2.1 Model description

Based on the process-based land surface model ORCHIDEE-MICT (Guimberteau et al., 2018), the branch ORCHIDEE-PEAT was developed explicitly to represent peatland-specific hydrological, carbon and area dynamics. ORCHIDEE-PEAT has been tested at both local and regional levels, and a detailed description was provided by Qiu et al. (2018, 2019). Here, we give only a brief outline of the model.

Peatland is represented as an independent sub-grid hydrological soil unit (HSU) that has a high saturated water content (0.9 m³/m³), a moderately high saturated hydraulic conductivity (2.45 × 10⁻⁵ m/s), and an above-surface water reservoir (≤ 10 cm) (Largeron, Krinner, Ciais, & Brutel-Vuilmet, 2018). The peatland HSU receives surface runoff from non-peatland HSUs containing grasslands, forests and bare soil in each grid cell of the model, and has zero bottom water drainage (Largeron et al., 2018).

Vegetation in each grid cell is discretized into plant functional types (PFTs). One PFT is used to represent bare soil in the grid cell. Twelve PFTs are used to represent non-peat vegetation, including eight tree PFTs, two natural grass PFTs (C3 and C4) and two crop PFTs (Krinner et al., 2005). There is only one peat-specific PFT in the present study, representing an average of all vegetation growing in the ecosystem (see dedicated discussion on the peat-specific PFT by Qiu et al., 2018). Parameters of the peat-specific PFT, that is, rooting depth and the maximal rate of carboxylation, were calibrated extensively against observed gross primary productivity (GPP) and net ecosystem CO₂ exchange (NEE) from 30 northern peatland sites, thus it is not applicable for tropical peatlands. In this study, the fractional coverage of non-peatland PFTs is calculated as functions of bioclimatic limitations by the dynamic global vegetation model (DGVM) of ORCHIDEE described by Krinner et al. (2005) and recalibrated by Zhu et al. (2015). The simulation of peatland area relies on a cost-efficient TOPMODEL scheme to calculate the flooded area fraction of each grid cell and on an algorithm to identify the flooding frequency and the water and carbon balance conditions that are suitable for development of peatland (Qiu et al., 2019). Using sub-grid-scale (pixel) topography information (given by the HYDRO1k dataset, https://www.usgs.gov/centers/eros/science) of a given grid cell, TOPMODEL calculates a topographic index for each pixel. This topographic index can represent how likely a pixel is to be flooded (“floodability”) given the mean water table of a land surface model grid cell containing many pixels. According to the floodability of all pixels in the grid cell, soil properties of the grid cell (a parameter that describes the decrease of water transmissivity with soil depth and was calibrated grid by grid in this study; Supporting Information Figure S1a takes a Siberian grid cell as an example to show that the simulated flooded area fraction for a given grid-cell mean water table position varies with this parameter) and the simulated grid-cell mean water table, TOPMODEL redistributes the grid-cell mean water table depth to pixels to delineate the extent of sub-grid area at saturation with a water table of zero (Beven & Kirkby, 1979). Owing to the high spatial resolution of the topographic information, the above calculation is computationally costly.

Based on the rationale of Stocker et al. (2014), a cost-efficient version of TOPMODEL was adapted in our model to reduce the computational cost; an asymmetric sigmoid function is used to approximate the empirical relationship between the flooded fraction of a grid cell and its mean water table position calculated at each time step by the land surface model. The asymmetric sigmoid function is calibrated for each grid cell (Supporting Information Figure S1b,c shows the simulated present-day peatland area fraction and peat soil carbon density with the function being globally uniform, could be contrasted with simulations in the main manuscript which used grid-by-grid calibrated functions) so that the simulated maximum flooded area fraction of the grid cell matches with the present-day wetland areas that are regularly inundated or subject to shallow water tables after excluding lakes and permanent water bodies (Tootchi, Jost, & Ducharne, 2019). Then a scheme of peatland initiation and development is included in the land surface model, to distinguish areas that are suitable from the simulated wetland extent to become a peatland. Conditions required for peat inception and persistence are; (a) a minimum flooding duration over a long period (at least Num flooded months during 30 years, with Num being the number of growing season months during the 30 years); (b) a positive summer water balance; and (c) a long-term positive carbon balance. In our model, expanding peatlands encroach carbon from non-peatland vegetation, while carbon of contracted peatlands is given to non-peatland vegetation. The model keeps track of contracted peatland area so that expanding peatland first expands into the area where peatland used to be.

