Warming of the global climate will lead to widespread permafrost thaw and degradation with impacts on ecosystems, infrastructure, and emissions that amplify climate warming (Oberman, 2008; Callaghan et al., 2011; Schuur et al., 2015). Permafrost-affected soils in the high northern latitudes contain 1300 ± 200 Gt of carbon, where ∼ 800 Gt C is preserved frozen in permafrost with ∼ 550 Gt C in the top 3 m of soil (Tarnocai et al., 2009; Hugelius et al., 2014). As permafrost thaws, organic matter frozen within permafrost will thaw and decay, which will initiate the permafrost carbon feedback, releasing an estimated 120 ± 85 Gt of carbon emissions by 2100 (Schaefer et al., 2014). The wide range of estimates of carbon emissions from thawing permafrost depends, to a large extent, on the ability of models to simulate present permafrost extent (Brown et al., 1997). For example, the simulated permafrost in some models is significantly more sensitive to thaw, with corresponding larger estimates of carbon emissions (Koven et al., 2013). Narrowing the uncertainty in estimated carbon emissions requires improvements in how land surface models (LSMs) represent permafrost thermal and carbon dynamics.

The active layer in permafrost regions is the surficial soil layer overlying the permafrost, which undergoes seasonal freeze–thaw cycles. Active layer thickness (ALT) is the maximum depth of thaw at the end of summer. LSMs used to estimate emissions from thawing permafrost typically assume that the frozen carbon is located in the upper permafrost above 3 m depth and below the maximum ALT (Koven et al., 2011; Schaefer et al., 2011; MacDougall et al., 2012). Thus, the simulated ALT determines the volume of permafrost in the top 3 m of soil, and thus the initial amount of frozen carbon. Consequently, any biases in the simulated ALT will influence the initial amount of frozen carbon, even if different models initialize the frozen carbon in the same way. In addition, the same thermal biases that lead to deeper simulated active layers lead to warmer soil temperatures, making the simulated permafrost more vulnerable to thaw and resulting in higher emissions estimates (Koven et al., 2013).

The surface organic layer (SOL) is the surface soil layer of nearly pure organic matter that exerts a huge influence on the thermodynamics of the active layer. The organic layer thickness (OLT) usually varies between 5 and 30 cm, depending on a balance between the litter accumulation rate relative to the organic matter decomposition rate (Yi et al., 2009; Johnstone et al., 2010). A recent model intercomparison study shows that LSMs need more realistic surface processes such as an SOL and better representations of subsoil thermal dynamics (Ekici et al., 2015). The low thermal conductivity of the SOL makes it an effective insulator, decreasing the heat exchange between permafrost and the atmosphere (Rinke et al., 2008). The effect of the SOL has been well presented in several modeling studies. For example, Lawrence and Slater (2008) showed that soil organic matter affects the permafrost thermal state in the Community Land Model, and Jafarov et al. (2012) discussed the effect of the SOL in the regional modeling study for Alaska, United States. Recently, Chadburn et al. (2015a, b) incorporated an SOL in the Joint UK Land Environment Simulator (JULES) model to illustrate its influence on ALT and ground temperatures both at a site-specific study in Siberia, Russia, and globally. In essence, the soil temperatures and ALT decrease as the OLT increases. Consequently, how (or if) LSMs represent the SOL in the simulated soil thermodynamics will simultaneously determine the initial amount of frozen permafrost carbon and the vulnerability of the simulated permafrost to thaw.

In this study we improved present-day frozen carbon stocks in the Simple Biosphere/Carnegie-Ames-Stanford Approach (SiBCASA) model to reduce biases in initial permafrost carbon stocks and improve the dynamics of future permafrost carbon release. To achieve this we introduce three improvements into the SiBCASA model: (1) improve the soil thermal dynamics and ALT, (2) improve soil carbon dynamics and build-up of carbon stocks in soil, and (3) initialize the older, frozen carbon using observed circumpolar soil carbon (Hugelius et al., 2014).

