Objective To investigate the ability of different modeling methods to characterize the relationship between rice agronomic traits and yield, construct a rice yield prediction model based on deep neural networks (DNN), and evaluate the application potential of deep learning methods for rice yield prediction and high-yield trait screening.

Method Based on the agronomic trait data of 12 traits from 442 rice hybrid combinations, a predictive model was constructed after standardization and of missing value imputation. Four algorithms were employed for comparative modeling: DNN, linear regression, random forest, and extreme gradient boosting (XGBoost). Model performance was evaluated using mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination (R²). The feature contributions of the models were analyzed using the SHapley Additive exPlanations (SHAP) method, while Pearson correlation analysis and difference significance tests were conducted to validate the statistical reliability of the interpretations.

Result The DNN model achieved the best performance on the test set with an R² of 0.947 6, MSE of 1 094.08, and MAE of 25.96. Although its absolute accuracy was slightly lower than that of the linear regression model (R²=0.956 7), cross-validation showed that the DNN exhibited smaller prediction fluctuations and demonstrated superior robustness. SHAP analysis revealed that filled grains per panicle, effective panicle number, 1 000-grain weight and panicle weight per plant were the core traits affecting yield prediction, with their contribution rankings highly consistent with correlation test results, consistent with the yield formation principle that rice yield was determined by “effective panicle number-grain number-1 000-grain weight” triad.

Conclusion Compared to traditional statistical models, the DNN model demonstrates superior resistance to interference and predictive robustness. The analysis framework combining DNN and SHAP offers unique advantages in visualizing nonlinear feature contributions and threshold effects, providing a new deeply interpretable tool for rice yield prediction and high-yield trait screening.