Soil water fluxes and storage are resolved by a multi-layer physically based soil hydrology scheme (Guimberteau et al., 2018). Decomposition and accumulation of peatland soil C are also parameterized as a multi-layer scheme (Qiu et al., 2019); soils are vertically discretizated into 32 layers (≤ 38 m, with exponentially coarser vertical resolution as depth increases). In each model layer, different base decomposition rates are prescribed for litter and soil carbon pools (three pools: active, slow and passive), and temperature and moisture modifiers are used to constrain decomposition in cold and/or dry/anoxic conditions. A downward movement of soil C is used to represent the upward build-up of peat C (Qiu et al., 2019).

Snow processes are simulated by snow scheme of intermediate complexity (Guimberteau et al., 2018; Wang et al., 2013); the snowpack is represented by three snow layers with variable snow density, conductivity and thickness. Thermal and water budgets are resolved inside the snowpack. Thawing, infiltration of melt water into the next lower layer and refreezing of liquid water are accounted for. Soil thermodynamics, energy exchanges between the soil and the atmosphere, are calculated based on a one-dimensional Fourier equation. Soil water freeze–thaw processes and the latent heat exchanges involved in soil water phase change, in addition to its impacts on soil thermal and hydrological properties, are represented (Guimberteau et al., 2018). Permafrost processes, such as palsa formation and subsidence or collapse of the peat soil column in the event of ground ice thaw, are not modelled. In our model, soil freezing and permafrost: (a) affect soil carbon profiles, because the root distribution of the modelled plants and soil carbon inputs are within the local active layer thickness (ALT); and (b) reduce vertical water infiltration and movements and affect thermal states of the soil, thus affecting plant water availability and soil carbon decomposition, because only liquid water can move within frozen or partly frozen soil, and the hydraulic conductivity, heat capacity and heat conductivity of the soil depend on the water content in the soil in the liquid and solid phases (Gouttevin, Krinner, Ciais, Polcher, & Legout, 2012).

2.2 Model experiments

Bias-corrected, gridded climate outputs of two GCMs from the Inter-Sectoral Impact Model Inter-comparison Project (ISIMIP2b; Frieler et al., 2017) are used in this study to force ORCHIDEE-PEAT v.2.0 over the NH (> 30° N) at 2° × 2° spatial resolution. Critical parameters of the cost-efficient TOPMODEL were calibrated for each of the 2° × 2° grid cells, based on 1 km × 1 km topographic information of sub-grid pixels (HYDRO1k) in the grid cell. The grid-by-grid calibration ensures that ORCHIDEE-PEAT behaves consistently across different spatial resolutions in reproducing the observed extent of wetland area for the present day (Supporting Information Figure S2a simulation at 1° × 1° resolution versus 2° × 2° simulations in the main manuscript). Among the four GCMs that have been used for the ISIMIP2b forcing data, the IPSL-CM5A-LR GCM has the fastest increase in global mean temperature (GMT) relative to pre-industrial levels and the GFDL-ESM2M the slowest. We thus selected these two end-member GCMs to maximize the uncertainty range of projected climate change for a given RCP, which will translate into a GCM-induced uncertainty in our projections of the carbon balance of peatlands. The climate change scenarios we tested include a strong mitigation pathway (RCP2.6), a no-mitigation pathway (RCP6.0) and a worst-case scenario (RCP8.5).

