VOLUME 26 JOURNAL OF CLIMATE 15JULY 2013

Estimating the -Carbon Climate Response in the CMIP5 Climate Models Using a Simplified Approach

ELEANOR J. BURKE AND CHRIS D. JONES Met Office Hadley Centre, Exeter, United Kingdom

CHARLES D. KOVEN Earth Sciences Division, Lawrence Berkeley National Laboratory, Berkeley, California

(Manuscript received 31 July 2012, in final form 5 November 2012)

ABSTRACT

Under , thawing permafrost may cause a release of carbon, which has a

on the climate. The permafrost-carbon climate response (gPF) is the additional permafrost-carbon made vulnerable to decomposition per degree of global temperature increase. A simple framework was adopted

to estimate gPF using the database for phase 5 of the Coupled Model Intercomparison Project (CMIP5). The projected changes in the annual maximum thicknesses (ALTmax) over the twenty-first century were quantified using CMIP5 soil temperatures. These changes were combined with the observed distri-

bution of soil organic carbon and its potential decomposability to give gPF.ThisestimateofgPF is dependent on the biases in the simulated present-day permafrost. This dependency was reduced by combining a ref-

erence estimate of the present-day ALTmax with an estimate of the sensitivity of ALTmax to temperature 21 from the CMIP5 models. In this case, gPF was from 26to266 PgC K (5th–95th percentile) with a radi- 2 2 ative forcing of 0.03–0.29 W m 2 K 1. This range is mainly caused by uncertainties in the amount of deeper in the soil profile and whether it thaws over the time scales under consideration. These results suggest that including permafrost-carbon within climate models will lead to an increase in the positive global carbon climate feedback. Under future climate change the northern high-latitude permafrost region is expected to be a small sink of carbon. Adding the permafrost-carbon response is likely to change this region to a source of carbon.

1. Introduction To help quantify the permafrost-carbon climate feed- back, we need to know how much carbon is vulnerable to Permafrost soils contain ;1672 Pg of organic carbon release, how fast it will be released, and whether it will (Tarnocai et al. 2009), much of which is permanently be released as (CO ) or (CH ). frozen and consequently relatively inert. Under in- 2 4 These processes are highly uncertain, spatially variable, creased temperature, permafrost degrades and a pro- and often hard to measure (Schuur and Abbott 2011). portion of this old permafrost-carbon will become more Burke et al. (2012) developed a simple framework to vulnerable to decomposition and subsequent release estimate the additional temperature increase caused by into the climate system. Like burning, this is permafrost-carbon loss that includes the impact of many an irreversible process over time scales of hundreds of of these uncertainties. They suggest that, had permafrost- years and it has the potential to cause a further increase carbon been included within the Hadley Centre climate in greenhouse gases in the atmosphere, leading to a pos- model HadGEM2-ES (note that all model names used itive carbon–climate feedback (Schuur et al. 2008). in this paper are expanded in Table 1), there would be an additional temperature increase of 0.028–0.368C (90% range). Schneider von Deimling et al. (2012) used an alternative simple approach, again included relevant Corresponding author address: Eleanor Burke, Met Office Hadley Centre, FitzRoy Road, Exeter, EX1 3PB, United Kingdom. uncertainties, and suggested there might be an additional E-mail: eleanor.burke@metoffice.gov.uk warming of 0.048–0.238C (68% range). MacDougall

DOI: 10.1175/JCLI-D-12-00550.1

Ó 2013 American Meteorological Society 4897 Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4898 JOURNAL OF CLIMATE VOLUME 26 et al. (2012), Koven et al. (2011), and Schaefer et al. a feedback from permafrost with any interactions, a 21 (2011) developed their land surface schemes to include second-order effect. The land feedback (gL;PgCK )is a simple representation of permafrost-carbon and quan- quantified by the change in the global land carbon per tified the permafrost-carbon lost, but they incorporated degree of global mean temperature change. As before, a more limited uncertainty assessment. MacDougall et al. negative values represent a loss of carbon from the land

(2012) coupled their land surface scheme with a global surface. Combining gL with gPF provides an estimate of of intermediate complexity and found the overall carbon–climate response including permafrost. an additional warming of 0.098–0.758Ccausedbythe permafrost-carbon climate feedback. In general, the 2. Models and methods permafrost-carbon climate feedback is not yet included a. Reference data within coupled earth system general circulation models (GCMs). Large-scale observations of the annual maximum of the

