Evaluating and improving APSIM's capacity in simulating long-term corn yield response to nitrogen in continuous- and rotated-corn systems

Process-based models are increasingly used to explain and predict crop yields and long-term changes in soil organic matter (SOM) and hence should be regularly evaluated for their accuracy. Currently, there is a knowledge gap of how well process-based models can estimate the economic optimum nitrogen rate (EONR) across environments and years. We evaluated the Agricultural Production Systems sIMulator (APSIM) software ability to simulate corn yield response to nitrogen (N) input and crop rotation in long term experiments. Furthermore, we explored causes for over/under prediction of the EONR. Measurements included crop yields from 14 long-term (N) fertilizer rate experiments representing major production regions in the U.S. Corn Belt and SOM distribution from seven long-term experiments. Corn yield response to N rate was analyzed with statistical models to estimate the EONR (386 N rate trials). The model successfully captured spatiotemporal patterns in observed crop yields across N rates in the soybean-corn (SC) system, but overpredicted crop yield by 5–25% at high N rates in the corn-corn (CC) system. This overprediction was partially resolved by adding algorithms in APSIM to account for the well-known continuous corn yield penalty. The improved model simulated yield response to N and the EONR with a model agreement of 0.93 and 0.66, respectively, across rotations. The lower accuracy in predicting EONR compared to crop yields was attributed to 1) inherent model error in simulating yields: for example, a 10% error in yield simulation of a single point can result in a 34% error in EONR; and 2) the inability of the APSIM model to fully generate the quadratic nature of corn yield response to N rate, which resulted in the linear-plateau model being selected in most cases. Forcing use of only a quadratic plateau or quadratic regression model to fit the simulated yields, which is the expected biophysical yield response to N, improved EONR prediction by 23% while the accuracy of the statistical model fit was minimally decreased (<1%). The APSIM model simulated SOM distributions after 20 years of cropping with an agreement index of 0.93. We believe that APSIM is well suited to supplement N research in the U.S. Corn Belt, recognizing identified limitations between yield estimation and N response determination. This research provided a solution for process-based models to cope with the continuous corn yield penalty, thoroughly evaluated APSIM, and identified research priorities towards increasing EONR prediction.