Introduction

Soil organic carbon (SOC) storage and nitrous oxide reduction emissions are key strategies for agricultural climate change mitigation but implementing regenerative practices on croplands at scale is challenging due to inherent field variability, numerous quantification approaches, practical and financial barriers, and vested interests limit data accuracy, reproducibility, and scalability. Uncoordinated, inconsistent carbon market initiatives reflect these challenges1, failing to attract significant producer participation and hindering confidence in environmental and financial outcomes. In short, the identification and quantification of climate benefits at a meaningful scale has been stymied by broad ranges and large uncertainties2.

Individual process-based ecosystem models developed for agriculture, more precisely called cropping system models (CSMs)3, have been used to investigate the effects of management and climate change on SOC and GHG emissions4,5,6 at regional to global scales (e.g., Refs.7,8). CSMs incorporate varying theoretical assumptions, often using limited empirical observations for model calibration and validation9,10. Owing to the different crops, management practices, and geographic regions for which they have been developed, different CSMs can provide divergent predictions for SOC change, particularly for those models applied to environments and scenarios, present and future, for which they have not been calibrated or developed11.

Research over the last several decades has shown that there is no silver-bullet crop and biogeochemical model12, although some have been more widely tested than others, and some specific combinations of models can perform better13. So, even if CSMs provide an attractive alternative to large-scale SOC and N2O measurement and monitoring programs, no one model can satisfactorily quantify its change across all combinations of soils, climates, crops, and agronomic practices at varying scales, despite the well-validated performance of individual models in specific contexts. For all these reasons, models practitioners and certificate providers still face the challenge of selecting the right model, appropriate to the specific conditions in which they are working14.

Climate modelers have addressed this problem with multi-model ensemble (MME) approaches (e.g., Ref.15), which have also worked well for predicting climate-induced crop productivity change4,16. MMEs capture differences in model structure as well as broader sets of environmental conditions and management practices than those for which individual models have been developed and tested. Consequently, they provide a substantial advantage for evaluating SOC change and GHG emissions in response to changes in management practices and climate, which could be particularly attractive for addressing measurement, monitoring, reporting, and verification (MMRV) challenges of current carbon credit frameworks1,17. At small scales, previous MMEs have been used to predict SOC change more accurately than individual models in croplands (e.g. Refs.4,13,18), grasslands (e.g., Ref.19), and bare soil (e.g., Ref.20), but have yet to be applied to large, landscape-scale questions despite the scientific and policy importance of mitigation interventions in agriculture21.

Soil sampling programs typically cannot meet MMRV challenges because of the cost of sampling and analysis, the need to adequately capture spatial variability, and the decadal or longer time frame required to detect credible SOC change22,23,24,25. Consequently, on-farm observations, unless conducted as part of funded scientific trials with scientific expert supervision, typically cannot reliably capture the magnitude of carbon dynamics, particularly the rate of SOC change26, leading to large uncertainties in SOC datasets27. These are major problems for calculating cropland SOC budgets and credits28, with important implications for crediting SOC gains17.

Ultimately, the global challenge of climate mitigation demands a new paradigm for evaluating SOC change. It should not be assessed in isolation, but in terms of its ability to offset major agricultural emissions—most notably, N2O29. A single practice that increases SOC does not necessarily represent the most effective management strategy if associated N₂O emissions are not also considered30,31.

This study investigates the large-scale effects of regenerative management practices on SOC and N2O dynamics using a multi-model ensemble (MME) approach. The ensemble includes widely used CSMs—APSIM, ARMOSA, CropSyst, DayCent, DSSAT, EPIC, SALUS, and STICS—and is applied across 40,000 unique locations across 934 counties in 12 US Midwest States. Estimates of SOC stock change will be used to (1) quantify the potential SOC accrual from the adoption of regenerative practices relative to a dynamic business-as-usual baselines, and (2) assess the climate mitigation potential of these practices by evaluating their capacity to offset N₂O emissions at scale.

We hypothesize that SOC gains will continue to offset N₂O emissions—with the extent of offset depending on site-specific interactions between pedo-climatic conditions and the practices adopted—ultimately resulting in a net positive climate mitigation effect from implementing regenerative agricultural practices.

Material and methods

Process-based cropping systems models (CSMs)

Eight different process-based CSMs were used to develop the MME (Table S1): Apsim32, Armosa33, Cropsyst34, Daycent35, Dssat36, Epic37, Salus38 and Stics39. Models were selected primarily based on their ability to reproduce interactions between soil, plant, management and the atmosphere to simulate crop yields and ecosystem processes, including SOC change and N2O emissions. Each model uniquely simulates and mathematically represents these processes (Table S2) and has been shown to appropriately represent biogeochemical cycles under well-validated field conditions representative of cropping systems and soils of the US Midwest cropland. All models operate with daily weather inputs and can simulate carbon and nitrogen cycling, SOC change, nitrous oxide emissions, water balance, and plant growth in response to different crop management practices including fertilization, tillage, and irrigation. A brief description of each model is provided in the supplemental material (Sect. 1).

Scaling the multi-model ensemble

The MME was scaled to cover the US Midwest cropland over a total of ~ 46 million hectares that includes large areas in the 12 states of Illinois, Indiana, Iowa, Kansas, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, and Wisconsin. Simulations were performed over 42 years (1980–2022) using the gridMET weather observations data (see ‘Data inputs’ section), with 1980 to 1990 used for setting initial conditions through a spin-up, and the remainder for data analysis.

