Integrating data assimilation, crop model, and machine learning for winter wheat yield forecasting in the North China Plain

Timely and reliable regional crop yield forecasting before harvest is critical for managing climate risk, adjusting agronomic management, and making food trade policy. Although various methods exist for crop yield forecasting, including process-based crop models and machine learning techniques, the potential of integrating these methods for early-season yield forecasts has not been well investigated. In this study, we proposed a hybrid framework for crop yield forecasting that firstly assimilated leaf area index and soil moisture into a crop model and then combined the data-assimilated crop model with machine learning techniques to improve the prediction skill further. The proposed framework was applied to winter wheat yield forecasting in the North China Plain during 2009–2015. We found that the assimilation significantly enhances wheat yield estimates, achieving additional ACC = 0.27, MAPE = 6.12 %. Incorporating weather forecasts enabled reliable winter wheat yield forecasts up to 1–3 months in advance, achieving ACC = 0.69, MAPE = 7.79 %. Furthermore, integrating the assimilated crop model with machine learning techniques improved the forecasting further, achieving ACC = 0.97 and MAPE = 1.74 %. The proposed framework for crop yield forecasting can be adapted to other crops and regions and has great potential in developing food security early warning system at a regional scale.