Comprehensive global climate impact assessment for crop yields

Climate change impacts on agriculture are subject to large uncertainties from a variety of sources. One of the most important sources of uncertainty is the uncertainty in the realization of climate change itself. In the absence of clear climate mitigation strategies and substantial uncertainties on population growth, economic development, technology and lifestyles, a very broad set of greenhouse gas emission scenarios has been developed to inform climate modeling. Climate models often differ in the spatial patterns of projected changes in particular with respect to changes in precipitation. The Coupled Model Intercomparison Project (CMIP5, CMIP6) provides a broad range of future climate change projections.

Crop models are often applied at selected sites or with global coverage, as in the Global Gridded Crop Model Intercomparison (GGCMI) of the Agricultural Model Intercomparison and Improvement Project (AgMIP). Global crop model applications have been shown to have some skill, but also add additional uncertainty, given that many processes cannot be calibrated properly for the lack of suitable reference data and because management information is largely absent (Müller et al., 2017).

However, already the computational power required to compute the comprehensive set of climate projections prohibits such applications. Instead, typically, small and largely random selections of climate scenarios are used to project impacts, such as agricultural crop yields. McSweeney and Jones (2016) find that a selection of 5 climate models as often applied, is insufficient to cover the range of projections in all regions.

Here we present initial results of a comprehensive global climate impact assessment for crop yields that explores the full range of the CMIP6 climate projection archive. For this, we use a set of 9 global gridded crop model emulators (Franke et al., 2019b) that were trained on a very large systematic input sensitivity analysis with up to 1404 global-coverage, 31-year simulation data sets per crop and crop model (Franke et al., 2019a). The training domain includes variations in atmospheric carbon dioxide (CO2) concentrations (4 levels from 360 ppm to 810 ppm), air temperature (7 levels from -1 to +6°C), water supply (8 levels from -50 to +30% and full irrigation), nitrogen fertilization (3 levels from 10 to 200 kgN/ha) and adaptation (2 levels: none and regained growing seasons) and thus represents an unprecedented rich data base for emulator training. The emulators, in form of grid-cell specific regression models with 27 coefficients, are computationally light-weight and can thus be applied to the full CMIP6 data archive.

We here present first results from this analysis, breaking down the different sources of uncertainty (emission concentration pathways, climate model, crop model). Results will help to interpret crop model simulations in general: the unstructured reduction of the uncertainty space from selecting a small number of climate scenarios by e.g. first availability and/or individual crop models has so far hampered to quantify the uncertainty in crop model projections.

 

Franke (2019a) Geoscientific Model Development Discuss, 2019:1-30.

Franke (2019b)  Geoscientific Model Development, submitted

McSweeney & Jones, (2016) Climate Services, 1:24-29.

Müller (2017) Geoscientific Model Development, 10:1403-1422.