It is difficult to obtain labels of samples for hyperspectral data. Few labeled samples usually lead to low classification accuracy. In view of this situation, an improved spatial and spectral constraint graph-based semi-supervised classification algorithm (SS-GSSC) is proposed. First of all, Euclidean distance combined with radial basis function (RBF) is used to construct the spatial similarity edge weight; Spectral correlation angle (SCA) is used to calculate spectral similarity weights; Then, the two kinds of weights are combined to the form of product to restrict the similarity measurement; Finally, the label propagation algorithm is used to predict the test data labels so as to obtain the classification results. Classification experiments on Indian Pines image and DC Sub image show that, compared with the previous classification algorithm, the algorithm designed by the authors can better eliminate the phenomenon of the existence of the same category map spot included in other categories of scattered points, and can achieve higher classification accuracy under the condition of less label points (25 per class).