Adaptation and evaluation of the CROPGRO-soybean model to predict regional yield and production

In spite of the availability of numerous crop-growth models, there has been limited experience in applying them to predict regional production and its variability. The main difficulty is a substantial mismatch between spatial and temporal scales of available data and crop simulation model input requirements. This study developed and tested an operational procedure to predict soybean (Glycine max L. Merrill) yield and production by linking the CROPGRO-soybean model with a low resolution regional database of weather, soils, management, and varieties. Historically observed census yields were detrended to remove effects of changes in technology and then aggregated to a scale of 0.5° cell (about a 50 km grid cell) g using an area-weighting approach. Spatial yield variability within a grid cell was simulated (?) using nine input combinations (3 varieties, 3 planting dates, 1 soil and 1 initial condition) which were averaged for comparison with aggregated census yield, in each cell and year. Yield bias was estimated by minimizing the root mean squared error (RMSE) between corrected y and g. The yield correction factor needs to be site-specific to account for spatial variations in constraints and management. Yield correction factor ranged from 0.40 to 0.50 in more than 75% of grid cells. When corrected, the success rate for the goodness of fit of y and g was ~100 and 80% for variance at the 95% confidence limit. The 17-year mean of actual yield was accurately predicted with a slope of 0.95, small intercept (?0.025) and R2 of 0.95. When validated, the prescribed factor test error was 14%, within the 16% guideline set by the Environmental Protection Agency (1982) as an acceptable criteria for a model to qualify for management application. Median RMSE were 15, 8, 32, 7 and 85% for 1991, 1992, 1993, 1994 and 1995, respectively. Years 1993 and 1995 were dominated by high water stress. We conclude that the grid-specific yield correction approach can effectively correct bias in simulated yields and accurately predict interannual variability using readily available inputs. Future steps are needed to incorporate procedures that account dynamically for yield susceptibility to pests and diseases. Testing and improvement of the model should continue to realize its potential.