What Data Layers are Important for Variable Rate Soybean Seeding Prescriptions?
Field Facts written by Ethan Smidt, John Gaska and Shawn Conley, Ph.D., Department of Agronomy; Jun Zhu, Ph.D., Department of Statistics and Department of Entomology; University of Wisconsin-Madison. - (University/DuPont Pioneer Research Summary1)
- Many growers have abundant spatial data from their soybean fields, but do not have a clear understanding of how to use these data in variable rate technology prescriptions to improve yields.
- Research was conducted at 22 locations over 2 years in Wisconsin to help identify which production factors are the best predictors of soybean yield.
- Across all locations, soil type, slope and other characteristics (collectively called “soil symbol”) were determined to be the most important yield predictors, regardless of year.
- When looking at fields on an individual basis, elevation and the soil sampling factors of phosphorus, potassium, organic matter and pH were the most important predictors.
- The results suggested that variable rate technology soybean prescriptions are useful in certain cases, but other factors are better predictors of soybean yield and should be analyzed and addressed first.
Growers are collecting many forms of spatial data for their fields, including yield, elevation and soils data. Highly accurate GPS systems along with advances in variable rate technology (VRT) are allowing growers to create and use variable rate planting prescriptions to optimize soybean yields and seed placement (Hoeft et al., 2000). As soybean seed prices continue to rise (USDA-ERS, 2014), growers are looking for ways to optimize seeding rates across their fields (Hoeft et al., 2000). However, growers and researchers alike feel there is an abundance of raw data but a shortage of methods and knowledge on how to use the data for advancements in precision agriculture (Bullock et al., 2007). To address these issues, a research study was initiated.
The objectives of this research were to:
- find the key measureable predictors determining soybean yield in Wisconsin, and
- use those predictors to create accurate, data-based variable rate technology prescriptions.
This study was conducted on a total of 22 sites between 2013 and 2014 as shown in Figure 1. Seeding rate prescriptions containing 3 unique rates were created prior to planting for each site as shown in Figure 2. The middle seeding rate was equivalent to the single rate each individual grower would have used in their respective field without variable rate technology capabilities and the high and low rates were targeted at ±30% from the medium rate. After planting, soil samples were taken at geo-referenced points and submitted for pH, organic matter, phosphorus and potassium levels. Soil survey and satellite imagery data were also obtained during the growing season to determine any possible relationships with soybean yield.
Figure 1. Research locations.
Figure 2. Example of seeding rate and soil type map.
Climate and Planting Factors
Soybean plant counts were taken at the same geo-referenced points used for soil sampling to verify the prescriptions were applied correctly. Relative emergence compared to the planted rate is shown for each field in Figure 3 and the 80% emergence level is highlighted by a dashed line. The 2013 growing season was more stressful, both early and late in the season, compared to 2014 (National Climate Data Center, 2015). As a result, some sites had noticeably low emergence. Discussions with the growers at these locations revealed that field conditions (soil moisture, temperature, etc.) and equipment (coulters, age of disc openers, etc.) were most likely to blame.
Figure 3. Average soybean emergence levels at each site based on initial seeding rates.
The average soybean yield for the 2013 sites was 52 bu/acre with individual field averages ranging from 37 to 68 bu/acre, and the pooled average for the 2014 sites was 55 bu/acre with individual fields yielding from 30 to 69 bu/acre on average. The “random forest method” (Breiman, 2001), a statistical algorithm, determined predictor importance in each data set, and the ranked results are found in Table 1. Soil symbola was ranked as the most important factor regardless of year.
Table 1. Random forest resultsb from 2013
and 2014 pooled data.
|2013 Most Important
|2014 Most Important
|Soil Symbola||Soil Symbola|
|Soil Phosphorus||Soil Phosphorus|
|Soil Organic Matter||Elevation|
|Available water supply
from 0 to 39 inches
|Soil Potassium||Soil Organic Matter|
aA soil symbol (or "map unit symbol”) is a descriptive label on a soil map. It gives information about soil type (or “series”) and other soil features such as slope, erosion, etc.
bResults of a statistical algorithm used to rank items in order of their importance to soybean yield.
Individual Field Results
The results from similar analyses for individual fields were, in general, quite different compared to the pooled dataset from the same growing season. The predictor rankings were averaged (value in parentheses) and elevation was the top predictor for soybean yield across both years (Table 2).
