The CERES-EGC model was used to simulate croplands in Europe. CERES-EGC is a process-based biogeochemical model in the soil–plant–atmosphere domain adapted from CERES (Jones and Kiniry, 1986). The model is designed to simulate C and N dynamics; heat transfer; and water exchanges from soil, plants and the atmosphere. It works at a daily time step designed to the field scale. Inputs require meteorological and management data as forcing variables and soil and crop data as factors. Meteorological data are constituted by daily minimum and maximum temperature, precipitation, global solar radiation, and wind speed. Management includes tillage, irrigation, fertilisation, information on sowing and incorporation of crop residues. Soil is divided into sublayers with specific depth, physical and chemical characteristics. Simulated crop species include maize (grain and fodder), soft wheat, durum wheat, rye, oat, barley, rapeseed, sorghum, sunflowers, pea, sugar beet and soybean, with the possibility of selecting specific varieties.

Soil C and N dynamics in the ploughed layer are simulated by means of the NCSOIL model (Molina et al., 1983; Nicolardot et al., 1994), which is a nested module in CERES-EGC. NCSOIL computes nitrification, immobilisation and mineralisation of N; the decomposition of soil organic matter (SOM) after incorporation of crop residues; and SOM formation. The module works with a series of specific pools, three pools for crop residues (easily fermentable carbohydrates, cellulose and lignin) and four endogenous pools (zymogenous and microbial biomass, active and passive humus), where CO2 is released from the decomposition of each pool. N uptake by plants is calculated through a specific supply–demand scheme depending on mineral nitrogen availability and root length density. CERES-EGC includes the model NOE (Hénault et al., 2005) for simulating N2O emissions from denitrification and nitrification processes in the topsoil (0–20 cm depth). Denitrification and nitrification are computed from a soil-specific potential rate limited by unitless factors related to soil water content, soil temperature and substrate content (nitrates, NO3, and ammonium, NH4, for denitrification and nitrification, respectively). Ammonia (NH3) volatilisation is calculated in a detailed module, while plant growth is simulated according to the crop-specific genetic potential and the photosynthetically active solar radiation absorbed by the canopy. Potential dry matter production is constrained by air temperatures, soil water availability and the N deficit.

PaSim is a biogeochemical process-based model able to simulate C, N and water dynamics in the plant–soil–atmosphere–livestock grassland system (Calanca et al., 2007). Five interacting sub-models of soil biology and physics, microclimate, vegetation, and grazing herbivores constitute the model structure. The model runs on a daily (or hourly) time step, and inputs require soil property data, management and meteorological characteristics (global solar radiation, minimum and maximum air temperature, relative humidity, wind speed, precipitation, and atmospheric CO2 concentration). The soil is described in six sublayers, allowing us to parameterise different soil depths with site-specific soil physical and chemical characteristics. Management includes grazing, mowing and N fertilisation. Grazing is considered a dairy or suckling system managed by grazing periods with specific stocking density and live weight. Indoor periods are not simulated. Vegetation cover is considered a homogeneous cover with a fixed legume fraction. The vegetation cover comprises the root system and three shoot compartments (laminae, sheaths and stems, and ears) divided into age classes. Soil C dynamics (based on the CENTURY model; Parton et al., 1994) are computed in five pools: a structural and a metabolic pool for fresh organic carbon (plant residues) and an active, a slow and a passive pool for the microbially processed organic carbon. Photosynthetic C is allocated in plant (root and shoot) and can be lost as CO2 by ecosystem respiration and as CH4 through enteric fermentation.

Soil N inputs are represented by atmospheric N deposition, symbiotic N2 fixation, mineral or organic fertilisation, animal faeces, and urine. These inputs, together with the nitrogen mineralised from the organic carbon pools, constitute the mineral N pool. N availability for plants is reduced by losses via processes of immobilisation, NO3 leaching, NH3 volatilisation, nitrification and denitrification. N2O emissions from nitrification and denitrification depends on substrate availability (NO3 or NH4). These emissions are modulated by factors controlling the effects of soil temperature and water content. Furthermore, the release of N2O produced in the soil towards the atmosphere is calculated with a resistance model in the rooting zone and plant canopy (Schmid et al., 2001).

CERES-EGC and PaSim were selected for this evaluation at the regional scale since they have been calibrated and evaluated in different conditions worldwide (Brilli et al., 2017; Ehrhardt et al., 2018; Sándor et al., 2018) and in Europe, i.e. France, Denmark, Germany, Italy, Sweden and the UK for CERES-EGC (Rolland et al., 2008; Lehuger et al., 2009; Wattenbach et al., 2010; Drouet et al., 2011; Lehuger et al., 2011; Goglio et al., 2013; Ferrara et al., 2021; Haas et al., 2021) and France, Germany, Hungary, Ireland, Italy, Portugal, Spain, the Netherlands and the UK for PaSim (Lawton et al., 2006; Calanca et al., 2007; Gottschalk et al., 2007; Vuichard et al., 2007; Ma et al., 2015; Sándor et al., 2016). These models are suitable to simulate a number of crops and rotations, mown or grazed grasslands, and the effects of management practices on plant–soil–atmosphere–livestock. Besides, they are able to simulate GHG emissions and the carbon budget at the field scale through the C assimilated from photosynthesis; C emitted into the atmosphere from autotrophic and heterotrophic respirations; C recycled (dung, plant residues) or introduced from external sources (fertilisers, soil improvers); and, finally, the C exported from the system by production activities. CO2 fertilisation was not simulated for croplands (see S4 in the Supplement). Furthermore, the two models used in this study do not represent potential impacts of air pollution and pest and disease effects on plant production.

