Durum Wheat Yield Prediction: A Machine Learning Framework Integrating Sentinel-2 Imagery and Meteorological Data

Accurate crop yield prediction provides essential information for strategic decision-making and precision agriculture, yet achieving reliable forecasts across unseen growing seasons remains challenging. The main objective of this study is to develop a two-stage data assimilation framework that integrates multi-scale environmental data to improve yield predictions and overcome the spatial–temporal autocorrelation limitations of standard machine learning (ML) models. Utilizing an eight-year continuous dataset (2018–2025) of durum wheat fields in Thessaly, Greece, this study integrates high-resolution Sentinel-2 multispectral imagery with macro-scale ERA5-Land meteorological variables. Eight ML algorithms are trained to predict yield at the pixel level. To test model generalization and prevent overfitting, the framework is evaluated using both standard random splitting and leave-one-year-out (LOYO) cross-validation. Concurrently, multiple linear regression (MLR) is utilized to select the most significant meteorological predictors from monthly temperature (maximum and minimum) and precipitation data, which are then integrated into the pixel-level predictions via additive and multiplicative late-fusion assimilation. The results demonstrate that, while standard random splitting produces high explained variance (R2 > 0.90), LOYO validation shows a predictive maximum of approximately 50% explained variance. The late-fusion assimilation slightly reduces interannual offsets, from an RMSE of 966 kg ha−1 to 902 kg ha−1. However, the R2 values remain static due to informational saturation. This study concludes that, while integrating regional climate data improves absolute annual yield magnitudes, securing reliable agricultural forecasts requires the integration of localized agronomic metadata, such as soil properties and field-specific management practices.