Report

Radiocarbon constraints imply reduced carbon uptake by soils during the 21st century

See allHide authors and affiliations

Science  23 Sep 2016:
Vol. 353, Issue 6306, pp. 1419-1424
DOI: 10.1126/science.aad4273

Abstract

Soil is the largest terrestrial carbon reservoir and may influence the sign and magnitude of carbon cycle–climate feedbacks. Many Earth system models (ESMs) estimate a significant soil carbon sink by 2100, yet the underlying carbon dynamics determining this response have not been systematically tested against observations. We used 14C data from 157 globally distributed soil profiles sampled to 1-meter depth to show that ESMs underestimated the mean age of soil carbon by a factor of more than six (430 ± 50 years versus 3100 ± 1800 years). Consequently, ESMs overestimated the carbon sequestration potential of soils by a factor of nearly two (40 ± 27%). These inconsistencies suggest that ESMs must better represent carbon stabilization processes and the turnover time of slow and passive reservoirs when simulating future atmospheric carbon dioxide dynamics.

Soil carbon is a dynamic reservoir that may increase substantially in size during the 21st century, as predicted by Earth system models (ESMs), thereby influencing the sign and magnitude of carbon cycle feedbacks under climate change (14). Under a high radiative forcing scenario (representative concentration pathway 8.5), changes in soil carbon estimated by different models vary from a loss of 20 Pg C to a gain of more than 360 Pg C (5). These models suggest that the global carbon inventory in mineral soils may increase by 30% or more over a time span of about two centuries. The multimodel mean soil carbon accumulation of 109 Pg C (5) represents about one decade of global fossil fuel emissions at current rates and 5% of cumulative fossil emissions by 2100 for this scenario (6). This soil carbon sink represents a negative feedback on CO2 emissions and, if robust, would slow the rate of climate change.

Still, there are substantial uncertainties in the soil carbon sink projected by ESMs (5). Rapid rates of carbon sequestration in ESMs contrast with findings from CO2 and warming experiments (7, 8), as well as multiple theoretical and observational constraints indicating slow (millennial) rates of soil organic carbon (SOC) accrual and turnover (914). Model uncertainty—as measured by intermodel spread—is high for soil carbon turnover time (τ) and exceeds the uncertainty estimated for carbon uptake through gross primary production (GPP) (15, 16).

In coupled model simulations, the relative sink strength (i.e., percentage change in soil carbon) depends on the responses of net primary production (NPP) and soil carbon dynamics to increasing atmospheric CO2 concentrations and, to a lesser extent, climate change (5). Elevated CO2 increases photosynthesis and NPP, which results in greater carbon inputs to soil pools with decadal or longer residence times. Carbon sequestration in soils reduces the buildup of CO2 in the atmosphere (the carbon-concentration feedback). On the other hand, elevated CO2 also warms the climate, which tends to accelerate soil carbon turnover and reduce carbon storage (the carbon-climate feedback) (17, 18). Although these feedbacks oppose one another, the carbon-concentration feedback is more than four times as high, on average, as the carbon-climate feedback in current ESMs at the global scale (3). Differences in the representation of elevated CO2 versus climate effects on ecosystem processes result in substantial variation in soil carbon sequestration estimates (19) (table S1).

Without a strong carbon-concentration feedback, ESMs would likely project smaller gains or larger losses of soil carbon over the 21st century. Our aim was to constrain the magnitude of the soil carbon–concentration feedback with soil radiocarbon observations. Radiocarbon content provides information about soil carbon turnover over centuries to millennia based on radioactive decay and over decades, based on inputs of 14C from atmospheric weapons testing (“bomb carbon”). Accurate carbon turnover times are important for ESM projections because pools with short turnover times rapidly adjust to increasing NPP, whereas pools with long turnover times (and, by inference, low rates of inputs) change only slowly, possibly beyond the time horizon of effective climate mitigation efforts. Therefore, inaccuracies in the representation of carbon turnover times will have consequences for the rate and magnitude of the carbon-concentration feedback simulated by ESMs. Here, we used Δ14C measurements at 157 sites across multiple biomes (Fig. 1 and table S2) along with carbon inventory data to constrain soil carbon dynamics in five biogeochemically coupled ESM simulations (esmFixClim1) from the Coupled Model Intercomparison Project Phase 5 (CMIP5) (20). In these idealized simulations, the atmospheric CO2 mole fraction starts at a preindustrial value of 285 parts per million (ppm) and rises at a rate of 1% year−1, thus quadrupling over 140 years. The biogeochemical components of each model experience the increasing trajectory of atmospheric CO2, whereas the atmospheric radiation submodels do not, limiting impacts solely to the direct effects of CO2 on plant physiology and thus enabling diagnosis of carbon-sink sensitivity to increasing CO2.

