Abstract: ‍ ‍Existing self-supervised learning algorithms are not capable of small-sample Synthetic Aperture Radar （SAR） image characterization. And it cannot fully meet the requirements of Automatic Target Recognition （ATR）. Therefore， this paper presents a small sample SAR ATR method based on twin self-supervised learning. Firstly， the positive and negative sample pairs are established in the unlabeled SAR data by data augmentation in the twin feature extraction network module. Secondly， in the contrastive learning head network and the feature redundancy reduction head network of the twin self-supervised learning module， the model parameters are jointly optimized by the unsupervised contrastive learning loss function and the feature information redundancy loss function. Thus， a pre-trained network with improved representation capability is obtained. Finally， the self-supervised pre-trained network weight is loaded into the downstream network， and small sample SAR image supervised recognition is performed for the downstream network through cross-entropy loss. The experiment results show that the mainstream accuracy rate of this method reaches 82.95% with only 3.13% training data for Moving and Stationary Target Acquisition and Recognition datasets. The proposed method gets improved representation capability in unlabeled data and effectively improve the overfitting problem of small sample SAR ATR.