Table 2. Average random forest resultsb from 2013
and 2014 individual field analyses.
|2013 Individual Field
|2014 Individual Field
|Elevation (1.55)||Elevation (2.00)|
|Soil Organic Matter (3.18)||Soil pH (3.09)|
|Soil Potassium (3.36)||Soil Potassium (3.27)|
|Soil Phosphorus (4.09)||Soil Organic Matter (3.45)|
|Soil pH (4.09)||Soil Phosphorus (3.82)|
bResults of a statistical algorithm used to rank items in order of their importance to soybean yield.
The commonly used soil sampling variables of organic matter, potassium, phosphorus and pH made up the rest of the top 5 predictors in both years. Soil symbol fell to 6th most important on average when looking at individual fields. The National Commodity Crop Productivity Index (NCCPI) was not determined to be an important predictor at any site.
Elevation and the soil sampling factors of soil phosphorus, potassium, organic matter and pH were the most important predictors when looking at fields on an individual basis.
Satellite Imagery and Quantile Regression Results
Satellite images were gathered from June to September for 2 sites in 2013, and for 3 sites in 2014. Early season (June) images showed no correlation to final soybean yield in either year. At both sites in 2013, the late-season (early September) NDVI values showed high correlation to yield (r values of 0.762 and 0.857). The 2014 sites showed less correlation overall, with the highest correlation appearing in the midseason (July/Aug) images at 2 sites (r values of 0.425 and 0.77) and the remaining site showing the highest correlation in September (0.486). Quantile regression was used to see if the seeding rate impacted yield across the yield ranges in each field. Only 4 of the 22 sites (18%) had a majority of the data points fall outside the linear regression, meaning the remaining sites had a consistent relationship between seeding rate and yield throughout the field. However, over 36% of the fields had a negative linear regression slope, which means that yield decreased as seeding rate increased.
Soil symbol was by far the most important variable for predicting soybean yield in both the 2013 and 2014 statewide pooled data sets. This could be useful for wide-ranging recommendations and statewide research. However, elevation and the soil sampling factors of phosphorus, potassium, organic matter and pH were the most important predictors when looking at fields on an individual basis. Since this type of analysis is possible for many growers and agronomists, these factors should be more useful for specific fields if the data are available.
NDVI and other aerial imagery data were unable to accurately predict soybean yield until mid- to late-summer and were more accurate during the 2013 growing season when many fields were exhibiting late-season stress. It also appears that scale is an important factor when determining the predictors best for characterizing soybean yield due to the differences between the pooled and individual data sets.
The pooled results can be used for general recommendations; however, if accurate data are available for specific fields, more accurate results would be likely and should be addressed in order of importance. In short, variable rate technology soybean prescriptions are useful in certain cases, but other factors are better predictors of soybean yield and should be analyzed and addressed first. A ‘one size fits all’ approach for creating the prescriptions is not recommended due to the numerous possible differences between fields.
Breiman, L. 2001. Random forests. Machine learning. 45:5-32.
Bullock, D.S., N. Kitchen, and D.G. Bullock. 2007. Multidisciplinary teams: A necessity for research in precision agriculture systems. Crop Sci. 45(5) 1765-1769.
Hoeft, R.G., E.D. Nafziger, R.R. Johnson, and S.R. Aldrich. 2000. Modern corn and soybean production. 1st ed. MCSP Publications, Champaign, IL
National Climate Data Center. 2015. Climate data online. Accessed on Mar 5, 2015.
Smidt, E., J. Gaska, S. Conley, and J. Zhu. 2015. What data layers are important for variable rate soybean seeding prescriptions? University of Wisconsin, Madison. Accessed on Sept. 15, 2015.
USDA-ERS. 2014. Soybean costs and returns. Accessed on Feb. 11, 2015.
1Research conducted as a part of the DuPont Pioneer Crop Management Research Awards (CMRA) Program. This program provides funds for agronomic and precision farming studies by university and USDA cooperators throughout North America. The awards extend for up to four years and address crop management information needs of DuPont Pioneer agronomists, Pioneer sales professionals, and customers. This Field Facts publication was adapted from the final report entitled "What data layers are important for variable rate soybean seeding prescriptions?" by Ethan Smidt, John Gaska and Shawn Conley, Department of Agronomy; and Jun Zhu, Department of Statistics and Department of Entomology; University of Wisconsin- Madison. This project was sponsored by DuPont Pioneer Agronomy Sciences and coordinated by Barry Anderson and Paul Carter with local support from DuPont Pioneer Field Agronomists Arnie Imholte, Bob Berkevich, and Aaron Prestemon.