Developing a machine learning and proximal sensing-based in-season site-specific nitrogen management strategy for corn in the US Midwest

Authors: LI, DAN - University Of Minnesota; MIAO, YUXIN - University Of Minnesota; Ransom, Curtis; FERNANDEZ, FABIAN - University Of Minnesota; Kitchen, Newell; BEAN, GREGORY - McCain Foods, Inc; CAMBERATO, JAMES - Purdue University; CARTER, PAUL - Farmer; FERGUSON, RICHARD - University Of Nebraska; FRANZEN, DAVID - North Dakota State University; LABOSKI, CARRIE - University Of Wisconsin; NAFZIGER, EMERSON - University Of Illinois; SAWYER, JOHN - Iowa State University; SHANAHAN, JOHN - Agoro Carbon Alliance

Submitted to: International Conference on Precision Agriculture Abstracts & Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 6/1/2022
Publication Date: 6/26/2022
Citation: Li, D., Miao, Y., Ransom, C.J., Fernandez, F., Kitchen, N.R., Bean, G., Camberato, J., Carter, P., Ferguson, R., Franzen, D., Laboski, C., Nafziger, E., Sawyer, J., Shanahan, J. 2022. Developing a machine learning and proximal sensing-based in-season site-specific nitrogen management strategy for corn in the US Midwest. Proceedings of the 15th International Conference on Precision Agriculture, June 26-29, 2022, Minneapolis, Minnesota. Available: https://ispag.org/proceedings/

Interpretive Summary: To improve nitrogen (N) fertilizer management in order to maximize corn yields and minimize environmental issues, new tools could be developed using machine learning techniques. The purpose of this analysis was to 1) evaluate four different machine learning algorithms for predicting grain yield using a variety of inputs from soil, weather, canopy reflectance sensors, and management information; and 2) use the best machine learning model to make in-season N fertilizer rate recommendations. Results showed that using some machine learning algorithms to predict grain yield using soil, weather, plant reflectance measurements, and management information can perform very well. Using the best machine learning model (R^2 of about 0.75) proved to be a promising method for recommending accurate N rates during the growing season. Further development is needed but a final product could help farmers better understand how much N to apply to optimize their grain yield. Further, these results show promise for helping finding N management strategies that will result in fewer related environmental issues that result from over applying N fertilizer.

Technical Abstract: Effective in-season site-specific nitrogen (N) management or precision N management (PNM) strategies are urgently needed to ensure both food security and sustainable agricultural development. One promising method is the use of active canopy reflectance sensor-based PNM strategies, which have been developed and evaluated in different parts of the world. However, recent studies have shown that sensor-based N recommendation algorithms developed in localized regions of the US generally did not perform well when used broadly across the US Midwest. While efforts have been made to improve these algorithms using soil, weather, and management information, they could still be improved upon. The objective of this research was to develop a machine learning-based in-season and site-specific N recommendation strategy by incorporating active canopy sensor data with soil, weather, and management information. Data consisted of 2,333 observations from 36 site-year N rate trials conducted over three years (2014-2016) in eight US Midwest states. At each site-year, there were 16 N rate treatments with different pre-plant and side-dress combinations. A portable active canopy reflectance sensor, RapidSCAN CS-45, was used to collect canopy reflectance at V6-V10 stages before a side-dress N application. Four machine learning algorithms [ridge regression (RR), random forest regression (RFR), extreme Gradient Boost regression (XGBR), and support vector regression (SVR)] were used to develop the corn yield estimation model. Models were trained, tested, and validated on a subset of the data containing 64, 20, and 6% of the data, respectively. The input variables included normalized difference vegetation index, normalized difference red-edge index, Maccioni index, Canopy chlorophyll content index, corn heat units, growing degree days, abundant and well-distributed rainfall, Shannon Diversity Index, precipitation, irrigation, pre-plant N rate, side-dressed N rate, seeding rate, previous crops, tillage practice and soil texture (clay, silt, and sand percentage). The R^2 for validation results of the four machine learning models ranged from 0.53 to 0.76 with the relative root-mean-square error (RRMSE) ranging from 15.62 to 11.07. Among the machine learning models, the XGBR model outperformed the other models with an R2 of 0.76 and an RRMSE of 11.07. Using the XGBR model, the economical optical N rate (EONR) was also estimated based on in-season simulated yield response to side-dress N application rates, with the predicted EONRs at 69% of site-years being within 45 kg/ha of the EONRs based on measured yield. These results show that a regional dataset can be used to derive a robust canopy reflectance sensors-based recommendation that works well across the region. More studies are needed to further improve this in-season N recommendation strategy and evaluate it under on-farm conditions.