The field experiment was conducted in six sites in northern China, including Hulunber, Suihua, Chifeng, Gongzhuling, Yulin, and Jiuquan ( Figure 1 ). The climatic and soil characteristics of these sites are given in Table 1 .

Field trials were conducted at each of the six experimental sites for five consecutive years during 2012-2017 to test the suitability and to measure the aboveground dry-matter yield of alfalfa. Field management was kept consistent across experimental sites. At each site, four parallel trial blocks of 1 m × 3 m were used to represent four replicates of alfalfa pastures; blocks were separated by a space of 0.5 m in width. Seeds were sown in 0.015-0.025 m soil depth and rows with a 0.2 m inter-row spacing. Seed density was kept at 18 kg ha^-1 for all sites. Diammonium phosphate was applied as the base fertilizer at a rate of 150 kg ha^-1. Alfalfa plants were harvested 3-4 times a year, as stipulated by the management calendar ( Table 2 ).

A quadrat of 0.5 m × 0.5 m was used at each cutting time and at harvest to determine biomass yield. The quadrat was laid down at three random points within each trial block. Alfalfa plants inside the quadrat were manually cut at 0.05 m above the soil surface. The plants were taken to the laboratory and oven-dried at 65°C for 24 hours or until a constant weight was reached. The dried plants were weighed to obtain the dry weight.

Three soil samples of 0-0.3 m depth were randomly taken using a soil drill (0.07 m inner diameter; 0.1 m height, New Landmark Instrument Co., Ltd., Beijing) near the quadrat. Samples from the same sampling point were mixed to form a composite sample. All samples were air-dried in the laboratory. SOC was determined using the potassium dichromate oxidation method (Patrick et al., 1996).

Soil gas fluxes were measured using the opaque static chamber method (Zhang et al., 2010). The static chamber system consisted of a stainless-steel frame (open top and bottom, 0.5 m in length × 0.5 m in width × 0.1 m in height) that was driven into the soil and a stainless-steel chamber (open bottom, 0.5 m L × 0.5 m W × 0.5 m H) that was placed tightly in the base groove during the sampling period. The frame was inserted into the soil to a depth of 0.1 m below the soil surface. A cover was placed on top during sampling times and removed afterward. A fan of 0.1 m in diameter was installed at the top of each chamber to generate turbulence when the chamber was closed. The external surface of each chamber was covered with white plastic foam to minimize the effects of direct radiative heating during sampling. Three replicate chambers were randomly established in each plot and used for simultaneous measurements. The headspace in each chamber was sampled at intervals of 0, 10, 20, and 30 minutes after the chamber was closed. The gas was transferred immediately into a pre-evacuated 0.05 litter airbag using a 0.06 litter plastic syringe (Hede Inc., Dalian, China). This sampling procedure was executed every 3 days during the growing season (June-October) from 2013 to 2015 in Hulunber. All measurements were taken between 9 and 11 a.m. The CH₄, N₂O, and CO₂ concentrations of the gas samples were analyzed in the laboratory using gas chromatography (Agilent 7890 A, Agilent Technologies Limited Co., USA).

Soil temperature and soil moisture were measured in situ in 0-0.1 m soil depth. Soil temperature was measured using a TJ2 soil temperature recorder (Beijing Hezhong Bopu Technology Development Co., Ltd, China). Soil moisture was measured using an HS2 portable soil moisture sensor (Decagon Devices, USA).

The DNDC model (Li et al., 2003) was employed to simulate alfalfa yield in five northern provinces in China for the period between 2025 and 2035. These provinces included Inner Mongolia, Heilongjiang, Jilin, Shaanxi, and Gansu ( Figure 1 ). The alfalfa pastures were assumed on marginal lands only. The marginal land SOC contents with and without alfalfa pastures were also simulated.