In this paper, the permafrost-carbon climate response active layer thickness (ALTmax) are unavailable. A refer- (gPF) is defined as the permafrost-carbon made vulner- ence estimate of the present-day ALTmax was obtained able to decomposition per degree of global temperature from the Joint UK Land Environment Simulator (JULES) increase. Negative values represent a loss of carbon land surface scheme (Best et al. 2011) driven by the Water from the land surface. This is a committed loss (i.e., the and Global Change (WATCH) forcing data (Weedon et al. amount that may ultimately be lost if the temperatures 2011) at a 28 resolution (Burke et al. 2013). Using this were to stabilize). The majority of other studies estimate reference model allows us to remove the considerable the amount lost over a specified time frame. This is a uncertainty associated with initial permafrost distributions realized loss. The committed definition of gPF was used across the CMIP5 models, while still allowing us to sample because it significantly reduces the dependence of gPF model uncertainty arising from other aspects of the CMIP5 on the rate of change of global mean temperature. The ensemble, such as the , arctic amplifica- amount of vulnerable permafrost-carbon can be quan- tion, and sensitivity of active layer deepening to warming. tified using the simple framework developed by Burke This version of JULES calculates the soil temperature et al. (2012). They took the change in the maximum thaw every 10 cm within the soil profile. A predecessor of the depth output from a global climate model, quantified the JULES land surface scheme the Met Office Surface Ex- amount of soil carbon in this depth range and, using an change Scheme (MOSES)/Top-Down Representation estimate of its quality, determined how much soil carbon of Interactive Foliage and Flora Including Dynamics was available for decomposition. Their approach is adop- (TRIFFID) is included within HadGEM2, which over- ted here to quantify gPF for the phase 5 of the Coupled estimates the permafrost extent (Table 1) and underesti- Model Intercomparison Project (CMIP5) ensemble of mates the difference between the air and soil temperatures global climate models. (Koven et al. 2013). However, the snow scheme within Other carbon climate feedbacks currently represented JULES is an updated multilayer snow scheme that in- within GCMs include the change in soil and vegetation creases the snow insulation effect and provides a more carbon in response to climate change. The Coupled realistic estimate of permafrost extent (Burke et al. 2013). Climate Model Intercomparison Project b. CMIP5 global climate models (C4MIP) generation of models shows a global release of land carbon under increasing temperature and an uptake The data analyzed here were obtained from phase 5 of of land carbon under increasing CO2 with a net positive the Coupled Model Intercomparison Project (CMIP5) feedback on to the global climate (Friedlingstein et al. multimodel data archive. These CMIP5 global climate 2006). In contrast, in the northern high latitudes, Qian models support the Intergovernmental Panel on Climate et al. (2010) showed that the majority of the C4MIP Change Fifth Assessment Report (IPCC AR5). The ex- models have a slight uptake of carbon in response to periments discussed include the ‘‘historical’’ experiments climate change. This suggests that, in this region and when from the mid-1800s to the present day and two future permafrost-carbon is not included, the increase in primary scenarios for the twenty-first century, called rcp45 and productivity and litterfall outweighs any increase in soil rcp85. Representative concentration pathways (RCPs) 2 respiration rates at the higher temperatures. Including 4.5 and 8.5 correspond to forcings of 4.5 and 8.5 W m 2 permafrost-carbon processes may change the total high- by 2100 respectively, and represent intermediate and high latitude terrestrial response from sink to source (Koven warming scenarios (Moss et al. 2010). Data from 17 of the et al. 2011). Here we make the simplifying assumption that CMIP5 multimodel ensemble members that provided the terrestrial carbon feedback at high latitudes can be split depth-resolved soil temperatures were used (Koven et al. into two terms: a land feedback without permafrost and 2013). This subset includes models with and without an

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4899

TABLE 1. Table showing the physical characteristics of the permafrost simulated by the different CMIP5 models. Each of the CMIP5 models was regridded to the 28 resolution of the JULES-WATCH reference data (Burke et al. 2013) and the area calculated. The models that fall near the observed range of between 12.2 and 17.0 million km2 (Zhang et al. 2003) are highlighted in italics. The items listed in bold are observations/reference data for comparison purposes.

Historical mean RCP4.5 mean PF RCP8.5 mean PF PF extent (million extent (million km2; extent (million km2; Model km2; 1995–2000) 2095–2100) 2095–2100) Beijing Climate Center Climate 2.5 1.2 0.7 System Model (BCC-CSM1-1) Canadian Centre for Climate Modelling 4.5 0.5 0.4 and Analysis Earth System Model 2 (CanESM2) Goddard Institute for Space Studies 7.3 2.8 1.3 Model E2-R (GISS-E2-R) Institut Pierre-Simon Laplace Coupled Model, 8.8 1.5 0.1 version 5a, medium resolution (IPSL-CM5A-MR) Community Climate System Model, version 4 (CCSM4) 9.9 2.6 0.6 Model for Interdisciplinary Research on 9.9 3.5 0.3 Climate–Earth System Model (MIROC-ESM) MIROC-ESM chemistry (MIROC-ESM-CHEM) 10.1 2.2 0.1 Max Planck Institute Earth System Model LR (MPI-ESM-LR) 10.3 2.1 0.9 Institute of Numerical Mathematics Coupled 10.9 4.7 1.7 Model, version 4 (INM-CM4) IPSL coupled model version 5a, low resolution 11.1 9.0 7.5 (IPSL-CM5A-LR) Norwegian Climate Centre Earth System 11.8 4.9 1.8 Model 1-M (NorESM1-M) Meteorological Research Institute Coupled General 14.4 9.4 5.2 Circulation Model version 3 (MRI CGCM3) Joint UK Land Environment Simulator with Water 16.1 — — and Global Change forcing data MIROC version 5 (MIROC5) 16.5 11.7 5.2 Observations 17.4 — — Geophysical Fluid Dynamics Laboratory Earth 24.1 18.5 13.2 System Model 2M (GFDL-ESM2M) Hadley Centre Global Environmental Model 26.0 15.4 7.6 version 2-ES (HadGEM2-ES) GFDL Earth System Model 2G (GFDL-ESM2G) 26.5 21.2 14.1 HadGEM version 2-CC (HadGEM2-CC) 28.5 18.8 9.5 interactive carbon cycle and those models with an in- c. Estimating the maximum active layer teractive carbon cycle do not include nutrient dynamics. thickness (ALTmax) The multimodel ensemble spans a range of climate sen- sitivities and contains several different representations If the permafrost-carbon feedback had been included of land surface processes. The CMIP5 models’ repre- within the CMIP5 models, the physical state of the sentation of permafrost dynamics is discussed in detail by permafrost would have been quantified directly from the both Koven et al. (2013) and Slater and Lawrence (2013). CMIP5 simulated soil temperatures with no corrections

For comparison purposes, the sensitivity of land car- for biases. Therefore one estimate of gPF uses these 21 bon (no permafrost) to climate change (gL; PgC K ) uncorrected monthly mean soil temperatures to di- was estimated from the climate models that include the agnose the thaw depth or active layer thickness (ALT). interactive carbon cycle. Specialized simulations were For each month and in each grid cell the ALT was de- used to derive gL in which the radiation code sees the fined as the deepest point in the soil column with soil increase in atmospheric CO2 but the carbon cycle does temperature at or above freezing (Koven et al. 2013). not (Friedlingstein et al. 2006). In other words, the CO2 The vertical resolution of the land surface schemes in is radiatively active and biogeochemically inert. In the many of the CMIP5 models is often poor, leading to CMIP5 experimental design ‘‘esmFdbck1’’ is such a set possible biases in the estimate of the active layer thick- of simulations with the atmospheric CO2 increasing by ness (see Burke et al. 2012 for details). The annual 21 1% yr . Data from these simulations are only available maximum of the active layer (ALTmax) was then defined for a small subset of the models. from these monthly values. Permafrost was assumed to