To assess the net greenhouse gas exchange (NGHGE) of the agro-ecosystems investigated, the contribution of the biogenic GHGs (CO2, N2O, CH4) is combined and normalised to grams of CO2 equivalent (g CO2 eq.) by using the relative global warming potential (γgas) at the 100-year time horizon (298 for N2O, 25 for CH4 and 1 for CO2; IPCC, 2018), following the approach presented by Soussana et al. (2007). 1NGHGE=NEP+γN2OFN2O+γCH4FCH4 The net ecosystem production (NEP) is the amount of organic C available for net ecosystem C storage, export or loss in an ecosystem, in terms of CO2. NEP represents the difference between the gross primary production – or photosynthesis – and the ecosystem respiration, which is the sum of the autotrophic respiration and heterotrophic respiration (HR); ruminant respiration from grasslands ecosystems is not accounted for in the HR term. Conventionally, a negative value of NEP indicates an uptake of CO2 by the system, whereas a positive value is a release towards the atmosphere.

The annual net greenhouse gas balance (NGB) is calculated on the basis of Ammann et al. (2020) by including the export of C by harvested biomass (crop yield, mowing and animal intake), the export as crop residues and the import of C by manure (organic fertilisers and the excreta from grazers). 2NGB=NGHGE+FC-harvest+FC-residues-FC-manure Since livestock do not graze throughout the whole year, their contribution to the carbon balance is represented by the intake of biomass, enteric fermentation (CH4) and C in excreta. Carbon emissions from farm operations (i.e. tractor emissions), erosion and leaching processes, fire, or off-farm emissions (i.e. fertiliser manufacture, barns) are not included in the C budget; the effects of volatile organic compounds and CH4 emissions from manure and from soil are considered negligible as well. Moreover, the C exported from animal production (body mass increase and milk production) is neglected in NGB calculation (e.g. Chang et al., 2015).

Results from this study confirmed that the effects of climate change, implying shifts in the temperature, precipitation and plant growing length among other factors, represent a serious drawback to plant production.

Air temperature. Our findings pointed out that the increase in air temperature during the climate scenarios was negatively correlated with productivity, leading to persistent reductions in biomass production in both grasslands and croplands. This behaviour is also confirmed by previous studies (e.g. Challinor et al., 2014; Lobell and Tebaldi, 2014; Olesen and Bindi, 2002; Zhang et al., 2017) and was more pronounced for the more pessimistic climate scenario (-0.15 and -0.29 t DM ha-1 yr-1 ∘C-1 for RCP4.5 and RCP8.5, respectively, in the 2050–2099 period). Effects on crop yields ranged from +5 % to -11 % for every degree (∘C) of rising air temperature from 2005–2100. This effect remained negative throughout most of the projected climate scenarios (-1 % and -5 %, for RCP4.5 and RCP8.5, respectively, in the period 2025–2099), as also reported by recent studies using modelling and multi-modelling approaches (e.g. Asseng et al., 2015; Bassu et al., 2014; Zhao et al., 2017; Yang et al., 2019). The extension of irrigable areas to all European croplands reduced the dependence of daily maximum and minimum air temperatures on crop production (Fig. S8). This leads to the assumption that even with access to water (no limitation on irrigation of the European cropping surfaces), biomass production will decline due to increasing air temperatures, as reported by Minoli et al. (2019). This can also be seen from the trend of biomass projections in Fig. 2, considering an increase in temperatures over time. Interestingly, grassland productivity assumed a less pronounced correlation with air temperature during climate scenarios compared to croplands (Fig. 12). The climatic scenario RCP8.5, characterised by a strong reduction in grassland production in the second half of the century, showed a significant negative correlation with minimum and maximum daily air temperatures (r=-0.6), whereas null correlation is observed for RCP4.5. Furthermore, crop yields were significantly correlated with minimum and maximum air temperatures (r=0.64 and r=0.57, respectively) compared to grasslands, which did not show such a dependence (r=0.1 for both minimum and maximum air temperatures), highlighting a greater sensitivity of the CERES-EGC model to air temperatures compared to PaSim.