The DNDC (DeNitrification DeComposition) model was first described by Li et al. (1992) as a biogeochemical simulation model for predicting N₂O, CO₂, and N₂ emissions from agricultural soils in the U.S. After many years of development, DNDC can now be used as a plot-scale, process-based model driven by climate, soil and crop parameters to simulate crop growth, greenhouse gas emissions, SOC and other soil properties not only in the crop but also in livestock and grassland systems (Levy et al., 2007; Gilhespy et al., 2014; Xu et al., 2018; Khalil et al., 2020). To address the unique characteristics of intensive cropping systems in China, Li et al. (2017) included typical Chinese cropping systems (e.g., double and triple paddy rice) into the DNDC model. In addition, the support for common Chinese cultivars of alfalfa was also included.

The DNDC consists of two modules. The first module simulates crop growth and C and N cycling in soil by using three sub-models of soil climate, crop growth, and decomposition. The second module simulates biogeochemical processes related to soil environmental factors by using sub-models of nitrification, denitrification, and fermentation. Crop growth is simulated based on daily accumulations of photosynthesis, respiration, C allocation, water and N uptake by crops (He et al., 2019). Both water and N uptake rely on several factors such as soil N distribution, soil moisture content, and root length, etc. Water utilization depends on potential transpiration linked with leaf area index and climatic conditions. Water stress is simulated when potential transpiration is relatively higher than normal or actual water supply (Zhang et al., 2017). The C pools are divided into four SOC pools, namely, plant residue (litter), microbial biomass, active humus, and passive humus. The litter pool was further divided into three sub-pools of very labile, labile, and resistant litter based on their different C/N ratios and decomposition rates. The DNDC model predicts SOC dynamics mainly by calculating C input (litter, crop residues, root exudates, or manure application) and C output (SOC decomposition or erosion). Litter incorporated into the soil is broken down by soil microbes, partly served energy of the microbes, and partly turns into microbial biomass. After the death of microbes, the microbial remains turn into an active humus pool by decomposition. Then, active humus can be further utilized by microbes to turn into a passive humus pool (Li et al., 2016).

A three-step procedure was followed in this study to simulate alfalfa yield and SOC: (1) the DNDC model was calibrated to each experimental site by adopting a set of finetuned biophysical parameters for alfalfa ( Table S1 ); (2) the model was validated using site-specific measurements of soil temperature, moisture, soil gas flux, alfalfa yield, and SOC content; and (3) regional forage production and C sequestration potentials of sown alfalfa pastures on marginal lands in the five northern provinces under three contrasting IPCC SSP climate scenarios between 2025 ad 2035 were simulated.

The DNDC model’s performance in predicting alfalfa yield and SOC was evaluated using the statistical metrics of mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE), which were obtained using Equations 1, 2 and 3, respectively:

where y is the field observed value; y ^   is the model-predicted value and N is the sample size; additionally, the mean absolute percentage accuracy (MAPA) was referred to as the overall prediction accuracy, which was derived from MAPA=100−MAPE .

Future climatic data were generated by the Beijing Climate Center’s Climate System Model 2 Medium Resolution version (BCC-CSM2-MR) (Wu et al., 2019) coupled with three contrasting Shared Socioeconomic Pathway (SSP) scenarios, namely, SSP1-2.6, SSP2-4.5, and SSP5-8.5 (Meinshausen et al., 2020). The BCC-CSM2-MR was a participating model in the Coupled Model Intercomparison Project Phase 6 (CMIP6) that supported the IPCC’s Sixth Assessment Report (Eyring et al., 2016). The model’s simulation results were provided at a grid resolution of 1.125 degrees or roughly 100 km in northern China. Descriptions of the three abovementioned SSPs are given in Table 3 . The BCC-CSM2-MR model-simulated variations in mean temperature and precipitation during the annual growing seasons of alfalfa (MST and MSP, respectively) in the five northern provinces from 2025 to 2035 are presented in Figure 2 .