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4900 JOURNAL OF CLIMATE VOLUME 26 be present in any given grid cell where ALTmax is shal- plausible range of sensitivities within the northern high- lower than 3 m. This approach provides an estimate of latitude permafrost zone. shallow permafrost and ties in with the observed distri- Time series of twenty-first century ALTmax were recon- bution of soil organic carbon content that is available for structed for each grid point where there is permafrost in the top 3 m of the soil. The value of ALTmax was defined thepresent-dayreferenceestimatebysamplingALTsensitivity annually for the permafrost region for the years between from the 5th–95th percentile range and combining it with b 1995 and 2100. The baseline ALTmax (ALTmax)isan the local annual mean near-surface air temperature b estimate of the present-day value; ALTmax was defined change for each CMIP5 model and each RCP scenario as the maximum of the monthly ALT during the period [Eqs. (1) and (2)]:

1995–2000. ALTmax was calculated for both the rcp45 5 2 and rcp85 simulations. Results using this set of ALTmax ALTmax(t, x) ALTmax(t 1, x) are denoted ‘‘uncorrected.’’ 1 ALTsensitivity[Tlocal(t, x) The CMIP5 soil temperatures have considerable errors 2 2 when compared with observations (Koven et al. 2013; Tlocal(t 1, x)], (1) Slater and Lawrence 2013), leading to biases in the di- 5 5 b agnosed present-day permafrost state. Therefore, an al- ALTmax(t 0, x) ALTmax(x), (2) ternative estimate of twenty-first-century values of

ALTmax was made by combining the reference estimate where Tlocal is the local temperature, t is the time, and x b identifies each grid cell that has permafrost in the present- of the present-day permafrost state (ALTmax)withan estimated rate of change of ALTmax with temperature day reference estimate of the active layer. Here, the sampled from the CMIP5 ensemble. For each CMIP5 CMIP5 ensemble was used to sample both the uncer- model and each RCP scenario the rate of change of tainty in the warming and the uncertainty in the response

ALTmax per degree (ALTsensitivity) was calculated using of ALTmax to this warming. Results using this derivation a linear regression fit between ALTmax and the local near of ALTmax are denoted ‘‘bias-substituted.’’ surface annual mean air temperature on a grid point by These two estimates of twenty-first century ALTmax grid point basis for all grid points where there is perma- can then be combined with the observed distribution of frost in the top 3 m. If ALTmax becomes greater than 3 m soil organic carbon content to provide two estimates of all subsequent times are excluded from the analysis. This the permafrost climate response. The uncorrected gPF analysis assumes there is a linear relationship between would have been obtained if the permafrost climate

ALTmax and local temperature. However, Koven et al. feedback were included in the current generation of cli- (2013) show that the sensitivity of ALTmax to local tem- mate models, and the bias-substituted gPF is an estimate perature tends to increase as temperatures approach the using the permafrost susceptibility to air temperature thawing point. This is emphasized by poor soil dis- derived from the CMIP5 ensemble referenced to the b cretization in some of the CMIP5 models. This assump- JULES-WATCH ALTmax. Both estimates have large tion means that the simulated permafrost thaws too early associated uncertainties. during the twenty-first century but has a smaller impact d. Estimating the permafrost-carbon climate on errors in the depth of ALT by the end of the max response (g ) twenty-first century. The regression fits result in a distri- PF bution of ALTsensitivity that samples uncertainty from the The observed distribution of the soil organic carbon different CMIP5 models. content (SOCC) was taken from the Northern Circum-

In reality, ALTsensitivity is also spatially variable and polar Soil Carbon Database (NCSCD; Tarnocai et al. dependent on, among other things, soil type, vegetation 2009). In general, soil carbon in the NCSCD is severely cover, and climatology. Because many of the CMIP5 undersampled, particularly at depths greater than 100 cm, models simulate a very low permafrost extent (Table 1, and the uncertainties in any estimates of SOCC are po- column 1; Koven et al. 2013; Slater and Lawrence 2013) tentially large. Burke et al. (2012) used the information and therefore indicate no permafrost in regions where provided by Tarnocai et al. (2009) to provide an estimate the reference data suggests permafrost exists, it is not of SOCC along with its uncertainties in the 0–100-, 100– possible to estimate this spatial distribution for the ob- 200-, and 200–300-cm depth ranges. The SOCC at 100– served permafrost extent. However, the 5th–95th per- 200 cm is assumed to be less than that at 0–100 cm and centile range of the distribution of ALTsensitivity (0.02– less again for 200–300 cm. The SOCC of the soil at 21 b 0.29 m K ) can be used as the upper and lower limits. depths shallower than ALTmax was assumed to be already While these values will not be representative of any active within the carbon cycle and is not considered here. particular model or region, they will encompass the During any year and for any grid cell that ALTmax is

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4901

b greater than ALTmax there is a quantifiable amount of in these realizations arise from potential errors in the b thawed permafrost carbon. A proportion of this thawed JULES-WATCH estimate of ALTmax, the range of carbon is assumed to be passive, very stable, and not re- ALTsensitivity discussed above, and the CMIP5 model leased over the time scale of this study. The remainder of and RCP specific local temperature change. Dankers the carbon is assumed to all decompose over time scales et al. (2011) suggested that JULES simulations of of about 0–200 years. This is defined as the carbon vul- ALTmax are slightly too deep when compared with ob- nerable to decomposition (Cvul) and is the amount of servations. Therefore the range of uncertainties on the b carbon that could contribute to the permafrost-carbon JULES-WATCH estimates of ALTmax was set to be- climate response. The amount of carbon that is passive is tween 50% too deep and 20% too shallow. For the bias- uncertain. Dutta et al. (2006) use laboratory incubations substituted simulations LHS sampling was used to sample on soils and estimate the passive soil carbon to be the parameters required to estimate the time series of

