https://orcid.org/0000-0002-0474-2760
Figures
Abstract
The exchange of carbon between the soil and the atmosphere is an important factor in climate change. Soil organic carbon (SOC) storage is sensitive to land management, soil properties, and climatic conditions, and these data serve as key inputs to computer models projecting SOC change. Farmland has been identified as a sink for atmospheric carbon, and we have previously estimated the potential for SOC sequestration in agricultural soils in Vermont, USA using the Rothamsted Carbon Model. However, fine spatial-scale (high granularity) input data are not always available, which can limit the skill of SOC projections. For example, climate projections are often only available at scales of 10s to 100s of km2. To overcome this, we use a climate projection dataset downscaled to <1 km2 (∼18,000 cells). We compare SOC from runs forced by high granularity input data to runs forced by aggregated data averaged over the 11,690 km2 study region. We spin up and run the model individually for each cell in the fine-scale runs and for the region in the aggregated runs factorially over three agricultural land uses and four Global Climate Models. We find that the aggregated runs systematically over-predict SOC compared to the fine-scale runs, with a mean difference of 7.2 tonnes C per hectare and a mean absolute error of 9.1% at the end of the 77 year simulation (2022–2099). We find large spatial variance in SOC across cells, reflecting variability in climate as well as other environmental drivers. We conclude that future research should focus on developing more high-granularity input datasets for SOC modeling, and we also reflect on the significant computational resources required to conduct fine-scale simulations.
Citation: Wiltshire S, Clemins PJ, Beckage B (2024) Granularity of model input data impacts estimates of carbon storage in soils. PLOS Clim 3(10): e0000363. https://doi.org/10.1371/journal.pclm.0000363
Editor: Ji Chen, Institute of Earth Environment, Chinese Academy of Sciences, CHINA
Received: November 15, 2023; Accepted: July 17, 2024; Published: October 24, 2024
Copyright: © 2024 Wiltshire et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The code used in this study is in a public GitHub repository at https://github.com/brianbeckage/SoilCarbonGrid. Input and simulation output data are in a public Zenodo repository at https://doi.org/10.5281/zenodo.11261491.
Funding: Support for this project was provided by the University of Vermont ARS Food Systems Research Center (SW, BB). In addition, Brian Beckage was supported in part by NASA grant 80NSSC20M0122 and by USDA National Institute of Food and Agriculture Hatch project 1025208 (BB). Patrick J. Clemins was supported by the National Science Foundation under grant OIA-1556770 (PC). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
The capture and storage of atmospheric carbon in agricultural soils is a strategy for mitigating anthropogenic climate change [1–3]. Estimates of the potential for agricultural soils to store carbon as soil organic carbon (SOC) vary widely, due to differences in modeling methodology and parameterization [4–9]. One potential source of uncertainty in SOC projections is the spatial granularity of the input data that force the model. For example, carbon storage in soils is sensitive to localized surface temperature and precipitation [10–13]; but climate projections are usually at the scale of tens to hundreds of square kilometers, aggregating and smoothing across fine-scale variations.
We examine the effect of input data granularity on projected SOC storage in agricultural soils in Vermont, USA using the Rothamsted Carbon Model (RothC) [14]. RothC is scale-agnostic: it computes SOC stocks on a per-unit-area basis using a set of environmental and land management drivers. The spatial granularity of input data determines the spatial scale at which SOC (in metric tonnes per Hectare [t/Ha]) is reported. Therefore, a model run can occur at any scale for which input data are available, from a single field to an entire region.
Projecting SOC sequestration at the landscape scale can be accomplished by extrapolating results from a set of small sites, although this assumes the selected sites are representative of the larger area [15]. Another approach is to divide a large study area into relatively-homogeneous sub-units with their own climatic and environmental data, and run the model for each area [6–8]. Studies that explore the impact of data extrapolation and spatial granularity on ecological systems modeling have found effects on the order of 10 to 20% error in estimated quantities [16–19].
