Comparing impacts of climate change and mitigation on global agriculture by 2050

Systematic model inter-comparison helps to narrow discrepancies in the analysis of the future impact of climate change on agricultural production. This paper presents a set of alternative scenarios by five global climate and agro-economic models. Covering integrated assessment (IMAGE), partial equilibrium (CAPRI, GLOBIOM, MAgPIE) and computable general equilibrium (MAGNET) models ensures a good coverage of biophysical and economic agricultural features. These models are harmonized with respect to basic model drivers, to assess the range of potential impacts of climate change on the agricultural sector by 2050. Moreover, they quantify the economic consequences of stringent global emission mitigation efforts, such as non-CO2 emission taxes and land-based mitigation options, to stabilize global warming at 2 °C by the end of the century under different Shared Socioeconomic Pathways. A key contribution of the paper is a vis-à-vis comparison of climate change impacts relative to the impact of mitigation measures. In addition, our scenario design allows assessing the impact of the residual climate change on the mitigation challenge. From a global perspective, the impact of climate change on agricultural production by mid-century is negative but small. A larger negative effect on agricultural production, most pronounced for ruminant meat production, is observed when emission mitigation measures compliant with a 2 °C target are put in place. Our results indicate that a mitigation strategy that embeds residual climate change effects (RCP2.6) has a negative impact on global agricultural production relative to a no-mitigation strategy with stronger climate impacts (RCP6.0). However, this is partially due to the limited impact of the climate change scenarios by 2050. The magnitude of price changes is different amongst models due to methodological differences. Further research to achieve a better harmonization is needed, especially regarding endogenous food and feed demand, including substitution across individual commodities, and endogenous technological change.


Introduction
The Paris Agreement on climate change aims to keep the increase in global mean temperature well below 2 • C above pre-industrial levels by the end of the century while safeguarding food security and recognizing the particular vulnerabilities of food production systems to the adverse impacts of climate change.
The agricultural sector is, on the one hand, directly affected by climate change due to altered weather conditions and resulting biophysical effects , Rosenzweig et al 2014. On the other hand, agriculture, forestry and other land use are responsible for almost a quarter of anthropogenic greenhouse gas (GHG) emissions , and reduction of emissions from agriculture is necessary to achieve the global climate change goals (Reisinger et al 2013, Gernaat et al 2015, Wollenberg et al 2016. In order to achieve ambitious climate mitigation targets, both CO 2 and non-CO 2 GHG emissions need to be reduced substantially. Furthermore, achieving the ambitious targets is conditional on the large-scale availability of negative emissions technologies, in particular carbon sequestration through afforestation and bioenergy systems connected to carbon capture and storage (BECCS), in the second half of the century (Clarke et al 2014, Anderson andPeters 2016). Without these technologies, even more modest stabilization would require substantially larger GHG emissions reductions in the medium term, further increasing the impacts of mitigation on the food system (Havlík et al 2015a). However, both afforestation and BECCS are very land use intense negative emissions technologies (Smith et al 2016) and therefore affect agriculture via the land markets. In this context an integrated multimodel assessment taking into account both the range of potential climate change impacts as well as climate change mitigation measures, and their interaction, is required to provide insights for effective and efficient policy decision making.
Until now, the quasi-totality of global agricultural sector assessments considered either climate change impacts or climate change mitigation policies. The assessment of impacts of climate change on agricultural production and food security has a long tradition (Rosenzweig and Parry 1994, Parry et al 2004, Nelson et al 2010. Recent work as part of the Agricultural Model Inter-comparison and Improvement Project (AgMIP) (Rosenzweig et al 2013) has examined (and narrowed) the differences between models in estimated impacts on agriculture of a certain level of climate change through systematic model inter-comparison (Nelson et al 2013, von Lampe et al 2014, Lotze-Campen et al 2014. This work focused on a single 'middle-of-the-road' Shared Socioeconomic Pathway (SSP 2, O'Neill et al 2014) and climate impacts for a single high-emission Representative Concentration Pathway (RCP 8.5, Vuuren et al 2011). Wiebe et al (2015) extended these analyses to three SSPs and three RCPs. The land use and agricultural sector implications of ambitious climate change stabilization policies, without taking into account the climate change impacts, were recently analysed in a multi-model setup within the SSP-RCP framework by Popp et al (2017). Havlík et al (2015a) analyzed in a single model consistent scenario setup both unmitigated climate change impacts on the agricultural sector and the effects of ambitious mitigation scenarios, however they ignored the interactions between these two dimensions. Finally, also in a single model study, Hasegawa et al (2015) represented the scenarios in an integrated way, where the mitigation scenario took systematically into account the residual climate change impacts. Their results show, similarly to the results from Havlík et al (2015a) that at least in the medium term, till 2050, the climate stabilization scenarios would have a more severe impact on food security than unmitigated climate change.
In this paper, we present a set of alternative scenarios by different models, harmonized with respect to basic model assumptions, to assess the range of potential economic impacts of climate change on the agricultural sector by 2050, as well as the economic consequences of stringent global emission mitigation efforts (e.g. bioenergy use, afforestation, reduction of methane and nitrous oxide emissions in agriculture) aiming to stabilize global warming at 2 • C by the end of the century under different SSPs. For this purpose, the analysis covers selected combinations of SSPs and RCPs. This approach allows us to assess the interplay of socioeconomic developments, climate change impacts and climate mitigation policies on the agricultural sector while taking into account the model related uncertainties. Agro-economic models can present quite different results when analyzing economic and related impacts of climate change on agriculture. As the focus is usually on the model results, it is often not clear if differences are due to the model specification (e.g. partial or general equilibrium models), model parameterization (e.g. supply, demand or trade elasticities), scenario assumptions (e.g. future economic, population and policy developments) or to data sources. Therefore, this paper not only focuses on the scenario results but also on the harmonization of key model inputs and the comparison of what is driving model results.

