Application of machine learning methodologies for predicting corn economic optimal nitrogen rate

Authors: QIN, Z - Dupont Pioneer Hi-Bred; MYERS, D - Dupont Pioneer Hi-Bred; RANSOM, C - University Of Missouri; Kitchen, Newell; LIANG, S - Dupont Pioneer Hi-Bred; CAMBERATO, J - Purdue University; CARTER, P - Dupont Pioneer Hi-Bred; FERGUSON, R - University Of Nebraska; FERNANDEZ, F - University Of Minnesota; FRANZEN, D - North Dakota State University; LABOSKI, C.A.M. - University Of Wisconsin; MALONE, B - Dupont Pioneer Hi-Bred; NAFZIGER, E - University Of Illinois; SAWYER, J - Iowa State University; SHANAHAN, J - Fortigen

Submitted to: Agronomy Journal
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/30/2018
Publication Date: 10/18/2018
Citation: Qin, Z., Myers, D.B., Ransom, C.J., Kitchen, N.R., Liang, S., Camberato, J.J., Carter, P.R., Ferguson, R.B., Fernandez, F.G., Franzen, D.W., Laboski, C., Malone, B., Nafziger, E.D., Sawyer, J.E., Shanahan, J.F. 2018. Application of machine learning methodologies for predicting corn economic optimal nitrogen rate. Agronomy Journal. 110(6):2596-2607. https://doi.org/10.2134/agronj2018.03.0222.
DOI: https://doi.org/10.2134/agronj2018.03.0222

Interpretive Summary: Nitrogen (N) fertilizer management is critical in corn production. However, determination of in-season N demand is challenging because the underlying relationships of genetics (G), environmental factors (E), and management (M), jointly impact in-season N availability and crop demand. The difficulty arises from the interactions of these factors, often represented as “GxExM”. An approach for exploring the complexities of these interactions and determining corn N fertilizer recommendations that has had little investigation is the use of computer-based machine learning (ML) methods. Machine learning uses algorithms to ‘learn’ the structure and behavior of underlying processes from large datasets. We used four ML models to explore the relationships of soil and weather to the target variable of corn economic optimal N rate (EONR). This was done using data obtained from across the U.S. corn belt in the 2014-2016 growing seasons, both when all N fertilizer was applied at planting and for split-applications with a small amount at planting and the rest at side-dress. The ML models examined explained up to as much as 50% of the variability in EONR. The ML model that consistently performed best for explaining EONR at both planting and side-dress application times was called “Gradient Boost Regression Trees”. This ML model was 10 to 60% better than the other ML approaches. Two calculated soil-based features [water-table adjusted available water capacity (AWCwt) and the ratio of in-season rainfall to AWCwt] were found helpful in predicting EONR. From this we concluded that improvement in defining in-season soil hydrology seems essential for more accurately modeling soil and fertilizer N availability and crop N need. We also found improved results when combining 10-year historical weather data with the weather from the growing season the EONR was being evaluated for. This investigation helps explain important GxExM interactions, and can be the basis of improved N fertilizer recommendations. If fertilizer recommendations can be better matched with crop need, farmers will profit more because fertilizer nutrients will be more efficiently used by crops. At the same time, this can help reduce N loss to the environment (e.g., water and air).

Technical Abstract: Determination of in-season N requirement of corn (Zea mays L.) is challenging due to the complicated interactions of genotype, environment (e.g., weather and soil, including their variability), and management. Machine Learning (ML) utilizing soil, management, and weather features created based on field measurements and using domain knowledge was used to model Economic Optimum Nitrogen Rate (EONR) for corn grown at 44 Corn Belt sites in three growing seasons. Two features, water table adjusted Available Water Capacity (AWCwt) and ratio of in-season rainfall to AWCwt (RAWCwt), were created to capture impact of soil hydrology conditions on N dynamics. Four ML models - Linear regression (LR), Ridge regression (RR), LASSO regression (LAR), and Gradient Boost Regression Trees (GBRT) were assessed and validated using ‘leave-one-location-out’ (LOLO) and ‘leave-one-year-out’ (LOYO) approaches. Algorithms developed with GBRT consistently outperformed other models in predicting EONR for when N was applied at planting or when split applied, with improved R^2 and Mean Absolute Error (MAE). When GBRT was used to predict EONR from split-applied N, weather features surrogated with 10 years’ historical weather data produced a robust model with MAE of 28.5 kg/ha MAE and R^2=0.42.