Considering peatland areas that respond dynamically to climate change has two main effects on the carbon balance of peatlands. For instance, in response to an increase of peat area, there will be: (a) a direct change in carbon storage in response to CO₂ fluxes in the new areas conquered by peat; and (b) an indirect C storage effect from the encroachment of previous vegetation type by peat. This second process is a gain of carbon by peat compensated by the loss of carbon from the previous vegetation, thus with no effect on the atmospheric CO₂ budget.

To clarify the role of each effect in the carbon storage of northern peatlands, we conducted factorial simulations without peatland (SimNN) and with variable (SimYY) or fixed (SimYN) peatland areas (Supporting Information Figure S3). The peatland module is deactivated throughout the simulation period in SimNN, a baseline simulation similar to conventional land surface models that do not represent peat. The peatland module is activated in both SimYY and SimYN between the early Holocene up to the year 2005, with variable peatland areas. After the year 2005, the SimYN keeps peatland area fixed at the same value in the future, whereas SimYY continues with variable climate-driven peatland area. The difference between SimYY and SimYN isolates the effect of variable areas alone.

Agricultural drainage on peatland and peat fires are not represented explicitly in the model; thus, they are not considered in the present study. But we masked grid cells that are dominated by cropland (according to the MIRCA2000 dataset; Portmann, Siebert, & Döll, 2010) or by leptosols (according to the WISE soil database; Batjes, 2016) from the output of the model (the same grid cells were masked in all the simulations irrespective of whether or not peatland occurs in those grid cells).

3.1 Simulated area of peatlands and C stock

3.1.1 Present-day patterns

The simulated present-day (1986–2005 mean) area of the northern peatlands is 5.7 and 5.8 million km² with the IPSL-CM5A-LR and GFDL-ESM2M historical climate forcing, respectively (Figures 1a and 2a). The simulated spatial distributions of peatland area are similar between these two simulations, except for West Siberia, where IPSL-CM5A-LR produces areas closer to observations (Figure 1b; Supporting Information Figure S4) and GFDL-ESM2M gives an overestimation (Figure 2b). Similar peat areas are found between the two GCMs, mainly because during the historical period, they are both bias-corrected to the 1979–2013 observational climate datasets (EWEMBI) (Frieler et al., 2017; Hempel, Frieler, Warszawski, Schewe, & Piontek, 2013). Compared with observation-based estimates of northern peatland area of 3.4–4 million km², both simulations overestimate peatland area, and this overestimation mainly occurs in Southeastern USA, the Canadian Arctic Archipelago, the Russian Far East and the China–Russia border (Supporting Information Figure S4).

The simulated present-day northern peatlands carbon stock is 511 and 470 PgC with the IPSL-CM5A-LR and GFDL-ESM2M climate forcing, respectively (Figures 3a and 4a). Both values are at the upper end of the range (270–540 PgC) given by inventories. Both peatland area and carbon stocks are found to be overestimated in Western and Central Europe, where peatlands have been converted extensively to agriculture or forest (Joosten, 2010).

3.1.2 Past and future trends

As shown in Figures 1a and 2a, the simulated area of northern peatlands increased historically and continues to increase in the future in SimYY simulations for all scenarios and GCMs. When the model is forced by the IPSL-CM5A-LR climate forcing, peatland area increases faster under RCP2.6 than under RCP6.0. With the GFDL-ESM2M climate forcing, similar increasing trends of peatland area are found in both scenarios. With the larger warming of the IPSL-CM5A-LR model, we found a decrease of peatland area in the future in the Southern USA, part of Western Canada, Western Europe and part of Northern Europe under both RCPs (Figure 1c and 1d). Other regions generally show a continuing expansion of peatland extent; in particular, over Northern Asia. With the smaller warming of the GFDL-ESM2M model, we found only a small reduction in peatland area in Western Europe and Southern USA, and an expansion in the Northwest Territories of Canada and West Siberia, for both RCPs (Figure 2c,d).