Model framework
For the analysis, five global economic models were employed (table 1, supplementary material S1 available at stacks.iop.org/ERL/13/064021/mmedia). Using a combination of integrated assessment (IMAGE), partial equilibrium (CAPRI, GLOBIOM, MAgPIE) and computable general equilibrium (MAGNET) models for the analysis ensures a good coverage of biophysical features on land availability, quality, and spatial heterogeneity, as well as cross-sectorial linkages through factor markets and substitution effects. The spatial resolution and the level of disaggregation of the agricultural sector are very different across the models, as both are functions of each model's history and original purpose. Furthermore, the employed models differ in a number of other characteristics. For instance, some of the models can be used to depict alternative levels of second-generation bioenergy production, while others have no explicit representation of bioenergy or focus on feedstock use for firstgeneration biofuels, electricity and/or heating (Wicke et al 2015). Three models have spatially explicit representations of bilateral trade flows, even if differing in the specific approach used (Berkum and van Meijl 2000, van Tongeren et al 2001. Food demand is endogenous in GLOBIOM, CAPRI and MAGNET by iso-elastic or CDE (constant differences of elasticities) demand functions and exogenous for MAgPIE. The IMAGE model is a global integrated assessment model that covers the human and earth systems and gets its agro-economic information by a linkage to MAGNET. Although the five models employed are state of the art and are frequently used for agroeconomic analyses, including the assessment of climate change impacts on agriculture, it has to be kept in mind that they are simplifications of reality and designed to illustrate complex processes. The models are theoretical constructs representing economic processes by a set of variables and quantitative relationships between them, using simplified assumptions and not able to specifically address all factor dynamics. Furthermore, behaviour is represented by structural parameters which are quantified using historical data and are often kept constant over time.

Scenarios-setup and assumptions
The experimental design to analyze the impact of climate change and climate change mitigation under three contrasting socioeconomic developments (SSP1/SSP2/SSP3) is outlined in   in order to stabilize global warming at 2 • C above pre-industrial levels (with no residual climate change impact). A comparison of scenario A and C gives the pure impact of mitigation policies on agriculture ('Mitigation' effect). Scenarios in row D add the climate impacts of the RCP2.6 (without CO 2 fertilization) to the scenarios in row C and thus allow to analyze the climate change impact in a 2 • C world ('CC RCP2.6' effect) 9 . Finally, scenarios D compared to scenarios A provide the 2 degree mitigation effect including the residual climate change impacts ('Mitigation + Residual CC'). While the effects 'CC RCP 2.6' and 'Mitigation' are used for diagnostic purposes, comparing the full climate change effects of 'CC RCP 6.0' and the full mitigation effects 'Mitigation + RCP 2.6' allows to directly evaluate the benefits/costs of climate stabilization compared to a world without climate policies. 9 For MAGNET, we run additional SSP_CC26 scenarios to identify more precisely the 'CC RCP 2.6'effect in comparison to the NoCC scenario as there are strong interaction effects between the mitigation and climate change effects.
The following sections briefly describe the underlying assumptions for the SSPs, climate change related crop yield impacts, and mitigation measures in the agricultural sector. Model inputs are partly harmonized but the specific implementation can differ between the models.

