Datensatz

Zusammenfassung

Context
Accurate cultivar-level yield predictions are crucial for ensuring agricultural resilience, but remain challenging due to complex Genotype × Environment (G×E) interactions and the limited integration of cultivar-specific information.
Objective
This study aims to develop and demonstrate the effectiveness of a novel crop-climate fusion framework that integrates cultivar-level performance with phenology-aligned Climate Factors (CF), to explicitly capture G×E interactions and improve yield predictions using Machine Learning (ML) models.
Methods
The Best Linear Unbiased Estimator (BLUE) was derived for 514 cultivars across three rice maturity groups using multi-environment cultivar trials (total 7136 observations) over 2017–2023. This indicator was coupled with 13 CF aligned to cultivar-specific phenological phases. ML models (RF, SVR and XGBoost) of different architectures were rigorously assessed in yield predictions for unseen cultivars through cultivar-independent train-test splits.
Results
Models combining BLUE and CF features consistently outperform those using crop traits or CF alone, achieving R² of 0.54–0.87 and normalized RMSE below 6% across maturity groups. While XGBoost exhibits a consistent advantage over RF and SVR, the ensemble mean model matches or slightly exceeds the best single model, showing the robustness across heterogeneous environments. However, the predictive accuracy varies markedly between years with notable decreases in low-yield years, underscoring the models' sensitivity to climatic anomalies and extreme events. SHAP analysis identifies BLUE as the most influential predictor, while higher baseline yield is linked to stronger panicle–sink formation and grain filling. Among climatic drivers, CF during the emergence-to-panicle initiation exert stronger influences on yield predictions. The relationships between CF and yields are complex and non-linear, with early- and medium-maturity cultivars mainly influenced by low-temperature and water availability indicators, whereas late-maturity cultivars are more affected by thermal ones.
Conclusion
The proposed crop-climate fusion framework improves cultivar-specific yield predictions by integrating cultivar-level and climate factors. BLUE serves as a stable baseline for yield predictions, whereas CF during critical phases modulate this baseline. These findings highlight the importance of linking stable cultivar performance with stage-specific climatic responses for regional rice yield forecasting and cultivar improvement.
Significance
The framework provides a practical method for accurate cultivar-level yield forecasting, while offering breeding-relevant insights into trait profiles and climatic sensitivities.