18% of the total. Following a literature review, Falloon ALTmax, the soil organic carbon distribution, and the et al. (1998) suggested that the passive pool could be proportion of decomposable organic carbon. much larger and range between 15% and 60%. This un- The contribution of each uncertain parameter–process certainty range was included in the analysis. to the range of gPF was determined as in Burke et al. The permafrost-carbon climate response (gPF; (2012) by splitting the parameter–processes into a set of 21 PgC K ) was quantified as bins and calculating the mean of gPF for each bin. This was then compared with the mean of gPF for all of the C g 5 vul , (3) simulations. If gPF is sensitive to a parameter–process PF DT there will be notable differences between the mean of

gPF in each bin and that for all of the simulations. This where Cvul is the mean decomposable soil organic car- can be quantified using the following equation: bon (in PgC) made vulnerable by 2095–2100 and DT is the mean change in global mean temperature (in K) N 2 2 1 (mi m) between 1995–2000 and 2095–2100. Any permafrost- S 5 å , (4) N 5 s2 carbon released will feed back onto the global mean i 1 temperature, which will be amplified in the northern where m is the mean of bin i, m is the mean of all the high latitudes. This process is not included here and its i simulations, and s is the standard deviation of all of the absence may change g . PF simulations. e. Uncertainty assessment f. Estimating the land carbon climate feedback (gL) Uncertainties in this simple framework are large and The sensitivity of land carbon (no permafrost) to climate 21 can arise from differences between the RCP scenarios, change (gL;PgCK ) was found using esmFdbck1 21 uncertainties in the derivation of the simulated twenty- where the atmospheric CO2 increases by 1% yr in first-century ALTmax, uncertainties in both the amount the following manner: of soil organic carbon present and how much of it is decomposable over the time scales of consideration, and DC deviations from the simplified linear analysis presented g 5 L , (5) L DT here. In the uncorrected estimates of gPF, ALTmax was defined once for every CMIP5 model and both RCP scenarios. Five hundred Monte Carlo simulations were where DCL (PgC) is the change in the mean total land carried out for each model and scenario with the soil carbon and DT is the change in the mean global mean organic carbon distribution and the proportion of pas- temperature (K). The change is defined for the mean of sive organic carbon randomly sampled from the range the first and last 5 years of the 140-yr simulation, which of plausible values using a Latin hypercube sampling starts at preindustrial CO2 and ends at 4 times preindustrial (LHS; McKay et al. 1979) strategy. In the bias-substituted CO2. These simulations do not include the confounding estimates of gPF, ALTmax was derived by combining effects of changes in land use, non-CO2 greenhouse gases, b the JULES-WATCH estimate of ALTmax with both aerosols, etc., and so provide a controlled experiment with ALTsensitivity and the location, model, and scenario- which to evaluate land carbon–climate interactions specific twenty-first-century near-surface air temperature (Arora et al. 2013). It is assumed that there are no inter- change defined as the ‘‘local’’ temperature change. This actions between the carbon currently within the GCM’s results in 500 realizations of the time series of ALTmax carbon cycle and any new carbon released from thawed for each model and scenario. Additional uncertainties permafrost.

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4902 JOURNAL OF CLIMATE VOLUME 26

3. Results For many of the models, the permafrost-carbon re- sponse is larger for RCP4.5 than for RCP8.5. RCP8.5 The mean CMIP5 permafrost extents during the pe- has a larger twenty-first century temperature increase riod 1995–2000 are highly variable and range between and therefore a deeper active layer by the end of the 2.5 and 28.5 million km2 (see Table 1, ordered by in- century. However, at the greater depths the observed creasing permafrost extent). The observed extent is esti- soil organic carbon content is lower and therefore the mated to be somewhere between 12.2 and 17.0 million km2 rate of increase of vulnerable carbon per degree tem- with the permafrost affected area estimated to be perature change is reduced at higher temperatures when 22.8 million km2 (Zhang et al. 2003). Very few models compared with lower temperatures. In some places this fall within this range, with the majority underestimating gives a nonlinear relationship between change in ALTmax the extent and a few significantly overestimating it. The and vulnerable soil organic carbon and causes a reduced models (NorESM1-M, MRI-CGM3, and MIROC5) that response for RCP8.5 compared with RCP4.5. However fall close to this range are highlighted in italics in Table 1. some models such as HadGEM2 have a similar values for b Koven et al. (2013) suggest that, in general, the differ- RCP4.5 and RCP8.5 because ALTmax is initially deeper ences between models and observations are dominated and the thawed permafrost releases soil organic carbon by the parameterization of surface exchange, the snow from the deeper levels irrespective of scenario. This scheme, and coupled thermal–hydrological dynamics highlights both the scenario dependence of this definition within the soil column rather than by errors in the cli- of gPF and the importance of observing the depth distri- mate model simulation of the local mean air temperature. bution of the soil organic carbon content. This was also demonstrated by Burke et al. (2013), who The models in Fig. 1 are ranked in order of increasing showed significant differences in the simulated present- mean gPF. This ranking is very similar to the ranking of day permafrost extent between two simulations identical present-day permafrost extent in Table 1. As might be except for their representation of snow. The simulated expected, models with little present-day permafrost changes in the permafrost extent in response to a chang- (and hence only a small amount of soil organic carbon) ing climate are also highly variable (Table 1). However, have a low gPF and vice versa. There is a strong re- all of the models simulate a loss of permafrost over the lationship between gPF and the total soil organic car- twenty-first century leading to additional thawed carbon, bon content in the top 3 m of the permafrost affected which could decompose and be released into the atmo- zone defined using the present-day extent for each sphere. As might be expected, this loss of permafrost is CMIP5 model (Fig. 2). Although the RCP8.5 scenario greater for RCP8.5 than for RCP4.5 because RCP8.5 has is shown, results for RCP4.5 are qualitatively similar. If a bigger increase in global mean temperature over the the permafrost-carbon climate feedback were included twenty-first century. within the current generation of climate models, biases Figure 1 shows the permafrost-carbon climate re- in the simulated permafrost extent would significantly sponse (gPF) for the CMIP5 model simulations and the bias the estimate of gPF (Figs. 1 and 2). two RCP scenarios using the uncorrected derivation of The error bars on the points in Fig. 2 show two standard