Precipitation. Results confirmed that rainfall has a significant positive effect for both crop production (r=0.41 and 0.13 for RCP4.5 and RCP8.5, respectively) and grassland production (r=0.26 in RCP8.5; r is null for RCP4.5). Compared to the historical period, a reduction in precipitation was predicted in the first half of the century for both scenarios (-2.1 mm yr-1 for RCP4.5 and -0.74 mm yr-1 for RCP8.5; p<0.01), whereas in the second half of the century rainfall increases in RCP4.5 (+1.2 mm yr-1; p<0.01) and decreases in RCP8.5 (-0.59 mm yr-1; p<0.01). This effect was more pronounced for low latitudes (-1.2 and -2.3 mm yr-1 for RCP4.5 and RCP8.5, respectively; p<0.01) compared to high latitudes where the rainfall tends to increase during the century (+0.26 and +0.1 mm yr-1 in RCP4.5 and RCP8.5, respectively, with respect to the historical period; p<0.5). This reduction in the cumulated precipitation will negatively affect the productivity with the climate change scenarios, as also confirmed by Hsu et al. (2012) for grasslands and by Olesen et al. (2011) for croplands.

Length of crop growing cycle. Apart from increases in temperature and reduction in precipitation, our simulation highlighted that crop yield is affected by the shortening of the length of the growing cycle, as confirmed by, for example, Bassu et al. (2014) and Tao and Zhang (2011). As detailed in our results, with a multi-model approach Bassu et al. (2014) predicted a general reduction in the growing cycle length for maize, especially in central Europe. A reduction from 6 to 22 d for maize cultivation in RCP4.5 and up to 8–29 d in RCP8.5 was also forecasted by de Souza et al. (2019) for Brazil conditions using DSSAT-CERES-Maize. Moreover, the consistent reduction in maize production observed with the climate scenarios in our study is most probably due to the shorter growing period (-8 to -22 d), characteristic of the spring crops. Concerning wheat, the magnitude of reduction in the length of growing cycle is consistent with the findings of Yang et al. (2019) for the Mediterranean area, who forecasted up to -26 d with the STICS model compared to -22 d of our simulations.

Our findings confirmed that climate change will have a regionally distributed impact (Howden et al., 2007; Challinor et al., 2014; Parry et al., 2005; Lobell and Tebaldi, 2014) even in the scenario that includes mitigation measures to offset climate change (RCP4.5), creating the opportunity to the design cropping systems with multiple crops in a year. Multiple cropping can represent a viable alternative in regions with long growing seasons and where water (rain or irrigation) and solar radiation are not limiting factors (Mueller et al., 2015; Waha et al., 2020), as well as where cardinal temperature requirements for crop and varieties are met. Furthermore, our study confirm that a certain number of actual crops and varieties could be cultivated in Europe, even in the worst climate projection. These crops could potentially yield higher production than today, especially at high latitudes, while an overall reduction in crop production is forecasted for low European latitudes.

Finally, the production levels of cropland and grasslands are in line with the available historical data (see Sect. 3.1.1) and the recent – albeit scarce – literature, making this study coherent and representative. Regarding the climatic projections, yields estimated with the DayCent model (Lugato et al., 2018) for the RCP4.5 scenario in the period 2015–2099 reported an average over Europe of 4.34 t DM ha-1 yr-1 (ranging from 3.69 to 4.90 t DM ha-1 yr-1), which is in line with our estimation of 4.49 t DM ha-1 yr-1 (ranging from 3.55 to 5.49 t DM ha-1 yr-1) predicted for the same period and climate projection. Assessing the effects of climate change in the European croplands and grasslands, our study can provide support for the identification of climate-smart practices. Among these, the modulation of crop sowing dates and the implementation of irrigation represent possible solutions in the short to medium term to prevent water stress (Lehmann et al., 2013).

Sowing date. Shifting sowing dates represents a promising adaptation to overcome yield drops (Olesen et al., 2012). Accordingly, our results showed that earlier sowing dates are expected for spring-sown crops under future climate scenarios compared to historical dates. Differences between historical and future sowing dates ranged from 0 to -5 d for both RCP4.5 and RCP8.5 scenarios approaching 2050, whereas at the 2100 horizon earlier sowing dates are predicted with differences of -5 and -7 d for RCP4.5 and RCP8.5, respectively. This evidence shows that climate change allows significantly more advanced sowing in Europe, as confirmed by the review of Tubiello and Rosenzweig (2008). For winter-sown crops, sowing dates are extended in a range from +5 to +9 d moving towards 2050 and to +13 d at the end of the century for RCP4.5. These increases were greater in RCP8.5, ranging from +7 to +13 d moving towards 2050 and reaching +19 d moving towards 2100. The extension of irrigation in all simulated crops in Europe had a negligible influence on the length of the crop cycles, as discussed by Minoli et al. (2019), despite an increasing demand of water over the course of the century.

