Machine learning approaches for wheat yield prediction integrating biophysical modeling and remote sensing: Effects of sample size, dimensionality, and transferability

Prediction of wheat yield plays a crucial role in mitigating the impacts of climate extremes and ensuring food security. While vegetation indices (VIs) derived from remote sensing have been widely used for this purpose, integrating biophysical variables, such as crop stress, simulated using a biophysical model has the potential to enhance model robustness. However, the role of integrating biophysical stress information in machine learning based yield prediction, particularly in terms of sample size efficiency, dimensionality reduction, and spatial transferability, remains insufficiently examined. In this study, we address this gap by systematically evaluating the impact of incorporating a Stress Index (SI) using three widely applied machine learning (ML) models Random Forest (RF), Gaussian Process Regression (GPR), and Extreme Gradient Boosting (XGBoost) to predict wheat yield across five Australian states. These models leveraged both Sentinel-2-derived VIs at the peak of greenness and a biophysical factor: the SI. The integration of the SI led to clear improvements in model performance. For XGBoost, the coefficient of determination (R²) increased from 0.65 to 0.73, and the RMSE decreased from 0.67 to 0.58. GPR showed an increase in R² from 0.66 to 0.70 and a reduction in RMSE from 0.76 to 0.60. RF reached an R² of 0.68 and an RMSE of 0.64 after including SI. Additionally, SI contributed to dimensionality reduction, lowering the number of required VIs from 11 to 6 without loss of accuracy. The inclusion of SI also improved spatial transferability. When tested in Western Australia, cross-regional RMSE decreased by 0.09 for XGBoost, 0.16 for GPR, and 0.13 for RF. These results highlight that integrating biophysical crop responses, such as SI, into ML model calibration can enhance the reliability and robustness of yield prediction using remote sensing data.