We used the SiBCASA model (Schaefer et al., 2008) to evaluate current soil carbon stocks in permafrost affected soils. SiBCASA has fully integrated water, energy, and carbon cycles and computes surface energy and carbon fluxes at 10 min time steps. SiBCASA predicts the moisture content, temperature, and carbon content of the canopy, canopy air space, and soil (Sellers et al., 1996; Vidale and Stockli, 2005). To calculate plant photosynthesis, the model uses a modified Ball–Berry stomatal conductance model (Ball, 1998; Collatz et al., 1991) coupled to a C3 enzyme kinetic model (Farquhar et al., 1980) and a C4 photosynthesis model (Collatz et al., 1992). It predicts soil organic matter, surface litter, and live biomass (leaves, roots, and wood) in a system of 13 prognostic carbon pools as a function of soil depth (Schaefer et al., 2008). The model biogeochemistry does not account for disturbances, such as fire, and does not include a nitrogen cycle. SiBCASA separately calculates respiration losses due to microbial decay (heterotrophic respiration) and plant growth (autotrophic respiration).

SiBCASA uses a fully coupled soil temperature and hydrology model with explicit treatment of frozen soil water originally from the Community Climate System Model, version 2.0 (Bonan, 1996; Oleson et al., 2004). To improve simulated soil temperatures and permafrost dynamics, Schaefer et al. (2009) increased the total soil depth to 15 m and added the effects of soil organic matter on soil physical properties. Simulated snow density and depth, and thus thermal conductivity, significantly influence simulated permafrost dynamics, so Schaefer et al. (2009) added the effects of depth hoar and wind compaction on simulated snow density and depth. Recent model developments include accounting for substrate availability in frozen soil biogeochemistry (Schaefer and Jafarov, 2015).

We spun SiBCASA up to steady-state initial conditions using an input weather data set from the modified Climatic Research Unit National Center for Environmental Prediction (CRUNCEP)

ftp://nacp.ornl.gov/synthesis/2009/frescati/temp/land_use_change/original/readme.htm

The dynamic SOL decreased the simulated ALT on average 50 % across the domain and allowed the model to simulate permafrost in discontinuous zones where it could not before (Fig. 1). The area of near surface permafrost simulated with the current version of the model is equal to 13.5 mil km2, which is almost 38 % greater than without the dynamic SOL (Schaefer et al., 2011). This area is closer to the observed area from the International Permafrost Association: 16.2 mil km2 (Brown et al., 1997). Simulated ALT less than 2 m covers about 92 % of the area in the new simulations (Fig. 1b) in comparison to 66 % of the area in the Schaefer et al. (2011) simulations (Fig. 1a). The previous version of SiBCASA could not simulate permafrost in many parts of the discontinuous zone with relatively warm climate. Adding the dynamic SOL essentially decreased the thermal conductivity of the surface soil, allowing SiBCASA to simulate permafrost where the mean annual air temperatures are close to 0 ∘C.

To illustrate the improvement of the simulated ALT with respect to the observed data, we compared simulated ALT with measured values from Circumpolar Active Layer Monitoring (CALM) stations. The CALM network is a part of the Global Terrestrial Network for Permafrost (Burgess et al., 2000). The monitoring network measures ALT either using a mechanical probe or a vertical array of temperature sensors (Brown et al., 2000; Shiklomanov et al., 2010). After matching up the CALM coordinates with the coordinates of previously simulated ALT (Schaefer et al., 2011), we excluded sites with no measurements or ALT greater than 3 m depth, ending up with 76 CALM stations. Figure 2 shows simulated vs. observed ALT for the 76 CALM sites. The current simulations have a higher resolution than Schaefer et al. (2011) simulations, which allowed us to reach a higher order of heterogeneity between measured and simulated ALTs. The Pearson's correlation coefficient, R, is negative and not significant for the Schaefer et al. (2011) simulations (Fig. 2a), but is positive and statistically significant for the current simulations assuming p < 0.05 (Fig. 2b). The dynamic SOL greatly improves the simulated ALT, but SiBCASA still tends to overestimate ALT.

Figure 3 illustrates the effect of the frozen soil restrictions on phenology and GPP at a single point in central Siberia. Before applying a frozen soil restriction, SiBCASA maintained fine roots even in winter, resulting in root growth all year with a peak in spring corresponding to simulated leaf out (Fig. 3a). Simulated GPP was restricted by liquid water availability and was closely tied to thawing of the active layer, resulting in a lag as high as 60 days between leaf out and start of GPP in spring. Restricting growth and GPP to when the soil is thawed essentially synchronizes all phenological events to occur at the same time (Fig. 3b).

