Assessment of Sugarcane Simulation Models and Their Ensemble for Yield Estimation in Different Regions, Cropping Systems, Seasons and Harvest Months in Brazil.

The estimation and forecasting of sugarcane yields are very complex, however, highly important for crop planning and decision-making. Recent studies have shown that the use of crop simulation models in an ensemble can reduce simulation uncertainties, together with a yield decline factor along successive ratoons. Having this in consideration, the aim of this study was to evaluate the performance of three sugarcane models, previously calibrated (FAO-AZM, DSSAT/CANEGRO and APSIM-Sugarcane), separately and in an ensemble to estimate sugarcane yield in different regions, cropping systems (rainfed and irrigated), seasons and harvest months in Brazil. A total of 635 data from sugarcane fields and common statistical indices in crop modelling were employed. Three yield data aggregations, given by the average, were considered: a) by region, system, growing seasons and harvest months (n = 90); b) by region, system and growing seasons (n = 42); and c) by region, system and harvest months (n = 24). In the first aggregation, that is important for monthly yield forecasting, all the models presented lower errors (MAE and RMSE lower than 14.5 and 20.0 t ha^-1, respectively), good accuracy (d > 0.92), precision (R² > 0.81) and efficiency (E > 0.77) to estimate sugarcane yield. The second aggregation, that is important for seasonal yield forecasting, all the models presented acceptable errors (MAE and RMSE lower than 19 and 26 t ha^-1, respectively), good accuracy (d > 0.89), precision (R² > 0.82) and efficiency (E > 0.72). In the final aggregation, which is important for seasonal harvest planning, all the models presented lower errors (MAE and RMSE lower than 8.5 and 10.6 t ha^-1, respectively), high accuracy (d > 0.96), precision (R² > 0.88) and efficiency (E > 0.87). In all conditions, ensemble was slightly better than single models. Future studies related meteorological and soil data aggregation and how much time before harvest the yield could be forecast is necessary.