The dry-matter yield of sown alfalfa pastures on marginal lands from 2025 to 2035 in northern China was projected by the DNDC model to fluctuate slightly around a level of 4.5 Mg ha^-1 in rainfed and 9.0 Mg ha^-1 in irrigated conditions ( Figure 6 ). The rainfed yield compared lower than the site-observed average yield of 6.6 Mg ha^-1, while the irrigated yield compared substantially higher, reflecting the fact that the site experiments of alfalfa trials involved only partial irrigation to avoid plant failure because of drought. A slightly decreasing tendency in alfalfa yield during 2025-2035 was captured by the DNDC model, confirming the yield pattern of Chinese cultivars as also revealed by experiments on the Loess Plateau (Jiang et al., 2006). Moreover, the projected alfalfa yield tended to be the highest with the SSP2-4.5 climate scenario, followed by SSP5-8.5 and SSP1-2.6. The average irrigated yield was projected as 9.19, 9.03, and 8.79 Mg ha^-1, whereas the average rainfed yield was projected as 4.71, 4.46, and 4.43 Mg ha^-1 under the three SSP scenarios, respectively, showing that the SSP-induced yield differences were much smaller than irrigation-induced differences. The DNDC model projected temporally consistent patterns for irrigated yield across the three SSP scenarios. However, the same pattern was hardly discernable for rainfed yield, especially between the SSP5-8.5 and SSP1-2.6 scenarios.

The SOC content in the marginal soil under sown alfalfa pastures in northern China was projected to increase at a consistently positive rate of 0.19 g kg^-1 yr^-1 across water regimes and climate scenarios between 2025 and 2035. The projected SOC levels under rainfed and irrigated conditions were averaged at 18.9 and 20.3 g kg^-1, respectively, corresponding to a net increase in SOC by 8.47 and 9.88 g kg^-1 above the fallow management. Moreover, the projected SOC increase was compared considerably lower than the site-observed SOC increase of 2.01 g kg^-1 yr^-1, showing that DNDC well represented the differences between plot trials and field production. The projected SOC content also displayed a clear high-low differentiation across the climate scenarios and water regimes from 2025 to 2035. The SOC content under irrigated alfalfa pastures tended to be the highest with the SSP2-4.5 scenario (14.00 g kg^-1), followed by SSP5-8.5 (13.61 g kg^-1) and SSP1-2.6 (13.05 g kg^-1) scenarios. It is worth noting that the SOC content under rainfed alfalfa pastures was the lowest with the SSP2-4.5 climate scenario (12.36 g kg^-1), compared to 12.57 g kg^-1 with SSP1-2.6 and 12.90 g kg^-1 with SSP5-8.5 scenarios, showing that the climate sensitivity of SOC content was well reproduced by the DNDC simulation results.

If half of the available marginal lands in northern China were utilized, the forage production from sown alfalfa pastures would increase from 0.83 million Mg in 2025 to 1.42 million Mg in 2035 under rainfed and from 2.08 million Mg in 2025 to 3.42 million Mg in 2035 under irrigated conditions ( Table S2 ). The alfalfa forage production variated only slightly across the SSP climate scenarios. The average rainfed alfalfa forage production was projected to be 1.11, 1.16, and 1.09 million Mg under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios, respectively, whereas the irrigated forage production was projected to be 2.69, 2.72, and 2.68 million Mg under the same scenarios. This corresponds to an increase in the livestock supporting capacity from 17.26 million stand sheep units (SSU, derived from livestock numbers using a conversion factor of 1 for sheep, 0.8 for goats, and 5 for cattle, horses, and camels (NBSC, 2021)) in 2025 to 29.26 million SSU in 2035 under the rainfed conditions. Likewise, the sown alfalfa pastures’ livestock-supporting capacity would increase from 45.07 million SSU in 2025 to 74.09 million SSU in 2035 under irrigated conditions. Effectively, due to the displacement of market demands by livestock production in captivity, the livestock stocking rate on natural grasslands in northern China would decrease by 0.29 and 0.77 SSU ha^-1 or 7.8% and 20.4% under rainfed and irrigated conditions in 2025. The grassland stocking rate was projected to decrease continuously during the 2025-2035 period. In 2035, the stocking rate would decrease by 0.50 SSU ha^-1 or 13.2% under rainfed and 1.26 SSU ha^-1 or 33.5% under irrigated conditions ( Figure 7 ). Full details on alfalfa pasture acreage and stocking rate evaluations are given in Table S2 of the Supplementary Material .

