Scene classification is a hot issue in the field of SAR image interpretation. Many SAR image interpretation tasks can be promoted with the development of highly credible scene classification methods. But the fussy steps of traditional methods and the imperious demands of labeled samples in deep learning-based methods restrict the effective feature learning in SAR scene classification. Hence, a self-supervised learning method based on Jigsaw puzzles is proposed to address the problems. Concretely, the Jigsaw puzzle reassembly of the SAR image block is firstly taken as the upstream task without manual labels. Once the correct spatial arrangement is obtained from it, the learned high-level feature from the upstream task is used as the pre-training model for the downstream task, which is then fine-tuned with only a few labeled samples to enhance the performance of the SAR scene classification task. Experimental results on 25-class real SAR scenes confirm the proposed method can greatly improve the scene classification performance than directly training the network with the same number of labeled samples used in the downstream task.