Abstract: Crop models, used to make projections of climate change impacts, differ greatly in structural detail. Complexity of model structure has generic effects on uncertainty and error propagation in climate change impact assessments. We applied Bayesian calibration to three distinctly different empirical and mechanistic wheat models to assess how differences in the extent of process understanding in models affects uncertainties in projected impact. Predictive power of the models was tested via both accuracy (bias) and precision (or tightness of grouping) of yield projections for extrapolated weather conditions. Yields predicted by the mechanistic model were generally more accurate than the empirical models for extrapolated conditions. This trend does not hold for all extrapolations; mechanistic and empirical models responded differently due to their sensitivities to distinct weather features. However, higher accuracy comes at the cost of precision of the mechanistic model to embrace all observations within given boundaries. The approaches showed complementarity in sensitivity to weather variables and in accuracy for different extrapolation domains. Their differences in model precision and accuracy make them suitable for generic model ensembles for near-term agricultural impact assessments of climate change.

Abstract: Regional agricultural yield assessments allowing for weather effect quantifications are a valuable basis for deriving scenarios of climate change effects and developing adaptation strategies. Assessing weather effects by statistical methods is a classical approach, but for obtaining robust results many details deserve attention and require individual decisions as is demonstrated in this paper. We evaluated regression models for annual yield changes of winter wheat and silage maize in more than 300 German counties and revised them to increase their predictive power. A major effort of this study was, however, aggregating separately estimated time series models (STSM) into panel data models (PDM) based on cluster analyses. The cluster analyses were based on the per-county estimates of STSM parameters. The original STSM formulations (adopted from a parallel study) contained also the non-meteorological input variables acreage and fertilizer price. The models were revised to use only weather variables as estimation basis. These consisted of time aggregates of radiation, precipitation, temperature, and potential evapotranspiration. Altering the input variables generally increased the predictive power of the models as did their clustering into PDM. For each crop, five alternative clusterings were produced by three different methods, and similarities between their spatial structures seem to confirm the existence of objective clusters about common model parameters. Observed smooth transitions of STSM parameter values in space suggest, however, spatial autocorrelation effects that could also be modeled explicitly. Both clustering and autocorrelation approaches can effectively reduce the noise in parameter estimation through targeted aggregation of input data. (C) 2015 Elsevier B.V. All rights reserved.