While others have examined the effects of data aggregation on model projections in various agro-ecological contexts, to date there has not been a specific focus on input data aggregation for SOC modeling. This study investigates the extent to which aggregation of input data can affect the projections of RothC, all else being equal. In the high-granularity runs, we run RothC on grid cells of ∼0.65 km2, corresponding to the spatial resolution of the climate input dataset [20]. In the low-granularity runs, input data are averaged across the 11,690 km2 study region. We run the model for three common agricultural uses (crops, hay, and pasture) with climate projections from four different Global Climate Models (GCMs). We evaluate the difference in estimated SOC that results from data aggregation to the landscape scale and also examine the variability of SOC stocks across individual cells within the study region.
Materials and methods
RothC model
RothC is a widely-used soil process model that can produce reliable SOC projections in non-waterlogged soils [6–9, 14, 15, 21–23]. We use a version of RothC ported to R [24]. RothC divides SOC into five pools based on mean residence time, which statistically correlate to measured SOC fractions in topsoil samples [25]. An advantage of RothC is its relatively minimal set of input parameters, focusing on climate, soil, and land management factors. Fig 1 shows the flow of SOC between RothC’s pools, as well as its required inputs. The code used in this study is in a public GitHub repository [26]; input and simulation output data are available by request from the authors.
This figure depicts the required input data and structure of SOC pools in the model. Adapted from the RothC manual [14]. Previously printed in [7].
Limitations of RothC include a focus only on SOC, and not above-ground C; an assumption that all water infiltrates [27]; no model of direct tillage or short-term priming effects [28]; and a soil model based solely on clay percent [14]. We use RothC’s default input distributions of 59% DPM, 41% RPM for plant material and 49% DPM, 49% RPM, 2% HUM for manure; and a topsoil depth of 30cm.
Climate data
Temperature and precipitation in this study are derived from a dataset with daily timesteps between 1950 and 2099, downscaled to ∼0.65 km2 cells within a region encompassing much of Vermont, USA [20]. The data leverage the Climate Multimodal Intercomparison Project’s CMIP5 ensemble, forced by Representative Concentration Pathway (RCP) 4.5, an intermediate scenario of future climate change. Winter et al. [20] statistically downscaled the 1/8° bias-correction with constructed analogs (BCCA) data using landscape topology [29], then applied additional bias-correction to the resulting cell data using historical NOAA Global Historical Climatology Network (GHCN) daily station observations [30]. We use temperature and precipitation projections from four GCMs to bound the range of likely regional climate change. The four GCMs capture variability in regional projections across climate models. GCM ccsm4 projects cooler and drier regional conditions, miroc-esm is warmer and drier, noresm1-m is warmer and wetter, and mri-cgcm3 is cooler and wetter. Fig 2 shows the projected average temperature and precipitation from this dataset for the study area.
Average annual temperature and precipitation across the four GCMs, using a ten-year rolling mean to better visualize the overall trend. Gray area is the standard deviation between GCMs, and red vertical line is the beginning of the simulation (January 2022).
Several steps are required to process the GCM data for use in our cell-based RothC model. First we convert daily timesteps to the monthly timesteps used by RothC. Since we are focused on agricultural soils, cells are filtered using NLCD data so that only cells within the study area containing crops, hay, or pasture are included [31]. This yields 17,991 cells with a total area of 11,690 km2, of which 2,793 km2 are in agricultural use (379 km2 crops, 1,238 km2 hay, and 1,176 km2 pasture).
Other input data
The remaining data procurement and preparation methods are similar to those used in our previous RothC studies [7–9]. Land use data are from the 2016 National Land Cover Database [31] and 2017 USDA Census of Agriculture [32]. Evapotranspiration is from NASA GLDAS remote sensing [33]. We initially attempted to use the gSSURGO database for soil data [34], but when we extracted data at high granularity, we found that it contains voids where no data exist for some cells. We overcame this using the POLARIS dataset, which downscales gSSURGO to a 30m resolution [35].
Monthly SOC inputs and soil cover state for each agricultural use (crops, hay, and pasture) were generated in consultation with USDA Extension experts. For simplicity, we assume just one management style for each agricultural use, based on the typical method used in Vermont. The derivation and values used for these land management inputs are given in [8].