Socioeconomic narratives
The SSPs represent different global futures, with narratives for future demographic and economic developments, lifestyle, policies and institutions, technology, and environmental protection (O'Neill et al 2017).
In this paper we focus on SSP1, SSP2 and SSP3 (tables 3 and 4). SSP 2 (Middle of the Road) is a pathway with modest overall growth in population and incomes, and a slow pace of overall trade liberalization. SSP 1 (Sustainability) features lower population growth, higher growth in per capita incomes, faster globalization and a higher environmental awareness (less waste, less meat consumption), while SSP 3 (Regional Rivalry) describes a more fragmented world with less international trade, higher population growth, a lower growth in per capita incomes and less environmental Sustainability A future pathway with low challenges for adaptation and mitigation. A pathway that makes relatively good progress towards sustainability, with sustained efforts to achieve development goals, while reducing resource intensity and fossil fuel dependency. Elements that contribute to this are an open globalised economy, rapid development of low-income countries, a reduction of inequality (globally and within economies), rapid technology development, low population growth and a high level of awareness regarding environmental degradation. More environmental awareness reduces food waste, the appetite for meat as well as making land use regulation stricter.

SSP2
Middle of the Road A business as usual scenario. In this world, trends typical of recent decades continue, with some progress towards achieving development goals, reductions in resource and energy intensity at historical rates, and slowly decreasing fossil fuel dependency. A world with only medium challenges for adaptation and mitigation.

SSP3
Regional Rivalry A world with high challenges for adaptation and mitigation. A world which is separated into regions characterized by extreme poverty, pockets of moderate wealth and a bulk of countries that struggle to maintain living standards for a strongly growing population. Regional blocks of countries have re-emerged with little coordination between them. Countries focus on achieving energy and food security goals within their own region. The world has de-globalized, and international trade, including energy resource and agricultural markets, is severely restricted. Population growth in this scenario is high as a result of limited improvements in education and low economic growth.  (2017). The level of technological change differs across SSPs and most models focus on yields. In GLOBIOM, IMAGE/MAGNET and CAPRI, SSP-related yields are a function of GDP, whereas in MAgPIE yields are endogenous depending on cost effectiveness compared with land expansion and a SSP-specific discount rate. In MAGNET, labour and capital productivity differs between SSP scenarios as these are calibrated to preserve scenario-specific GDP growth rates (see Robinson et al 2014 and supplementary material S1 for information on the general treatment of technological change in the models).

Climate change related crop yield impacts
We rely on a representative selection of climate change impact scenarios on crop and grassland yields. The selection is based on data on climate change impacts on crops yields from global gridded crop models (GGCM) for different climate scenarios (Rosenzweig et al 2014). The climate scenarios are bias-corrected implementations (Hempel et   RCP2.6 and RCP6.0 have been selected for their representativeness of the 2 degree mitigation and the no mitigation policy under SSP2, respectively, at the end of the 21st century (van Vuuren et al 2011, Riahi et al 2017). However, it has to be noted that they are not distinctively different in 2050 (the time horizon in this paper), as in 2050, GHG concentrations of RCP2.6 are still close to peak concentration levels whereas RCP6.0 has still relatively low GHG concentrations in 2050, so that the radiative forcing of RCP2.6 and RCP6.0 are quite similar in 2050. RCP8.5 was not considered as even without climate change mitigation, none of SSP1-3 is likely to reach a GHG concentration that is high enough to be compatible with RCP8.5 (Riahi et al 2017). While GHG concentration levels and climate impacts on crop yields on average will further increase beyond 2050 (Rosenzweig et al 2014), several of the economic models are not well prepared to cover scenarios beyond 2050. Hence, this analysis was confined to 2050.