Irrigation. Water demand has been shown to increase by +6 % during the first half of the century, to slightly decrease in the second half for RCP4.5 (-2 %) and to increase again for RCP8.5 (+23 %). These variations are in line with the results of the multi-model approach used by Wada et al. (2013) analysing the uncertainty in the response of different hydrological models over Europe. These authors showed a decrease in water demand for irrigation moving towards 2100 in Europe of <5 % for RCP4.5 and a rise of >20 % for RCP8.5. Furthermore, from our study we observed that water demand assumes a strong regional variation in Europe, with low latitudes needing 227 mm yr-1 on average over the historical period (mean 1985–2004), which is an order of magnitude higher than mid-latitudes (29 mm yr-1) and high latitudes (9 mm yr-1). These proportions between the latitudes remained unvaried over the course of the century, whereas middle and high latitudes displayed a +20 % increase in the evapotranspiration deficit approaching 2050 (mean 2030–2049) compared to the historical period, in both climate scenarios. This phenomenon observed for low latitudes is strictly related to climate perturbation (i.e. a strong increase in air temperature and reduction in rainfall), which increased crop water demand (Olesen et al., 2011). Furthermore, the increase in water demand even in middle and high latitudes confirms that irrigation needs to be supplied even for the crops that are now commonly rainfed (e.g. spring and winter soft wheat, spring barley, sunflowers, rapeseed). By 2100, the volumes of water needed for European croplands were largely reduced to below the quantities observed during the historical period, especially for low latitudes. These findings underline that even with high availability of irrigation water, the reduction in the crop growing cycle for the actual crop varieties – which sharpens towards the end of the century – is a more decisive factor to determine drops in crop yields. This is more evident for grain maize, the most water-demanding crop (Fig. 3), which needs an additional +35 mm yr-1 (average over Europe) to support production by 2050 compared to the historical period. Approaching 2100 water demand for grain maize remains identical to the historical period for RCP4.5, while it is increased (+25 mm yr-1) for RCP8.5. Conversely, water demand for winter soft wheat remained constant throughout the century for both RCP4.5 and RCP8.5 scenarios, whereas i_RCP4.5 and i_RCP8.5 scenarios confirmed an increasing water demand of about 50 mm (average over Europe; Fig. 3), as also confirmed by Yang et al. (2019) for the Mediterranean regions.

N2O. The estimation and the projection of N2O emissions in the historical and the climate change scenarios improved upon previous studies over Europe. Lugato et al. (2017) estimated averaged emissions ranging from 1.18 to 2.63 kg N-N2O ha-1 yr-1 in the period 2010–2014 for both cropland and grassland production systems with the DayCent model. In comparison with Lugato et al. (2017), we found similar results for the Mediterranean latitudes (about 1 kg N-N2O ha-1 yr-1), while we predicted significantly lower emissions for central Europe (1.1 kg N-N2O ha-1 yr-1, this study), as well as at higher latitudes (0.96 kg N ha-1 yr-1, this study), compared to the 3 kg N-N2O ha-1 yr-1 forecasted by Lugato et al. (2017). Indeed, lower emissions at high latitudes were also observed by other studies (e.g. World Bank, 2021; Eurostat, 2017; Stehfest and Bouwman, 2006; Wells et al., 2018). Other research studies in the field were also within the range of our results; e.g. Reinds et al. (2012) estimated emissions ranging from 1.1 to 2.4 kg N-N2O ha-1 yr-1 for arable lands for the year 2000, and de Vries et al. (2011) estimated 0.27 and 0.38 Mt N-N2O yr-1 for fertilisers and manure and from grazing, respectively. Recently estimation by Eurostat (2017) reported values of 0.39 Mt N-N2O yr-1 (184.8 Tg CO2 eq.) for the year 2015 based on a lower-tier methodology, while our study reports a lower value equal to 0.17 Mt N-N2O yr-1 (80 Tg CO2 eq.) for the same year. Based on global inventories, Tian et al. (2020) reported emissions from European agriculture on the order of 0.51 Mt N-N2O yr-1 for the decade 2007–2016, which are significantly higher than those found in this present study (0.17 Mt N-N2O yr-1) for the same period. In addition, the estimation by Tian et al. (2020) also included manure management and aquaculture and suffers from high uncertainties given by the quality of the data and statistics used as input and, foremost, by the use of default emission factors. Regarding climate projection studies, Lugato et al. (2018) quantified N2O emissions for croplands in the RCP4.5 scenario, reporting losses of 1.81 and 1.77 kg N-N2O ha-1 yr-1 for the first and the second part of the century, respectively. These estimations were comparable to, although slightly higher than, the emissions for croplands issued from our study, both for the first part (1.53 ± 0.23 kg N ha-1 yr-1) and for the second part (1.66 ± 0.28 kg N ha-1 yr-1) of the century.

