On-farm soybean seed protein and oil prediction using satellite data

Soybean [Glycine max L. (Merr.)] seed composition is receiving increased attention among farmers, agronomists, and commodity traders. Increasing the ability to predict seed quality traits such as protein and oil at the field level before harvest will provide a competitive ability to segregate quality and create an economic advantage to position the production at both domestic and global markets. Therefore, this study aims to use remote sensing satellite data to spatially predict soybean seed protein and oil concentrations at the field level before harvest time. The dataset consisted of 47 fields located in Kansas and Iowa, United States, from the 2019 to 2021 seasons. Six machine-learning approaches (ElasticNet, Random Forest, XGBoost, LightGBM, CatBoost, and an ensemble) were tested evaluating different vegetation indices and spectral bands to predict before harvest seed protein and oil concentrations from satellite imagery. The optimal timing for training prediction models was identified within a week after the peak of the green chlorophyll vegetation index, with different spectral indices and bands of importance for each seed quality component. The XGBoost outperformed the rest of the algorithms for both seed quality traits. Overall, models reported an absolute error of 1.80 % for protein and 1.04 % for oil concentrations. Our research describes a pipeline that combines on-farm data, open access satellite imagery, an intensive use of spectral bands, and machine learning to forecast seed quality before harvest. Future research guiding crop management interventions should be directed to i) integrating major drivers of spatial variation of seed quality traits such as soil and weather data, and ii) exploring satellite data-fusion approaches and iii) assesing alternatives models such as deep learning methods.