Location: Adaptive Cropping Systems LaboratoryTitle: Cultivar coefficient stability and effects on yield projections in the SPUDSIM model
Submitted to: Agronomy Journal
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/10/2019
Publication Date: 2/7/2020
Citation: Fleisher, D.H., Haynes, K.G., Timlin, D.J. 2020. Cultivar coefficient stability and effects on yield projections in the SPUDSIM model. Agronomy Journal. 2020:1-16.
Interpretive Summary: Mathematical models of crop growth and development are often used to evaluate ways to improve farm management. They have also been used in more recent efforts on food security issues. Because all models are incomplete descriptions of the real world, model users need to understand what the limitations are when using these tools. Otherwise, decisions made using these models may give unexpected results. One way to evaluate how good these models are is to evaluate how effectively they can be calibrated. Two years of field data in Maryland and Maine were obtained and used to explore how the accuracy of an existing potato model may be affected when calibrating it for these different locations. This study showed the model performed equally well in either state regardless of the difference in calibration method. Further analysis demonstrated that yield predictions under future climates in these regions may vary slightly depending on calibration method, but not significantly enough to be of practical concern. The study indicated areas of needed improvement in the model where its description of the real world could be improved so as to give better predictions. These results are useful for growers, scientists and policy planners interested in using decision support tools to make better informed decisions regarding cultivar selection and farm management in the United States Northeastern region.
Technical Abstract: Process level crop models are formulated to accurately simulate plant growth and development. Calibration refers to the process where values for a subset of model parameters, usually referred to as genetic or cultivar coefficients, are obtained. Because model predictions are inherently biased due to simplifications in the system of equations that represent real-world phenomenon, these calibrated parameters are just as likely to compensate for limitations in model structure as well as reflecting true phenotypic characteristics. This confounding of genetics with components of the production environment limits model accuracy when conducting various assessments including climate change impacts. One approach to testing the extent of this confounding is to evaluate changes in calibration parameters and model prediction results when calibration is conducted cross- versus within- locations. We evaluated the calibration stability of the potato model SPUDSIM for two cultivars grown in two contrasting locations in the U.S. using two calibration methods, location specific (R1) or cross location (R2). Differences between phenology coefficients under the R1 method ranged from 4 to 17 percent between the two locations depending on cultivar. A wider range was observed for growth coefficients, which likely reflected over-coupling of canopy expansion rate within the model structure as well as an over-sensitivity with to temperature and photoperiod. R2 method results for phenology coefficients gave values within the range for that obtained for both cultivars in R1. However, canopy expansion rate was almost 50 percent larger for one cultivar. Despite these differences, independent validation year results were similar for R1 and R2 with respect to end-of-year yields (17 percent error or less) and yield-RMSE (less than 36.1 grams per plant). Projected climate change impacts for 2030, 2040, 2050, and 2080 decades indicated R2 over-predicted yields by 5 percent or less as compared with R1, with a maximum discrepancy of 4.4 grams per plant. Consistent differences between R1 and R2 calibrated models in rates of yield decline per decade were not observed for cultivar, location, or climate scenario, but averaged -0.59 grams per plant per year for R1 and -0.67 for R2. These results suggest calibration for these cultivars and locations was relatively stable for the SPUDSIM model.