Empirical SOM measurements for each agricultural land use and location are from an internal UVM Soil Laboratory dataset. These data are converted to SOC% using the Van Bemmelen method [36], and finally to t/Ha SOC using bulk density from the POLARIS data [35].
Spinup routine
A spinup is necessary so that each cell begins in an equilibrium state that corresponds to environmental and management conditions. The spinup sets the assumed monthly below-ground plant C input levels such that, when those inputs, together with known manure and plant residue inputs, are continued over an arbitrarily-long timeframe, total SOC eventually converges on an empirical real-world target (here, from the UVM Soil Lab dataset). The spinup routine uses an optimization procedure to iteratively adjust assumed below-ground C inputs until the target is met. This type of iterative spinup approach is considered best practice for soil process modeling [37], and has been used widely in similar studies [7, 8, 22, 38]. Throughout the spinup period, we use 30-year normals (monthly averages between 1992–2022) for precipitation and temperature for each cell [20]. Climate and all other input data do not change year-on-year over the 750-year spinup duration, allowing the model to reach equilibrium. The inert organic matter (IOM) fraction, which remains static throughout a RothC simulation, is calculated using the Falloon method [39].
Simulations
Our simulations project how each cell’s SOC stocks change over time in response to climate change while the same land management is continued. The simulation runs from 2022 to 2099. The fine-scaled simulations are forced by unique soil properties and monthly temperature and precipitation projections for each cell. While soil properties vary between cells, they remain static over the simulation period. The monthly below-ground plant C inputs, which are calculated during the model spinup, are also unique for each cell, reflecting the variability in empirical SOC measurements across the study region. We conduct separate runs for crops, hay, and pasture for each cell that contains those uses. Land management parameters corresponding to the three agricultural uses are held constant across cells and across time. RothC returns data in the form of SOC (t/Ha) per month in each of its five pools over the simulation period. We calculate total SOC as a sum of these five pools.
We also conduct a series of aggregated simulation runs at the landscape scale. For these runs, climate and soil data from all cells containing the target agricultural use are averaged, and these mean values are used to drive the regional model simulations. The climate data still change over time according to the climate projections described above, but reflect monthly regional averages rather than projections for individual cells. The model is spun up using aggregated data, and then simulation runs for each agricultural use and GCM are conducted.
The fine-scale simulations are computationally intensive due to the high number of model runs. For each of the three land uses, we run the model in each cell containing that land use four times (once for each GCM), yielding 71,964 total runs. To accomplish this, we obtained time on the Cheyenne supercomputer at the National Center for Atmospheric Research (NCAR) [40]. Using the supercomputer’s parallel processing capabilities greatly reduces the computational time for the entire set of simulations.
Analysis
We compare projections of SOC from the fine-scale runs to the aggregated runs and also examine the variability across cells. For each GCM and agricultural use, we report the total SOC in the final year of the simulation for the lowest and highest individual cell, the mean across all cells, and the results for the aggregated runs. We also show the distribution of SOC across cells in each year of the simulation using timeseries density plots, and in the final year using histograms. Finally, we examine the change in SOC storage between the initial and final year.
Results
The projected SOC storage at the end of the 77 year simulation (2022–2099) was consistently overestimated across GCMs and land uses when using aggregated versus high-resolution input data (Fig 3, Table 1). The aggregated runs overpredicted SOC compared to the mean of the fine-scale runs by 6.3% to 15.1% (4.8 to 12.0 t/Ha) across the four GCMs and three agricultural land uses. The average difference was 7.2 t/Ha, or a 9.1% mean absolute error (MAE).
Lines show the cell with the highest average SOC over the run, the mean of all cells, the cell with the lowest average SOC, and the results when the model was run with input data aggregated over the study area. Colors show distribution of SOC across cells in each year as share of runs passing through each bin. Data are smoothed by year.