Restricting growth and GPP to when the soil is thawed delayed the onset of plant photosynthesis in spring in permafrost-affected regions. Introduction of the thawed root fraction in the model reduced GPP primarily in early spring. To illustrate the difference between unconstrained and restricted root growth (Fig. 3), we ran the model for 10 years for both cases. The difference between unconstrained and restricted root growth resulted in an overall ∼ 9 % reduction in annual GPP for the entire permafrost domain, nearly all of which occurred in spring.

To illustrate soil carbon distribution with depth we selected three representative areas: a continuous permafrost area corresponding to tundra type biome above the Arctic Circle, an area in the boundary of continuous and discontinuous permafrost corresponding to the boreal forest biome, and an area near the south border of the discontinuous permafrost corresponding to poorly vegetated–rocky areas. We calculated the mean and standard deviation of the carbon density distribution with depth for 200 grid points around each of the three selected locations. Simulated typical carbon densities from the selected locations are shown in Fig. 4. All profiles shown in Fig. 4 show a similar pattern: a 20–30 cm SOL with reduced carbon content at the bottom of the active layer. The SOL and permafrost carbon content matches observed values (Harden et al., 2012), but carbon content near the bottom of the active layer does not, most likely because our model does not include cryoturbation processes.

The decrease in ALT resulting from a dynamic SOL increases the volume of permafrost in the top 3 m of soil, greatly increasing the initial amount of frozen permafrost carbon in the simulations. Schaefer et al. (2011), without the dynamic SOL, assumed a uniform permafrost carbon density of 21 kg C m-3, resulting in a total of 313 Gt of permafrost carbon at the start of their transient run (Fig. 5a). To compare with the results of Schaefer et al. (2011), we initialized the permafrost carbon using the same assumed uniform carbon density and ran SiBCASA to steady-state initial conditions (Fig. 5b). Assuming the same uniform carbon density, the current version with the dynamic SOL results in a total of ∼ 680 Gt C compared to 313 Gt C in Schaefer et al. (2011). The dynamic SOL effectively doubled the volume of permafrost in the top 3 m of soil and the amount of simulated frozen carbon.

Initializing SiBCASA with the observed spatial distribution of permafrost carbon from the NCSCDv2 resulted in ∼ 560 Gt C of carbon stored in permafrost after spin-up, close to the observed value ∼ 550 Gt C in the top 3 m of soil (Hugelius et al., 2014). This does not mean that after the spin-up-simulated permafrost carbon stocks exactly matched the NCSDCv2 data. In discontinuous zones, for example, if the model simulated permafrost, it tended to produce a deeper ALT and thus less permafrost carbon than the NCSCDv2. Assuming a uniform permafrost carbon density does not account for the spatial heterogeneity in permafrost carbon and overestimates the total amount of permafrost carbon compared to the NCSCDv2 (680 Gt C vs. 550 Gt C, see Fig. 6a and b). The spatial correlation between simulated and observed permafrost carbon is 0.63 when initializing with the NCSCDv2 (Fig. 6c), compared with a spatial correlation of 0.12 for the uniform permafrost carbon density. The amount and spatial distribution of permafrost carbon significantly improves when initializing with NCSCDv2.

Failure to simulate soil carbon in southeast Canada and southwest Siberia (see Fig. 6b and c) is attributed to deep ALT. These areas correspond to the peat lands. Our model uses the Harmonized World Soil Carbon Database (HWSD) (FAO et al., 2009) to initialize soil texture and related thermal properties. Deep layers of peat have low thermal conductivities, providing an ideal condition for permafrost existence. However, the HWSD does not address peat lands in southeast Canada and southwest Siberia.

The overestimation of soil organic carbon (SOC) in central Siberia results from coupling between GPP and ALT. The dynamic SOL and rooting depth strengthens the feedback between GPP and ALT (Koven et al., 2009). Higher GPP produces greater litter fall, which increases the input soil carbon at the surface and results in a thicker SOL. The dynamic SOL changes the properties of the near surface soil, resulting in a shallower ALT and cooler soil temperatures. The dynamic rooting depth accounts for a shallower ALT and modulates GPP accordingly. The cooler soil temperatures slow microbial decay and increase the carbon accumulation rate, which in turn increases the SOL and reduces ALT further. Eventually, this feedback results in the development of a peat bog. The changes we describe here indicate that SiBCASA can simulate the dynamics of peat bog development, but the model does not yet include a dynamic vegetation model to account for conversions between biome types, such as boreal forest to peat bog.