Management scenarios and dynamic baselines

Eight management scenarios were simulated to evaluate SOC dynamics, four (1–4) with full nitrogen (FN, Table 1) fertilization and four (5–8) with reduced N fertilization (RN, Table S3). These scenarios constitute business-as-usual (BAU) and regenerative agricultural practices and crop rotations across the US Midwest, reflecting those most frequently studied in the literature at test site locations (see 'Experimental sites with observed data’ section). Crop rotations include maize and soybean with or without a cover crop (Table S4) and with either conventional tillage or continuous no-till (Table S5). These scenarios allow comparisons of SOC change and N2O emissions due to different management practices over the long-term at specific locations and at multiple scales using a dynamic baseline, a time and space dependent reference that accounts for natural and management-driven variations in SOC levels.

Table 1 Scenarios simulated using a multi-model ensemble (MME) approach, representing different combinations of management practices involving tillage and cover cropping.
Full size table

Data inputs

Simulations were conducted at 1/24th degree spatial resolution as determined by gridMET weather observations40. Each grid (approximately 4 km2) represents an individual simulation-based unit (with a geolocated centroid) labeled with a unique identifier (UID). Grids were extracted for the 12 Midwest states involved in the analysis. The cropland area (maize or soybean grown for at least one year between 2008 to 2022) within each grid was calculated, as provided (30 m resolution) by the USDA National Agricultural Statistics Service Cropland Data Layer41. The percentage of cropland area per grid was calculated by dividing the cropland area by the total cropland area summed across the 12 Midwest states. Then with grids ordered from smallest to largest percent cropland, the percentages were summed to calculate the cumulative percent of cropland area. We selected UIDs that when combined covered 90% of the cropland area (Fig. S1). Selected UIDs were then categorized into discrete latitude bins to enable more straightforward identification and homogenous extraction of the final number of UIDs to be run. Within each bin, the main soil textures were identified, and their total area (km2) determined. Forty thousand total UIDs (balancing the computational power required and area representability) were then chosen across the region using a weighted random selection process, where the number of grids selected within each latitude and bin-texture combination was proportional to the area of cropland within the bin and the area covered by each texture. This selection process gave a higher probability for grids with a larger crop area to be selected.

For each UID, daily variables of maximum temperature, minimum temperature, precipitation accumulation, downward surface shortwave radiation, wind velocity, maximum relative humidity, and minimum relative humidity were downloaded from gridMET for 1980 to 2022.

Soil data for model inputs were taken from the USDA Natural Resources Conservation Service Gridded Soil Survey Geographic (gSSURGO)42 at 30 m spatial resolution. Variables extracted included bulk density, sand/silt/clay content, stone content, organic matter, and pH for each layer in the soil profile, using the most representative component percentage from the gSSURGO database. From these data, the following variables were calculated: organic carbon, total nitrogen, drained upper limit and lower limit43, and saturated hydraulic conductivity44. The dominant soil type within each weather grid was selected to represent that specific weather grid, excluding missing or shallower than 30 cm soil profiles. For each CSM, the SOC pools were initialized to match the mass of the initial SOC stock, with each CSM default value adopted to set their percentage distribution following the45 procedure.

Annual planting and harvesting dates for maize and soybean for 1980–2022 were estimated from USDA NASS state-level weekly progress reports, spatially interpolated with county-level annual crop production area. Daily progress of planting or harvesting was linearly interpolated between the weekly progress reports to find when 50% of the state was planted or harvested. State centroids were computed, weighted by county-level crop production. The planting or harvesting dates were spatially interpolated between the weighted state centroids using a generalized additive model (GAM). The model is defined as:

$$PD=s\left(Lat,Lon\right)+ti\left(Lat*year\right)+ti(Lon*year)$$
(1)

where PD is the predicted date, s(Lat, Lon) is the two-dimensional smoothing function with latitude and longitude as covariates, and ti(Lat*year) and ti(Lat*year) are tensor functions of the random effect of year and latitude and longitude, respectively. The model was applied to the weather grid centroids for each year, predicting planting and harvesting dates for each weather grid. Missing annual county-level production data was filled using the average across all available years. When all data was missing (e.g., soybean progress reports for 1979), the long-term average centroid and progress values were used in the model.

Nitrogen (N) fertilizer application amounts for maize were based on the USDA ARMS survey as reported in46. The average N rate was used to represent each state, except for Wisconsin where the maximum reported N rate was used to account for the application of manure not included in the reported range of N rates, as indicated by the relatively low range.For Kansas and Nebraska, where data were not available, the same rate as a nearby state (Missouri and Iowa, respectively) was applied (Table S6).

Crop maturity regions identification and cultivars parametrization

Crop maturity regions (MR) were identified to better represent crop cultivar extent across the US Midwest. These regions were based on six latitude bins each for maize and soybean, with each bin determined by the crop growing degree days (GDD) accumulated at maturity. Initially, the GDD data for maize for each UID were derived from47 and for soybean, computed from weather data (according to NASS planting and harvest dates). Then, the distinctions between each MR for each crop were computed by calculating the quantile-based breaks for the desired number of bins (6). These breaks were then used to categorize all UIDs into a distinct maturity region group (Fig. S2). The final GDD value for each MR was computed as the median of each group (Table S7).

For each CSM, crop cultivars were then parametrized using the GDD value according to the MR. In addition, to ensure consistency across CSM, harvest index (HI) and crop base temperature were set to standard values. Defaults for HI were initially set to 0.5 and 0.4 for maize and soybean, respectively (some CSMs then recalculated these HI values throughout the simulation period), and for crop base temperature were 8 °C and 10 °C for maize and soybean, respectively. Thus, this rationale was intended to set common main crop parameters across models to simulate crop yield (Figs. S3 and S4).

