In-season prediction of corn yield and economic optimum nitrogen rate using stacking regression

Authors: LI, DAN - University Of Minnesota; WANG, CHONGYANG - Guangdong Academy Of Agricultural Sciences; MIAO, YUXIN - University Of Minnesota; FERNANDEZ, FABIAN - University Of Minnesota; Ransom, Curtis; CHEN, SHUISEN - Guangdong Academy Of Agricultural Sciences

Submitted to: European Conference on Precision Agriculture Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 6/29/2025
Publication Date: 6/29/2025
Citation: Li, D., Wang, C., Miao, Y., Fernandez, F., Ransom, C.J., Chen, S. 2025. In-season prediction of corn yield and economic optimum nitrogen rate using stacking regression. In: Precision agriculture '25. Proceedings of the 15th European Conference on Precision Agriculture, June 29-July 3, 2025, Barcelona, Spain. p. 1357-1364.

Interpretive Summary: Farmers struggle to determine the right amount of nitrogen fertilizer for their corn fields. Traditional methods for predicting yield and nitrogen needs are often inaccurate, as they don't consider field conditions like soil, weather, and farming practices. In this study, we developed and tested different machine learning methods to solve this problem. We used data from 49 experiments conducted across the U.S. Corn Belt between 2014 and 2016 to evaluate three machine learning models: random forest regression, transfer learning regression, and stacking regression. The stacking regression model performed the best, accurately predicting corn yield and the optimal nitrogen rate. By combining data from multiple sources and using advanced machine learning models, researchers can offer a promising approach to help farmers use nitrogen fertilizer more efficiently. This would improve yields and reduce environmental impacts.

Technical Abstract: Accurate in-season estimation of the yield and yield response to nitrogen (N) fertilizer is a crucial step for precision nitrogen management. Efforts have been made to improve the prediction accuracy of the economic optimum N rates (EONR) by integrating soil, weather, and management information with statistical and machine learning (ML) methods. The objective of this study was to evaluate different ML models for in-season prediction of corn yield and its response to N application and develop an active canopy sensor and multi-source data fusion-based stacking regression model for in-season prediction of EONR across the US Midwest Corn Belt. Forty-nine site-years of N rate experiments across an array of soil, weather, hybrid, and management conditions of the US Corn Belt conducted from 2014 to 2016 were used for this research. Treatments included four replications of N fertilizer rates between 0 and 315 kg N ha-1 applied either all at planting or split where 45 kg N ha-1 was applied at planting and the remaining N fertilizer was applied at the V9±1 corn developmental stage. The performance differences of eight ML algorithms (Random Forest Regression (RFR), Support Vector Regression (SVR), Neural Network Regression (NNR), eXtreme Gradient Bossting Regression (XGBR), CatBoost Regression (CBR), Stacking Regression (STR), Transfer Learning Regression (TLR), and Long Short-Term Memory (LSTM) algorithms were evaluated for predicting corn grain yield and yield response to N application. The Shapley Additive explanation (SHAP) was employed to explain the ML models. Results show that RFR, TLR, and STR exhibited high R^2 values and low errors across all datasets in corn grain yield estimation, demonstrating robustness and good generalization capabilities, while Tree-based algorithm, SVR and NNR showed signs of over-fitting. Sidedress N rate, Preplant N rate, and Normalized Difference Red Edge (NDRE) were key variables influencing corn yield, with different models handling these variables differently, leading to variations in their influence. The STR model performed the best in EONR estimates, showing superior predictive capabilities (R^2 =0.89, RMSE=24.42 kg N ha^-1). It is concluded that integrating multi-source data, such as active sensor data, soil properties, weather variables, and management practices using stacking regression is a promising strategy to predict corn yield and EONR for in-season N management across the US Midwest Corn Belt.