Fig. 1 Locations of radiocarbon soil profiles used to constrain ESM soil carbon mean ages and turnover times (N = 157).

The carbon-weighted Δ14C to a depth of 1 m is denoted with the color shade of each symbol. A summary of the location, sample year, and reference for each site is provided in table S2.

Total initial soil carbon in the ESMs was not significantly different from the total amount in the top meter of the Harmonized World Soil Database (HWSD) (Fig. 2, A and B) for four of the five models [P > 0.05, except for the Community ESM (CESM), P = 0.03]. Therefore, we compared ESM-derived Δ14C to observations derived from soil profiles to a 1-m depth. The carbon and 14C patterns of the soil profiles we used were similar to those reported in a recent synthesis paper (21), and we used some of the same profiles in our analysis.

Fig. 2 Site and global values of soil carbon inventory, Δ14C, and mean age from observations and reduced complexity models.

(A and B) SOC content of the original ESMs. (C and D) The Δ14C of the reduced complexity model optimized to the original ESMs. (E and F) Corresponding mean age (G and H) The Δ14C of the 14C-constrained reduced complexity models. The left column shows the values of the models sampled at the locations of the individual soil profiles; the right column shows the global model distribution. Data from profile sites and the HWSD represent carbon content in the top 1 m of soil; data from ESMs are the total carbon stock. The star denotes the mean; the + symbol denotes outliers beyond the 25th and 75th percentiles.

Comparing ESM outputs with 14C observations requires a model analysis approach because most ESMs do not yet explicitly simulate Δ14C in soils, and no ESMs had reported turnover times for soil carbon pools. Therefore, we used a reduced complexity (RC) model to approximate soil carbon dynamics in each ESM. This approach allowed us to estimate the turnover times and Δ14C associated with the carbon pools in different ESMs (table S3). Where possible, we used a three-pool RC model (with fast, slow, and passive pools) to simulate carbon and 14C dynamics. A multipool structure is essential because radiocarbon observations show that soil carbon fluxes (NPP inputs and heterotrophic respiration) exchange mainly with short-lived pools, whereas carbon stocks are dominated by long-lived pools (12, 18, 22, 23). The three-pool RC model had five parameters representing turnover times of fast, slow, and passive pools (τfast, τslow, and τpassive) and transfer coefficients (rf and rs) that regulated carbon flow from the fast to slow, and slow to passive pools (fig. S1). We used a two-pool RC model for Geophysical Fluid Dynamics Laboratory model ESM2M (GFDL-ESM2M) because it represents soil carbon with two pools (24) and for Hadley Global Environment Model 2-ES (HadGEM2-ES) because it reported carbon for two pools (table S4). The two-pool RC model had three parameters, representing τfast, τslow, and rf (fig. S1). After verifying that the RC model was a good approximation of each ESM based on minimization of root-mean-square error, we used the RC models to simulate Δ14C values at each grid cell, with observed atmospheric Δ14C for the past 50,000 years as a boundary condition and accounting for radioactive decay (see supplementary materials).

We used an inverse analysis to determine the RC model parameters that were most consistent with our Δ14C data set. In the inversion, we adjusted the parameters described above to match both the total carbon and radiocarbon constraints. With these constraints, turnover time and carbon input rate for each pool were coupled such that an increase in turnover time required a compensatory decline in inputs (fig. S2). RC parameters derived from the inversion were subsequently used to assess consequences of 14C constraints for the carbon-concentration feedback.

All ESMs projected an increase in soil carbon over 140 years with a multimodel mean of 32% (Table 1). This increase was primarily driven by increasing carbon inputs to soil under the quadrupling of CO2 (table S3), as temperature increased by only a small amount (mean ± 1 SD was 0.52° ± 0.68°C) for this set of biogeochemically coupled simulations. CESM showed the smallest soil carbon increase (6.3%), primarily because of low litter inputs relative to other ESMs (table S3). For this time period and set of model runs, storage in soil carbon accounted for 42 ± 17% of the total accumulation of carbon in the terrestrial biosphere.