Agricultural GHG mitigation
Agriculture is the largest contributor to the global anthropogenic non-CO 2 GHG emissions of methane and nitrous oxide, accounting for about 10%-12% of total global GHG emissions . Despite the importance of agricultural non-CO 2 emissions, their mitigation has received somewhat less attention than the land-based mitigation potential of CO 2 (e.g. bioenergy production, afforestation and reduced emissions from deforestation and forest degradation). The non-CO 2 emission sources and mitigation measures covered in the models are CH4 emissions from (i) enteric fermentation, (ii) rice production, (iii) animal waste management, and (iv) on-field burning, and N 2 O emissions from (i) manure excreted on pasture, range and paddock, (ii) cropland fertilization (mineral fertilizer and manure applications), (iii) animal waste management, and (iv) agricultural waste burning. The model specific implementation of these non-CO 2 emissions and measures are described in table A.2. Next to these non-CO 2 mitigation options we include REDD (reduced emissions from deforestation and forest degradation), afforestation and bioenergy as land-based CO 2 mitigation options that affect agriculture by competing on the land market. The mitigation scenarios are implemented in all scenarios via a carbon price. A carbon price implemented on direct non-CO 2 emissions from the agricultural sector leads to adoption of more GHG efficient production systems and dedicated technologies but also to reduction of agricultural production because it indirectly increases production cost. Carbon price implemented on CO 2 emissions from land use and land use change fosters reduction of deforestation and further afforestation and hence also contributes to higher production cost through increased land rents. The carbon price implemented on fossil fuel related CO 2 emissions leads to increased demand for biomass for bioenergy production, and also contributes to competition for land. Finally, the tax for residual emissions is paid by producers, which transforms then into higher producer\market prices because of the increased cost. The producers can transfer only part of the tax cost to consumers because the price elastic demand. All these dynamics lead to rising food prices which reduce food consumption. All models represent endogenously mitigation of non-CO 2 emissions, however some models do not represent the full land use and hence approximate the impacts of the carbon price on land use by land use projections from other models (MAG-NET), and except for MAGNET and IMAGE, the models do not represent directly the energy system, and hence use biomass demand projections from other models to simulate the substitution of fossil fuels by bioenergy. The positive income effect of tax income recycling or increased agricultural income is not considered by most models (only MAGNET takes these effects into account).

Agricultural non-CO 2 emissions
Mitigation measures strongly reduce agricultural non-CO 2 emissions by about 40%-45% in CAPRI, IMAGE (MAGNET) and MAgPIE (figure 1), with methane and nitrous oxide emissions being reduced by 50% and 30%, respectively. As IMAGE and MAgPIE use the same marginal abatement cost curves (Lucas et al 2007), the relative reduction in both models is similar, though slightly higher in IMAGE. In both models, the relative reduction is comparable across the different SSPs, as in all SSPs much of the mitigation potential is already applied early due to fast increasing carbon taxes. The mitigation effort in CAPRI is similar in all SSPs as the same emission taxes and the same assumptions regarding mitigation technologies are applied across SSPs. Emission reduction is much smaller in GLOBIOM than in the other three models, and differs across SSPs, with SSP3 showing the lowest reduction. This is because mitigation in this implementation of GLOBIOM is mostly based on GHG efficiency improvements through changes in production system composition and production relocation across regions, both mediated through prices, but not via technological mitigation measures.
MAgPIE neglects price-mediated food demand shifts, and therefore, for example, also the pricing of methane emissions does not lead to consumption changes for livestock products, which dampens production decreases and hence limits related emission reductions in the mitigation scenarios. A reduction of global feed demand is however possible via trade of livestock products from regions with higher feeding efficiencies. In IMAGE, technological mitigation measures are combined with system-wide effects due to GHG pricing (calculated via MAGNET). In CAPRI, the decline in agricultural non-CO 2 emissions is similar to the decline in IMAGE and MAgPIE as the same reference (Taylor et al 2012) has been used for mitigation effects in non-European regions. CAPRI has a quite detailed non-CO 2 mitigation modelling for Europe, but the global results are dominated by other regions. CO 2 emissions from land use change (LUC) are strongly deceasing in most mitigation scenarios and in some scenarios even become negative (notably SSP1) due to avoided deforestation (REDD), and afforestation (IMAGE and MAgPIE). As this indicator is not available from all the models, CO 2 emissions related to land use change CO 2 can be better indicated by the agricultural area expansion (see below).