Examining data from the final simulation year (2099), SOC stocks in the fine-scale runs varied substantially across individual cells in the study region. Fig 4 shows the systematic bias in the distribution of individual cell SOC projections relative to the aggregated runs. The largest overprediction of SOC resulting from aggregation (in terms of error) was 173.5%, where the SOC stock for the individual cell was 34.3 t/Ha and the aggregated run was 93.7 t/Ha, a difference of 59.5 t/Ha. This occurred for pasture under the miroc-esm GCM. The largest underprediction of the aggregated method compared to an individual cell was 28.4 t/Ha, a 21.7% error. Differences in SOC estimates resulted from spatial variation across cells in soil, temperature, and precipitation, as well as temporal variation in the climate variables due to climate change. Fig 5 plots the input data and SOC simulation results spatially.
Histograms of SOC across cells in the last year of the simulation (2099), for each GCM and land use for the high-resolution simulations. Solid black line plots cell average. Dashed red line plots study-area aggregated value.
Left column shows bulk density (g/cm3) and clay percent, input data common to all runs in each cell. Top row shows temperature (°C) in 2022 (GCM average) and then in 2099 for each GCM. Middle row shows precipitation (mm/day) in 2022 and 2099 for each GCM. Bottom row shows SOC (t/Ha) in 2022 and 2099 for each GCM averaged across the three agricultural land uses. Vermont boundary map is from https://www.census.gov/geographies/mapping-files/time-series/geo/carto-boundary-file.html and is in the public domain as described at https://ask.census.gov/prweb/PRServletCustom?pyActivity=pyMobileSnapStartArticleID=KCP-4928.
Exploring the difference between SOC in 2022 vs. the 2099 projection, storage typically declined over time with climate change. Reductions in mean SOC were more pronounced in the fine-scale model runs compared to the aggregated runs (Fig 3, Table 2). The average SOC loss across all land uses and GCMs in the aggregated runs was 6.7 t/Ha (7.1% loss), compared to an average loss across all fine-scale runs of 13.3 t/Ha (14.1% loss). SOC losses varied by GCM and land use: in general, pasture showed the highest loss in terms of t/Ha, followed by hay and crops, although the percent loss was similar across land uses. The miroc-esm GCM had the largest decline in SOC (-11.2% aggregated vs. -22.5% fine-scale avg.), followed by noresm1-m (-11.6% aggregated vs. -17.0% fine-scale avg.), mri-cgcm3 (-3.3% aggregated vs. -9.4% fine-scale avg.), and ccsm4 (-2.1% aggregated vs. -7.7% fine-scale avg.). We note, however, that SOC storage increased in a some cells under specific combinations of climate change, soil characteristics, and land use. The maximum per-cell increase in SOC occurred under ccsm4 for pasture (10.7 t/Ha more SOC, representing a 9.6% increase). The maximum per-cell decrease occurred under miroc-esm for pasture (35.1 t/Ha less SOC, a 30.7% decrease).
Discussion
This study investigated the impact of aggregating data from ∼0.65 km2 cells to a study area encompassing 11,690 km2 (with 2,793 km2 in farmland) using the RothC model. We found that aggregated runs systematically overpredicted SOC compared to the mean of fine-scale runs. The mean absolute error resulting from aggregation was 9.1%, and the maximum error on a per-cell basis was 173.5%.
While additional research is required to investigate whether the trend we find here holds for other study areas, the potential for systematic bias resulting from data aggregation points to the importance of running SOC models at as fine a spatial scale as feasible when projecting SOC dynamics at the landscape level. Further, the variance in our results suggests that studies that aggregate data over a large regional area cannot be relied upon by individual land managers hoping to understand the potential for SOC storage at the field level.
Our results are consistent with other efforts to characterize the effect of spatial granularity on error in agro-ecological models. Hoffman et al. [19] evaluated the error associated with climate and soil data granularity on crop yield simulations from 14 models, finding that soil data aggregation up to 100 km caused MAEs in yield projections up to 15% depending on the model, with climate data aggregation having a smaller effect. Others have evaluated the effect of data aggregation on crop yield projections, finding MAEs of up to 18% [16, 17]. Dalsgaard et al. [18] explored the effects of temporal and spatial aggregation of input data on SOC in Norwegian forests, finding that projected SOC varied by up to 10% depending on data scale, and with temporal scale more impactful than spatial scale.
