Winter wheat yield prediction using UAV imagery and machine learning: case studies in Sweden and Morocco

Sammanfattning

To address global food security challenges caused by population growth and climate change, accurate prediction of crop yield is essential for sustainable agriculture. Winter wheat, a key global crop, faces production instability in regions like Sweden and Morocco due to extreme weather conditions. This study employed unmanned aerial vehicle (UAV)-based multispectral imagery and digital surface models (DSMs) data derived from UAV data, integrated with three machine learning models, Random Forest (RF), Support Vector Machine Regression (SVR), and Extreme Gradient Boosting (XGBoost), to forecast winter wheat yield at high spatial resolution in experimental fields in Sweden (humid climate) and Morocco (arid to semi-arid climate). Prediction results revealed that the grain-filling growth stage was the best stage for wheat yield prediction in both sites, with the SVR model demonstrating best performance (Sweden: R²=0.88, MAE=0.71 t/ha, RMSE=0.96 t/ha; Morocco: R²=0.83, MAE=0.42 t/ha, RMSE=0.56 t/ha). Four spectral band reflectance and ten vegetation indices (VIs) showed strong correlations with yield during the heading/flowering and grain-filling stages in both regions. Combing relative wheat height information enhanced prediction accuracy during early growth stages (jointing /booting stage and heading/flowering stage) but introduced uncertainty in the grain-filling stage. Cross-regional model transferability was limited, with better results when using the larger Swedish dataset to predict yields in Morocco than vice versa, underscoring the role of dataset size and yield variability. These results confirm the potential of UAV-based remote sensing combined with machine learning for precise, within-field winter wheat yield predictions, providing practical insights for improving agricultural strategies. The study emphasizes the importance of phenological timing, spectral features, and sufficient data volume, laying a foundation for future advancing cross-regional yield prediction methodologies.