Multi-model ensemble computation

Daily simulation using the eight CSMs over 42 years across 40,000 UIDs in the US Midwest cropland generated a large dataset of more than 4 billion data values per output metric. For each model and UID, the output data was aggregated first by year and then grouped into a unique dataset. Then, for each management scenario, the yearly data were grouped by county. Outliers were detected using the interquartile range (IQR) method. This function calculates the range between the first (25th percentile) and third (75th percentile) quartile and defines outliers as data points that fall below the first quartile minus 1.5 times the IQR or above the third quartile plus 1.5 times the IQR. Once outliers were removed, the MME median (e-median) was used as the estimator of the ensemble model simulations.

When aggregating data by county, a weighted mean was used to account for the varying cropland area extents associated with each UID within the county.

$$Weighted\ mean= \frac{{\sum }_{i=1}^{n}{w}_{i}*{x}_{i}}{{\sum }_{i=1}^{n}{w}_{i}}$$
(2)

where \({x}_{i}\)​ is the value of the ith UID value and​ \({w}_{i}\) is the weight associated with the ith observation. Each weight is assigned based on the cropland area of each UID (i.e., grid). Equal weight was adopted across models. Except for Figs. 1 and 5, where single UID means are used, scenario outputs are reported throughout the manuscript as mean and standard deviation at the county level. The delta SOC change for the MME was calculated as the mean of each annual SOC change.

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Representation of the dynamic baseline concept and the mean MME annual SOC change and annual N2O emissions difference between scenarios. (A) The counterfactual scenario dynamic baselines (red), as well as the regenerative practice lines (green), account for yearly climate variability and site-specific soil properties, while the static baseline reflects the assumption of steady-state conditions. Three different generic cases are provided as examples of potential errors if changes are compared to static baseline and not the dynamic baseline. (B) MME annual soil organic carbon change (Mg ha-1 yr-1, 0–30 cm) and (C) annual N2O emissions (kg ha-1 yr-1) as the difference between Scenario 4 (NT FN CC, no-till with full nitrogen fertilization and cover crops) and Scenario 1 (CT FN, conventional till with full nitrogen fertilization). Each UID included in the study (n = 40,000) is reported as a single-colored dot. The gray density histograms represent the distribution of UIDs values across the US Midwest cropland Maps were created using RStudio v2025.05.0.496 https://www.r-project.org.

The global warming potential (GWP), reported as climate mitigation was quantified following the equation:

$$Climate\ mitigation \left[Mg\ {CO}_{2}-eq\ {ha}^{-1}{yr}^{-1}\right]= {GHG}_{{N}_{2}O}-{GHG}_{\Delta SOC}$$
(3)

where GHGN2O represents the C2O equivalent coming from N2O emission48, and GHGΔSOC the C2O equivalent coming from the carbon sequestered or lost in the soil.

R software was used for statistical analysis and graphical visualization.

Multi-model ensemble validation

To test the MME performance in reproducing SOC change, we evaluated its accuracy and uncertainty in comparison with every individual CSMs. The MME and each CSM were evaluated comparing the simulated SOC stock and annual SOC change values for different management practices (i.e., experimental treatments) with observed (measured) values from 17 experimental sites (see 'Experimental sites with observed data’ section ). We tested MME and CSMs without performing a specific calibration for the soil-related parameters. Our goal was to demonstrate that the MME is suitable for upscaling over large areas where site specific soil parameter calibration is not feasible. Therefore, we rely primarily on the crop yield parametrization described in the section 'Crop maturity regions identification and cultivars parametrization’ for the upscaling, and on the following section for the experimental sites. Lastly, MME uncertainties were also evaluated on the upscaled simulations across the US Midwest.

Experimental sites with observed data

To ensure consistent, high quality observed data sets were used for MME validation, experimental sites were selected from the literature when five data conditions were present: [1] initial soil conditions (prior to experimental treatments), including SOC stock, [2] multiple years of crop yield data, [3] a SOC time series (with either SOC stock values or BD and C% available for SOC stock calculation), [4] detailed management descriptions and ancillary information, and [5] the presence of maize or soybean, grown either continuously, or in a two-year rotation, or within an extended rotation with other crops (e.g., wheat and alfalfa). Sites were also intentionally selected to cover a wide range soil types and management, including conventional tillage, reduced tillage, no-till, different fertilization rates, and the presence of cover crops. Sampling depth up to 30 cm was preferred, but in 10 locations the depth of sampling was less (i.e., between 10 and 25 cm). In total, 17 experimental locations, with 52 site treatments, were located in North America (16 in the USA and 1 in Canada) (Fig. S5).

To simulate these sites, weather data was extracted from gridMET for US sites and NASA Power49 for Canadian sites. Soil data were preferentially and primarily sourced from the publications, but if unavailable, gSSURGO data at the closest site were used. Water hydraulic soil properties were recalculated using pedotransfer functions as reported above. If not reported, initial soil water content was set equal to the drained upper limit (DUL) and initial inorganic nitrogen was set to 40 kg N ha-1 for the entire soil profile, with ammonium and nitrate ratios assigned based on each model’s default. Table S8 reports a full description of all sites and treatments.