Our study brought up several challenges inherent to this type of fine-grained computational simulation. Firstly, fine-scale data are required to drive the simulations. While high-resolution remote sensing raster data are typically available to characterize land use and soil drivers, fine-scale climate projections have not been calculated for all regions. We were fortunate to have access to the downscaled climate projection dataset [20], as the project would not have been possible otherwise. Additionally, some of the GIS data that had previously been sufficient when we were working at larger scales [7–9] proved to be lacking when we attempted to geoprocess them to map onto the <1 km2 cells. Specifically, the gSSURGO database had voids in which no data was available for a subset of cells, necessitating a shift to the downscaled POLARIS soil dataset [35]. One of the primary challenges to more accurate SOC projections is a lack of fine-scaled input datasets. Therefore, efforts to facilitate better collection and processing of such data should be prioritized going forward.
Finally, simulating SOC storage across our study area at such a fine spatial scale required significant computational resources. Using a typical laptop, the 71,964 model runs would have taken roughly a year to compute if the machine ran continuously, whereas this simulation was possible in just a few hours on the NCAR supercomputer by leveraging parallel computation [40]. This suggests that powerful, shared computing resources are an essential tool as we strive to make better projections and decisions about climate change.
In any complex simulation problem, the average of the function is not necessarily the function of the average. We find that this adage holds here, as the full study area runs using spatially-aggregated inputs systematically differ from the mean of individual cells across all GCMs and land uses. Our results suggest that emphasizing a localized approach to ecological modeling should be prioritized wherever predictive accuracy is of prime importance and sufficient computational resources are available. This is especially true in the case of modeling climate impacts, since it is imperative that we develop an accurate understanding of coming shifts in the climate system, allowing policymakers to respond appropriately.
Limitations and future research
This study set out to answer a relatively simple but important research question: all else being equal, to what extent can changing the scale of aggregation of input data affect the results of SOC models at the landscape scale? We have shown that in this context aggregation of input data introduced a systematic bias on the order of 10%. Future research is required to determine whether the magnitude and/or direction of bias from aggregation found here holds in other study areas. Further, it was not within the scope of this study to analyze the sensitivity of the error to each specific factor driving the model, although this would also be a valuable avenue for further research. Finally, a limitation of this study stems from the fact that comprehensive timeseries data on SOC change is not available for our study area, which limits the possibility of conducting a thorough validation procedure for the model.
Conclusion
Spatial aggregation of input data resulted in an overestimation of the potential for soils to store carbon and mitigate climate change for agricultural land in Vermont. The effect of temperature, precipitation, soil properties, and land management on decomposition rates is complex and nonlinear, meaning that data aggregation can affect results in unexpected ways [5, 18, 41]. We found large spatial variation in SOC storage, resulting from interactions between fine-scale variability in the parameters forcing the model over time. Aggregating input data to the landscape scale obfuscates these relationships and may introduce a systematic bias in SOC projections.
Acknowledgments
We thank Dr. Joshua Faulkner and Dr. Juan Alvez from the University of Vermont United States Department of Agriculture Extension Service for their expert advice to parameterize the model.
References
- 1. Lal R. Sequestering carbon in soils of agro-ecosystems. Food Policy. 2011;36:S33–S39.
- 2. Kane D. Carbon sequestration potential on agricultural lands: a review of current science and available practices. National Sustainable Agriculture Coalition Breakthrough Strategies and Solutions, LLC. 2015; p. 1–35.
- 3. Paustian K, Lehmann J, Ogle S, Reay D, Robertson GP, Smith P. Climate-smart soils. Nature. 2016;532(7597):49–57.
- 4. Lal R. Carbon sequestration. Philosophical Transactions of the Royal Society B: Biological Sciences. 2008;363(1492):815–830.
- 5. Luo Z, Feng W, Luo Y, Baldock J, Wang E. Soil organic carbon dynamics jointly controlled by climate, carbon inputs, soil properties and soil carbon fractions. Global Change Biology. 2017;23(10):4430–4439.
