Optimal Nitrogen Rate Can Be Predicted Using Average Yield and Estimates of Soil Water and Leaf Nitrogen with Infield Experimentation

Abstract

Core Ideas Nitrogen fertilizer decisions can be improved when on‐farm experimentation is combined with historical yields and the sensing of crop N and soil water status. Crop simulation and machine learning can be combined to gain insight into the key drivers that influence an N decision. In spring wheat crop, an N‐minus strip can provide advanced warning of an impending N stress. ABSTRACTCrop nitrogen (N) fertilizer decision aids traditionally provide an N fertilizer recommendation that requires information about the supply of N from the soil and the likely demand of N by the crop. These are often complex to implement. To simplify, we take elements of multiple N decision philosophies and, by way of simulation and machine learning, determine which variables need to be estimated to inform N decisions for rainfed wheat that are near optimal. Simulations were conducted using the APSIM (Agricultural Production Systems Simulator) crop wheat model for sites across the Australian grain belt for fields containing test strips where 0 or an additional 80 kg N ha−1 of fertilizer was added. The random forest analysis of 59 separate soil and plant variables within the 1.2 million simulations demonstrated that the long‐term mean yield was the most important variable to predict the optimal N rate for a particular site and season. When long‐term mean yield was used to predict the optimal N regime for a given season alone, the RMSE of the predicted N rate was 53.9 kg N ha−1. The RMSE declined to 21.8 kg N ha−1 when additional information was included in the Random Forest analysis. Leaf N concentration and soil water status to depth enhanced the N prediction, especially when this information was collected from an infield N experiment with an N‐minus, standard, and N‐rich strip of N.