To ensure appropriate biomass feedback from the crop residues and roots, each CSM was calibrated at each experimental site only against the observed crop yield data without involving any soil-related parameters. All the selected CSMs are designed to accurately predict crop yields, root and residue biomass (which includes total above-ground biomass minus the grain yield) because of their critical importance for modeling long-term SOC dynamics. The RRMSE (Relative Root Mean Square Error, Eq. (4)) was used to evaluate individual CSMs calibration performance (Table S9). Most of the time, total biomass of cover crop was not reported, and default models’ values were used to set its productivity. Data_S1, in the supplemental material, reports for each CSM and location, with the full set of parameters used, their default values, and the final values obtained from the calibration process.

$$RRMSE\left(\text{\%}\right)= 100*\frac{\sqrt{\frac{{\sum }_{i=1}^{n}{({P}_{i}-{O}_{i})}^{2}}{n}}}{\overline{O} }$$
(4)

where \({P}_{i}\) and \({O}_{i}\) are predicted and observed data respectively, n the number of P/O pairs and \(\overline{O }\) the mean of observed data.

Multi-model ensemble and individual model accuracy and uncertainty analysis

Before evaluating the MME and each CSM’s performance in terms of accuracy and uncertainty, we compared the measured delta SOC change values from the 52 experimental site treatments with ranges from a recent meta-analysis on regenerative practices50 to identify values that were inconsistent with the literature range, i.e., for no-till (− 0.5–1.22 Mg ha-1 yr-1), conventional tillage (− 0.47–0.5 Mg ha-1 yr-1), and reduced tillage (− 0.97–0.65 Mg ha-1 yr-1). To do so, we calculated the weighted mean delta SOC change for the measured data following the same mathematical form of Eq. (2) (where \({x}_{i}\) is the ith measured delta SOC change and \({w}_{i}\) its weight based on time elapsed since the previous SOC observation), and noted for each site*treatment whether it was consistent with the data range (Table S10). For the accuracy and uncertainty analyses described below, individual paired comparisons between simulated and observed data were included only when observed weighted mean annual delta SOC change was consistent with the literature ranges reported above. For the MME and individual CSMs instead, the weighted mean annual delta SOC change was calculated as the average of the annual SOC changes over the simulated time period, with weights applied to account for shorter durations in the initial or final years of the simulation.

To evaluate MME and CSM accuracy, a composite score was developed by combining two metrics: the Savage score51, and the weighted standard deviation (s) of model ranks. The absolute value of the prediction errors (PE, Eq. (5)), computed as the difference between predicted (i.e., simulated) and measured delta SOC change was used to rank every jth model accuracy within each ith site*treatment combination.

$$PE=\frac{CF}{0.5}{*(P}_{j(i)}-{O}_{i})$$
(5)

where Pj(i) and Oi are the predicted and measured values of the jth model and ith site*treatments combination respectively, and CF is a correction factor that considers measurement uncertainties13,52.

Each site*treatment combination ranking was then transformed into a Savage score. The total Savage score (Si) for each model, calculated as the sum of its scores across all site*treatment combinations, was used to rank model accuracy, with greater weight assigned to models demonstrating consistently high performance. Rank variability, represented by the weighted standard deviation of ranks (srank), quantified the stability of model rankings. The weight (1/rank) of the standard deviation was included to favor models that consistently fluctuate among top-ranking positions, rather than those that exhibit variability in lower-ranking positions.

Both metrics were normalized to a 0–1 scale and, for each model, the composite score (CS) was computed as:

$$CS=\left(0.5* {S}_{j}\right)+0.5*(1-{s}_{rank\left(j\right)})$$
(6)

where Sj is the normalized sum of the Savage-scores assigned to the jth model and srank(j) the weighted standard deviation of the jth model rank, while 0.5 ensures the same weight to both metrics. This approach ensured that the highest-ranked models demonstrated both superior accuracy and low variability.

To assess uncertainty, we used the mean squared error of prediction (MSEP) and its decomposition in prediction squared bias (\({bias}^{2}\)) and prediction variance (\(var\)) as reported by Maiorano et al.53 on the raw SOC stock variable. This metric was used to quantify MME uncertainty in comparison with individual CSMs. Since no calibration was performed against SOC data, measured data from experimental sites were used as the prediction dataset.

$${bias}^{2}=\frac{1}{N}{\sum }_{i=1}^{N}{\left({Y}_{i}-\frac{1}{M}{\sum }_{m=1}^{M}{\widehat{Y}}_{m, i}\right)}^{2}$$
(7)

where N is the number of treatments, \({Y}_{i}\) is the measured variable for the treatment i, \({\widehat{Y}}_{m, i}\) is the simulated variable for the treatment i by the m CSM (or MME). The predicted variance is the variance of the values simulated by the population of m models (1 when individual CSMs are tested) averaged across treatments:

$$var=\frac{1}{N}{\sum }_{i=1}^{N}{var}_{M}({\widehat{Y}}_{i})$$
(8)

where N is the number of treatments, \({\widehat{Y}}_{i}\) is the simulated variable for the treatment i. The estimate of MSEP can then be obtained as:

$$MSEP={bias}^{2}+var$$
(9)

We further evaluated how the uncertainty estimates changed with the number of models involved using the coefficient of variation of the multi-model ensemble e-median over the upscaled dataset in the US Midwest. To do so, we performed a bootstrap calculation (i.e., random sampling with replacement) for each value of M (number of models in the ensemble) from 1 to 8. For each ensemble size M we extracted 10,000 bootstrap samples of M models with replacement. This bootstrap sampling was repeated N times equal to the number of scenarios (1–4), and C times equal to the number of counties involved (934), considering single UID 30-year overall C gain/loss values within each county. The total number of samplings with replacement exceeded 35 million. The variation in the e-median across the bootstrap samples due to random model selection was computed with the coefficient of variation (CV):

$$CV({\widehat{y}}_{e-median,{M}{\prime}})=\frac{1}{N}\frac{1}{C}{\sum }_{i=1}^{N}{\sum }_{j=1}^{C}\left(100*\frac{{sd}_{B}({\widehat{y}}_{e-median,i}^{M})}{{mean}_{B}({\widehat{y}}_{e-median,i}^{M})}\right)$$
(10)

where \(CV({\widehat{y}}_{e-median,{M}{\prime}})\) is the coefficient of variation of e-median for the model ensemble of size M, \({sd}_{B}\) and \({mean}_{B}\) are the standard deviation and the mean of the B (number of bootstrap sample) e-medians of the model ensemble of size M for the ith scenario (from 1 to 4) and the jth county.