Simulated total soil C stock of northern peatlands also increased historically and continues increasing in the future (Figures 3a and 4a). If future peatland area is held constant from 2006 to 2099 (SimYN), the simulated peatland soil C stock with IPSL-CM5A-LR and GFDL-ESM2M forcing under RCP2.6 increases by 23 and 28 PgC, respectively, with the largest increases being found over Eastern North America, Western Europe, West Siberia and the China–Russia border (Figures 3c,d and 4c,d). The predicted increase in peatland soil C stock under RCP6.0 is similar to that under RCP2.6, indicating that with fixed peat area, changes in carbon stocks are not very sensitive to climate change in our peat model. In comparison, a strong increase of peatland C stock is predicted with a variable peatland area (SimYY). This increase of peat carbon is not because more CO₂ is removed from the atmosphere, but rather because expanding peatlands encroach carbon from non-peatland vegetation.

In SimYN, from the present day (1986–2005 average) to the end of the century (2080–2099 average), the total northern soil carbon storage (peatland + non-peatland) is projected to increase by 62 PgC under RCP2.6 with IPSL-CM5A-LR, with a 23 PgC increase in peatland and a 39 PgC increase of non-peatland biomes (Supporting Information Table S1). This shows the remarkable contribution of peatlands to future soil carbon storage increase in the NH (37.1%) despite the small fraction of area they occupy (8.8%). In contrast, in SimYY for the same GCM and RCP scenario, peatland expands over non-peatland biomes and encroaches their carbon, leading to a total northern soil C storage increase of 75 PgC, split into a 85 PgC increase in peatland and a 10 PgC decrease in non-peatland vegetation (Supporting Information Table S1). Thus, although the total change in soil carbon is similar between SimYY and SimYN, its attribution to peat versus non-peat is very different. Figure 5 displays the partitioning of the change of northern peatland soil C stock from the present day to the end of the century in SimYY between: (a) C sequestered from the atmosphere into the present-day peatland; and (b) C encroached from non-peatland biomes and C sequestered from the atmosphere by the new peatland areas.

From 1861 to 2005, we estimated that the simulated peatland soil C storage increased by 79 (Figure 3a; IPSL-CM5A-LR) and 72 PgC (Figure 4a; GFDL-ESM2M), which is attributed to C encroached from non-peatland biomes when peatland expands and C is sequestered by peatland from the atmosphere; 14.5 PgC sequestration from the atmosphere by peatland (Section 3.2; peatland annual NEE over 1861–2005 being −0.10 PgC/year) and 64.5 PgC from encroachment of peatland on non-peatland biomes with the IPSL-CM5A-LR climate forcing, and 16 PgC sequestration from the atmosphere by peatland (Section 3.2; peatlands annual NEE over 1861–2005 being −0.11 PgC/year) and 56 PgC from encroachment of peatland on non-peatland biomes with the GFDL-ESM2M climate forcing.

3.2 Historical and future peatland net ecosystem exchange

Simulated annual NEE (a negative NEE value indicates a CO₂ sink) of all northern peatlands from 1861 to 2005 ranges from −0.02 to −0.18 PgC/year with the IPSL-CM5A-LR forcing and from 0.01 to −0.25 PgC/year with the GFDL-ESM2M (Figure 6). The long-term average NEE sink of atmospheric CO₂ is −0.10 (IPSL-CM5A-LR) and −0.11 PgC/year (GFDL-ESM2M), which matches published estimates of −0.07 to −0.1 PgC/year (Clymo, Turunen, & Tolonen, 1998; Frolking et al., 2011; Gorham, 1991). The spatial distribution of simulated present-day peatland NEE is shown in Figure 7a (IPSL-CM5A-LR) and in Figure 7d (GFDL-ESM2M). Peatlands in the Hudson Bay Lowlands, West Siberia and across the China–Russia border appear as hotspots for CO₂ sequestration in both figures. To evaluate the magnitude of simulated NEE, we compared model output with eddy covariance (EC)-measured NEE at 30 northern peatland sites (Qiu et al., 2018) with simulated values extracted from grid cells corresponding to these sites. The model shows good performance in capturing observed monthly mean NEE across these peatland sites (Supporting Information Figure S5).

