This study evaluated the potential of Sentinel-2-derived NDVI (Normalized Difference Vegetation Index) for predicting within-field variability of winter wheat grain yield in central Lithuania during the 2024 growing season. Reliable within-field yield prediction remains challenging in regions with heterogeneous soils and limited region-specific models, particularly in northeastern Europe. Grain yield data were obtained from combine harvesters equipped with GPS yield monitoring across 13 fields with a total area of 283.6 ha. NDVI values were calculated for four half-monthly periods from March to May, corresponding to key phenological stages (from tillering to spike emergence). Spatial and temporal variability in NDVI–yield relationships was observed, with early May consistently showing the strongest correlations (r up to 0.49), particularly in lower-fertility fields, indicating its critical role in yield prediction. Machine learning models (Random Forest, XGBoost, and Deep Neural Networks), along with linear regression, were applied to predict yields based on NDVI from four growth stages. Random Forest achieved the highest predictive accuracy (MAE = 0.951 t/ha), outperforming the other models. The model also showed the highest correlation with observed yields (Pearson r = 0.717), indicating strong agreement between predicted and measured values. Feature importance analysis confirmed NDVI from 1 to 15 May as the most influential predictor across all models. Predicted yield maps closely matched observed patterns, with the largest discrepancies near field edges due to combine harvester effects. These findings highlight the utility of mid-season NDVI for precise estimation of within-field grain yield variability and demonstrate that Random Forest models can effectively capture the NDVI–yield relationship, particularly under heterogeneous field conditions.
Okupska et al. (Sun,) studied this question.