The overall amount of permafrost carbon is less than that calculated assuming a uniform frozen carbon distribution. It is important to note that the SOL, ALT, and the permafrost thickness are the same for both cases (Fig. 6a and b). This is due to the fact that in both cases soil carbon is added in the permafrost layer below the active layer. Consequently, the ALT does not change between simulations, and the volume of permafrost in the top 3 m of soil does not change as well. The smaller permafrost carbon stock simulated for the nonuniform case is mainly due to the fact that we did not initialize frozen carbon in regions where it is not present according to the NCSCDv2, such as the Brooks Range in Alaska.

The dynamic SOL insulates ALT from air temperature, allowing SiBCASA to simulate permafrost in many discontinuous permafrost regions where it could not before, consistent with previous results where changes in thermal properties associated with the presence of soil organic matter cooled the ground (Lawrence and Slater, 2008; Yi et al., 2009; Ekici et al., 2014; Chadburn et al., 2015a, b). In addition, our work confirms findings by Koven et al. (2009) showing that including SOL dynamics in the model improves agreement with the observed permafrost carbon stocks. However, to better simulate known permafrost distribution in the discontinuous permafrost zone, it is important to know the exact OLT. Unfortunately, in situ measurements of OLT are scarce and essentially lacking in most areas of continuous and discontinuous permafrost.

To investigate the influence of the environmental factors on ALT further, we looked at the relationship between ALT and near surface air temperature (NSAT), soil wetness fraction (SWF), downwelling long-wave radiation (DLWR), and snow depth (SD). The simulated ALT is most influenced by NSAT and soil SWF, with a slightly smaller influence by DLWR, and nonlinearly influenced by SD (Fig. 7). To show the influence of the NSAT, we averaged two early fall months over 10 years. The areas with deep simulated ALT correspond to annual NSAT > 1 ∘C in southwest Siberia and NSAT > 5 ∘C in southeast Canada with a statistically significant correlation of 0.62 (Fig. 7a). DLWR showed a similar, but slightly weaker relationship with ALT, with higher DLWR values in southeast Canada and southwest Siberia and statistically significant correlation of 0.45 (Fig. 7b). Figure 7c shows maximum simulated snow depth calculated over the last 10 years of the steady-state run. Zhang (2005) indicates that SDs of less than 50 cm have the greatest impact on soil temperatures. Our results show no correlation between SD and ALT, but the effects of snow on ALT are less obvious and depend on different physical processes, such as wind, snow metamorphism, and depth hoar formation (Sturm et al., 1997; Ekici et al., 2015; Jafarov et al., 2014). We also observe high SWF in southwest Siberia and southeast Canada (see Fig. 7d) where SiBCASA simulates deep ALT with a statistically significant correlation of 0.68, suggesting that wet soils modulate the insulating effects of the SOL (Lawrence and Slater, 2008). This work does not address the impacts of fire on soil thermodynamics and recovery from fire, both of which are strongly influenced by the changes in the SOL (Jafarov et al., 2013). Studies show that wildfires and climate change could substantially alter soil carbon storage (Yuan et al., 2012; Yi et al., 2010). In the current version of the model the topsoil carbon stays in the system and provides resilience to permafrost. However, in reality, the upper SOL could be removed by fire, which would alter soil thermal properties and perturb permafrost carbon stability.

This work shows that the dynamic organic layer directly improves the distribution of carbon in soil, as well as indirectly through the improved ALT. Initialization of the carbon according to the NCSCDv2 map allowed us to better match simulated soil carbon with the observed carbon distribution. Restriction of the root growth within the thawed layer prevented artificial accumulation of permafrost carbon. Our model developments improved both the total amount and the spatial distribution of simulated permafrost carbon. The total permafrost carbon increased from 313 to 560 Gt C, compared to the observed value of 550 Gt C, and the spatial correlation with the observed distribution increased from 0.12 to 0.63. These improvements indicate the importance of including these developments in all land surface models.

In addition, most of the LSMs calculate soil properties based on prognostic soil carbon and soil texture from HWSD. We found that HWSD does not include thermal properties of peat lands, which resulted in inaccurate modeling of the ALT at the southern boundaries of the permafrost domain in Canada and Russia.