Global agricultural production developments
In general, global agricultural production in SSP1 is less than in SSP2, which in turn is less than in SSP3 (see IMAGE, MAGNET and MAgPIE results in figure S.3.1). Following the SSP storylines and their implementation in the models, this indicates that the demand for agricultural products is more influenced by population developments and assumptions about waste and dietary preferences together than by assumptions on GDP developments. CAPRI exhibits the opposite trend, indicating that GDP developments are a stronger driver than population and that the implementation of dietary changes has been more conservative than in the other models. In SSP3, MAGNET\IMAGE assume additional changes from 2020-2050, including a 33% waste and food losses increase, 20% higher meat consumption and 10% higher food import taxes, which increase demand and therefore also agricultural production. Also MAg-PIE assumes that waste and food losses and livestock consumption for a given per-capita income are higher in SSP3 and lower in SSP1, relative to SSP2. These features are less pronounced in GLOBIOM, which only considers a slower reduction in waste and food losses compared to SSP2 and SSP1. In this paper the further emphasis is not on the SSP results as they are dealt with in other papers (e.g. Nelson et al 2013, Popp et al 2017 and we focus on the climate and mitigation impacts using the four comparisons identified in section 2.2. Figure 2 shows that the impact of climate change RCP6.0 and RCP2.6 on global agricultural production (primary crop and livestock) is negative with a range of 0.5%-2.5%. The impact is only slightly higher in RCP6.0 than RCP2.6, which is due to the selection of median scenarios as they actually imply rather similar yield impacts of the two RCPs by 2050. The 'Mitigation' column shows that in all SSPs and all models the mitigation measures result in negative impacts on global agricultural production that are larger than the negative climate change impacts. The combined effect of mitigation costs and climate change effects as shown in the 'Mitigation + RCP2.6' column is more negative on agricultural production than the no-mitigation scenario 'CC RCP 6.0'. The gain in reduced negative climate impacts of RCP2.6 compared to RCP6.0 (compare first two columns) is too small to compensate for the negative impact of the mitigation measures. While from the literature it can be expected that RCP2.6 is more favorable for agricultural production than RCP6.0 in the long run beyond 2050 (Rosenzweig et al 2014), this does not hold in this study as radiative forcing in 2050 is quite similar. The selection process to identify a representative scenario of biophysical crop yield shocks is based on the globally aggregated change in yields. However, GCM projections can differ substantially, especially in the spatial patterns of climate change (McSweeney and Jones 2016) and also GGCMs differ in their response to individual drivers and the represented agricultural system (Folberth et al 2016). For simulations of land-use changes under two cases of similar radiative forcing (RCPs in 2050), the spatial configuration of climate change impacts may thus be quite prominent in overall results. The spatial pattern of projected changes has not been considered in the selection of the median impact scenarios, which were selected on globally aggregated yield changes. Even if we had covered a large number of crop yield simulations with our economic model ensemble, it is very likely that the effects and their uncertainty bands in 2050 would have been similar. Our selection of SSPxRCP and GGCM combinations may have contributed to the finding that the costs of agricultural GHG mitigation under 'Mitigation + RCP2.6' are dominating over any climate related benefits for agricultural production compared to the pure climate change scenario 'CC RCP6.0' by 2050. It is very likely that this comparison would change later in the century.
In general, the additional cost of agricultural mitigation measures reduces production, most notably for rice and especially ruminant meat, in most models, which can be explained by the high GHG intensity of these two products. For most models the production of non-ruminants also decreases except for CAPRI, which observes an increase in production of some commodities (dairy and non-ruminants) 10 Symbols represent the value for total global agricultural production which is the total of all primary crop and livestock sectors within a model. The bars represent the mean value across models. as consumers shift from the more GHG intensive ruminant meat to non-ruminant meat. From a technical perspective this is driven by higher cross price elasticities for CAPRI than for MAGNET and the other models do not include cross price elasticities. Cross price elasticities vary significantly in the literature, and there are few comprehensive studies available. Therefore, these are adjusted within CAPRI and MAGNET to calibrate the demand system. In MAGNET the elasticities are not implemented directly in the model but used in the calibration process of the CDE (constant differences in elasticities) demand parameters (see Woltjer et al 2014).