Results

Regenerative practices impact on SOC change and N2O emissions

The evaluation across ~ 46.2 million hectares of US Midwest cropland using MME results grouped by counties shows that in comparison with a dynamic baseline of conventional till without cover crops (CT), combined adoption of no-till and cover crops (NT CC; Scenario 4; Table 1) increases SOC accrual annually by 16.4 Tg C (on average 0.36 ± 0.12 Mg C ha-1 yr-1, Fig. 1B), while increasing the N2O emissions by on average 0.16 kg ha-1 yr-1 (Fig. 1C). This rate of SOC stock increase is approximately twice that of when either no-till (NT; Scenario 3) or cover crops (CT CC; Scenario 2) are practiced independently on conventional till (i.e., 8.9 Tg C or 0.19 ± 0.13 Mg C ha-1 yr-1 for NT and 8.5 Tg C or 0.18 ± 0.14 Mg C ha-1 yr-1 for CC, Fig. 2). Similarly, under no-till, the introduction of a cover crop (NT CC) increases SOC accrual by 7.5 Tg C (on average 0.16 ± 0.10 Mg C ha-1 yr-1, Fig. 2C). See figures S6 and S7 for a full set of management comparisons. When N fertilizer is reduced to 75% of the BAU rate (Scenarios 5 vs. 1, 6 vs. 2, 7 vs. 3, and 8 vs. 4), SOC stock changes are negligible, averaging −0.02 Mg C ha-1 yr-1, while averaged across the RN scenarios, N2O emissions decrease by 15% (−0.22 ± 0.45 kg ha-1 yr-1), and with cover crop adoption, increase by 11% (0.15 ± 0.51 kg ha-1 yr-1) and 12% (0.16 ± 0.45 kg ha-1 yr-1) under conventional and no-till management (Fig. S8), respectively. See figures S9 to S11 for a full set of county-level annual delta SOC and N2O emissions per scenario.

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Effect of different regenerative practices on county-based MME mean annual SOC change. (A) Effect of no-till on county annual soil organic carbon (SOC) change (Mg ha-1 yr-1, 0–30 cm) shown as the difference between Scenario 3 (NT FN, no-till with full nitrogen fertilization) and the baseline Scenario 1 (CT FN, conventional till and full nitrogen fertilization); (B) effect of cover crop on county annual soil organic carbon (SOC) change (Mg ha-1 yr-1, 0–30 cm) shown as the difference between Scenario 2 (CT FN CC, conventional till with full nitrogen fertilization and cover crop) and the baseline Scenario 1 (CT FN, conventional till and full nitrogen fertilization); (C) effect of cover crop under no-till on county annual soil organic carbon (SOC) change (Mg ha-1 yr-1, 0–30 cm) shown as the difference between Scenario 4 (NT FN CC, no-till, full nitrogen fertilization and cover crop) and the baseline Scenario 3 (NT FN, no-till with full nitrogen fertilization). Baselines serve as counterfactual benchmarks for comparison with different scenarios. Individual counties are represented by their mean weighted values, with weights accounting for the agricultural area percentage of each county’s UID.

Pedo-climatic variables and SOC change

Using MME analysis, locations with higher initial carbon stocks (see Fig. S12) typically show SOC loss (up to −0.43 Mg C ha-1 yr-1 in sandy soils under CT; Fig. 3). Conversely, soils with low initial carbon stocks typically show SOC increases under both CT and NT, particularly with cover crops (CT CC and NT CC). Larger, consistent SOC accrual rates were apparent across soil textures and initial carbon stock when no-till and cover crops (NT CC) were present together, with similar trends emerging for RN scenarios (Fig. S13). When individual CSMs were evaluated, changes in pedo-climatic variable inputs (initial SOC stock, clay content, and temperature and precipitation; Fig. 4) show no consistent impact on annual SOC change. As initial SOC stock increases, a decrease in the soil’s capacity to store more C is observed (CSMs 1, 3, 4, 5, and 6), but some show variable (CSMs 7 and 8) and no (CSM 2) response to initial SOC stocks. Increasing clay content beyond ~ 30% shows an increase (CSMs 1 and 8), no change (CSMs 2, 4, and 7), or a decrease (CSM 3) in SOC stock. Soils at higher latitude show lower annual SOC change in CT scenarios but were inconsistent under NT management.

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

MME mean annual SOC change rates across scenarios based on different soil textures and initial SOC stock levels across 46 M ha. Annual soil organic carbon (SOC) change (Mg ha-1 yr-1, 0–30 cm) across scenarios (1 to 4, from left to right), soil texture (Sandy, Loam and Clay), and initial SOC stock level (Low, < 40 Mg C ha-1; Medium, 40–80 Mg C ha-1; and High, > 80 Mg C ha-1). Annual SOC change values derived from all UIDs mean across the US Midwest cropland. All scenarios use maize-soybean crop rotation.

