Automatic calibration of DSSAT and APEX models for maize yield simulation and economic optimum nitrogen rate determination

Abstract

Multiple crop growth models are being used to simulate the impact of different factors on crop yield and support farmers for optimizing crop management. However, accurate and efficient calibration of these models is a challenging and critical step in their practical applications. The objectives of this study were to 1) evaluate an automatic model calibration strategy; and 2) compare the performance of DSSAT and APEX models for simulation of maize (Zea mays L.) growth, plant nitrogen (N) uptake, yield in response to different N application rates and the estimation of the economic optimum N rate (EONR). Detailed data collected from eight site-years of N experiments conducted from 2014 to 2016 in Minnesota and Wisconsin, USA were used in this research. Observed and simulated maize yield responses to N rate were fitted with regression models to estimate the EONR. The results indicated that both DSSAT and APEX models performed well at Minnesota sites in maize yield simulation, with manual calibration achieving R² of 0.78–0.95, root mean square error (RMSE) of 0.2–1.4 t ha⁻¹ , and normalized root mean square error (NRMSE) of 2 %–17 %. Automatic calibration using a model-independent data assimilation (MIDA) optimizer yielded robust yield prediction results (R² = 0.66–0.95; RMSE of 0.3–1.4 t ha−1, and NRMSE of 3–17 %). In addition, both DSSAT and APEX models performed well in simulating yield at different N rates and EONR for both preplant and split N application scenarios, although high prediction errors were observed at some site-years. The independent test using Wisconsin data further confirmed the high accuracy of yield prediction for both models using the two calibration strategies (R2: 0.79–0.88). While both models overestimated biomass at maturity during manual calibration, automatic calibration using the MIDA optimizer significantly reduced prediction errors, particularly for DSSAT (RMSE: 0.7–3.5 t ha⁻¹; NRMSE: 3 %–17 %). Plant N uptake at maturity was reasonably well simulated using the DSSAT model (R²: 0.66–0.92), but the APEX model exhibited low and variable R² values (0.00–0.91), highlighting the need to improve plant N uptake simulation of the APEX model. Significant inconsistencies were observed between the two models' simulations of biomass and plant N uptake at the tasseling growth stage. It is concluded that MIDA optimizer-based automated model calibration can effectively optimize crop growth model parameters, mitigating the risk of local optima. Both DSSAT and APEX models performed well for simulating maize yield, yield response to N rates and EONR, but improvements are needed for the APEX model for simulating biomass and plant N uptake. More research is needed to further evaluate the automatic calibration tool for calibrating crop growth models under on-farm conditions to support their wider practical applications.