In order to improve the accuracy of seed selection and shorten the cultivation period of cultivars, the improved random forest algorithm is used to construct the evaluation model of the history data of wheat breeding. Before training the classifiers, the improved SMOTE algorithm is used to improve the non-balance of the training samples. After the training of the base classifiers, we test every classifier's performance and delete bad classifiers to realize the screening of the base classifier in random forest. The experimental results show that the proposed algorithm has achieved good results in wheat germplasm evaluation, which can help to breed varieties.