Deployment-oriented evaluation of stage-specific information for rice yield prediction across years, sites and...

Reliable rice yield prediction requires machine-learning models with inputs that match the intended decision point and evaluation schemes that reflect the target deployment setting. Random record-level cross-validation remains common in crop-yield machine learning, but it can overestimate performance when models are transferred to future seasons, new sites or new site-year environments.

We developed a scenario-aware machine-learning framework using 3,204 yield records from 168 cultivars evaluated across 10 sites, 7 years and 67 site-year environments in southern China. Five modeling feature sets were compared across random five-fold cross-validation, leave-one-year-out validation, forward-year prediction, leave-one-site-out validation and leave-one-site-year-out validation. The feature sets represented whole-season weather, stage-specific climate information, leakage-controlled cultivar yield BLUP, field-survey traits and heading-stage updates.

Random five-fold cross-validation produced the highest apparent accuracy, with the best field-survey-enhanced model reaching R² = 0.844 and RMSE = 340.59 kg ha^-1. Deployment-oriented performance was lower. The same information scenario achieved a macro mean R² of 0.359 in forward-year prediction, 0.290 in leave-one-site-out validation and 0.131 in leave-one-site-year-out validation. Stage-specific climate information and leakage-controlled cultivar potential improved prediction before field-survey traits were available, whereas field-survey traits provided the strongest later-stage accuracy. Reduced training size lowered accuracy and increased unseen-cultivar burden. Climate-risk environments showed risk-specific error shifts rather than a uniform model failure.

This study supports a deployment-oriented framework for rice yield prediction, rather than a general claim of universal out-of-sample improvement. Its main value is to show how prediction skill, feature utility and model interpretation change when validation is aligned with realistic temporal, spatial and site-year deployment settings.