Assuming that the extant peatland will neither expand nor contract (SimYN) in the future, the strength of peatlands as a CO₂ sink in Western Canada, Western Europe and the China–Russia border is predicted to decrease under both RCPs with the IPSL-CM5A-LR climate forcing (Figure 7a–c). The decrease under RCP6.0 is stronger than under RCP2.6, with some peatlands in the above regions predicted to turn from present-day CO₂ sinks to future CO₂ sources as a result of simulated soils that are drier and warmer than at present under RCP6.0 (Supporting Information Figure S6a,c). In contrast, although CO₂ sink strengths of peatlands in Western Canada and the China–Russia border are also predicted to decrease with the GFDL-ESM2M forcing, these regions remain CO₂ sinks in the future (Figure 7d–f). The GFDL-ESM2M forcing projects higher precipitation and lower temperature (especially in summer) in Western Canada and the China–Russia border than the IPSL-CM5A-LR forcing (Supporting Information Figures S7 and S8), producing a smaller increase in soil temperature and wetter-than-present soil conditions (Supporting Information Figure S6b,d). Peatlands in Western Europe, however, are predicted to turn from present-day CO₂ sinks to future CO₂ sources with the GFDL-ESM2M forcing under RCP6.0 (Figure 7f), owing to drier and warmer peat soils in the future.

Simulated peatland annual NPP, HR and NEE are shown in Figure 6. Peatland NPP and HR both increased from 1861 to 2005. From 2006 to 2099, the model gives different results for the two RCPs. The NPP and HR keep on increasing until the end of the century under RCP6.0, but they increase until middle of the century and then stabilize under RCP2.6. Calculated as the difference between HR and NPP (a negative NEE value being a CO₂ sink), the annual NEE of northern peatlands became more negative from 1861 to 2005; that is, an intensified CO₂ sink. In the future, the strength of the peatland CO₂ sink is predicted to increase until the middle of the century before decreasing, in both RCPs. The annual mean NEE of northern peatlands is −0.18 PgC/year under both RCPs with the IPSL-CM5A-LR forcing and −0.21 PgC/year with GFDL-ESM2M, from 2006 to 2099. When spatio-temporal changes of peatland area are modelled (SimYY), the difference in annual mean NEE of all northern peatlands from SimYN simulations is negligible (Supporting Information Figure S9), although SimYY-predicted peatlands in Western Canada and Western Europe are larger CO₂ sources than in SimYN (Supporting Information Figure S10).

3.3 Impact of peatlands on northern terrestrial carbon balance

To quantify the impact of peatland on future (2006–2099) northern terrestrial carbon balance, we first look at the carbon storage effect of peatland with fixed future peatland areas. Figure 8a,b shows that northern peatlands alone remove a cumulative amount of atmospheric CO₂ of 17–20 PgC in the future (range from the two GCMs and two RCPs). The CO₂ sequestration by peatlands is similar between RCP2.6 and RCP6.0, but non-peatland biomes are larger CO₂ sinks under RCP6.0.

