Multi-environment field testing is an important way to select optimize maize yield and yield stability varieties. However, because of its high cost, it has gradually become a challenge in plant breeding. The combination of field sparse testing and genome-wide prediction method can be used to predict untested phenotypes, reduced the effort and cost on field testing. In this experiment, 244 inbred lines of natural populations were planted in Shunyi, Beijing and Mishan, Heilongjiang in 2022 and 2023. Six agronomic traits were studied, including days to anthesis, plant height, ear height, ear length, kernel number per row and ear row number. The effects of four different models (Single, Across, M×E and R-norm), two different cross-validation schemes (CV1 and CV2) and three different training sets sampling ratios (0.5, 0.7 and 0.9) on the prediction accuracy were compared. The results showed that the average prediction accuracy of the six agronomic traits was 0.67, 0.58, 0.50, 0.33, 0.33 and 0.48. The average prediction accuracy of the Single model, Across model, M×E model and R-norm model was 0.36, 0.52, 0.53 and 0.53 for each trait. In CV1, the average prediction accuracy of each model in six traits ranged from 0.19 to 0.65, and in CV2, the average prediction accuracy ranged from 0.47 to 0.89. The comparison of different training set sampling ratios shows that the improvement of the proportion of the training sets has limited improvement in the prediction accuracy of different traits in different models, and the maximum is only 0.05. The results show that the CV2 training set can be used to form a scheme and include phenotypic data from multiple environments in the prediction model to provide good prediction accuracy for multi-environment prediction.