ALTmax. This is the response that might have been ob- deviations both in the total soil organic carbon content tained if permafrost-carbon were included directly in the and in gPF. The uncertainties in gPF for any particular global climate models—the majority of which have large model are driven by uncertainties in the distribution and biases in their simulation of permafrost extent. The decomposability of soil organic carbon. For the models box represents the inter quartile range (25%–75%), the with the small permafrost extents, uncertainties are low mean of gPF is the horizontal black line within the box, because there is little soil carbon present initially in the and the whiskers represent the 5th–95th percentile range. permafrost zone. The vertical lines in Fig. 2 show the Points outside this range are show with dots. There is mean of the observed estimate of soil organic carbon a wide spread in gPF with values ranging from 23to within the permafrost zone where the permafrost zone is 2 2 2123 PgC K 1 for RCP4.5 and from 21to280 PgC K 1 defined by either the International Permafrost Association (both 5th–95th percentile) for RCP8.5. The three (IPA) permafrost extent or the JULES-WATCH perma- models with permafrost extents similar to that observed frost extent. This falls toward the center of the range of (NorESM1-M, MRI-CGM3, and MIROC5) are high- model simulations. Also shown on Fig. 2 is a linear least lighted in gray. Using just these models to estimate gPF squares regression fit between the initial soil organic 21 changes the lower limit of gPF from 21to217 PgC K carbon content in the permafrost region and gPF with the (5th percentile, both RCP scenarios included), which intercept set to zero. The thin dashed–dotted lines are significantly increases the minimum likely positive the confidence intervals for this fit. For RCP8.5 the R2 feedback. value is 91%. Using this relationship, the confidence

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4903

FIG. 1. Permafrost-carbon climate response (gPF) estimated using the uncorrected ALTmax for the (top) RCP4.5 and (bottom) RCP8.5 scenarios. That is gPF that would have been cal- culated if permafrost-carbon were included within the CMIP5 climate models. The boxes represent the 25th–75th quantile, the whiskers the 5th–95th quantile, and the points the out- liers. The models shaded in gray are the three models that have present-day permafrost similar to that observed. intervals for two standard deviations, and the observa- spatially variable and dependent on CMIP5 model and tions as a constraint on gPF, gPF could be between 218 its associated uncertainties need to be considered. 21 and 263 PgC K (Fig. 2). The confidence in the re- The other estimate of gPF uses the bias-substituted gression fit for RCP4.5 (not shown) is lower, leading to ALTmax. In this case, gPF is independent of biases in the 21 awiderrangeofgPF —between 25and2105 PgC K . CMIP5 estimates of present-day permafrost extent. The These latter estimates assume that all of the models simu- 5th–95th percentile range of the frequency distribution 21 late the same rate of loss of organic carbon as a function of of gPF is from 26to266 PgC K (Fig. 3). The two dif- temperature and the errors are solely driven by uncer- ferent RCP scenarios are pooled. This range is contained tainties in the initial soil organic carbon content. However, within the spread of values shown in Fig. 1. The median 2 the rate of change of the active layer with temperature is value of 222 PgC K 1 falls toward the lower end of

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4904 JOURNAL OF CLIMATE VOLUME 26

FIG. 2. The relationship between gPF calculated using the uncorrected ALTmax for the RCP4.5 scenario and initial soil carbon content. The mean of the estimate of the initial soil carbon content for the JULES-simulated permafrost extent is shown by the vertical dark gray line and that of the IPA observations is shown by the vertical light gray line. Also shown is the

linear regression (forced through zero) between the initial soil organic carbon content and gPF (dashed dark line). the frequency plot. This is because the distribution is cannot be ruled out. The 5th–95th percentile range skewed with a long tail of stronger feedbacks (larger of carbon made vulnerable to release by 2100 is between negative values). Although these larger permafrost feed- 11 and 135 PgC for RCP4.5 and between 18 and 181 PgC backs have a relatively small likelihood of occurrence, they for RCP8.5. This is considerably lower than the estimate

FIG. 3. Frequency distribution of gPF estimated using the bias-substituted ALTmax. Both RCP scenarios are pooled. The value of gPF is estimated using the reference estimate of b present-day ALTmax found from JULES-WATCH. The 5th–95th percentile range is shaded light gray.

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4905 by Harden et al. (2012) who found between 108 and 706 TABLE 2. The relative contributions of the listed uncertainties to PgC may thaw by 2100 under RCP8.5. the overall uncertainty in gPF calculated using the uncorrected As defined in this paper, g is the committed change ALTmax from Fig. 1. These contributions are in percent and the PF total contribution is 100%. by the end of the twenty-first century. At this time not all of the thawed and decomposable soil carbon will have Relative contribution to gPF decomposed, but it is assumed that it will eventually calculated using the uncorrected Uncertainty ALT (Fig. 1) (%) decompose if the climate were to stabilize at that global max mean temperature level. An alternative definition of RCP scenario 3 the permafrost-carbon response, used by, for example SOCC 0–100 cm 2 SOCC 100–200 cm 10 Schneider von Deimling et al. (2012), is the amount of SOCC 200–300 cm 18 permafrost-carbon lost per degree temperature change— Decomposable proportion 3 arealizedgPF. This realized gPF can be estimated for of thawed carbon HadGEM2-ES using the methods developed by Burke CMIP5 model 64 et al. (2012). A comparison of the committed gPF with the realized gPF for HadGEM2-ES and the RCP4.5 scenario suggests that by the end of the twenty-first century ap- RCP scenarios in Fig. 1, they contribute a relatively small proximately half of the vulnerable carbon will have been amount to the overall uncertainty. lost. Therefore, using this as a guide, a rough estimate of Figure 4 explores the uncertainties by the end of the the committed response from Schneider von Deimling twenty-first century for the bias-substituted estimates of et al. (2012) is very similar for both RCP4.5 and RCP8.5 gPF. The black dashed line in Fig. 4 represents the mean 2 and between 212 and 238 PgC K 1 (both 68% range). for the whole ensemble (17 000 members) and the gray

This is comparable with the committed gPF estimated band is two standard deviations wide and centered on 2 here (212 to 244 PgC K 1 also 68% range) with the mean. The black lines and error bars represent the a similar uncertainty range. Using the Special Report on means and standard deviations for subensembles that Emissions Scenarios (SRES) A1B high emissions sce- have been grouped by roughly equally sized bins around nario, Schaefer et al. (2011) estimated a realized loss of the values shown. Both the standard deviations and 110 6 40 PgC and Koven et al. (2011) a realized loss of means change depending on subensemble (Fig. 4). In

