Soybean plays an important role in the Brazilian agricultural scenario due to its large contribution to national crop production and exports. This study developed a municipal-level soybean yield model for the Brazilian Midwest by integrating Sentinel-2 imagery, climate data, and XGBoost. Soybean areas were identified from 2019/2020 to 2021/2022 using land cover maps. For each municipality, average values of spectral bands and vegetation indices from Sentinel-2, climate data, and historical soybean yield were extracted. Six models were trained (70% of data) and tested (30% of data) based on days after sowing (DAS): 30, 60, 90, 120, 150 and 180 DAS. The best performing model, the Soy Yield Model (SYM), selected at 150 DAS, achieved an R2 of 0.72 and RMSE of 301.52 kg/ha. When applied to independent states, SYM maintained robust performance (R2 from 0.34 to 0.76 and RMSE from 168.31 to 491.17 kg/ha). Precipitation, solar radiation, and water deficit were the most influential climate variables, while NDRE and the red-edge, SWIR, red, and NIR bands were the most important spectral features. Results showed the potential of integrating Sentinel-2 and climate data for accurate, scalable and early-season soybean yield prediction at municipal level, providing a practical framework for regional crop monitoring and yield estimation.
Pereira et al. (Mon,) studied this question.