Our study highlighted that crop type is a significant determinant of the EFs of fertilisers, with most of the cereals having a low EF (barley, fodder maize, soft spring wheat and rapeseed; mean of 1.1 %) and pulses, soybean and potato a high one (mean EF of 3.1 %) during the 1985–2004 integration period. The highest EF for leguminous crops indicates that the management of fertilisation for these crops or for the rotation itself can be improved. Finally, information about crop-specific EFs turns out to be useful to design improved crop successions and to compile emission inventories (Myrgiotis et al., 2019). However, our results were higher than the 1 % default value defined by the IPCC guidelines for the N applied to agricultural soils, mainly because we considered only the N applied as fertiliser. Anyway, this default factor shows large uncertainties at local to regional scales due to the scarcely captured dependence on spatial diversity of the management, pedoclimatic, and soil physical and biochemical conditions (Leip et al., 2011; Reay et al., 2012; Shcherbak et al., 2014; Cayuela et al., 2017), which, however, are considered in our study. We observed that N2O emitted from croplands had a significant (p<0.05) and positive correlation with rainfall (r=0.47), as well as minimum and maximum air temperatures during the historical period (Fig. 12). The correlation with the minimum and maximum air temperatures increased significantly depending on the climatic scenarios (r>0.5 for RCP4.5, and r>0.9 for RCP8.5; Fig. 12), while the relation to rain became negative for RCP8.5 (r=-0.32). This trend inversion is probably connected to the strict dependency of N2O emissions on the length of the crop growing period rather than the yearly cumulated rainfall, which can occur outside of the cultivation period, as also stated by Shcherbak et al. (2014). Accordingly, the correlation from N2O and the irrigation amount, which occurs during the cultivation period, rose in the climate scenarios (r=0.23 and r=0.59 for RCP4.5 and RCP8.5, respectively). The rise in the projected temperature in the climate scenarios displays a latitudinal impact with N2O emissions, which is directly correlated at middle and high latitudes for croplands (r>0.5 for RCP4.5, and r>0.9 for RCP8.5) and at low latitudes for grasslands (r>0.45 for RCP8.5 and r>0.75 for RCP4.5). Precipitation in the mild climate has a direct influence on N2O emissions (r=0.25) at mid-latitudes for both production systems and also at high latitudes for croplands. Precipitation is anticorrelated with the N2O emissions in the RCP8.5 scenario at middle and high latitudes for croplands and at low latitudes for grasslands. Moreover, N2O emissions from cropland and grasslands were both positively correlated with soil clay content (r>0.5, p<0.01; data not reported) for values lower than 32 %, as higher clay content can promote complete denitrification (Weitz et al., 2001).

CH4. Methane emissions were mainly concentrated in the European regions with the highest density of grazing animals, as also observed by Vuichard et al. (2007). The values of the emissions simulated for the historical period (6.71 ± 0.4 kg C-CH4 ha-1 yr-1) were in the range of the experimental trials from central European grasslands (-2 to 108 kg C-CH4 ha-1 yr-1) discussed by Hörtnagl et al. (2018) and lower than the findings of Soussana et al. (2007), who reported typical emissions of 41 kg C-CH4 ha-1 yr-1 with animal densities comparable to our study. Our simulations were slightly lower than the simulations of Chang et al. (2015), which found emissions in a range of 18.7 ± 7.9 kg C-CH4 ha-1 yr-1 (period 1961–2010) over Europe with the ORCHIDEE-GM model, and lower than Vuichard et al. (2007) with an average of 108 kg C-CH4 ha-1 yr-1 (period 1994–2003) using the PaSim model but with a higher stocking rate. CH4 emissions decreased towards the end of the century, especially in the scenario RCP8.5 (-9 % compared to the historical period), due to reduced biomass productivity of grasslands that lessened the intake of animals (Fig. 2) and the stocking density, which declined to -8 % compared to RCP4.5 in the last decade of the century. Reduction in the stocking density was also foreseen by Chang et al. (2015). With regard to climatic changes, the rise in temperatures and the reduction in rainfall could directly act on reducing protein content and forage digestibility (process not simulated at the moment), possibly leading to reduced N2O losses from manure and urine in pastures. However, this mechanism could be offset by an increase in CH4 losses (Wilkinson and Lee, 2017).

Nitrogen use efficiency (NUE). We observed an increase in the NUE for the European croplands, especially for the mild climate change projection. In fact, compared to the historical period, in the RCP4.5 scenario there is a reduction in the correlation between N2O and the other N losses (NO3 and NH3) and crop yield (Fig. 12). At the same time, there is an intensification of dependence on the N dose. Indeed, with a constant amount of N applied in the rotations over the simulated years, both NO3- and NH3 losses were reduced over the century (data not reported) and crop yields increased until – at least – 2050. This indicates a potential increase in the NUE. Our findings were supported by the study of Kanter et al. (2016), who observed an increase in the NUE by 2050 related to the forecasted increasing yields. The improvement in NUE indicates a key factor to reduce negative environmental effects and mitigate GHG emissions (Maaz et al., 2021). UNEP (2013) indicated that NUE improvement could reduce N2O emissions by more than 30 % by 2050 in the RCP8.5 scenario. On the other hand, our results indicate that in the RCP8.5 scenario the correlation between N2O emissions and the N dose is lost, and a significant negative correlation (p<0.05) between yield and other nitrogen losses took place, indicating a reduction in NUE. This lack of a relationship is most probably connected to the interannual variability in N2O emissions in the strongest climate scenario and in the second part of the century. Higher NUE is typical of low latitudes in Europe, which benefit from generally higher yield and lower N losses compared to the middle and high latitudes (Sutton et al., 2011). Improving actual agronomic practices to increase crop yield or reduce reactive N losses has a direct and positive impact on the NUE (Lassaletta et al., 2014; Myrgiotis et al., 2019). In this context, irrigation represents a fundamental intensification practice to counteract the effects of climate change on crop production. In our case, the extension of the irrigation to all cropping systems in Europe significantly decreased N losses (-5 % for NO3 leaching and -4 % for NH3 emissions, on average, for both i_RCP4.5 and i_RCP8.5 scenarios) and contributed to increasing crop yields (Fig. 2), leading to a potential increase in NUE.