62 6 7 PgC by 2100. These can be roughly compared with general, the bins that result in gPF with the smallest the approximate realized carbon loss shown here of be- magnitudes have narrower uncertainties and vice versa. tween 9 and 90 PgC (5th–95th percentile) by 2100 for This is because for these smaller permafrost-carbon RCP8.5. responses there is little carbon made vulnerable and therefore less potential uncertainty on the total amount. Uncertainty assessment Compared with Fig. 1 there are now small differences Uncertainties in Fig. 1 are caused by our uncertain among the CMIP5 models. The permafrost-carbon re- knowledge of the profile of soil organic carbon content, sponse remains larger for RCP4.5 than for RCP8.5. Both the spatial distribution of soil organic carbon content, of these differences arise because the rate of increase of the decomposability of the soil organic carbon content vulnerable carbon per degree temperature change is defined by the proportion of soil organic carbon that is nonlinear with temperature. It is generally lower at the active, and the model- and RCP-specific dependence of larger temperature changes and deeper ALTmax.Schneider ALTmax on temperature change. The relative contribu- von Deimling et al. (2012) found similar values for both tion of each of these uncertainties to the overall un- RCP scenarios because they did not explicitly represent certainty for gPF derived using the uncorrected ALTmax the profile of soil carbon within the permafrost, which was estimated using Eq. (4) and is shown in Table 2 for the causes the notable differences here. end of the twenty-first century. The continuous parameters Table 3 shows the relative importance of the eight were divided into four equally sized bins. There are only uncertainties addressed here. As in Table 2 they were two bins representing the two RCP scenarios. Differences calculated by dividing the continuous parameters into in the number of bins considered have a relatively minor four equally sized bins. The SOCC below 100 cm now impact on the relative contributions. Any differences dominates the uncertainties in gPF. This causes ap- between the CMIP5 models dominate the overall un- proximately 59% of the overall uncertainty, with slightly certainty in Fig. 1, contributing 64% to the total spread. more from the shallower soil than the deeper soil. Al-

The other main contributors are uncertainties in the soil though the sensitivity of ALTmax to local temperature b organic carbon content at depths below 100 cm. Despite change (ALTsensitivity) and the potential biases in ALTmax there being obvious systematic differences between the were both given a relatively wide range of values, they

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4906 JOURNAL OF CLIMATE VOLUME 26

FIG. 4. Sensitivity of the Monte Carlo simulations to the most important model parameters–processes. The black dashed line represents the mean for the whole ensemble and the gray band is two standard deviations wide and centered on the mean. The black lines and error bars represent the means and standard deviations for subensembles that have been grouped by roughly equally sized bins around the values shown. The parameters/processes are sorted in order of decreasing important from the (top left) most important ones to the (bottom right) least important ones. only contribute ;8% of the overall uncertainty each. The larger overall contribution from the uncertainty in the RCP scenario, proportion of carbon that is decomposable, SOCC compared with the smaller contributions from and the temperature sensitivities of the CMIP5 models the diagnosis of the changes in ALTmax suggests that each contribute 4% or less to the total uncertainty. The combining observational constraints with the relationship

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4907

TABLE 3. The relative contributions of the listed uncertainties to esmFdbck1. Globally there is a loss of carbon in re- the overall uncertainty in g calculated using the bias-substituted PF sponse to climate change with gL between 211 2 ALTmax from Fig. 3. These contributions are in percent and the and 2128 PgC K 1. However, in the northern high- total contribution is 100%. latitude permafrost region all of the models except Relative contribution GFDL-ESM2M show a slight uptake of carbon. This is to gPF calculated using the because the increase in net primary productivity and bias-substituted litterfall outweighs the increase in in this Uncertainty ALT from Fig. 3 (%) max region at increased temperatures. These results are RCP scenario 4 qualitatively similar to those found for the C4MIP SOCC (0–100 cm) 12 SOCC (100–200 cm) 32 models. These gL are realized values rather than com- SOCC (200–300 cm) 27 mitted values. The committed gPF for RCP4.5 was di- Decomposable proportion 4 vided by 2 in order to approximate the realized gPF of thawed carbon found by Burke et al. (2012) and qualitatively compare CMIP5 temperature sensitivity 3 gPF with a realized gL. In general gPF is smaller than gL. Sensitivity of ALTmax to global 8 mean temperature change However, its inclusion will increase the positive carbon– b climate feedback already suggested by g . Although Biases in JULES-WATCH ALTmax 9 L for the present day smaller than gL, gPF is likely to be large enough to change the northern high latitudes from a sink to a source of carbon in the future. It should be noted that between SOCC in the permafrost region and g shown PF this analysis neglects any interactions between the land in Fig. 2 may be applicable when estimating g . PF carbon cycle and the permafrost-carbon cycle. However, the thawed permafrost-carbon will impact the carbon 4. Comparison with the land carbon climate cycle in the northern high latitudes. The models shown in response (gL) Table 4 do not include nitrogen cycle feedbacks. Friedlingstein et al. (2006) found that the mean Arneth et al. (2010) suggest that nitrogen cycle climate feedbacks may well halve gL. In addition, release of ni- land carbon sensitivity gL (without permafrost) was 2 279 PgC K 1 across the C4MIP generation of carbon trogen from the thawed permafrost may stimulate pro- cycle models under the high SRES A2 scenario for ductivity in the in the northern high latitudes. This may the period up to 2100 with a range of between 220 result in some uptake of the thawed permafrost-carbon 2 and 2177 PgC K 1. This represents a loss of carbon and by the vegetation. results in a positive carbon climate feedback. This loss is dominated by extraboreal regions (Boer and Arora 5. Comparison with other biogeochemical 2010) with the majority of models suggesting a net up- feedbacks take in the northern high latitudes of up to 17 PgC by

2100 (Qian et al. 2010). Table 4 summarizes gL found To compare it with other feedbacks, gPF was converted using Eq. (5) from the available CMIP5 models for to an equivalent by multiplying by f

TABLE 4. The land carbon climate response for the global land surface and for the northern high latitudes permafrost region defined using the JULES-WATCH simulated present-day permafrost extent for the 1% scenario. Also shown is the 5th–95th percentile range of the estimated realized permafrost-carbon climate response for the RCP4.5 pathways. The value for gL for the C4MIP ensemble is taken from Friedlingstein et al. (2006).