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

MME mean annual SOC change response to variations in initial SOC stock, clay content, and latitude across 46 M ha. Each line, represented by a unique color, corresponds to an individual model, with the black dotted line representing the MME median. They represent individual model annualSOC change (Mg ha-1 yr-1, 0–15 cm) response to variations in initial SOC stock (Mg ha-1), clay content (%), and latitude (°). The different scenarios (1 to 4 from top to bottom) are displayed at the right end of each row. Scenario acronyms: CT (conventional till), NT (no-till), FN (full nitrogen), CC (cover crop).

Climate mitigation potential

The climate mitigation potential shown in Fig. 5 reflects the extent to which SOC changes can offset N₂O emissions, based on the comparison between scenario 4 and scenario 1. In this comparison, no-till combined with cover crop (scenario 4) resulted in an average climate mitigation potential of 1.2 Mg CO₂-eq ha⁻1 yr⁻1 across the Midwest. This regenerative practice emits up to 2.5 Mg CO₂-eq ha⁻1 yr⁻1 less than its conventional tillage counterpart, although the magnitude of this benefit varies by geographic location (Fig. 5).

Fig. 5
Fig. 5The alternative text for this image may have been generated using AI.
Full size image

Climate mitigation (Mg CO2-eq ha-1 yr-1) of Scenario 4 (NT FN CC, no-till with full nitrogen fertilization and cover crops) compared to Scenario 1 (CT FN, conventional till with full nitrogen fertilization). Climate mitigation accounts for N2O emission and SOC change as reported in Eq. (3). Each UID included in the study (n = 40,000) is reported as a single-colored dot. The gray density histograms represent the distribution of UIDs values across the US Midwest cropland.

When each scenario is evaluated individually for its ability to offset N₂O emissions (Fig. S14), none consistently achieve full offset. Instead, the climate mitigation effect is highly dependent on site-specific pedo-climatic conditions. The mitigation potential tends to increase along a management gradient—from conventional tillage without cover crops to no-till with cover crops—highlighting the cumulative benefits of stacking regenerative practices.

Discussion

Standardizing SOC quantification

Adopting regenerative management practices can reduce the carbon footprint of row-crop agriculture, but to what degree and whether it is meaningful as part of climate mitigation efforts is debated54. A major issue in determining impact is the lack of standardization in quantifying SOC stock change55; estimates for individual fields based on re-inventorying, even after extended periods of time and at high sampling densities, can be variable and inaccurate, with gains or losses in SOC stock often determined even when no change has occurred26. These challenges are consistent with our findings from comparisons of SOC changes from individual long-term experimental treatments (Table S8) with meta-analysis ranges50, where 38% (20 out of 52) of the peer-reviewed data fell outside the range, some showing ‘unrealistic’ rates of change (e.g., 8 Mg C ha-1 yr-1; Table S10). The likelihood for spurious values is increased when SOC change is determined using consecutive short-term sampling (e.g., 4–5 years apart), typical with current C market approaches for paying producers, where contract lengths of only one year are common1. Short-term contracts that eschew consideration of long-term SOC trajectories, risk exposure to measurements that are strongly impacted by seasonal trends (e.g., 'saw-tooth’ dynamics). While acknowledging high quality SOC field measurements are needed to better understand SOC dynamics and help to initialize models, their use in determining SOC stock change is problematic for short-duration payment requirements, and effectively unfeasible for larger-scale projects.

Upscaling regenerative practices

By using high spatial resolution data aggregated to the US Midwest cropland, MME SOC change values are applicable and consistent with observations at multiple scales. As a management option, combining cover crops and no-till shows the greatest promise (~ 16.4 Tg C yr-1) averaged over the total area investigated. With cover crop introduction, the MME SOC increase of 0.17 Mg C ha-1 yr-1 (~ 8.0 Tg C yr-1) across tillage types, is consistent with the meta-analyses of Peng et al.56 and McClelland et al.57 who found increases of 0.24 Mg C ha-1 yr-1 and 0.21 Mg C ha-1 yr-1, respectively. Over similar total areas of US cropland, potential SOC accrual suggests that MME provides a reasonable balanced estimate between non-conservative values that use IPCC factors (58, 13.3 and 17.7 Tg C yr-1 for no-till and cover crop adoption, respectively), and conservative evaluation that uses satellite imagery and longitudinal surveys (59, 5.3 Tg C yr-1 for cover crop adoption).

Even without site-specific validation, cumulative annual N₂O emissions estimated by the MME align with values reported in the literature or with IPCC-based estimates (see Table S11). For example, cover crop adoption increases N2O emissions by on average ~ 0.15 kg ha-1 yr-1 across tillage managements (Fig. S8), an increase consistent with the variable results reported by Basche et al.60. Prior multi-model studies have found significant uncertainty and divergence of N2O emissions across models and have adopted different approaches that omit agricultural management and crop growth61. Although we make no claim, an MME approach that includes these factors may be well suited to ‘filter’ out data outliers, often reported in single-model studies due to the challenges of accurate simulation, especially on daily timescales.

Estimating N₂O emissions and changes in SOC is essential to fully evaluate the global warming potential (GWP, expressed as Mg CO2-eq ha-1 yr-1) of individual agricultural practices31. Balancing SOC sequestration against N₂O emissions provides a more accurate assessment of the climate mitigation potential, specifically by determining whether increases in SOC stock, when they occur, are sufficient to offset associated N₂O emissions completely12.

