Wheat yield prediction for sustainable food security using Sentinel-2 and machine learning

Accurate pre-harvest wheat yield estimation is critically important for sustainable agricultural policies. In this study, 13 vegetation indices derived from Sentinel-2A/B images (10 from the literature and 3 derived for this study), temperature, precipitation, and sunshine duration data from the ERA-5 dataset, and ground measurements were used. Pearson correlation analysis was applied to determine the most suitable period for yield estimation, and the most effective factors were identified using the Decision Tree (DT) model. Considering these factors, yield estimation was performed using the Multiple Linear Regression (MLR) model. The results showed that the factors most affecting wheat yield were NDVI, WDRVI, RDVI, and Red Edge 1 in April; Derived Index 3 in December; and LAI, average temperature, and total precipitation in March. The heading-flowering period has been determined as the most suitable time for yield estimation. There is a high degree of agreement between the actual and predicted yields (R = 0.92, R² = 0.86).