The contribution of soil heterotrophic respiration to the
boreal–Arctic carbon (CO
Warming in the northern high latitudes (
Pronounced changes have occurred in the northern high latitudes, especially
during the shoulder seasons. Satellite remote sensing data sets over the past
several decades indicate reductions of 0.8–1.3 d decade
Previous studies reported that the combination of warming and a longer
snow-free season has led to widespread greening and enhanced vegetation
productivity in the northern latitudes, especially during the early growing
season (Aurela et al., 2004; Humphreys and Lafleur, 2011; Buermann et al.,
2013; Pulliainen et al., 2017). However, a detailed understanding of how
soil respiration and other belowground processes respond to climate
variability, especially during the cold season, remains elusive. Soil
respiration is mainly the product of respiration by roots (autotrophic) and
soil decomposers (heterotrophic), while it is generally difficult to
partition soil respiration into the heterotrophic and autotrophic components
(Phillips et al., 2017). In this study, we focus on the heterotrophic
component of soil respiration and assume it is the dominant component of
total soil respiration in northern ecosystems during the cold season due to
root dormancy (Tucker et al., 2014; Hicks Pries et al., 2015). Surface
warming and a longer snow-free season are associated with earlier soil
thawing and deeper active layer thickness (ALT) in permafrost regions, which
can result in enhanced soil respiration and reduced annual net carbon uptake
(Lund et al., 2012; Yi et al., 2018). Moreover, ALT deepening in permafrost
regions will likely lead to a longer zero-curtain period (i.e., soil
temperature persists around 0
Landscape-level processes can affect the amount and age of soil carbon
released to the atmosphere (Hobbie et al., 2000). An important feature of
boreal–Arctic landscapes is strong surface heterogeneity, driven by
relatively fine-scale microtopographic variability of the order of 0.1–10 m (Zona et al., 2011; Kumar et al., 2016; Grant et al., 2017a, b),
which can influence coarser landscape-level behavior. However, current
large-scale models generally operate at scales of 10–100 km and are too
coarse to resolve finer-scale surface heterogeneity and its influence on
active layer dynamics and soil carbon decomposition (Yi et al., 2015; Tao et
al., 2019). Satellite or airborne remote sensing can provide information on
land surface heterogeneity across large extents and may provide critical
constraints on model predictions of regional active layer changes, soil
carbon, and permafrost vulnerability. Therefore, the objective of this study
was to develop a process-based permafrost carbon model mainly driven by
satellite remote sensing data. The model was designed at an intermediate
scale (
The remote-sensing-driven permafrost model (RS-PM), developed by Yi et al. (2018, 2019), was coupled with a terrestrial carbon flux (TCF) model (Yi et al., 2015) to investigate the climate sensitivity of carbon fluxes across Alaska (Fig. 1), with a particular focus on the shoulder season. The soil decomposition model in the original TCF model was revised in this study to account for vertical soil carbon transport in order to better simulate the depth-dependent soil carbon distribution and respiration fluxes. The RS-PM model simulates soil temperature and changes in soil liquid water content due to soil freeze–thaw transitions along the soil profile, using remote-sensing-based land surface temperature (LST), snow cover information, and total soil moisture content as key model forcing. The RS-PM outputs were then used as inputs to the carbon model and as constraints on both the vegetation productivity and soil respiration. A brief description of the modeling framework is described here, with a focus on the revised soil decomposition model, while a detailed description on the RS-PM model is provided in the Supplement.
Flow diagram describing the modeling procedure and main input data sets used in this study. The terrestrial carbon flux model has two components, including the light use efficiency algorithm for vegetation productivity estimates and a soil decomposition model for soil heterotrophic respiration estimates. The main equations used for each modeling component are referenced in the appropriate modeling box.