For RCP2.6, the overall (peatlands + non-peatlands) northern biomes are predicted to sequester 146 (IPSL-CM5A-LR) and 194 PgC (GFDL-ESM2M) from the atmosphere from 2006 to 2099 without peat (SimNN), and the sequestration is reduced by 2 (IPSL-CM5A-LR) and 23 PgC (GFDL-ESM2M) when peatland is included (SimYN) (Figure 8a,b). Under RCP6.0, the cumulative future northern NEE is a sink of 164 (IPSL-CM5A-LR) and 229 PgC (GFDL-ESM2M) without peat (SimNN), and the sink strength is reduced by 4 (IPSL-CM5A-LR) and 24 PgC (GFDL-ESM2M) when peatland is included (SimYN). In summary, adding peatland in a model leads to smaller CO₂ uptake from the atmosphere by the overall northern biomes. This is because although more carbon is stored in the northern soils owing to the anaerobic peat carbon decomposition in SimYN than in SimNN, carbon accumulated in the northern vegetation biomass of SimNN is larger than in any of the model runs where peatland is represented (Figure 8c,d). The choice of a GCM has more impact on the reduction of northern hemisphere CO₂ sink strength by peatland than the choice of an RCP scenario; the decrease in biomass carbon gains owing to the inclusion of peatland is almost compensated by the increase in soil carbon gains with IPSL-CM5A-LR, whereas the increase in soil carbon gains when peatland is included is not great enough to compensate the decrease in biomass carbon gains with the cooler and wetter GFDL-ESM2M (Figure 8).

Next, we look at the northern terrestrial carbon storage with variable peatland areas. With the IPSL-CM5A-LR climate forcing, SimYY predicted 13 (RCP2.6) to 17 PgC (RCP6.0) more overall soil carbon accumulation than SimYN. With the GFDL-ESM2M climate forcing, 8 (RCP2.6) to 9 PgC (RCP6.0) more carbon is accumulated for SimYY than for SimYN (Figure 8). In contrast, overall biomass carbon gain for SimYY is 7–9 PgC (range from both RCPs and GCMs) smaller than for SimYN. In summary, the expansion of peatland area in the future (SimYY) always results in increase of soil carbon gains and decrease of biomass carbon gains of the NH biomes. For both GCMs and RCPs, the decrease in biomass carbon gains induced by an increase in peatland area is compensated by the increase in soil carbon gains.

The data in Figure 8 show that there is always a larger increase in northern vegetation biomass carbon under RCP6.0 than under RCP2.6. IPSL-CM5A-LR always predicts smaller soil carbon uptakes under RCP6.0 than under RCP2.6, whereas GFDL-ESM2M always predicts larger soil carbon uptakes under RCP6.0 than under RCP2.6.

4.1 The fate of northern peatlands under rapid climate change

Previous studies, most of them conducted at the site level or in a small region, show that the stability and developmental trajectory of peatland ecosystems are divergent in response to climate warming; from severe degradation accompanied by a reduction in the C sink capacity of the peatland (Borge, Westermann, Solheim, & Etzelmüller, 2017; O’Donnell et al., 2012) to continued or enhanced C accumulation (Swindles et al., 2015; Turetsky et al., 2007). Occupying the drier end of peatlands distributional range, Western Canadian peatlands are expected to undergo a severe to extremely severe impact of climate warming (Kettles & Tarnocai, 1999; Tarnocai, 2006). Our model indeed predicts a notable lateral contraction and switch from CO₂ sinks to sources of peatlands in Western Canada attributable to climate warming and drying with the IPSL-CM5A-LR climate forcing, under the RCP6.0 climate change scenario. These peatlands are affected by climate change to a lesser extent under the lower warming scenario. The GFDL-ESM2M climate forcing, which projects a wetter and cooler climate in the future than the IPSL-CM5A-LR, preserves the peatland areas of Western Canada and their CO₂ sink function, in both RCP scenarios. In Western Europe, shrinking of peatlands occurs in all of the four combinations of RCPs and GCMs, however. Yet, western European peatlands are predicted to become CO₂ sources only under the stronger warming scenario, RCP6.0. The two largest peatland complexes in the Western Siberian lowlands (WSL) and the Hudson Bay Lowlands (HBL) do not respond as strongly to warming as do Western Canadian and Western European peatlands; with both area and the CO₂ sink strengths of peatland in WSL and HBL being predicted to increase in all four combinations.