Regarding grasslands, we observed a weak relationship between N2O emissions and applied doses of N as fertiliser. This is mainly due to the calculation of the doses of N, which is a function of the specific animal load, the legume fraction, mowing events, and the available quantity of mineral and/or organics fertilisers for each simulation unit. During the historical period, N2O emissions were positively correlated with NO3 and NH3 losses and negatively correlated with production, representing a potentially low NUE. Moreover, N2O emissions in grassland are anticorrelated with CH4 emissions. CH4 emissions are rather positively related to biomass production and livestock intake. Therefore, poor biomass production could potentially lead to increased N2O emissions, due to low NUE, and thus decreased CH4 losses, due to low livestock intake. Surprisingly, N2O emissions in grasslands were weakly correlated with meteorological variables, especially minimum and maximum air temperatures, whereas a relation to rain and solar radiation is noticeable for RCP4.5 and is not evident for RCP8.5. As observed for croplands, the relation between N2O and NH3 emissions is positive, especially for the RCP8.5 scenario, indicating a potential reduction in the NUE.

NEP. The NEP represents a simple indicator of carbon storage potential since it does not account for C removal in terms of yield, animal intake or crop residues. Our results concerning croplands confirmed a net potential storage of C during the historical period (-3403 ± 214 kg C-CO2 ha-1 yr-1) and are directly comparable with those of Kutsch et al. (2010), who observed fluxes of -2400 ± 1130 kg C-CO2 ha-1 yr-1 based on field measurements in multiple sites in Europe. During climate scenarios, a significant decline in C uptake was predicted for the northern cropping systems (British Isles, Scandinavian Peninsula) and the Mediterranean area. This is most probably due to the increase in soil heterotrophic respiration caused by climatic factors and to a potential reduction in NEP (Fig. S9), respectively, as also reported by Kirschbaum (1995). Further decreasing values of NEP (towards a carbon stock) are evident in central and north-eastern Europe, especially in the first part of the century. A substantial increase in NEP in croplands was predicted towards the end of the century for the RCP8.5 scenario. This increase is most probably related to low levels of heterotrophic respiration (i.e. microbial respiration due to decomposition processes of soil organic matter) associated with partial soil coverage (e.g. no cover crops) of the simulated crop successions, as reported by Emmel et al. (2018).

Lower average NEP values were observed in grassland systems compared to croplands. This is related to the continuous removal of biomass by grazers; the generally higher SOC content in the topsoil; long-term land use (Morais et al., 2019); and the higher heterotrophic respiration that characterises these soils, especially when extensively managed (Bahn et al., 2008). This evidence has also been described by Chang et al. (2015), who simulated an average of -570 ± 210 kg C-CO2 ha-1 yr-1 between 1961 and 2010 for Europe, slightly lower in absolute value than the mean value simulated by our study of -622 ± 62 kg C-CO2 ha-1 yr-1 in the historical period (Table 1).

In general, areas where the heterotrophic respiration is enhanced by climatic drivers or by a high amount of SOC would lead to lower NEP values (Chang et al., 2017). This is the case for the north-east of France and the British Isles, while for the Scandinavian Peninsula and north-east Europe, which are characterised by low C and low heterotrophic respiration, NEP reached higher values. These results underline that in view of expected increasing productivity by 2050, storing additional (new) carbon will be more challenging in areas with high SOC levels (Hassink and Whitmore, 1997), mainly due to high levels of heterotrophic respiration. Finally, grasslands remained a potential sink for C during the historical period, which is in line with experimental measurements performed in the last 2 decades, e.g. -2470 ± 670 kg C-CO2 ha-1 yr-1 reported by Soussana et al. (2007) and from -910 to -17 830 kg C-CO2 ha-1 yr-1 reported by Hörtnagl et al. (2018).

N2O emissions were able to offset (reduce) the C sequestration potential of croplands. Offsets were on the order of 5.4 % (184 kg C-CO2 ha-1 yr-1) for the historical period and the first part of the century and rose up to 6.1 % and 7.5 % in the second part of the century for RCP4.5 and RCP8.5, respectively. The extension of irrigation to all European arable lands reduced these gaps thanks to the increased values of NEP (offsets of 5.4 % and 7.1 % for i_ RCP4.5 and i_RCP8.5). Even though few data are available in the literature regarding the CO2 storage potential for croplands (Emmel et al., 2018), our results confirmed that croplands may act as a potential sink of C when C exports by harvest are neglected (Buysse et al., 2017; Ceschia et al., 2010). N2O and CH4 emissions in grassland systems were able to offset NEP during the historical period by 17 % and 1 %, respectively. These results are partially compatible with the studies reported by Soussana et al. (2010), who displayed offsets over Europe of 34 % and 10 % for N2O and CH4, respectively. During climate projections, offsets rise to 22 % for N2O and 1.2 % for CH4 moving towards 2050 for both RCP4.5 and RCP8.5. In the second part of the century N2O emissions offset the potential carbon sequestration by 26 % and 52 % for RCP4.5 and RCP8.5, respectively, while the offset potential of CH4 ranged between 1.2 % and 1.9 % for RCP4.5 and RCP8.5, respectively.