The RS-PM follows the prototype of a detailed permafrost hydrology model
(Rawlins et al., 2013; Yi et al., 2015) but has a flexible structure
designed to use satellite remote sensing data as the key model drivers and for model parameterization. The RS-PM uses a numerical approach for simulating soil freeze/thaw (F/T) and temperature profiles down to 60 m below the surface, using 23 soil layers with increasing layer thickness at depth. The model also accounts for the effects of seasonal snow cover evolution, organic soil, and soil water phase change on soil F/T processes. Satellite-based LST and snow cover time series data were used as model drivers. Soil thermal properties were parameterized using soil moisture data from the Soil Moisture Active Passive (SMAP) Level 4 (L4) data assimilation system (Reichle et al., 2017). RS-PM validation using in situ measurements shows favorable model accuracy for ALT (mean
We coupled the RS-PM and TCF models to represent the influence of permafrost
active layer processes on net ecosystem CO
Our soil decomposition model uses multiple litter and soil organic carbon
(SOC) pools to characterize the progressive decomposition of fresh litter to
more recalcitrant materials, which include three litterfall pools and three
SOC pools, with relatively fast turnover rates, and a deep SOC pool, with slow turnover rates (Thornton et al., 2002). The litterfall carbon inputs were first allocated to the three litterfall pools and then transferred to the SOC pools through progressive decomposition. In a previous study (Yi et al., 2015), the litterfall and SOC pools were arbitrarily distributed at
different soil depths within the top 3 m of soils to account for
depth-dependent differences in litterfall and soil organic matter substrate
quality. However, in this study we model the profile of the carbon pools by
introducing a vertical dimension (
The main RS-PM inputs include LST, snow cover properties, and soil moisture
from global satellite and reanalysis data products. LST and soil moisture
records from the MODIS 8 d composite data set (MOD11A2; Wan and Hulley,
2015) and SMAP L4 9 km daily surface (5 cm depth) and root zone (0–1 m depth) products (L4SM, Reichle et al., 2017) were used to define the model boundary conditions and parameterize soil thermal properties (Yi et al., 2018). MODIS 500 m snow cover extent (SCE) data (MOD10A2; Hall and Riggs, 2016) were used to downscale snow depth and density data from the Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) (
Other ancillary data sets included the 30 m National Land Cover Database
(NLCD) 2011 (Jin et al., 2013), 50 m SOC estimates for Alaska (to 1 m depth; Mishra et al., 2017), and the global 9 km mineral soil texture data
developed for the SMAP L4SM algorithm (De Lannoy et al., 2014). The dominant
NLCD land cover type within each 1 km pixel was used to define the modeling
domain, with open water and perennial ice and snow areas excluded (Fig. 2).
The SOC inventory data were used to define the organic fraction of the top 10
model soil layers (
The Alaskan land cover map and the location of in situ sites used for model validation. The land cover types are aggregated from the 30 m NLCD map (Jin et al., 2013), while the following land cover classes were used in the model simulations: developed and barren land, forest, scrub/shrub, grassland/herbaceous, croplands, and wetlands. The percentage of each land cover type is provided alongside the color bar legend labels.
A dynamic litterfall allocation scheme, based on the satellite NDVI time series, was used in Yi et al. (2015) to account for litterfall seasonality. We revised this scheme for the present study to incorporate a vertical distribution of root turnover, which is required by the soil decomposition model. The total litterfall was partitioned into aboveground (leaves and woody components) and belowground (mostly fine roots) litterfall using prescribed ratios for each biome type (Table S1 in the Supplement). A constant turnover rate for each 8 d composite period was assigned to the woody components of litterfall. The turnover rates of the other components of litterfall, i.e., leaves and fine roots, were calculated based on the annual time series of MODIS NDVI, with more litterfall generally allocated during the latter half of the year. The belowground litterfall was distributed through the rooting depth based on a vertical root distribution profile (Jackson et al., 1996). The maximum root depth in permafrost areas was limited to the maximum thaw depth. Then, the total litterfall at each depth was first allocated to the three litterfall pools according to the substrate quality of each litterfall component, i.e., labile, cellulose, and lignin fractions, and then transferred to the SOC pools through progressive decomposition. Table S1 provides the main parameters of the TCF model for each biome type, which were largely consistent with the prior study (Yi et al., 2015).
We used four Alaska eddy covariance (EC) tower sites with multilayer soil
temperature or moisture measurements to evaluate the simulated carbon fluxes
and temperature sensitivity of ecosystem respiration. Table 1 lists the
relevant site characteristics. The Atqasuk site (US-Atq) is about 100 km
south of Utqiaġvik on the Alaskan North Slope and consists of a mixture
of tussock tundra and shrubs with some sedges and sandy soils (Davidson et
al., 2016; Arndt et al., 2019). The Ivotuk site (US-Ivo) is about 300 km
south of Utqiaġvik in the northern foothills of the Brooks Range and is
characterized as a mixed tussock tundra/moss composition on a gentle slope
(Arndt et al., 2019). Soil temperature measurements were available at 5, 15, and 30 cm at US-Atq and 5, 15, 30, and 40 cm at US-Ivo, with full annual cycles recorded in 2014 and 2015. The two boreal forests sites (US-Prr and US-Uaf) are located near Fairbanks, Alaska, and dominated by mature black spruce forest (Ueyama et al., 2014; Ikawa et al., 2015). The leaf area index is
Characteristics of the eddy covariance tower sites used for model validation. Note: Tsoil – soil temperature; SM – soil moisture.
We used two regional data sets to evaluate the model performance during the
cold season. Daily snow depth and soil temperature measurements at SNOwpack TELemetry (SNOTEL) sites across Alaska (
We used the Natali et al. (2019b) in situ winter soil CO
For all of the site comparisons, the model was run using the 1 km spatial input data sets described in Sect. 2.2, and the model outputs from the 1 km grid cell encompassing each validation site were extracted. For the winter flux comparison, 1 km grid cells with biome types inconsistent with the local in situ sites were removed prior to the comparison.