Table 1 Global soil carbon stocks and carbon uptake for CMIP5 models that experienced a quadrupling of atmospheric CO2 from a preindustrial value of 285 ppm over a period of 140 years.

View this table:

Both two- and three-pool RC models reproduced the global carbon dynamics of the original ESMs (fig. S3 to S5 and table S5). The τfast across all RC models was less than 20 years, whereas τslow varied from 40 to 600 years (fig. S6) with a multimodel mean of 212 ± 104 years. The mean τpassive for the three-pool RC models from CESM, Institut Pierre Simon Laplace model (IPSL), and the Meteorological Research Institute model (MRI) was 1185 ± 123 years (Table 1 and fig. S7). Using the RC model parameters estimated at each grid cell within an ESM, we calculated the expected Δ14C. The resulting global average Δ14C for 1995 (median sample year of site profiles) from the RC models was significantly higher than the mean of the observations [–6.4 ± 64 per mil (‰) versus –211 ± 156‰] (Fig. 2, C and D (P < 0.001). Δ14C values from RC models approximating ESMs with passive pools were more negative (–53 ± 35‰) but still significantly higher than the observations (P < 0.001). Converting these Δ14C observations into mean age for the soil profile yielded an estimate of 3100 ± 1800 years for the observed soil carbon integrated to 1 m and 430 ± 50 years for the ESMs (Fig. 2, E and F). These results indicated that the ESMs did not have enough old carbon that had experienced substantial levels of radioactive decay; concurrently, the models assimilated too much bomb 14C.

The 14C-derived mean ages indicate that organic carbon soil is often thousands of years old (1214, 21), which is an order of magnitude older than suggested by ESM turnover parameters. This discrepancy is likely a consequence of incomplete representation of key biogeochemical processes and difficulties in developing accurate parameterizations for soil carbon at a global scale. Most ESMs do not account for stabilization mechanisms whereby mineral interactions and aggregate formation protect soil organic matter from decomposition over centuries to millennia (13, 2528). Moreover, first-order decay, as represented in ESMs, may not capture the response of mineral-stabilized carbon to changes in soil moisture, temperature, and other conditions (2931). In addition, some ESM turnover parameters are based on laboratory incubation studies, which are often biased fast compared with in situ decomposition rates (32). Finally, this set of ESMs did not explicitly resolve vertical differences in soil organic matter dynamics, which may cause underestimation of turnover times in deep soils with large carbon stocks (21, 25, 33, 34).

Because the turnover times derived from ESMs were inconsistent with 14C observations, we optimized the turnover parameters by fitting our RC models to the observations. We could then run the optimized RC models to reevaluate soil carbon storage for the transient 1% year−1 simulations. For this inverse approach, we optimized RC model parameters in each grid cell containing an observation site (Fig. 2, G and H, and figs. S8 and S9). We optimized the τ of the slowest pool and the corresponding transfer coefficient into this pool based on the 14C observations while holding soil inputs and τ for the faster pools at their ESM-derived values. The size of the slowest carbon pool was constrained by optimizing the turnover time and the transfer coefficient together using both 14C and total carbon. Consequently, the optimized RC model had about the same total carbon stock as the original ESM, thereby maintaining consistency with carbon inventory data. This optimization approach yielded τslow values of 3700 ± 2800 years for GFDL and 3500 ± 1300 years for HadGEM (using two-pool RC models), which were 16 to 17 times as long as the turnover times derived from the original ESMs.

For ESMs that included a passive pool, the optimization process yielded three distinct outcomes. For CESM, which has the largest passive pool (73% of soil carbon), the optimized τpassive was 4500 years, which was 3.7 ± 1.5 times as long as τpassive derived from the original model (Table 1). IPSL has a smaller fraction of passive carbon (46%) and therefore required a greater τpassive (16,500 years) to obtain agreement with the observed Δ14C. For MRI, the passive pool size was too small (only 13% of soil carbon) to bring Δ14C into alignment with the profile observations even after parameter optimization (fig. S10 and table S5). To adjust for MRI’s potential bias in the passive pool size, we optimized rf together with τpassive and rs to allow for simultaneous changes in slow and passive pool sizes. The resulting RC model for MRI was able to match observations (Fig. 2, G and H) with a passive pool fraction of 48% (see Methods and table S5). These results indicated that increasing the size and turnover time of the passive pool in ESMs would improve agreement with 14C-based mean age estimates.

