Coupling crop growth models and machine learning for scalable winter wheat yield estimation across major wheat regions in China

Abstract

The temperate continental climate, with its combination of ample sunlight and relatively low humidity during key growth stages, provides favorable conditions for wheat cultivation, reducing disease risks and supporting efficient grain filling. However, spatial heterogeneity in soil properties, crop varieties, and agronomic practices complicates large-scale yield prediction. This study focuses on China, the world's largest wheat producer, to develop an adaptable method for estimating yields in major wheat-growing regions. We coupled the WOFOST model with HYDRUS and CASA to simulate crop growth, soil moisture, and carbon cycling, assimilated remote sensing data for model calibration, and trained a machine learning model using multi-modal data to predict regional-scale winter wheat yields. The model, combining a Bagging Regressor with total above ground production (TAGP), gross primary productivity (GPP), and evapotranspiration (ET), achieved optimal performance (R = 0.83, RRMSE = 0.12) and reduced prediction errors compared to Global-WheatYield4km, with an average error of 1.92%. The model also reliably predicted yields across varying conditions, with discrepancies under 1,000 kg/ha in most counties. This integrated approach enhances yield prediction stability and precision, offering a solid foundation for regional-scale agricultural management.