The NGB, calculated by subtracting the other non-gaseous C fluxes (i.e. exports by harvest and crop residues, imports by manure) from NGHGE, indicated that European agricultural surfaces are a net C source. The most important components that determined these losses were the C exports – yield (FC-harvest) and crop residues (FC-residues) – which varied proportionally to the NEP in the various climatic projections; i.e. the lower the NEP, the lower the yields. For both cropland and grasslands, CO2 storage potential (estimated from NEP) provided the largest term in the net greenhouse gas exchange (NGHGE), as also confirmed by Jones et al. (2016). Non-CO2 GHGs, despite being high especially in the RCP8.5 scenario towards the end of the century, have a minor impact on differentiating the two climatic scenarios, although they represent an important component in the overall carbon balance at the European scale (see Table 2). The values of NGB highlight that inputs of C into the system, such as organic fertilisers (two-thirds of the component FC-manure), or actions aimed to recycle a portion of biomass in the field, such as crop residue management, are essential to improve the overall C budget to move towards net storage, as also reported by Ceschia et al. (2010) and Buysse et al. (2017). Moreover, our findings show that the contribution of the exported crop residues corresponded roughly to the whole carbon deficit in Europe. Therefore, crop residue could play a key role in land-based mitigation of anthropogenic emissions, as also reported by Stella et al. (2019) and Haas et al. (2022). This is in line with the “4 per 1000” initiative (Rumpel et al., 2019) promoting the maintenance of soil fertility as a key to achieve GHG mitigation strategies. In addition to the spatial diversity observed in the European agricultural area, the achievement of this goal depends on the complexity of the rural, economic and political structure of the territories (Amundson and Biardeau, 2018). Local policies can be supported by simulation tools such as those used in this study, bearing in mind that their effectiveness can be affected by the omission of large variances given by varied characteristics of small extents (see Sect. 4.5). Finally, irrigation management extended to all European cropping land led to slightly higher C deficits because it is able to increase the stored C (NEP =+3 %), but it also increases the removal of crop residues (FC-residues) and the emission of non-CO2 GHGs (up to +4 %).

The extension of field-scale models to regional scales faces several challenges associated with the representation of the systems under study, which can affect the confidence of the model outputs (Challinor et al., 2014; Folberth et al., 2019).

Input data. The input requirement for dynamic crop and grassland models for large and heterogeneous areas is difficult to fulfil (Therond et al., 2011). While soil and climate inputs are directly available from European databases at different spatial resolutions, details on crop and grassland management (e.g. type and number of inputs, timing of operation, tillage system, crop varieties) are less readily available and represent an important source of uncertainty (Molina-Herrera et al., 2016). In our assessment we used a dataset for cropland constituted by statistical data of crop rotations resulting from a spatial distribution of crops at the NUTS2 scale and by crop succession likelihoods, on a high-resolution scale (1 km × 1 km). This dataset does not include details about intercropping, cover crops, management or crop growth parameters. The absence of plant phenological development data, for instance, is a relevant source of uncertainty in regional assessments (Minoli et al., 2019). This information defines crop growth and the length of the growing season, influencing biogeochemical cycles at different scales and becoming key for future projections. To deal with this lack of information, we calculated crop-specific sowing and fertilisation dates as a function of climate (Ramirez-Villegas et al., 2015), together with the uses of different crop varieties following a latitudinal gradient to fulfil the thermal unit need, N doses and the crop-specific residue management, aiming to reduce the uncertainty in input data (Hansen and Jones, 2000). Furthermore, the use of two different crop rotations per simulation unit attempted to cover a range of uncertainties existing below the spatial resolution of 0.25∘, which, however, cannot be assumed to be fully covered by the range of setups presented here. Another limitation of this and other regionalised studies is the deviation in the representation of the quantities within the administrative units, which is related to the scarcity of management data with a fine spatial resolution. For example, the amounts of fertiliser to be distributed in cropping systems which are provided at a regional level show little heterogeneity within the boundaries of the region itself and can mark a sharp transition between adjacent regions. Regarding model parameterisation, in the present work the CERES-EGC model used fixed parameters issued from a calibration over different sites in Europe (Lehuger et al., 2010, 2011; goodness of fit, R2 is 0.59 to 0.76 for NEP; error of prediction reduced by 6 %–40 % for N2O compared with the model's standard parameters). Grasslands, as previously reported, were simulated with a parameter set resulting from a multi-site calibration for a network of European grasslands (i.e. flux tower network; see Ma et al., 2015; goodness of fit, R2 is 0.4 to 0.9). Likewise, PaSim follows adaptive management based on climate. Since the information concerning the input data is already the result of a scaling process, an uncertainty analysis concerning the input data is not suitable here (Hansen and Jones, 2000).

