Year-independent prediction of rice grain protein content using machine learning with agronomy-aligned multi-year field data.

Rice grain protein content exhibits substantial inter-annual variability due to interactions between management practices and environmental conditions, making robust prediction under realistic field conditions challenging. This study evaluated the year-independent predictability of rice grain protein content using multi-year field data and machine learning approaches, with a focus on data structuring and validation strategies. Models were trained using both plot-level and replicate-mean datasets, and evaluated using leave-one-year-out cross-validation to simulate prediction under unseen growing seasons. Replicate aggregation consistently improved predictive robustness by reducing plot-level variability, while year-independent validation provided a more conservative and realistic assessment of model generalization. Despite this stringent framework, predictive performance remained relatively stable, and residual diagnostics indicated no clear systematic bias across protein levels or nitrogen application rates. Anomaly-based analysis further suggested that deviations from year-variety baselines could be predicted with meaningful accuracy, indicating that relative variations beyond dominant temporal effects were effectively captured. These results suggest that aligning data structuring and validation strategies with agronomic experimental design may be important for reliable year-independent prediction. The proposed framework may provide a transferable approach for integrating machine learning with agronomic knowledge to support crop quality prediction and adaptive nitrogen management under variable field conditions.