Performance of 13 crop simulation models and their ensemble for simulating four field crops in Central Europe (2021.0)
Kostková M., Hlavinka P., Pohanková E., Kersebaum K., Nendel C., Gobin A., Olesen J., Ferrise R., Dibari C., Takáč J., Topaj A., Medvedev S., Hoffmann M., Stella T., Balek J., Ruiz-Ramos M., Rodríguez A., Hoogenboom G., Shelia V., Ventrella D., Giglio L., Sharif B., Oztürk I., Rötter R., Balkovič J., Skalský R., Moriondo M., Thaler S., Žalud Z., Trnka M.
The Journal of Agricultural Science, 159 (1-2), 69-89
doi:10.1017/s0021859621000216
Abstract
AbstractThe main aim of the current study was to present the abilities of widely used crop models to simulate four different field crops (winter wheat, spring barley, silage maize and winter oilseed rape). The 13 models were tested under Central European conditions represented by three locations in the Czech Republic, selected using temperature and precipitation gradients for the target crops in this region. Based on observed crop phenology and yield from 1991 to 2010, performances of individual models and their ensemble were analyzed. Modelling of anthesis and maturity was generally best simulated by the ensemble median (EnsMED) compared to the ensemble mean and individual models. The yield was better simulated by the best models than estimated by an ensemble. Higher accuracy was achieved for spring crops, with the best results for silage maize, while the lowest accuracy was for winter oilseed rape according to the index of agreement (IA). Based on EnsMED, the root mean square errors (RMSEs) for yield was 1365 kg/ha for winter wheat, 1105 kg/ha for spring barley, 1861 kg/ha for silage maize and 969 kg/ha for winter oilseed rape. The AQUACROP and EPIC models performed best in terms of spread around the line of best fit (RMSE, IA). In some cases, the individual models failed. For crop rotation simulations, only models with reasonable accuracy (i.e. without failures) across all included crops within the target environment should be selected. Application crop models ensemble is one way to increase the accuracy of predictions, but lower variability of ensemble outputs was confirmed.