Calibration of models. To fulfil the task of calibration over large areas, data representing the spatial and temporal variation in models' parameters are required. Although both models have been calibrated and verified with direct observations under various pedoclimatic and management conditions at the field scale, comprehensive studies aimed at calibrating these and other models with spatially extensive time series are still scarce (Balkovič et al., 2013; Lehuger et al., 2010; Lugato et al., 2010; Vuichard et al., 2007). Data aggregation over the same extent can be used to assess model representations, even if they do not represent the field-scale conditions for which the models have been originally calibrated (Lugato et al., 2017; Therond et al., 2011; van der Velde et al., 2009), exposing models to a broader range of conditions (e.g. weather and soil characteristics). Indeed, dealing with lacking and heterogeneous input data requires different procedures of downscaling and upscaling for the different data types, which potentially contribute to feeding the uncertainty in the representation. Consequently, projecting responses of regional models under future climate scenarios requires careful understanding of input and model uncertainty (Asseng et al., 2013; Challinor et al., 2009). This is the reason why the two periods of temporal aggregation considered in the present study, historical and climate scenarios, provide outcomes with different levels of confidence. In the historical period, results are obtained based on the spatial aggregation of real (statistical) data by means of models parameterised with current soil, climate and management conditions. The outcomes of the climate scenarios deal with the uncertainty related to the sensitivity of the model parameters and their algorithms to climate variables, which is expected to be different due to the diverging intensities of the two climate projections and the different conditions in the near and the long term (2100). For this reason, direct comparisons between the two aggregation periods should be made but with caution.

Model validation. Data quality and availability also prevent the validation of regional-scale models, even if the literature reports some effort (Challinor et al., 2009; Faivre et al., 2004; Niu et al., 2009). Comparing the model outputs with statistical data aggregated at the regional scale (production) allowed us to obtain indications about the magnitude of simulated variables at the same spatial extent. Furthermore, assessing the ranges of the model outputs, e.g. yield, with measured data and over Europe (R2=0.92, p<0.01, for croplands, Fig. S1; R2=0.68, p<0.05, for grasslands, Fig. 1), as well as other modelling interpretations (even if grounded on different approaches), contributed decidedly to increasing the reliability of our estimations.

The literature includes similar studies aiming to estimate crop and/or grassland production, GHG emissions, and carbon storage at the European scale. Lugato et al. (2014) created a database, based on EU statistics, of soil, climate and land use for grassland and cropland to simulate SOC with the CENTURY model, which worked at a monthly time step. This dataset was subsequently enhanced with the direct soil observation network over Europe (LUCAS – Land Use and Cover Area frame Survey) to estimate N2O emissions from cropland with the DayCent model, at a daily and long-term resolution (Lugato et al., 2017). Finally, these authors used an effective method of filling and scaling the simulated results by means of a meta-model, although adding a potential additional source of uncertainty. Regarding grasslands, Chang et al. (2015) used the ORCHIDEE-GM model (which contains the management equations derived from PaSim) to simulate the greenhouse gas balance in Europe, later extending their analysis with climate change scenarios (Chang et al., 2017). Blanke et al. (2018) improved the LPJ-GUESS global model to be applied to grasslands and to estimate the carbon and nitrogen balance with future climate scenarios. Compared to the studies mentioned above, our work combined two state-of-the-art models specific to the systems under study, producing both separate and joint results on croplands and grasslands. The novelty of this work is also based on the use of dynamic management with specific crop rotations, instead of exploring one crop or a few crops (e.g. Sansoulet et al., 2014; Yang et al., 2019), and the work was conducted at a finer spatial resolution than, for example, that of Vuichard et al. (2007) with the PaSim model or Leip et al. (2008) with DNDC and agrees with the resolution used by Chang et al. (2015) for grassland. Recent studies at an ever finer spatial scale have been proposed over France with the STICS model (Launay et al., 2021). Albeit aggregated in gridded simulation units, our work considered a variety of pedoclimatic conditions over Europe and is not based on an extrapolation of a few points or on a single European area (Ceschia et al., 2010; Kutsch et al., 2010; Myrgiotis et al., 2019; Soussana et al., 2010).

Finally, knowing and controlling the sources of uncertainty from regional applications could be a key to improving decision support tools for the design of policies. In this context, providing a range of possible outcomes, the application of multi-model ensemble (Ehrhardt et al., 2018; Martre et al., 2015; Sándor et al., 2018; Rosenzweig et al., 2013) at a regional scale could represent a valuable tool to tackle this source of uncertainty. Increasing the spatial resolution of the input dataset we used (e.g. weather and management data) could also represent a key to further reduce uncertainties from input data in future large-scale applications (Folberth et al., 2019; Hoffmann et al., 2016; Stella et al., 2019).