Integrating multi-source remote sensing and crop growth model for phenology-driven wheat yield prediction in the...

Accurate yield prediction is critical for food security, yet traditional methods relying on ground surveys and fixed temporal windows lack scalability and precision. To address these limitations, we proposed an integrated framework for large-scale wheat yield prediction. Based on a calibrated Agricultural Production Systems sIMulator (APSIM) model, we first generated physically consistent simulation data as alternative training labels to reduce dependence on scarce ground observations. A random forest model was then developed incorporating APSIM-generated data, climate variables, and temporal phenology to model wheat yield across the North China Plain.

The results demonstrated that the model using a dynamic temporal window significantly outperformed that with a fixed temporal window in both accuracy and interannual stability, confirming the importance of phenological synchronization for model robustness. At the county scale, the dynamic window model achieved a coefficient of determination (R²) of 0.54 and root mean square error (RMSE) of 1056 kg ha^-1, compared to 0.48 and 1179 kg ha^-1 for the fixed window. At the site scale, the dynamic window model achieved an R² of 0.40 and RMSE of 1270 kg ha^-1, compared to 0.34 and 1412 kg ha^-1 for the fixed window. Furthermore, the spatial pattern of predicted yield using the dynamic temporal window aligned well with county-scale statistical data, with aggregated provincial total yield prediction achieving exceptionally high accuracy (R² = 0.94, RMSE = 2.27 Mt).

This framework provides a comprehensive solution for reliable yield prediction in data-scarce regions, offering robust technical support for smart agricultural management and early warning systems for food security. © 2026 Society of Chemical Industry.