On-farm soybean yield estimation using earth observation data and machine learning models

Abstract

Soybean ( Glycine max (L.) Merr.) is a multi-purpose crop, with yields that have increased significantly since 2000s in the United States. Given the complex mechanisms involved in vegetation growth, soybean yield estimation is determined by many factors, including weather, topographical attributes (TA), management practices, and cultivars. Conventional methods for estimating soybean crop maturity and yields are labor-intensive and include biases. Emerging methodologies for estimating crop maturity and yields that leverage remotely sensed data are increasingly gaining prominence. The objectives of this study were: (1) to identify the most relevant features derived from multi-source Earth observation data such as digital elevation model (DEM) based topography, vegetation indices (VIs), and soil indices (SIs)—for enhanced soybean yield estimation, and (2) to systematically evaluate the performance and robustness of extreme gradient boosting (XGBoost) model, in estimating on-farm soybean yields using “Leave a field out (LOFO) cross validation (CV)” approach. To achieve these objectives, we applied the XGBoost model to estimate soybean yields across three agricultural fields in South Dakota, USA, using data from the 2019 and 2021 growing seasons. The ML model estimated yields based on high-resolution PlanetScope satellite multispectral imagery and a DEM. The multispectral PlanetScope data were obtained for six soybean growth stages, ranging from early germination through maturity, and DEM was used to calculate the site’s topographic features. Soybean yield data were obtained from a GPS/yield montor equipped combine. The XGBoost model had the highest R 2 (0.54) at the Edmunds and Miner trial fields and the lowest RMSE (574 kg ha -1 ) and MAE (455 kg ha -1 ). A SHAP analysis showed that DVI exhibited the largest SHAP value range (approximately −1 to +1). Spatial maps showed differences in prediction error between Edmunds in 2019 and 2021. Over and underestimation are clearly visible through maps. These findings indicate that ML models can successfully estimate yields and underscore the importance of appropriate model selection. Accurate and timely on-farm yield estimation is a valuable tool that enables growers to make scientifically informed crop management and marketing decisions.