Galmarini, S., Solazzo, E., Ferrise, R., Srivastava, A. Kumar, Ahmed, M., Asseng, S., Cannon, A.J., Dentener, F., De Sanctis, G., Gaiser, T., Gao, Y., Gayler, S., Gutierrez, J.M., Hoogenboom, G., Iturbide, M., Jury, M., Lange, S., Loukos, H., Maraun, D., and Moriondo, M.
Crop models are essential tools for assessing the impact of climate change on national or regional agricultural production. Starting from meteorology, soil and crop management, fertilization and irrigation practices, they predict the yield of specific crop varieties. For long term assessments, climate models are the source of primary information. To make climate model results usable in a specific time frame context, bias adjustment (BA) is required. In fact, climate models tend to deviate from day-to-day values of the physical parameters while conserving the climate variability signal. BA brings the climatic signal to the actual values observed in a specific location and period, and to be representative of a specific period in absolute terms. BA techniques come in different flavours. The broadest categorization is univariate and multivariate methods. Multivariate methods adjust the variables considering possible cross-correlations while univariate methods treat the variables one by one without accounting for possible dependence on one another. The hypothesis tested in this paper is that since crop models require as input climate variables that are in most of the cases cross-correlated, the multi-variate bias adjustment of the latter is likely to improve performance compared to univariate bias adjusted climate model results. To verify this hypothesis, 14 BA methods were applied to 9 variables from 8 climate models at 21 locations across Europe and Northern Africa for a period of 5 years. Twelve crop models, from the AgMIP Wheat community, were run using the climate model results. All crop models, except one, were restarted at every growing season. The crop models were also run using the AgMERRA re-analysis. The latter were used as reference to compare the results when using the other climate models treated with the various sets of bias-adjustment methods. The results show that multivariate BA treatment should be preferred to univariate ones. The error obtained by comparing crop simulation obtained with AgMERRA with those obtained with multivariate bias-adjusted climate prediction is systematically lower. The error reduction varies as a function of the variable, the location, the crop model, and the climate model though the tendency is for smaller errors when multivariate methods are used to treat the latter. The results are attributed to the nature of crop models and the fact that multivariate methods consider more adequately the correlation existing between the meteorological variables. The study shows the importance of considering the nature of a model and the selection of input data that best suited to the former. In this case the improvements produced when using multivariate data appears to be significant especially in the light of the variety of crop models used and the similar response obtained and it is therefore recommended. [Display omitted] • The impact of climate data multivariate bias-adjustment methods versus univariate on crop model results was estimated. • Crop model results improved when input data was treated using multivariate methods compared to univariate methods. • Multivariate methods maintain the variables correlation as required by crop models. • This result is attributed to the parameterized nature of the participating crop-model formulations. • Climate, bias adjustment and the crop modelling communities joined efforts in this unprecedented collaboration. [ABSTRACT FROM AUTHOR]