The SOC density (SOCD) in marginal soils under sown alfalfa pasture in northern China was projected to increase significantly at an average rate of 0.08 kg m^-2 year^-1, from 7.34 kg m^-2 in 2025 to 8.17 kg m^-2 in 2035 ( Figure 8 ). It is worth noting that the DNDC model also projected positive effects of fallow management on SOCD. The SOCD of fallowed marginal lands in northern China was projected to increase from 3.83 kg m^-2 in 2025 to 4.48 kg m^-2 in 2035. Taken together, the sown alfalfa pasture on marginal lands in northern China may cause a net increase in SOCD at magnitudes ranging between 3.51 kg m^-2 and 3.69 kg m^-2 during 2025-2035. The projected SOCD varied slightly among the SSP climate scenarios. The highest SOCD level (7.68 kg m^-2) was found in rainfed alfalfa pasture under the SSP5-8.5 scenario, whereas in irrigated alfalfa pasture, the highest SOCD (8.40 kg m^-2) was found under the SSP2-4.5 scenario with an overall standard deviation of 0.82 kg m^-2.

Previous research (Li et al., 2016; Khalil et al., 2020) found that pasture systems were less involved in DNDC model development and, consequently, the number of model applications to pasture systems was much less than to, e.g., cropland systems. To overcome the hurdle of insufficient DNDC modeling expertise for alfalfa pastures in northern China per se, we adopted a field experiment-based calibration approach in modeling alfalfa pasture’s forage production and C sequestration potentials. We collected daily data on soil moisture, soil temperature, and soil greenhouse gas flux for at least two alfalfa growing seasons to calibrate the DNDC model. We also collected yield and SOC data from six experimental sites between 2012 and 2017. These efforts were well rewarded by the obtained accuracy of the modeling results. We obtained an overall prediction accuracy of 90.73% for SOC and 92.41% for alfalfa’s dry-matter yield, aggregated from the six experimental sites across northern China. We also obtained high goodness-of-fit levels between the model-predicted and field-observed SOC and yield. The coefficient of determination (R²) was evaluated as 0.85 for SOC and 0.88 for yield. In general, our modeling results achieved higher accuracy than most of the DNDC model applications in China and elsewhere during the past five years ( Table 4 ). Moreover, the model reproduction of the observed variability patterns in space and time for both SOC and yield was well done, as shown by the specific scatter plots ( Figures 4B, D, F ; Figures 5C, D ), which were rarely provided in previous DNDC model applications, in addition to time-series plots ( Figures 4A, C, E , 5A, B ).

In comparison, Han et al. (2017), for example, simulated cropland SOC in China with the DNDC model and obtained a satisfactory prediction accuracy of 84.73% using country-wide soil and crop management data, which corresponded to a mean absolute percentage error (MAPE) of 15.27%. Likewise, Zhang et al. (2017) modeled the 0-20 cm SOC of the wheat-maize system in the North China Plain based on a 30-year experimental record and reached a prediction accuracy of 90.11%. Moreover, Khalil et al. (2020) simulated the SOC density in the 0-15 cm soil layer of a ryegrass pasture in the U.K. using the DNDC model and managed to keep the prediction error lower than 16%, a level comparable to Han et al. (2017). However, the coefficient of determination (R²) for Khalil et al.’s simulation results (0.49) was substantially lower than that of Han et al. (0.69), showing that DNDC is less capable of accounting for field variabilities in pasture soil C than in cropland systems. This accountability discrepancy is in part due to the low quality of input data, as elaborated by Li et al. (2016). In the wheat-maize system studied by Li et al. (2016), straw retention was the major C input into the soil. A rough estimation of the straw retention rate at the regional scale would naturally lead to low accountability of spatial variability in the modeling results. This was in sharp contrast to this study where all calibration data were experimentally collected at the plot scale.

Much attention has been given to yield prediction in almost all modeling efforts including DNDC. Although yield prediction was not the primary focus of the DNDC model (Li et al., 1992), the DNDC model was more and more frequently applied in yield prediction in recent years. For example, Jiang et al. (2021) simulated the yield of paddy rice in China and obtained a high prediction accuracy of 98.79%. Jarecki et al. (2018) simulated the yield of the maize-alfalfa system in Canada, which achieved an accuracy of 95.45%. Although the yield prediction accuracy of this study was slightly lower than these two examples, the yield prediction accuracy of this study was higher than all the rest studies listed in Table 4 . It is worth noting that the high prediction accuracy of this study was achieved for a pasture system that was not as well supported by the DNDC model as for a cropland system. Moreover, the high prediction accuracy of this study was achieved for northern China, where DNDC modeling expertise was much poorer than, e.g., in Canada (He et al., 2019).

