Evaluation of the CROPGRO-Soybean model for assessing climate impacts on regional soybean yields

Crop models have been used in many studies of crop production and risks at a field scale. Whereas crop models simulate yield for a unit area with known inputs, many studies, such as those that assess climate change impacts, require information at broader spatial scales. Little has been done to evaluate the ability of crop models to predict yield variability associated with climate variability over large areas. The purpose of this study was to evaluate different methods for using the CROPGRO-Soybean model to predict yields under different climate conditions at an aggregate scale. Soil and weather data (24 years) were obtained at a 0.5 degrees grid scale for the southeastern U.S. Data from eight grid cells were used to calibrate eleven different methods using 17 years of data, and model predictions were evaluated using data from seven independent years for the same grid cells. These crop model-based methods were compared with simple multiple linear regression of yields vs. rainfall at the different sites and with predictions based on mean historical yields, independent of climate. Root mean square errors of prediction (RMSEP) were computed to evaluate uncertainty in yield estimates for the different methods. Overall, root mean square errors at different locations ranged from 68 to 383 kg ha(-1) for the calibration data set, and increased to 174 to 530 kg ha(-1) for independent validation years. Model predictions of absolute yields were biased, resulting in a mean RMSEP of 1,487 kg ha(-1). When historical mean yields in each grid cell were used to remove bias, RMSEP was reduced to 367 kg ha(-1), averaged across locations. RMSEP values were between 23% and 29% of observed mean yields for different model-based approaches. The purely statistical approach, regressing yields vs. monthly rainfall, had lower RMSEP (266 kg ha(-1)) than the model-based results, 21% of mean observed yields. The best model-based predictions were obtained by accounting for bias and rainfall during planting and harvesting months in a combined model-regression approach. The use of historical mean yields in each grid cell was inferior to all adjusted model-based results. Bias correction of model-predicted yields are needed to achieve accuracies that are similar to those obtained by statistical methods alone.