Identifying Field Attributes that Predict Soybean Yield Using Random Forest Analysis

Until recently, soybean Glycine max (L.) Merr. fields were often seeded at a single rate. Advances in GPS and variable rate technology (VRT) are allowing growers to use variable rate planting prescriptions to optimize yields and input costs. This study was conducted to find the key predictors for characterizing soybean seed yield from commonly collected precision agriculture data layers. Research was conducted in 11 unique fields both in 2013 and 2014 in Wisconsin and all 22 site‐years were following corn Zea mays (L.). Seeding rate, soil sampling, yield, and soil survey data were gathered from each site for analysis. A statistical procedure used in the determination of key prediction parameters, random forest analysis, was used and identified soil map unit as the most important variable when predicting soybean seed yield for the pooled data sets in both 2013 and 2014. The next most important factors were, in order of importance, soil P, soil organic matter, soil available water supply in the upper 100 cm, soil K, and elevation in 2013 and soil P, elevation, soil K, soil organic matter, and soil available water supply in the upper 150 cm during 2014. Individual field random forest analyses determined elevation was the most important predictor of soybean yield, on average, for 2013 and 2014 followed by organic matter, K, P, and pH in 2013 and pH, K, organic matter, and P in 2014. Core Ideas Commonly accessed soybean yield predictors can vary depending on scale. Soil pH, organic matter, P, and K are good predictors of soybean yield in Wisconsin. Random forest and decision tree analyses are useful and accurate analysis tools.