- 6. Morais TG, Teixeira RFM, Domingos T. Detailed global modelling of soil organic carbon in cropland, grassland and forest soils. PLOS ONE. 2019;14(9):e0222604.
- 7. Wiltshire S, Beckage B. Soil carbon sequestration through regenerative agriculture in the US state of Vermont. PLOS Climate. 2022;1(4):e0000021.
- 8. Wiltshire S, Beckage B. Integrating climate change into projections of soil carbon sequestration from regenerative agriculture. PLOS Climate. 2023;2(3):e0000130.
- 9. Wiltshire S, Grobe S, Beckage B. A Historically Driven Spinup Procedure for Soil Carbon Modeling. Soil Systems. 2023;7(2):35.
- 10. Davidson EA, Trumbore SE, Amundson R. Soil warming and organic carbon content. Nature. 2000;408(6814):789–790.
- 11. Reichstein M, Rey A, Freibauer A, Tenhunen J, Valentini R, Banza J, et al. Modeling temporal and large-scale spatial variability of soil respiration from soil water availability, temperature and vegetation productivity indices. Global Biogeochemical Cycles. 2003;17(4).
- 12. Davidson EA, Janssens IA. Temperature sensitivity of soil carbon decomposition and feedbacks to climate change. Nature. 2006;440(7081):165–173.
- 13. Bradford MA, Wieder WR, Bonan GB, Fierer N, Raymond PA, Crowther TW. Managing uncertainty in soil carbon feedbacks to climate change. Nature Climate Change. 2016;6(8):751–758.
- 14.
Coleman K, Jenkinson DS. RothC-26.3—A Model for the turnover of carbon in soil. In: Powlson DS, Smith P, Smith JU, editors. Evaluation of Soil Organic Matter Models. Berlin, Heidelberg: Springer Berlin Heidelberg; 1996. p. 237–246. Available from: http://link.springer.com/10.1007/978-3-642-61094-3_17.
- 15. Barančíková G, Halás J, Gutteková M, Makovníková J, Nováková M, Skalský R, et al. Application of RothC model to predict soil organic carbon stock on agricultural soils of Slovakia. Soil and Water Research. 2010;5(No. 1):1–9.
- 16. Easterling WE, Weiss A, Hays CJ, Mearns LO. Spatial scales of climate information for simulating wheat and maize productivity: the case of the US Great Plains. Agricultural and Forest Meteorology. 1998;90(1):51–63.
- 17. Folberth C, Yang H, Wang X, Abbaspour KC. Impact of input data resolution and extent of harvested areas on crop yield estimates in large-scale agricultural modeling for maize in the USA. Ecological Modelling. 2012;235-236:8–18.
- 18. Dalsgaard L, Astrup R, Antón-Fernández C, Borgen SK, Breidenbach J, Lange H, et al. Modeling Soil Carbon Dynamics in Northern Forests: Effects of Spatial and Temporal Aggregation of Climatic Input Data. PLOS ONE. 2016;11(2):e0149902.
- 19. Hoffmann H, Zhao G, Asseng S, Bindi M, Biernath C, Constantin J, et al. Impact of Spatial Soil and Climate Input Data Aggregation on Regional Yield Simulations. PLOS ONE. 2016;11(4):e0151782.
- 20. Winter JM, Beckage B, Bucini G, Horton RM, Clemins PJ. Development and Evaluation of High-Resolution Climate Simulations over the Mountainous Northeastern United States. Journal of Hydrometeorology. 2016;17(3):881–896.
- 21. Falloon PD, Smith P, Smith JU, Szabo J, Coleman K, Marshall S. Regional estimates of carbon sequestration potential: linking the Rothamsted Carbon Model to GIS databases. Biology And Fertility Of Soils. 1998;27(3):236–241.
- 22. Falloon P, Smith P. Simulating SOC changes in long-term experiments with RothC and CENTURY: model evaluation for a regional scale application. Soil Use and Management. 2002;18(2):101–111.
