based on manifold learning. Firstly, the SIFT features of the image are extracted, and the manifold learning is applied to feature dimension reduction. Finally, the support vector machine is used for training and recognition. Experimental results show that the classification accuracy of glaciers, buildings and beaches is effectively improved on the experimental data of Satellite, NWPU and UC Merced, reaching about 85%. For remote sensing images of desert, rock and water, the classification accuracy is improved by about 10%. In summary, the data based on manifold learning can maintain the topological structure in the original high-dimensional space through the dimension reduction algorithm. Similar feature points can effectively aggregate, which prevents the "dimensional disaster", reduces the calculation amount and guarantees classification accuracy.