Bringing turnover time and carbon transfer parameters into agreement with 14C observations had considerable consequences for the magnitude of the carbon-concentration feedback. Using the 14C-based parameters, we conducted global transient simulations with each of the five RC models. These simulations showed that the soil as a whole (specifically the slow and passive pools) stored much less carbon in response to increasing levels of atmospheric CO2, primarily as a consequence of reduced flow into the slow or passive pool. The soil carbon sink decreased from 32 ± 18% to 18 ± 12% (Table 1), corresponding to an absolute sink reduction of 170 ± 127 Pg C (Fig. 3). The magnitude of the soil sink reduction varied widely across the different models; those with larger and older passive fractions at the onset of the transient simulation (Table 1) generally had smaller sink reductions.

Fig. 3 Absolute change in SOC content from the reduced complexity model fit to the original ESM (bars with white background) and the estimate obtained by applying the 14C constraint to the reduced complexity model (bars with gray background).

The estimates on the right side show the total carbon content (sum of fast, slow, and passive) averaged across all the models, before and after applying the radiocarbon constraint.

To assess the robustness of these sink reductions, we conducted a series of sensitivity experiments (see supplementary materials). We found that the sink reduction imposed by constraining the models with 14C observations was robust to (i) turnover times optimized specifically for different biomes, (ii) spatial variation and magnitude of soil carbon stocks, and (iii) variations in Δ14C across measurement sites (Table 2 and table S6). Sink reductions declined by a factor of 2 when the models were fit to an inventory that was 50% larger than the HWSD data set, suggesting that if soil carbon pools were larger in ESMs, 14C-imposed sink reductions would be lower (35). Last, we used our RC model approach to analyze four fully coupled ESM runs (1pctCO2) to address potential interactions between the carbon-climate and the carbon-concentration feedback. Constraints imposed by 14C still reduced the sink by at least 40% on average (fig. S11 and table S7) in the fully coupled simulations (see supplementary materials).

Table 2 Summary of sensitivity experiments.

View this table:

We conclude that CMIP5 ESMs underestimated the mean age of soil carbon, especially for slow-cycling pools. By adjusting the turnover times of slow and passive pools to bring the models into alignment with 14C observations, the potential for future soil carbon sequestration declined by 40 ± 27%. Although long-lived soil carbon pools consistent with old 14C ages imply a similar potential for carbon storage at steady state, the time scale required to reach equilibrium is too long to mitigate the potentially damaging climate effects of rising CO2 concentrations during the 21st century (fig. S2). These findings emphasize the need to incorporate 14C and other diagnostics into ESM development and evaluation. In addition, models require better representation of long-term mechanisms of soil carbon stabilization such as organic matter-mineral interactions. Considered together with potential nutrient limitation of NPP inputs to soil (36), our analysis suggests that the carbon-concentration feedback may be weaker in the 21st century than currently expected from ESMs. Therefore, a greater fraction of CO2 emissions than previously thought could remain in the atmosphere and contribute to global warming.

Supplementary Materials

www.sciencemag.org/content/353/6306/1419/suppl/DC1

Materials and Methods

Sensitivity Analysis Results

Fully Coupled Simulation Analysis

Figs. S1 to S11

Tables S1 to S7

References (37102)

References and Notes

  1. Acknowledgments: We thank C. Hatté for sharing her compilation of published 14C profiles. We received funding support from the Climate and Environmental Sciences Division of Biological and Environmental Research (BER) in the U.S. Department of Energy Office of Science. This included support from the Regional and Global Climate Modeling Program to the Biogeochemical Cycles Feedbacks Science Focus Area and several grants from the Terrestrial Ecosystem Science Program (DESC0014374 and DE-AC02-05CH11231). J.W.H. serves as chair of the science steering group for International Soil Carbon Network (http://iscn.fluxdata.org); however, the work of this publication reflects efforts on behalf of the U.S. Geological Survey Scientist Emeritus Program, which provided IT and infrastructure support. The model simulations analyzed in this study were obtained from the Earth System Grid Federation CMIP5 online portal hosted by the Program for Climate Model Diagnosis and Intercomparison at Lawrence Livermore National Laboratory (https://pcmdi.llnl.gov/projects/esgf-llnl/).
View Abstract

Navigate This Article