When combined with a dynamic baseline, this approach further clarifies the actual climate mitigation potential of a practice by explicitly comparing it to a realistic counterfactual scenario. Figure 5 illustrates the climate mitigation potential of Scenario 4 (NT FN CC: no-till with full nitrogen fertilization and cover crops) relative to Scenario 1 (CT FN: conventional tillage with full nitrogen fertilization). Even when accounting for N₂O emissions, adopting no-till plus cover cropping results in an average net saving of 1.2 Mg CO₂-eq ha⁻1 yr⁻1 compared to the conventional tillage baseline. Although direct comparison with the baseline remains essential, a separate assessment of Scenario 4 itself (Fig. S14) reveals that SOC accrual alone is not always sufficient to fully offset N₂O emissions. Nevertheless, Scenario 4 still provides, on average, a modest negative balance (computed as in equation 3) of approximately 30 kg CO₂-eq ha⁻1 retained annually.

A potential improvement to Scenario 4 involves reducing the nitrogen fertilizer application rate by 25%, as explored in Scenario 8 (Table S3). This comparison becomes extremely useful when evaluating whether potential yield reductions (due to decreased N supply) are justified by the corresponding reductions in N₂O emissions. In other words, the analysis helps clarify whether the expected reduction in N₂O emissions would sufficiently compensate for the likely decrease in SOC accrual due to reduced crop residues returned to the soil. Results show that the additional mitigation potential is minimal (~ 10 kg CO₂-eq ha⁻1 yr⁻1) when comparing Scenario 8 to Scenario 4. Furthermore, average maize and soybean yields under Scenario 8 are lower by approximately 150 kg ha⁻1 (± 390 kg ha⁻1) compared to yields obtained under optimal fertilization (Scenario 4). Thus, identifying the best overall management practice requires an economic justification that balances yield response against fertilizer cost savings. A straightforward economic analysis could clarify whether cost savings from reduced nitrogen fertilizer inputs adequately compensate for potential yield reductions, at least until farmers receive direct economic incentives for additional CO₂ abatement efforts. Whether industry initiatives or emerging carbon markets will support such incentives remains to be determined and lies beyond the scope of this analysis.

In addition to the broad-scale assessment across the entire Midwest region, we believe further analysis conducted within homogeneous pedo-climatic zones (such as Major Land Resource Areas, MLRAs) would provide even greater value. Although beyond the immediate objectives of the present study, such localized analyses could significantly influence agricultural decision-making. Instead of solely pursuing maximum yield production, management practices could better reflect realistic land-use potential, optimizing the trade-offs between achievable crop yields and associated greenhouse gas emissions.

Multi-model ensemble performance

When tested against long-term experimental trials across the US Midwest, characterized by substantial differences in soil types and agricultural management, the MME approach provides higher accuracy, and lower uncertainty in estimating SOC change (Table 2). Its performance reduces the biases associated with short-term sampling and the use of an individual CSM. In comparison with individual CSMs, MME best simulates changes in SOC across the experimental sites and treatments, with the highest overall composite score of 0.948 (Table 2). The MME Savage score of 39.3, indicates that it frequently ranks among the top models in terms of accuracy (i.e., lowest absolute prediction error) and has the lowest weighted standard deviation of the absolute prediction error rank (2.07; Table 2). Uncertainty assessment using MSEP and its decomposition into squared bias and variance components, shows MME consistently exhibits the lowest uncertainty, evident when compared with the full CSM population (Fig. S15) and with individual CSMs (Fig. S16). Thus, an estimate from an individual CSMs that is ‘best’ at one field may not be so at another, often for reasons that cannot readily be determined; individual CSMs have varying constraints and needs and are better suited for specific contexts14.

Table 2 Comparison of the performance of individual CSMs and the MME across measured experimental sites and treatments.
Full size table

When upscaled to the US Midwest, the coefficient of variation for the delta SOC change e-median decreases as the number of models increase, from 99% with one CSM to 36% with eight CSMs (Fig. S17). The notable disagreement among individual CSMs at large scale for the same scenarios (Fig. S18–S22), underscores the quantitative lottery of individual CSMs selection for a particular application. Thus, MME provides a more consistent and reliable output of long-term SOC change across multiple scales in agreement with prior studies13,20.

Dynamic baselines and flexible scales

Inconsistent and contradictory outputs highlight the ongoing risk of using a single CSM to provide the robust accounting needed to support agricultural carbon market projects across broad regions. A large cause of this incongruency is the known and somewhat intractable challenge of determining additionality through the choice of a suitable BAU baseline: the counterfactual scenario that would have occurred without management intervention. This choice can have a significant influence on the degree of SOC change, and, by extension, on how many carbon credits are generated; if inaccurate, too many or too few may be issued, calling into question legitimacy30. Indeed, only carbon that is additionally stored in soils or that is additional compared to a BAU scenario can be relevant for climate change mitigation62.

Usually, projected baselines are a ‘best guess’ manufactured scenario that considers steady-state historical conditions for a specific area (Fig. 1A) without accounting for within season climate variability and site-specific soil properties. Attributing SOC change to management using a static baseline is problematic because it is prone to errors if SOC changes in the counterfactual BAU scenario are not quantified. The use of process-based models avoids this issue by using dynamic baselines that are temporally and spatially adjusted to capture the impacts of fluctuating environmental variables. This allows quantification of the yearly SOC change difference between an adopted regenerative management practice and its dynamic counterfactual baseline to provide a more credible estimate of carbon gained or lost. Dynamic baseline scenarios were determined for every UID to allow for direct comparisons at the same location. This is essential for evaluating practice-level impacts on SOC accrual under the same pedo-climatic conditions, as a direct comparison with a business-as-usual (BAU) scenario is often unavailable or impractical on a field basis. For instance, when a farmer transitions from conventional tillage to no-till and measures SOC levels at the time of conversion, comparing SOC changes under no-till after 5 or 10 years against the initial SOC value (when conventional tillage ended) fails to account for what SOC levels would have been had conventional tillage continued on that field. Only through modeling counterfactual scenarios via dynamic baselines can realistic SOC differences between treatments be determined. These clarifications emphasize the importance of dynamic baselines and improve the overall clarity of our methodology and findings. MMEs can explore large pedo-climatic variability with a regionally consistent framework. Units of scale are flexible – they can be an agroecological zone with similar soils, weather, and agricultural practices, such as USDA Major Land Resource Areas63 or a political jurisdiction such as county or state64. Irrespective, estimates of SOC change can be quantified relative to a dynamic baseline for the scale chosen, and the results reported in readily accessible lookup tables, with the MMEs potential to generate and retrieve this information a valuable service, when reference values are needed for carbon credit payments or a regional policy document. As an example, the choice of county scale as the baseline jurisdictional unit allows their relatively uniform distribution across the US Midwest to capture soil and climate variability, and their administrative structure to provide a practical, credible framework for investment and implementation of regenerative practices for GHG abatement.