Many studies (Zhang et al., 2015; Gao et al., 2016; Kemp et al., 2020) have suggested that overgrazing was one of the most important drivers of grassland degradation in China. A recent survey found that nearly 90% of the grasslands in northern China experienced varying degrees of degradation (Kemp et al., 2020) and these degraded grasslands were overgrazed by 27-89% (Zhang et al., 2014). Previous research suggested that grazing intensity control should be used as the primary means of grassland conservation (Zhang et al., 2015). In most areas in northern China, a 50% reduction in grazing intensity was advocated. However, it remained uncertain how these grazing intensity reduction targets can be realized in practice, especially at the regional scale. Although livestock rearing in captivity has been proposed as a pragmatic measure to displace market pressures away from the grasslands (Chen et al., 2017), it relies on imported forage which has complex implications for the domestic food market. In this study, we demonstrated through DNDC model simulation that sown alfalfa pasture on marginal land has the potential to effectively lower the grazing intensity in northern China by, on average, 8-13% under climate change in the near-term future. We also demonstrated that the grazing intensity on natural grasslands in northern China could be reduced by one-third (20-33%) if irrigated alfalfa pasture had been regionally implemented. This represents significant progress in grassland protection due to multiple reasons.

Firstly, sown alfalfa pasture on marginal land represents a practical solution based on the integrated use of land resources. Since the beginning of the 21^st century, a range of policy reforms on the ecological protection of grassland resources has been instituted in China. These included the Grassland Eco-compensation Program, the Farmer’s Professional Cooperative Law, and so on, in addition to the well-known Natural Grassland Restoration Program and the Returning Grazing Land to Grassland Program (Robinson et al., 2017). Collectively under these institutions, marginal lands, which were previously under agricultural cultivation, were set aside on fallow. In 2020, there were 4,000 km^-2 of marginal land under fallow management or being regarded as a wasteland in northern China (NBSC, 2021). In this study, we considered a 50% use of these marginal land resources for sown alfalfa pasture in 2025. From 2026 to 2035, an annual expansion in sown alfalfa pasture areas of 5% was considered ( Table S2 ).

Secondly, sown alfalfa pasture represents the single most effective technical measure at the regional scale which has the potential to, if implemented on 100% of the marginal land area, lower the grazing intensity on northern China grasslands by at least 25 (rainfed) to 66% (irrigated) in changing climate conditions. Effectively, this will eliminate overgrazing on about half of the degraded grasslands in northern China, based on a comparison to the evaluation results of Zhang et al. (2014).

Thirdly, sown alfalfa pasture in combination with marginal lands in northern China also has social significance for herder households. A lower grazing intensity does not necessarily mean lower household income, because productivity per head of animal usually increases with decreasing herd size (Zhang et al., 2015). More importantly, a decreased herd size will hopefully spare time for education and training (Gao et al., 2016), which is much needed for herder households across the region.

Grasslands store approximately one-third of the global terrestrial C. On the one hand, improvements in grazing management and biodiversity restoration have the potential to deliver natural climate benefits in grasslands through soil C sequestration (Yao et al., 2015; Li et al., 2022). On the other hand, however, widespread grassland degradation poses significant threats to the soil C pool by enhanced soil C mineralization (Ye et al., 2010; Abdalla et al., 2018). The critical role of soil C in global climate change regulation has not only been recognized by scientists but also by policymakers. Notably, at the 21st session of the United Nations Framework Convention on Climate Change, which was held in Paris in 2015, an ambitious target of a 0.4% rate of the annual increase in global SOC stock was proposed under the Agenda for Actions (Soussana et al., 2019). This target of a 0.4% increase in global SOC stocks, codenamed “4 per Mille”, was evaluated as the amount of C sequestration required per year to offset anthropogenic greenhouse gases emissions of 8.9 Giga tons (Minasny et al., 2017). In this study, significant SOC increases were observed in experimental plots under sown alfalfa pastures at all 6 sites across northern China. In 2012, the site averaged SOC was measured at 11.4 g kg^-1 when alfalfa seeds were sown. SOC increased to 12.1 g kg^-1 in 2013, which corresponded to a 5.9% increase. The annual rate of SOC increase was measured at 9.8% in 2014, 10.6% in 2015, 15.8% in 2016, and 31.2% in 2017, which were all more than 10 times higher than the 0.4% threshold, showing alfalfa pasture’s significantly positive effects on soil C sequestration.