Approximate realized

Northern high bias-substituted gPF for RCP4.5 21 21 21 Model Global gL (PgC K ) latitudes gL (PgC K ) [5th–95th percentile (PgC K )] HadGEM2-ES 240 13 24to239 GFDL-ESM2M 2128 210 24to257 MPI-ESM-LR 266 2 23to240 CanESM2 263 1 23to233 BCC-CSM1–1 247 6 24to254 IPSL-CM5A-LR 216 0 23to238 NorESM1-ME 211 1 23to246 C4MIP range 220 to 2177 23to10 —

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 4908 JOURNAL OF CLIMATE VOLUME 26

(Arneth et al. 2010; Gregory et al. 2009), where f is development. Bog formation might sequester some of 2 2 0.0049 W m 2 PgC 1 and represents the linear approx- the thawed permafrost-carbon and render it inert. Other imation of the increase in radiative forcing with increased additional missing processes include formation of ice atmospheric concentrations. Pooling all models for wedges within the permafrost, , the pres- the RCP4.5 scenario and estimating the 5th–95th ence of soils, mosses as a vegetation type, and the percentile range gives a gPF between 0.03 and impact of nutrient availability on vegetation productivity. 22 21 0.45 W m K . Similarly, for RCP8.5 gPF is between Time scales of decomposition and release of carbon to the 22 21 0.03 and 0.25 W m K . As discussed previously, gPF atmosphere are not considered in this paper. Rates of tends to be slightly lower for the RCP8.5 scenario. change will impact the amount of carbon lost from the Arneth et al. (2010) estimated permafrost-carbon feed- northern high latitudes. For a low rate of thaw the eco- backs to be similar in magnitude to those estimated here. system is more likely to respond and utilize the thawed Arneth et al. (2010) showed that other feedbacks such nutrients and carbon. If the thaw is quick, the carbon and as wetland , ozone emissions, and nutrients are more likely to be washed into aquatic sys-

fire are likely to be smaller than gPF. This suggests the tems and lost. Field experiments are regularly carried out permafrost-carbon climate feedback is likely to be the to develop understanding of many of these processes lo- largest biogeochemical feedback not currently included cally and in different (e.g., Schuur et al. 2009), in coupled earth system GCMS. but increasingly upscaling methods are required to im- prove their representation at the larger scales, such as that of a GCM grid cell. 6. Discussion and conclusions The current generation of coupled earth system GCMs The permafrost carbon–climate feedback is not yet represented by the CMIP5 model ensemble have large included within the coupled earth system models used in biases in their estimate of present-day permafrost extent the CMIP5 analysis. Therefore the technique discussed (Koven et al. 2013; Slater and Lawrence 2013). These here provides a basic tool for quantifying the permafrost- biases need to be significantly reduced while including the carbon response and its uncertainty with respect to dif- permafrost-carbon climate feedback within global cli- fering model structural and parametric elements. This mate models. Therefore, future work should focus on paper uses a simple approach to provide a preliminary improving the accuracy of the coupling between the air estimate of the carbon made vulnerable to decomposi- temperature and the soil temperature in the northern tion through permafrost thawing per degree change in high latitudes and fully evaluating the land surface global mean temperature (gPF) for the CMIP5 global schemes using methods such as those discussed in Burke climate models. This is a committed response. Several et al. (2013) and Koven et al. (2013). This will help char- estimates of gPF are presented that suggest gPF is likely acterize the spatial variability in the active layer thickness to have a positive feedback on climate and is less than and its sensitivity to a changing climate. 2 2 2 298 PgC K 1 or 0.48 W m 2 K 1. After correction for An assessment of the impact of uncertainties on the present-day biases in permafrost extent and active layer estimate of gPF demonstrates the importance of knowing thickness, the best estimate of gPF is between 26and the amount of soil carbon in the soil below the present- 2 2 2 266 PgC K 1 or between 0.03 and 0.29 W m 2 K 1 (5th– day maximum thaw depth. Our knowledge of this can be 95th percentile range). This estimate is dependent on significantly improved through improved observational- future emissions scenario because vulnerable organic based analyses, such as that of Harden et al. (2012). carbon decreases with depth and is therefore non- All of the Monte Carlo simulations discussed here linearly related to global mean temperature change. show a loss of permafrost-carbon in the future. This loss One of the main limitations of this study is that it uses is likely to be large enough to change the high latitudes highly simplified models to simulate the carbon cycle. In from a potential sink to a source of carbon under future particular, it neglects interactions between the carbon climate change. The permafrost-carbon climate feed- currently within the GCM’s carbon cycle and any new back is probably the biggest biogeochemical feedback carbon released from thawed permafrost. In addition, not currently included within coupled earth system the GCMs do not simulate nutrient dynamics and cannot GCMs. represent, for example, the impact of nitrogen availability on the carbon cycle. Northern high-latitude ecosys- Acknowledgments. The work described in this paper tems are often nutrient limited and nutrients such as was supported by the Joint DECC/Defra Met Office nitrogen contained within the permafrost may stimulate Hadley Centre Climate Programme (GA01101). The au- primary productivity once the permafrost thaws (Keuper thors acknowledge the financial support by the European et al. 2012). Another neglected process is thermokarst Union FP7-ENVIRONMENT project PAGE21 under

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC 15 JULY 2013 B U R K E E T A L . 4909

