Within-season crop yield prediction in the Vojvodina Province (Serbia) using Sentinel-2 satellite vegetation indices

The main objective of this study was to develop and evaluate machine learning-based models capable of providing within-season predictions of crop yield. To achieve this, we developed algorithms based on Sentinel-2 satellite vegetation indices of plant health and moisture content, at the field level in the Vojvodina Autonomous Province (Serbia) for maize, wheat, soybean and sunflower. Yield data collected during several growing seasons (2017–2024) were used to train Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost) algorithms and to develop regression models to predict yields at least two months prior to harvest. Model performance was assessed using monthly mean 3-month and 2-month datasets and statistical metrics R2, RMSE, MAE, rRMSE and Bias. RF consistently achieved the best predictive performance across all crops, except for soybean, where the results of XGBoost matched RF’s performance (R2=0.90, MAE = 0.26 t/ha) while other models also showed strong performance (R2 ≥0.86). For maize, RF and XGBoost performed similarly well (R2 up to 0.82, MAE to 0.93 t/ha), whereas SVM showed lower accuracy. Wheat was predicted with moderate performance using RF (R2=0.73, RMSE = 0.4 t/ha) and sunflower exhibited the best predictability also with RF (R2=0.62). SHAP feature analysis showed that mid-season variables consistently provided the greatest contribution to yield prediction across all crops. Differences between the 2-month and 3-month windows are minor, conforming to the reliability of within-season predictions, with RF providing the most robust and stable results overall.