To improve the accuracy of cold resistance classification of wheat, which further provide references for hybrid combination selection. The experimental data of 3 049 national wheat varieties were used in this study, and the cold resistance classification of wheat was predicted by using convolutional neural network combined with circulating neural network（CNN-RNN）. The results showed that this method has state-of-the-art performances with an accuracy of 73.28% and a Kappa coefficient of 0.5956. Meanwhile, to reduce the interference of experimental sample imbalance in the accuracy of this study, SMOTE technology was used to achieve the sample balance. After oversampling, the accuracy and kappa coefficient of CNN-RNN were improved by 7.67% and 0.02, respectively. It indicated that the combination of the above methods could effectively improve the accuracy and consistency test coefficient of wheat cold-resistance prediction, which could be applied to wheat cold-resistance classification prediction,thereby shorten the traditional breeding cycle.