Developing an innovative stacking ensemble machine learning and multi-source data fusion-based precision nitrogen management strategy for corn

Abstract Purpose The primary goal of this research was to develop an innovative in-season nitrogen (N) recommendation strategy for corn ( Zea mays L.) using stacking ensemble machine learning (ML) and multi-source data fusion. Methods Forty-nine site-years of N rate experiments conducted across the U.S. Corn Belt were used to evaluate the performance of five individual ML algorithms (Random Forest Regressor (RFR), Support Vector Regressor (SVR), Extreme Gradient Boosting Regressor (XGBR), CatBoost Regressor (CBR), and Multi-Layer Perceptron (MLP)) and stacking regression (STR) for in-season corn yield prediction under different preplant and split N application conditions using the active canopy sensor data along with genetics, environmental and management information. These models were further evaluated for their prediction of yield responses to sidedress N application rates and in-season estimation of site-specific economic optimal N rate (EONR) across the U.S. Corn Belt. Results The results indicated that the stacking model performed consistently well across all datasets for corn yield prediction, demonstrating robustness (R 2 > 0.85 for validation dataset). Preplant N rate, Sidedress N rate and normalized difference red edge (NDRE) were identified as key variables for predicting corn yield. For EONR estimation, the stacking regression 2 model (STR2) using RFR, SVR, XGBR, CBR, and MLP as base estimators and linear regression as the meta estimator performed the best for the full dataset (R 2 = 0.82 and root mean square of error (RMSE) = 27.50 kg N ha − 1 ). Conclusion It is concluded that the stacking regression and multi-source data fusion framework is a promising strategy for in-season site-specific corn yield prediction and sidedress N recommendation.