Abstract: Crop models are important tools for impact assessment of climate change, as well as for exploring management options under current climate. It is essential to evaluate the uncertainty associated with predictions of these models. We compare two criteria of prediction error; MSEPfixed, which evaluates mean squared error of prediction for a model with fixed structure, parameters and inputs, and MSEPuncertain(X), which evaluates mean squared error averaged over the distributions of model structure, inputs and parameters. Comparison of model outputs with data can be used to estimate the former. The latter has a squared bias term, which can be estimated using hindcasts, and a model variance term, which can be estimated from a simulation experiment. The separate contributions to MSEPuncertain(X) can be estimated using a random effects ANOVA. It is argued that MSEPuncertain(X) is the more informative uncertainty criterion, because it is specific to each prediction situation.

Abstract: Croplands and grasslands are agricultural systems that contribute to land–atmosphere exchanges of carbon (C). We evaluated and compared gross primary production (GPP), ecosystem respiration (RECO), net ecosystem exchange (NEE) of CO2, and two derived outputs – C use efficiency (CUE=-NEE/GPP) and C emission intensity (IntC= -NEE/Offtake [grazed or harvested biomass]). The outputs came from 23 models (11 crop-specific, eight grassland-specific, and four models covering both systems) at three cropping sites over several rotations with spring and winter cereals, soybean and rapeseed in Canada, France and India, and two temperate permanent grasslands in France and the United Kingdom. The models were run independently over multi-year simulation periods in five stages (S), either blind with no calibration and initialization data (S1), using historical management and climate for initialization (S2), calibrated against plant data (S3), plant and soil data together (S4), or with the addition of C and N fluxes (S5). Here, we provide a framework to address methodological uncertainties and contextualize results. Most of the models overestimated or underestimated the C fluxes observed during the growing seasons (or the whole years for grasslands), with substantial differences between models. For each simulated variable, changes in the multi-model median (MMM) from S1 to S5 was used as a descriptor of the ensemble performance. Overall, the greatest improvements (MMM approaching the mean of observations) were achieved at S3 or higher calibration stages. For instance, grassland GPP MMM was equal to 1632 g C m−2 yr-1 (S5) while the observed mean was equal to 1763 m-2 yr-1 (average for two sites). Nash-Sutcliffe modelling efficiency coefficients indicated that MMM outperformed individual models in 92.3 % of cases. Our study suggests a cautious use of large-scale, multi-model ensembles to estimate C fluxes in agricultural sites if some site-specific plant and soil observations are available for model calibration. The further development of crop/grassland ensemble modelling will hinge upon the interpretation of results in light of the way models represent the processes underlying C fluxes in complex agricultural systems (grassland and crop rotations including fallow periods).

Abstract: Ensembles of process-based crop models are increasingly used to simulate crop growth for scenarios of temperature and/or precipitation changes corresponding to different projections of atmospheric CO2 concentrations. This approach generates large datasets with thousands of simulated crop yield data. Such datasets potentially provide new information but it is difficult to summarize them in a useful way due to their structural complexities. An associated issue is that it is not straightforward to compare crops and to interpolate the results to alternative climate scenarios not initially included in the simulation protocols. Here we demonstrate that statistical models based on random-coefficient regressions are able to emulate ensembles of process-based crop models. An important advantage of the proposed statistical models is that they can interpolate between temperature levels and between CO2 concentration levels, and can thus be used to calculate temperature and [CO2] thresholds leading to yield loss or yield gain, without rerunning the original complex crop models. Our approach is illustrated with three yield datasets simulated by 19 maize models, 26 wheat models, and 13 rice models. Several statistical models are fitted to these datasets, and are then used to analyze the variability of the yield response to [CO2] and temperature. Based on our results, we show that, for wheat, a [CO2] increase is likely to outweigh the negative effect of a temperature increase of +2 degrees C in the considered sites. Compared to wheat, required levels of [CO2] increase are much higher for maize, and intermediate for rice. For all crops, uncertainties in simulating climate change impacts increase more with temperature than with elevated [CO2].