Inter-comparison of performance of soybean crop simulation models and their ensemble in southern Brazil

Crop simulation models can help scientists, government agencies and growers to evaluate the best strategies to manage their crops in the field, according to the climate conditions. Currently, there are many crop models available to simulate soybean growth, development, and yield, with different levels of complexity and performance. Based on that, the aim of this study was to assess five soybean crop models and their ensemble in Southern Brazil. The following crop models were assessed: FAO – Agroecological Zone; AQUACROP; DSSAT CSM–CROPGRO–Soybean; APSIM Soybean; and MONICA. These crop models were calibrated using experimental data obtained during 2013/2014 growing season in different sites, sowing dates and crop conditions (rainfed and irrigated) for cultivar BRS 284, totaling 17 treatments. The crop variables assessed were: grain yield; crop phases; harvest index; total above-ground biomass; and leaf area index. The calibration was made in three phases: using original coefficients from modelś default (no calibration); calibrating the coefficients related only with crop life cycle phases; and calibrating all set of coefficients (below and above the soil). The results from the models were analyzed individually and in an ensemble of them. The crop models showed an improvement of performance from no calibration to complete calibration. Crop phases were estimated efficiently, although different approaches were used by the models. The estimated yield had RMSE of 650, 536, 548, 550 and 535kgha−1, respectively, for FAO, AQUACROP, DSSAT, APSIM and MONICA, with d indices always higher than 0.90 for all of them. The best performance was obtained when an ensemble of all models was considered, reducing yield RMSE to 262kgha−1. The same tendency for ensemble being best was observed for leaf area index. The harvest index was the crop variable with the poorest performance. In general, the results showed that an ensemble of completely calibrated models were more efficient to simulate soybean yield than any single one, which was also observed when testing this procedure with independent data.