In this paper, a semi-supervised learning method called semi-supervised sparse multi-manifold embedding (S3MME) was proposed for the classification of hyperspectral image. S3 MME exploits both labeled and unlabeled samples to adaptively find neighbors of each sample from the same manifold by using an optimization program based on sparse representation, which constructs an appropriate graph to characterize the manifold structure. Then it tries to extract discriminative features on each manifold in low dimensional space such that the data points in the same manifold become closer. The overall classification accuracies of the proposed method can reach 84.91% and 89.74% on PaviaU and Salinas hyperspectral data sets respectively, which significantly improves the classification of land cover compared with the conventional methods.