To project the future extent of peatlands, bioclimatic envelope models and ecological niche models have been used. Oke and Hager (2017) predicted that the areas of suitable climate in North America for Sphagnum peatland should increase in the future up to 2050. Progressive losses of suitable climatic area for peatlands in Western Europe in the future were also predicted by bioclimatic models (Gallego-Sala & Prentice, 2013; Jones, Donnelly, & Albanito, 2006). Although bioclimatic models can highlight peatland regions at risk under projected climate change, they cannot resolve changes in peat C decomposition and accumulation. Chaudhary et al. (2017) used the process-based DGVM model LPJ-GUESS with dynamic multi-layer peatland and permafrost dynamics across 180 randomly selected grid points. Simulated peat C accumulation rate among these points was interpolated to assess pan-Arctic peatland C accumulation under past, present and future climate. Their results showed that under the RCP8.5 scenario, peatland in Europe and Scandinavia, Central and Eastern Canada and European Russia would turn into C sources by 2099, whereas they will be stronger sinks in Alaska, Western and Northern Canada, Siberia and the Russian Far East. Wania et al. (2009) chose six peatland sites from Alaska, Canada, Sweden and Russia and conducted a suite of sensitivity experiments using the LPJ dynamic vegetation model to explore the influences of future climate conditions on vegetation dynamics and changes of NPP, HR and NEP. With future climate from the ECHAM5 GCM and a CO₂ concentration of 780 ppmv, their model projected that NEP of sites from Alaska, Quebec, Sweden and the Russian Far East will increase following the trajectory of future climate conditions, the site from the Western Siberia will stay stable, but NEP of the Western Canada site will decrease.

Our model predicts a retreat of permafrost zones and an increase of active layer thickness (ALT) in the future (Supporting Information Figure S11). As permafrost thaws, soil drying attributable to increased drainage and concomitant decrease of wetland area and an increase of CO₂ emissions were reported by some studies (Avis, Weaver, & Meissner, 2011; Lawrence, Koven, Swenson, Riley, & Slater, 2015), whereas other studies suggested wetter soils and an increase of C accumulation (Swindles et al., 2015). Peat soil water content in northern Russia is predicted to increase by our model, because precipitation is projected to increase in that region (Supporting Information Figure S6); peatland extent is predicted to increase accordingly (Figures 1 and 2). However, warmer soil temperature drives faster peat C decomposition, and therefore, peatlands in northern Russia are predicted to become weaker CO₂ sinks in the future (Figure 7). It is worth noting that the thawing of permafrost might expand the rooting zone, increase plant-available N in soils and induce stimulation of plant production and an alteration of species composition (Keuper et al., 2012). Additional understanding of the impact of release of N at the thaw front on plant production, decomposition and species composition is needed before parameterizing it into the model.

4.2 Peat C accumulation and areal expansion to mitigate climate change

Field studies and laboratory experiments show that the CO₂ fertilization effect on plant growth should be smaller in peatlands than in other northern ecosystems. On the one hand, some moss species were found not to be stimulated by elevated CO₂ (Berendse et al., 2001; Toet, 2006). On the other hand, the effects of CO₂ on growth of vascular plants in peatland might be limited by low nitrogen availability in oligotrophic peatlands (Hoosbeek et al., 2001). Mosses are not represented explicitly in our model, but parameters for the peatland PFT were calibrated against ecosystem productivity of 30 northern peatland sites (Qiu et al., 2018), making it feasible to predict the ecosystem-level response of northern peatland to environmental changes.

Under the warmer and drier climate of the IPSL-CM5A-LR climate projections, the increase in northern soil carbon accumulation as a result of peatland expansion and the consequent decrease in aerobic decomposition (SimYY versus SimNN) exceeds the reduction in biomass carbon accumulation in the future. The reverse result is predicted under the cooler and wetter climate of the GFDL-ESM2M climate projections. Although the CO₂ fertilization effect on plant biomass might reach a saturation point, the ecosystem respiration will keep on increasing as a result of future warming and drought (Peñuelas et al., 2017). By conserving carbon in anoxic soils, peatlands can dampen the impacts of warming and drought on C sinks.