- 23.
United Nations Food and Agriculture Organization. Measuring and modelling soil carbon stocks and stock changes in livestock production systems—A scoping analysis for the LEAP work stream on soil carbon stock changes. Food & Agriculture Org.; 2019.
- 24. Sierra CA, Mueller M, Trumbore SE. Models of soil organic matter decomposition: the SoilR package, version 1.0. Geoscientific Model Development. 2012;5(4):1045–1060.
- 25. Zimmermann M, Leifeld J, Schmidt MWI, Smith P, Fuhrer J. Measured soil organic matter fractions can be related to pools in the RothC model. European Journal of Soil Science. 2007;58(3):658–667.
- 26.
Wiltshire S, Beckage B. Github repository for R Code used in this paper; 2023. Available from: https://github.com/brianbeckage/SoilCarbonGrid.
- 27. Jenkinson DS, Andrew SPS, Lynch JM, Goss MJ, Tinker PB. The Turnover of Organic Carbon and Nitrogen in Soil [and Discussion]. Philosophical Transactions: Biological Sciences. 1990;329(1255):361–368.
- 28. Kuzyakov Y, Friedel JK, Stahr K. Review of mechanisms and quantification of priming effects. Soil Biology and Biochemistry. 2000;32(11):1485–1498.
- 29. Brekke LD, Barsugli JJ. Uncertainties in projections of future changes in extremes. Extremes in a Changing Climate: Detection, Analysis and Uncertainty. 2013; p. 309–346.
- 30. Menne MJ, Durre I, Vose RS, Gleason BE, Houston TG. An overview of the global historical climatology network-daily database. Journal of atmospheric and oceanic technology. 2012;29(7):897–910.
- 31.
Dewitz J. National Land Cover Database (NLCD) 2016 Products: U.S. Geological Survey data release; 2019.
- 32.
United States Department of Agriculture, National Agricultural Statistics Service. 2017 Census of Agriculture; 2017.
- 33. Rodell M, Houser P, Jambor U, Gottschalck J, Mitchell K, Meng CJ, et al. The global land data assimilation system. Bulletin of the American Meteorological Society. 2004;85(3):381–394.
- 34.
United States Department of Agriculture, Natural Resources Conservation Service. Gridded Soil Survey Geographic (gSSURGO) Database for the Conterminous United States; 2020.
- 35. Chaney NW, Wood EF, McBratney AB, Hempel JW, Nauman TW, Brungard CW, et al. POLARIS: A 30-meter probabilistic soil series map of the contiguous United States. Geoderma. 2016;274:54–67.
- 36.
Rollett A, Williams J. 2018-19 Soil Policy Evidence Programme: Review of best practice for SOC monitoring. Soil Policy & Agricultural Land Use Planning Unit; Land, Nature and Forestry Division; Department for Rural Affairs; Welsh Government; 2019.
- 37. Klumpp K, Coleman K, Dondini M, Goulding K, Hastings A, Jones MB, et al. Soil Organic Carbon (SOC) Equilibrium and Model Initialisation Methods: an Application to the Rothamsted Carbon (RothC) Model. Environmental Modeling & Assessment. 2017;22(3):215–229.
- 38. Bolinder MA, Janzen HH, Gregorich EG, Angers DA, VandenBygaart AJ. An approach for estimating net primary productivity and annual carbon inputs to soil for common agricultural crops in Canada. Agriculture, Ecosystems & Environment. 2007;118(1-4):29–42.
- 39. Falloon P. How important is inert organic matter for predictive soil carbon modelling using the Rothamsted carbon model? Soil Biology and Biochemistry. 2000;32(3):433–436.
- 40.
Computational and Information Systems Laboratory. Cheyenne: HPE/SGI ICE XA System (University Community Computing). Boulder, CO: National Center for Atmospheric Research; 2019. https://doi.org/10.5065/D6RX99HX
- 41. Paustian K, Larson E, Kent J, Marx E, Swan A. Soil C Sequestration as a Biological Negative Emission Strategy. Frontiers in Climate. 2019;1.