Land use
Cropland area generally increases when moving from SSP1 over SSP2 to SSP3 (figure S.3.2). This is due to higher demand for land due to the higher production levels (as described in 3.1), lower exogenous yields (as these are GDP-dependent and GDP declines from SSP1 over SSP2 to SSP3), and a low endogenous response of yields. In the case of MAg-PIE, also climate unrelated land protection policies explain differences between the SSPs. Climate change impacts (RCP2.6 and RCP6.0) increase cropland area in IMAGE\MAGNET, MAgPIE and CAPRI, whereas cropland area decrease in GLOBIOM. For the former four models lower crop yields (see , table A1) and an inelastic food demand induce the higher land use. For GLOBIOM the negative impact on cropland is because grassland is relatively favored by climate change compared to crops, which leads in some regions to a small shift in the livestock production systems towards more grazing and less reliance on feed crops (Havlík et al 2015b). In all models except CAPRI, cropland area (i.e. land used for non-energy crops) decreases about 4%-7% due to mitigation measures (part (a) of figure  3). The decrease is caused by less available cropland due to avoided deforestation or afforestation and demand for bioenergy. In CAPRI, GHG mitigation was exclusively achieved by non-CO 2 emissions and especially decreased ruminant production, which released grassland and allows for a limited expansion of cropland.
Mitigation measures, in particular carbon price and bioenergy expansion, result in a decrease of about 7%-10% in grazing area in the GLOBIOM, IMAGE and MAGNET models (part (b) of figure 3). Grassland decreases more than cropland, because land is allocated (with imperfect substitution) according to its rental price: cultivating crops gives higher returns to land than ruminants as the latter are more GHG intensive. Therefore, the decrease in available land impacts more on the ruminant sector where in addition partial substitution of grass by grains is possible. In CAPRI this effect is not reflected as afforestation is not specifically considered. The decrease in SSP1 is higher in GLOBIOM due to the assumption of faster transition possibility from grass based ruminant systems to larger reliance on concentrate feeds.

Producer prices
Compared to 2010, producer prices in real terms are lower in SSP1 in all scenarios, whereas they are stable or higher in SSP2 and further increase in SSP3 (figure S.3.4). Important drivers on the production side are lower yields in SSP3 than in SSP2 and SSP1. The lower prices in SSP1 are caused by supply side factors such as higher yields and higher labour productivity (MAGNET), which increase supply, in combination with demand side factors that lower demand such as lower food waste and less meat consumption. Price changes are rather small in GLOBIOM and CAPRI, intermediate in MAgPIE and rather big in MAGNET. In MAgPIE, producer prices are higher in SSP3 due to increased production costs as a result of a much higher population, more restricted trade and augmented costs for additional technological change. The higher price effects in the MAGNET model can mainly be explained by lower labor productivity growth, driven by lower GDP growth in SSP2 and especially SSP3 than in SSP1. As labor costs have a substantial share in total agricultural production costs, the labor productivity effect together with the yield effect imply that production costs are much lower in SSP1 than in SSP2, and much higher in SSP3. In addition to the labor productivity effect also land prices are an important driver of producer prices in MAGNET (van Meijl et al 2006). Figure 4 shows that climate change increases agricultural producer prices in almost all cases due to lower crop yields that restrict supply, which also leads to increased feeding costs for livestock production.
The climate change impacts are more pronounced in MAgPIE and MAGNET. Land prices play a major role in determining producer prices, and as by 2050 land is scarce, especially in the SSP3 scenario, climate change induced lower yields imply a rapid increase in land prices. Additionally, the price-inelastic demand in MAgPIE does not buffer food prices.
Mitigation efforts lead to a higher increase in agricultural prices than the climate change effect by 2050. For crop prices the impact is more pronounced in MAGNET and MAgPIE than in CAPRI and GLO-BIOM. As implemented in CAPRI and GLOBIOM, mitigation has almost no impact on crop prices, because the demand for feed crops decreases as a result of reduced livestock production due to the tax on livestock emissions. As implemented in MAGNET, the higher impact of mitigation is caused by the lower land availability for agriculture due to afforestation and demand for energy crops. Lower land availability for agriculture leads to an increase in land prices and therefore also food prices. The land pressure is highest in SSP3 and therefore also the impact of mitigation efforts on producer prices is highest in SSP3. For MAgPIE the combination of additional demand for bioenergy crops, non-CO 2 emission taxes and completely inelastic food demand leads to increasing crop prices in the mitigation scenarios. Regarding livestock producer prices, mitigation measures lead to higher price increases for livestock products than for crops, because livestock is more emission intensive and emission taxes, therefore, increase livestock production costs relatively more than crop production cost.