Alfalfa pasture’s positive effects on SOC were not only observed at the site scale but also confirmed by the DNDC model simulations at the regional scale. The marginal land SOC under sown alfalfa pastures in northern China was projected to increase at the rate of 1.31% in 2026, 1.81% in 2027, and 0.93% in 2035, the last year of the simulation period, under the rainfed conditions. All these rates are significantly higher than the 0.4% threshold stipulated by the 4 per Mille Initiative. Under the irrigated conditions, even higher SOC increase rates were projected ( Figure 9 ), reflecting the primary influence of C input quantities – both from the aboveground biomass through the litterfall pathway and from the belowground biomass through the root pathway – on soil C sequestration (Chen et al., 2021; Li et al., 2022; Xu et al., 2022).

We would like to indicate that, in addition to C sequestration in marginal lands, sown alfalfa pastures also result in invaluable ecosystem gains via, e.g., grazing intensity control. Previous research found that increasing gazing intensity on global grasslands was associated with decreasing trends of SOC and increasing trends of soil bulk density (Lai and Kumar, 2020). This signifies that lessening grazing intensity is itself an efficient grassland conservation measure that may produce significant indirect benefits for soil fertility conservation and soil water retention. We must also indicate that the so-called restoration of grassland productivity in northern China that was suggested by recent satellite data (Chen et al., 2017) and hailed as a proven achievement of past grassland conservation programs was both superficial and early mature. What the satellite data showed was the growth of aboveground biomass. The more endurable restoration, such as a replenished SOC stock, was hidden from satellite observations (Ma et al., 2017) and is yet to accomplish.

Uncertainties in future climates represent a major challenge to modeling ecosystem functions (Parker, 2013). One solution to improve the certainty of the modeling outcomes is to adopt an ensemble modeling scheme so that the uncertainty of the modeling outcome can be statistically obtained, as showcased by IPCC-AR6 (IPCC, 2022). In this study, however, we bypassed this heavy-duty option of ensemble modeling but opted to use three contrasting climate scenarios. A high degree of certainty was achieved in the simulated alfalfa yield and SOC, as suggested by the LSD tests applied to these results ( Figure 10 ). The LSD tests showed that alfalfa yields were not significantly different across the three SSP scenarios, meaning that alfalfa yield in northern China in the near-term future will be highly likely within the simulated yield range. Although a seemingly more complex picture was depicted by the simulation results for SOC under sown alfalfa pastures, the message that the results conveyed was quite clear. Our climate future will be much more likely along a sustainable development path (SSP1-2.6) than pretending unawareness of climate change (SSP5-8.5). There will be many people taking a middle road (SSP2-4.5), but the difference between this road and SSP1-2.6 was not statistically significant. This means that the marginal land’s SOC content under alfalfa pasture in northern China from 2025 to 2035 will be most likely within the predicted SOC range with the SSP1-2.6 and SSP2-4.5 scenarios.

It is well known that irrigation water is scarce in northern China. The situation will likely not turn better from 2025 to 2035 in the five northern provinces considered here ( Figure 2 ). The consideration to include an irrigation scenario in this study was to provide an option for areas where water resources were more abundant and hence a tradeoff among competing uses had economic importance, as in the case of Gansu province where irrigated alfalfa was practiced (Jiang et al., 2006). The statistically significant difference between irrigated and rainfed alfalfa forage yield in Figure 10 confirmed that irrigation will still be an economic incentive in local water management decisions. An additional ecosystem benefit that irrigated alfalfa will produce is that it significantly increases the soil’s carbon sequestration potential by 20%.