Estimating DSSAT Cropping System Cultivator-Specific Parameters Using Bayesian Techniques
Crop models are highly useful for simulating crop and soil processes in response to variations in climate and management. However, if one wishes to simulate a crop's performance in a specific soil and climate for a particular set of management inputs, cultivar-specific parameters (CSPs) are needed because of the genetic variations among cultivars of any crop. In this chapter, we summarized methods that have been used to estimate CSPs for the CERES and CROPGRO-based models in the Decision Support System for Agrotechnology Transfer (DSSAT) cropping system model. We primarily described a Bayesian parameter estimation procedure (the Generalized Likelihood Uncertainty Estimation, or GLUE) for use in estimating CSPs in DSSAT. The procedure is simple to use, requiring only that users select a crop, a cultivar, and the data for use in the estimation procedure from a list of data available for that cultivar in the DSSAT system. Results are displayed for users to view and copy to the standard cultivar file in DSSAT for the crop involved. The procedure does require a large number of model runs; we recommend 6000 but users can optionally change this number. Two cultivars, ‘Prisma’ maize (Zea mays L.) and ‘Williams’ soybean [Glycine max (L.) Merr.] were selected to demonstrate the performance of DSSAT GLUE program. For Prisma maize, two experiments conducted in Zaragosa, Spain in 1995 and 1996 were selected; for Williams soybean, three experiments individually conducted in Wooster, OH and Gainesville, FL, were selected for the demonstration. Results showed that the GLUE method performed better than the arbitrary default CSPs and slightly better than the hand-calibrated CSPs in simulating these maize and soybean cultivars when using one time measurements, such as phenology dates, final dry matter yield, maximum leaf area index, and grain yield. For example, in the Prisma maize experiments in Zaragosa, Spain in 1995 and 1996, the average relative absolute error (RAE) values between the simulated and measured output variables were only 3 and 8%, respectively, while they were between 4 and 10% for hand-calibrated CSPs and above 16% for the default CSPs.