Application of Machine Learning Methodologies for Predicting Corn Economic Optimal Nitrogen Rate

Authors

Zhisheng Qin, Dupont PioneerFollow
D. Brenton Myers, Dupont Pioneer
Curtis J. Ransom, University of Missouri
Newell R. Kitchen, USDA, Agricultural Research ServiceFollow
Sang-Zi Liang, Dupont Pioneer
James J. Camberato, Purdue University
Paul R. Carter, Dupont Pioneer
Richard B. Ferguson, University of Nebraska-LincolnFollow
Fabian G. Fernandez, University of Minnesota
David W. Franzen, North Dakota State University--Fargo
Carrie A. M. Laboski, University of Wisconsin-Madison
Brad D. Malone, Dupont Pioneer
Emerson D. Nafziger, University of Illinois at Urbana-Champaign
John E. Sawyer, Iowa State University
John F. Shanahan, Fortigen

Document Type

Article

Date of this Version

2018

Citation

Agron. J. 110:2596–2607 (2018)

Comments

Agron. J. 110:2596–2607 (2018)

Abstract

Determination of in-season N requirement for corn (Zea mays L.) is challenging due to interactions of genotype, environment, and management. Machine learning (ML), with its predictive power to tackle complex systems, may solve this barrier in the development of locally based N recommendations. The objective of this study was to explore application of ML methodologies to predict economic optimum nitrogen rate (EONR) for corn using data from 47 experiments across the US Corn Belt. Two features, a water table adjusted available water capacity (AWC_wt) and a ratio of in-season rainfall to AWC_wt (RAWC_wt), were created to capture the impact of soil hydrology on N dynamics. Four ML models— linear regression (LR), ridge regression (RR), least absolute shrinkage and selection operator (LASSO) regression, and gradient boost regression trees (GBRT)—were assessed and validated using “leave-one-location-out” (LOLO) and “leave-one-year-out” (LOYO) approaches. Generally, RR outperformed other models in predicting both at planting and split EONR times. Among the 47 tested sites, for 33 sites the predicted split EONR using RR fell within the 95% confidence interval, suggesting the chance of using the RR model to make an acceptable prediction of split EONR is ~70%. When RR was used to test split EONR prediction with input weather features surrogated with 10 yr of historical weather data, the model demonstrated robustness (MAE, 33.6 kg ha^–1; R² = 0.46). Incorporating mechanistically derived hydrological features significantly enhanced the ability of the ML procedures to model EONR. Improvement in estimating in-season soil hydrological status seems essential for success in modeling N demand.