Assessing uncertainty and complexity in regional-scale crop model simulations

Crop models are imperfect approximations to real world interactions between biotic and abiotic factors. In some situations, the uncertainties associated with choices in model structure, model inputs and parameters can exceed the spatiotemporal variability of simulated yields, thus limiting predictability. For Indian groundnut, we used the General Large Area Model for annual crops (GLAM) with an existing framework to decompose uncertainty, to first understand how skill changes with added model complexity, and then to determine the relevant uncertainty sources in yield and other prognostic variables (total biomass, leaf area index and harvest index). We developed an ensemble of simulations by perturbing GLAM parameters using two different input meteorology datasets, and two model versions that differ in the complexity with which they account for assimilation. We found that added complexity improved model skill, as measured by changes in the root mean squared error (RMSE), by 5–10% in specific pockets of western, central and southern India, but that 85% of the groundnut growing area either did not show improved skill or showed decreased skill from such added complexity. Thus, adding complexity or using overly complex models at regional or global scales should be exercised with caution. Uncertainty analysis indicated that, in situations where soil and air moisture dynamics are the major determinants of productivity, predictability in yield is high. Where uncertainty for yield is high, the choice of weather input data was found critical for reducing uncertainty. However, for other prognostic variables (including leaf area index, total biomass and the harvest index) parametric uncertainty was generally the most important source, with a contribution of up to 90% in some cases, suggesting that regional-scale data additional to yield to constrain model parameters is needed. Our study provides further evidence that regional-scale studies should explicitly quantify multiple uncertainty sources.