Calibration and validation of APSIM–Maize, DSSAT CERES–Maize and AquaCrop models for Ethiopian tropical environments

Abstract

Process-based crop models are popular tools to quantify the impact of changes due to climate or crop management. Accurate simulation of crop production for different agro-ecological conditions using an individual crop model remains challenging due to different sources of uncertainty. Studies with ensembles of crop models can give valuable information about model accuracy and uncertainty, but such studies are limited in tropical environments, including Ethiopia. Therefore, the aim of this study was to compare the performance of the outputs of three individual crop models and their ensemble mean. We calibrated three different crop models, namely, APSIM-maize, AquaCrop and DSSAT CERES-maize and evaluated them separately and in a multimodel ensemble approach using four maize varieties (BH546, BH661, Jibat and MH140) grown under rainfed conditions. Model input data were collected from field experiments conducted at three sites (Ambo, Bako and Melkassa) during the 2017/2018 crop growing season. The experiments were laid out in a randomised complete block design using a plot size of 10 m × 10 m. The crop models were calibrated using measured data from the Bako and evaluated with independent datasets from the Ambo and Melkassa. The calibration parameters used in each of the three crop models studied enabled accurate simulation of flowering, maturity, canopy cover (AquaCrop) and grain yield against measured data. Evaluation of the models indicated that APSIM-maize and DSSAT CERES-maize accurately simulated days to flowering and maturity with root mean square error (RMSE) values ranging from 1.73–4.09 and 1.66–5.36 days, respectively. However, the DSSAT CERES-maize model over-estimated the maturity period of late-maturing varieties at Ambo. The AquaCrop model accurately simulated maize canopy cover for all varieties studied with a RMSE of less than 10.8% and a high index of agreement (d) of 0.95. The simulated grain yield agreed fairly well with the measured data, with normalised RMSE ranging from 13–19%, 1–4% and 1–17% for APSIM, AquaCrop and DSSAT maize models, respectively. However, the APSIM model underestimated yield for all maize varieties at Ambo (RMSE of 1.14 t ha−1 and d-value of 0.50). The best performance was obtained when an ensemble of all models was considered, which reduced the RMSE values for grain yield to 0.35 t ha−1 at Ambo and 0.41 t ha−1 at Melkassa. Furthermore, the ensemble mean reduced the normalised RMSE by 8% while increasing the d-value to above 0.90 for both evaluation sites. On the other hand, the ensemble results were quite similar for grain yield simulated using the AquaCrop model. It is concluded that model ensembles reduced model uncertainty and improved simulation output accuracy compared to the outputs of individual models in tropical environments.