4.3 Outlook

The area of northern peatlands is simulated to have increased by 0.6 and 0.8 million km² from 1861 to 2005 and is predicted to increase over the coming century by 8 and 10% with the IPSL-CM5A-LR and GFDL-ESM2M forcing, respectively. The simulated expansion of the area of peatlands, which is in contrast to the conventional wisdom that the area of peatlands has remained relatively stable during the last millennium (Yu et al., 2010), could be attributed mainly to the following response in our model: with greater soil water-holding and retention capacity, a higher peatland HSU fractional area increases the simulated inundation area by raising the grid cell mean water table level, thus inducing a positive feedback on the area of peatland (Stocker et al., 2014). In the real world, the lateral expansion of peatland depends greatly on the slope and topography of the basin, with greater lateral expansion rates being observed when peatland spreads over flatter terrains of a basin than when the expansion is towards the steeper margin of the basin, when it slows down (Loisel, Yu, Parsekian, Nolan, & Slater, 2013). However, in our model, the inundated fraction of a grid cell is calculated by TOPMODEL according to topographic information (HYDRO1k) of 1 km × 1 km pixels in the grid cell, thus effects from margins being steeper than the centre of a basin could not be represented. Moreover, although the HYDRO1k represents modern-day topography, the accumulation of peat over time is a process of “basin infilling”, which smoothens over the basin surface over long time-scales. The use of modern-day topography to calculate the extent of peatland during the Holocene might thus overestimate Holocene peatland expansion rates for some regions, exacerbating the positive feedback of peatland HSU on area in our model.

The overestimation of the area of northern peatlands might cause biases on the simulated peatlands NEP. We estimate this impact crudely by masking the model result with an observation-based peatland map [meaning that only grid cells that have peatland in PEATMAP (Xu et al., 2018) are accounted for]; the simulated present-day northern peatlands area after masking is 4.4 (IPSL-CM5A-LR) and 4.6 million km² (GFDL-ESM2M), closer to previous estimates of 3.4–4 million km². These peatlands are predicted to sequester 13 (IPSL-CM5A-LR under both RCP2.6 and RCP6.0) and 16 PgC (GFDL-ESM2M under both RCP2.6 and RCP6.0) over 2006–2099 (Supporting Information Figure S12).

Under the worst-case scenario, RCP8.5, with spatio-temporal dynamics of peatland area being calculated by the model SimYY, the model predicts a significant reduction in area in the above peatland regions (Supporting Information Figure S13). Peatlands in Western Canada (with the IPSL-CM5A-LR climate forcing), Southeastern USA (IPSL-CM5A-LR), the China–Russia border (IPSL-CM5A-LR) and Southern and Western Europe (IPSL-CM5A-LR and GFDL-ESM2M) are predicted to become significant sources of CO₂ by the end of the 21st century (Supporting Information Figure S14). From 2006 to 2099, a cumulative amount of 6 (IPSL-CM5A-LR) to 17 PgC (GFDL-ESM2M) is removed from the atmosphere by northern peatlands. By the end of the century, northern peatlands are predicted to be either a CO₂ source (IPSL-CM5A-LR) or near neutral (GFDL-ESM2M) under RCP8.5 (Supporting Information Figure S15).

In this study, we use the ORCHIDEE-PEAT v.2.0 LSM to explore the impact of northern peatlands on the carbon balance of northern terrestrial ecosystems, but C losses from peatlands as methane and dissolved organic carbon are not included in the model simulation.

Fires and drainage are two most important disturbances to peatland, causing substantial emissions of CO₂ from peatlands (Benscoter & Wieder, 2003; Carlson et al., 2016). Although neither of them is investigated in this study, there is no doubt that peatland protection and restoration are important climate change mitigation strategies (Leifeld & Menichetti, 2018).