Developing a Machine Learning and Proximal Sensing-based In-season Site-specific Nitrogen Management Strategy for Corn in the US Midwest
1D. Li, 2Y. Miao, 3 . G. Fernández, 1N. R. Kitchen, 4C. . Ransom, 5G. M. Bean, 6 . E. Sawyer, 7J. J. Camberato, 1 . R. Carter, 9R. B. Ferguson, 10D. W. Franzen, 11C. A. Laboski, 12E. D. Nafziger, 1J. F. Shanahan
1. Guangzhou Institute of Geography, Guangzhou, Guangdong, 510070 China
2. Precision Agriculture Center, Department of Soil, Water, and Climate, University of Minnesota, St. Paul, MN 55108, USA
3. Precision Agriculture Center, Department of Soil, Water, and Climate, University of Minnesota, St. Paul, MN 55108, USA;
4. USDA-ARS Cropping Systems and Water Quality Research Unit, Columbia, MO 65211 USA
5. McCain Foods, Oakbrook Terrace, IL 60181 USA
6. Iowa State University, Ames, IA 50011 USA
7. Purdue University, West Lafayette, IN 47907 USA
8. Corteva Agriscience, 7100 NW 62nd Ave., P.O. Box 1000, Johnston, IA 50131 USA
9. University of Nebraska, Lincoln NE 68583 USA
10. North Dakota State University, Fargo, ND 58108 USA
11. University of Wisconsin-Madison, Madison, WI 53706 USA
12. University of Illinois, Urbana, IL 61801 USA
13. Soil Health Institute, Morrisville, NC 27560 USA
Effective in-season site-specific nitrogen (N) management strategies are urgently needed to ensure both food security and sustainable agricultural development. Different active canopy sensor-based precision N management strategies have been developed and evaluated in different parts of the world. Recent studies evaluating several sensor-based N recommendation algorithms across the US Midwest indicated that these locally developed algorithms generally did not perform well when used broadly across the US Midwest. Efforts have been made to improve the prediction of economically optimum N rates (EONR) using soil, weather, and management information with statistical and machine learning methods. However, these efforts only resulted in R2 ≤ 0.57 and therefore could still improve. The objective of this research was to develop a machine learning-based in-season and site-specific N recommendation strategy by predicting yield as a response to N rates by incorporating soil information, early season weather conditions, and preplant and sidedress N application information. Data from 36 site-year N rate trials conducted over three years (2014-2016) in eight US Midwest states were used. At each site-year, there were a total of 16 N rate treatments with different preplant and sidedress combinations. A portable active canopy sensor, RapidSCAN CS-45, was used to collect canopy reflectance at V6-V10 stages before a sidedress N application. Three machine learning algorithms (Random Forest, Extreme Gradient Boost, and Light Gradient Boosting Decision Tree) were used to develop corn yield estimation models. A total of 1,492 data points were used to develop the models, 467 data points were used to test performance of the models, and the remaining 374 data points were used to validate performance of the models. Input variables included normalized difference vegetation index (NDVI), corn heat units, growing degree days, abundant and well-distributed rainfall, Shannon Diversity Index, precipitation, irrigation, drainage, pre-plant N rate, side-dress N rate, seeding rate, tillage practice, and soil texture (clay, silt, and sand percentage). The preliminary results indicated that NDVI-based in-season estimate of yield (INSEY) could only explain 3% of corn yield variability at best, with root mean square error (RMSE) being 2693 kg ha-1 for validation. The validation results of the three machine learning methods were similar (R2=0.76-0.77) with the RMSE ranging from 1277 to 1319 kg ha-1,which were all significantly better than the INSEY results.The yield models were used before sidedressing (around the V9 stage) to predict corn yield responses to a series of side-dress N rates at small increments (e.g. 10 kg ha-1), and the results were used to determine EONR. More analyses are being performed and the results will be presented at the conference.