Discussion and conclusions
For this paper, common scenarios on climate change and mitigation options were assessed with five agroeconomic models. Model inputs were harmonized by using the same projections for population and GDP growth, SSP narratives, as well as relative biophysical crop yield changes due to climate change. Scenario results are relatively consistent across SSPs (SSP1, SSP2 and SSP3) and climate scenarios (RCP2.6 and RCP6.0 with and without mitigation policies in place), despite the fact of using models with some significant structural differences. The overall trends of the 12 scenarios are very similar and the few 'outliers' can be explained by structural model characteristics or different scenario implementation choices. The scenario results highlight vulnerabilities, changes in production, area and price effects in the global agricultural sector. In most models global agricultural production is lowest in SSP1 and highest in SSP3, which indicates that the demand for agricultural products is more influenced by the SSPrelated assumptions on changes in population, waste and food losses, and dietary preferences than by the assumptions on GDP developments. So far as median assumptions on yield shocks are taken, the impact of climate change on agricultural production by 2050 is negative but relatively small at the aggregated global level. The impact is fairly similar between the two CO 2 concentration pathways RCP6.0 and RCP2.6 as the radiative forcing in 2050 and therefore yield impacts are similar. In contrast, as crop model results have shown, climate impacts will increasingly differ between RCP2.6 and RCP6.0 after 2050. In addition, GGCMs are typically capable to reproduce interannual yield variability in regions with high inputs and stable management conditions, but currently cannot account for other sources of yield variability such as pests and diseases that may also respond to climate change, ozone damage, or direct heat stress (Müller et al 2017). Climate change impacts are projected to become more severe in the second half of the 21st century. For instance, Porter et al 2014 found that while only about 2% of available climate change impact projections for crop yield foresee a drop in yields by more than 50% for the period 2050-2069, almost 20% of the projections foresee such a strong decline in yields for the period 2090-2109. The negative effects of climate change can also be underestimated because only the trend in climate change impacts has been considered ignoring the likely increase in extreme weather events. Although it is projected that the negative effects of climate change will increase over time, our conclusions that the effect on agriculture of mitigation is stronger would probably hold even if moving the time horizon to 2080 and considering the strong climate change scenario RCP8p5. These conclusions are consistent with Havlík et al (2015a) who considered such scenarios in a single model framework. However, the purpose of our study is not to evaluate the full costs and benefits of climate change mitigation over the long term but rather to highlight potential challenges related to it in the medium term.
The modelled GHG emission mitigation measures have a negative impact on primary agricultural production for all SSPs across all models. In terms of reduced global agricultural production, the impacts of mitigation policies are larger than the negative impacts due to climate change effects in 2050. However, this is partially due to the limited impact of the climate change scenarios by 2050. Our analysis finds that a mitigation strategy that takes into account residual climate change effects (Mitigation + RCP2.6) has a negative impact on agricultural production relative to a no-mitigation strategy and stronger climate impacts (CC RCP6.0). In line with the production results, by 2050 climate impacts affect global agricultural prices less strongly than ambitious mitigation policies across the models in this study. The price impact is higher in the livestock sector because livestock production is more emission intensive and higher emission taxes directly increase livestock production costs. The magnitude of the producer price changes is very different between the models, which is mainly due to differences in the general model set-up (especially treatment of technological change and price responsiveness of demand) and assumptions on mitigation measures (e.g. carbon pricing). This analysis is a further step towards a better understanding of economic impacts of climate change and mitigation on the global agricultural sector and how they are reflected in the agro-economic models. The agro-economic models used in this study show shortcomings that should be improved for future research, especially regarding sound cross-price elasticities (all models), endogenous demand systems (MAgPIE) and an endogenous approach for the process of technological change. While all models largely agreed to the broad SSP and mitigation storylines, the specific implementation is not homogeneous across models, so that more work needs be done to increase consistency for a better comparison of model results. Moreover, we only present results at the global level, and further research is needed to identify vulnerabilities, adaptation and mitigation strategies for regional agricultural sectors.
Disclaimer: The views expressed are purely those of the authors and may not in any circumstances be regarded as stating an official position of the European Commission or the other institutions involved.