Opportunities and limitations for multi-model ensemble application

As part of its effort to reduce GHG emissions, the US has promoted programs to enhance soil carbon sinks65. The expectation is that emissions reductions will occur through the adoption of regenerative practices on cropland66, with market-based interventions best positioned to achieve this67. Voluntary offset and inset initiatives predominate in the US agricultural sector; however, carbon credits generated by agriculture constitute only ~ 1% of the total volume transacted68. Despite very high rates of producer carbon market awareness, participation rates are very low69, in large part a reflection of the barriers that currently hinder engagement, primarily high enrollment costs driven by quantification, recordkeeping, and data analysis, scale and aggregation limitations, and verification, which are insufficiently compensated for by C change payments67,70. Cultural and social factors can also deter participation, including confusion about the assortment of available market programs1,17,71.

Use of an MME accounting tool could alleviate many producer participation barriers, ease onerous MMRV requirements, provide confidence in SOC stock change quantification at different scales, and in doing so address several strategic US policy priorities72. At the initial stage, producers with eligible fields would not be required to provide detailed agronomic data to participate, thereby minimizing record-keeping burdens, auditing requirements, and potential privacy concerns. However, when such data are available, they can be incorporated to enhance the accuracy and performance of the MME by improving the quality of model inputs.

In effect, an MME approach provides multi-scale, pre-run, practice-based, dynamic baselines for determining emission factors with low uncertainty levels. MMEs may be a practical, intermediate approach for payment for practice and payment for output mechanisms. As a decision support tool, MMEs can also allow producers or project developers to quickly screen lands targeted for regenerative practices to determine their economic and program viability67.

Of the few registered agricultural land management (ALM) projects, most have generated low density (per hectare) and low volume (i.e., small area) credit numbers when compared to other sector project types68,73. Unless aggregation (combining individual fields or small projects) to form large scale projects is allowed, producers, particularly those with limited resources or opportunity, find cost barriers insurmountable. While aggregation in varied design is allowed in current agricultural protocols and results in verification cost reduction, it does not reduce quantification costs, which scale with the area enrolled.

Similarly, approaches and accounting rules have been recommended to address issues of permanence, leakage, and additionality, inherent with small-scale, individual ALM projects73. However, aggregated projects where management interventions are accounted for at larger scales reduce these concerns74. With sufficiently large spatial and temporal scales, additionality can be demonstrated through reduced emissions below recent regional baseline trends. Larger project area coverage and longer project times can alleviate practice reversal risks, unintentional or otherwise, by pooling across broader scales and stakeholders. Large programs can also mitigate leakage risks by addressing underlying economic drivers of emissions and can reduce jurisdictional risk by directly accounting for all emissions shifts inside the administrative unit64. Measuring BAU baselines on a project-by-project basis is not economically feasible as emissions reductions are too small relative to costs. Technology-adjusted historical baselines designed in accordance with regional dynamics seem a superior option75, and one that MMEs are well-suited to provide.

The present study comes with some limitations and possibilities for future work. Many process-based models simulate SOC dynamics across defined depth layers, yet the representation of vertical SOC redistribution—particularly under no-till systems—remains a challenge. Most models in this ensemble simulated the full soil profile, but one model was limited to 20 cm. While topsoil (0–30 cm), especially under no-till management, accounts for ~ 80% of the total management effects76, it may miss subsoil contributions.

We also recognize that in many models, the effect of tillage on SOC is treated empirically, often modifying decomposition rates without explicitly simulating changes in soil physical properties. Nevertheless, especially in no-till, SOC stock stratification is mainly due to crop rootsatification, something that all the models fully represent. Bioturbation, which promotes SOM redistribution due to biotic activity, but most of the models do not include it yet.

Overall, the MME approach helps mitigate individual model biases, but its reliability still depends on the quality of input data. While large-scale projects can help harmonize datasets, local-scale agronomic details (e.g., tillage intensity, residue management, fertilization methods) remain crucial for accurate single CSM testing and calibration.

Despite the uncertainties, our study represents an important step forward, as it is the first to examine prospective responses on practice adoption at such scale and resolution.

Conclusions

MME results generate low uncertainty estimates of climate mitigation for common management practices that can be used in C market accounting at multiple scales. This removes the exposure of market actors to accusations of model selection and reduces the uncertainty of practice change outcomes. With MMEs, the fine granularity and consistent framework of input data across a large area allow for a flexible choice of scale for model output aggregation. This and the availability of practice-based dynamic baselines, along with decision support capabilities, can allow MMEs to reduce barriers to carbon market participation, and ameliorate long-standing concerns related to additionality, leakage, and permanence. In doing so, MME addresses several policy priorities and has a potential role in enhancing adoption of regenerative agriculture at scale.