Aiming at the problem of poor model generalization performance caused by the scarcity of labeled samples in the target domain, a few-shot hyperspectral image classification method based on cross-domain mixup and self-supervised learning(FSC-CMS) is proposed. First, few-shot learning is used to extract meta-knowledge from the source domain that is more beneficial to target domain classification. Secondly, apply Mixup technology to few-shot learning, perform feature-level Mixup on the query sets of the source and target domains, expand the distribution of the target domain data through the source domain data, increase the diversity of the target domain data, and thus improve the generalization performance of the model. Finally, the few-shot learning process is constrained through selfsupervised learning in the target domain to obtain a more robust feature representation, thereby alleviating the over-fitting problem of the model. A large number of experiments were conducted on two public hyperspectral datasets. Compared with existing mainstream methods, the average accuracy of the proposed method increased by more than 3.2% and 3.6% respectively.