Contract GA282700. Performing the MOHC CMIP5 Thawing permafrost increases plant-available nitrogen in simulations was supported by the European Commis- subarctic peatlands. Global Change Biol., 18, 1998–2007, sion’s 7th Framework Programme, under Grant Agree- doi:10.1111/j.1365-2486.2012.02663.x. Koven, C. D., B. Ringeval, P. Friedlingstein, P. Ciais, P. Cadule, ment 226520, COMBINE project. We acknowledge the D. Khvorostyanov, G. Krinner, and C. Tarnocai, 2011: Per- World Climate Research Programme’s Working Group mafrost carbon–climate feedbacks accelerate global warm- on Coupled Modelling, which is responsible for CMIP, ing. Proc.Natl.Acad.Sci.USA,108, 14 769–14 774, and we thank the climate modeling groups for producing doi:10.1073/pnas.1103910108. and making available their model output. For CMIP the ——, W. J. Riley, and A. Stern, 2013: Analysis of permafrost thermal dynamics and response to climate change in the U.S. Department of Energy’s Program for Climate Model CMIP5 earth system models. J. Climate, 26, 1877–1900. Diagnosis and Intercomparison provides coordinating MacDougall, A. H., C. A. Avis, and A. J. Weaver, 2012: Signif- support and led development of software infrastructure icant contribution to climate warming from the permafrost in partnership with the Global Organization for Earth carbon feedback. Nat. Geosci., 5, 719–721, doi:10.1038/ System Science Portals. ngeo1573. McKay, M. D., R. J. Beckman, and W. J. Conover, 1979: A com- parison of three methods for selecting values of input variables REFERENCES in the analysis of output from a computer code. Techno- metrics, 21, 239–245, doi:10.2307/1268522. Arneth, A., and Coauthors, 2010: Terrestrial biogeochemical Moss, R. H., and Coauthors, 2010: The next generation of scenarios feedbacks in the climate system. Nat. Geosci., 3, 525–532, for climate change research and assessment. Nature, 463, 747– doi:10.1038/ngeo905. 756, doi:10.1038/nature08823. Arora, V. K., and Coauthors, 2013: Carbon-concentration and Qian, H., R. Joseph, and N. Zeng, 2010: Enhanced terrestrial car- carbon–climate feedbacks in CMIP5 Earth system models. bon uptake in the northern high latitudes in the 21st century J. Climate, in press. from the Coupled Carbon Cycle Climate Model Intercom- Best, M. J., and Coauthors, 2011: The Joint UK Land Environment parison Project model projections. Global Change Biol., 16, Simulator (JULES), Model description—Part 1: Energy and 641–656, doi:10.1111/j.1365-2486.2009.01989.x. water fluxes. Geosci. Model Dev., 4, 677–699, doi:10.5194/ Schaefer, K., T. Zhang, L. Bruhwiler, and A. P. Barrett, 2011: gmd-4-677-2011. Amount and timing of permafrost carbon release in response Boer, G. J., and V. K. Arora, 2010: Geographic aspects of tem- to climate warming. Tellus, 63B, 165–180, doi:10.1111/ perature and concentration feedbacks in the carbon budget. j.1600-0889.2011.00527.x. J. Climate, 23, 775–784. Schneider von Deimling, T., M. Meinshausen, A. Levermann, Burke, E. J., I. P. Hartley, and C. D. Jones, 2012: Uncertainties in V. Huber, K. Frieler, D. M. Lawrence, and V. Brovkin, 2012: the global temperature change caused by carbon release from Estimating the permafrost–carbon feedback on global warm- permafrost thawing. Cryosphere, 6, 1063–1076, doi:10.5194/ ing. Biogeosciences, 9, 649–665, doi:10.5194/bg-9-649-2012. tc-6-1063-2012. Schuur, E. A. G., and B. Abbott, 2011: Climate change: High risk ——, R. D. Dankers, C. D. Jones, and A. J. Wiltshire, 2013: of permafrost thaw. Nature, 480, 32–33, doi:10.1038/ Evaluating changes in near-surface permafrost during the 20th 480032a. century with the JULES land surface model. Climate Dyn., in ——, and Coauthors, 2008: Vulnerability of permafrost carbon to press. climate change: Implications for the global carbon cycle. Dankers, R., E. J. Burke, and J. Price, 2011: Simulation of perma- Bioscience, 58, 701–714, doi:10.1641/B580807. frost and seasonal thaw depth in the JULES land surface ——, J. G. Vogel, K. G. Crummer, H. Lee, J. O. Sickman, and T. E. scheme. Cryosphere, 5, 773–790, doi:10.5194/tc-5-773-2011. Osterkamp, 2009: The effect of permafrost thaw on old carbon Dutta, K., E. A. G. Schuur, J. C. Neff, and S. A. Zimov, 2006: release and net carbon exchange from tundra. Nature, 459, Potential carbon release from permafrost soils of northeastern 556–559, doi:10.1038/nature08031. Siberia. Global Change Biol., 12, 2336–2351. Slater, A. G., and D. M. Lawrence, 2013: Diagnosing present and Falloon, P. D., P. Smith, K. Coleman, and S. Marshall, 1998: Esti- future permafrost from climate models. J. Climate, in press. mating the size of the inert organic matter pool from total soil Tarnocai,C.,J.G.Canadell,E.A.G.Schuur,P.Kuhry,G.Mazhitova, organic carbon content for use in the Rothamsted carbon and S. Zimov, 2009: Soil organic carbon pools in the northern model. Soil Biol. Biochem., 30, 1207–1211. circumpolar permafrost region. Global Biogeochem. Cycles, 23, Friedlingstein, P., and Coauthors, 2006: Climate–carbon cycle GB2023, doi:10.1029/2008GB003327. feedback analysis: Results from the C4MIP model intercom- Weedon, G. P., and Coauthors, 2011: Creation of the WATCH parison. J. Climate, 19, 3337–3353. forcing data and its use to assess global and regional reference Gregory, J. M., C. D. Jones, P. Cadule, and P. Friedlingstein, 2009: crop evaporation over land during the twentieth century. Quantifying carbon cycle feedbacks. J. Climate, 22, 5232–5250. J. Hydrometeor., 12, 823–848. Harden, J. W., and Coauthors, 2012: Field information links per- Zhang, T., R. G. Barry, K. Knowles, F. Ling, and R. L. Armstrong, mafrost carbon to physical vulnerabilities of thawing. Geo- 2003: Distribution of seasonally and perennially frozen ground phys. Res. Lett., 39, L15704 doi:10.1029/2012GL051958. in the Northern Hemisphere. Permafrost, M. Phillips, S. M. Keuper, F., P. M. van Bodegom, E. Dorrepaal, J. T. Weedon, J. van Springman, and L. U. Arenson, Eds., Swets and Zeitlinger, Hal, R. S. P. van Logtestijn, and R. Aerts, 2012: A frozen feast: 1289–1294.

Unauthenticated | Downloaded 09/27/21 06:19 AM UTC