The permafrost carbon model was run at 1 km resolution and 8 d time step
from 2001 to 2017. The model domain encompassed the majority of the Alaska
land area (
Correlation analysis was used to examine the sensitivity of soil freeze-up
and carbon fluxes to snow cover changes and other environmental variables
across Alaska. We first calculated the onset of land surface freeze based on
the MODIS LST data, which was defined as the center date of the 8 d period
at which the mean LST during three consecutive 8 d periods dropped below 0
Finally, we used the gradient boosting regression (GBR) method to quantify
the contribution of selected environmental variables to the annual carbon
fluxes. The GBR method consists of a sequence of models, and each
consecutive model is developed based on the errors of previously added
models (Friedman et al., 2000). The above model-simulated annual carbon fluxes from
2002 to 2017 were used to train and evaluate the GBR models. We chose the
following nine contributing environmental factors or predictors of annual
carbon fluxes during the model fitting, including summer (June–August) NDVI,
annual freezing and thawing index, mean annual downward solar radiation,
root zone soil moisture during the thaw season, snow offset and onset, mean
snow depth averaged from January to March (representing annual maximum snow
depth), and snow depth during the early snow season (from October to
November). The GBR method was implemented using the sklearn package in
Python 2.7. The following method was used to determine the relative
importance of each predictor to the model predictive performance. We first
ran the model using all nine predictors, and the model results were referred to as the baseline simulation (
Previous studies have evaluated the performance of the RS-PM model in reproducing regional ALT patterns over the Alaskan domain (Yi et al., 2018) and the zero-curtain period in Arctic Alaska (Yi et al., 2019). Here we focus on assessing the model's capability of representing snow insulation effects and ecosystem carbon fluxes, particularly during the cold season.
The relationship between the normalized temperature amplitude difference
between surface air and 20 cm depth soil conditions (
Comparison of the snow insulation curve derived from in situ
measurements and model simulations at the Alaskan SNOTEL sites. The dark
line is drawn using the following parameters presented in Slate et al. (2017):
The model simulations showed overall favorable agreement with tower-based
8 d composite carbon fluxes at the two tundra sites (Fig. 4), including
strong correlation (
Model-simulated carbon fluxes and temperature sensitivity of
ecosystem respiration at two tundra sites (US-Ivo and US-Atq). “GPP1 obs”
and “GPP2 obs” represent the gross primary productivity (GPP) estimates derived using tower-based net ecosystem CO
At both sites, abrupt decreases in the model-simulated GPP and the net
carbon uptake occur during the peak growing season (Fig. 4a, c), which was
mainly due to imposed low-minimum temperatures and associated LUE reductions
defined by the MODIS nighttime LST observations. The largest GPP reductions
during the peak season were generally caused by very low nighttime LST,
which may have large uncertainties in cloudy sky conditions. In addition,
there is also large uncertainty imposed from the NEE partitioning method,
with different methods resulting in large differences (up to more than 1 g C m
The model-simulated soil temperatures showed overall good correspondence
with the in situ measurements over the soil profile (
The model-simulated carbon fluxes were also comparable to the in situ data
at the two boreal forest sites (Figs. 5 and S5). The model showed a
slight underestimation of GPP and the relationship between ecosystem respiration (Reco) at the US-Uaf site, with a respective mean bias of
Comparisons of model-simulated carbon fluxes with
tower-based estimates
The model-simulated ecosystem respiration showed a broadly similar response
to surface soil temperature during the cold season (October to April) relative to the in situ winter flux synthesis data from the larger Alaskan domain (Fig. 6). The temperature sensitivity of the winter flux shown here is generally similar to the temperature sensitivity curve at the two tundra sites (Figs. 4b and S2) when ecosystem respiration mainly consists of soil
respiration. The model indicates a rapid decrease in soil respiration as
soil temperature and unfrozen water content decrease. The in situ data
collected using chambers and the diffusion method show a similar response
pattern to the model; however, the EC data show large scattering in the
respiration temperature response and evidence of large winter carbon fluxes
when surface soil temperatures drop below
Effects of soil temperature on CO
The seasonal cycle of model-simulated carbon fluxes and the soil
heterotrophic respiration (Rh) from different soil depths averaged across
Alaska and within different permafrost regions is shown in Fig. 7. The model
simulations indicate that both GPP and Rh peak in July, while Rh persists
well into the cold season. There is a notable difference in the timing of
the Rh seasonal peak from different soil depths, with a longer temporal lag
for deeper soil layers. Figure 7c compares the seasonality of the Rh
fraction from different soil depths, averaged for regions with different
permafrost probability, using an ancillary permafrost map (Pastick et al.,
2015; Fig. S8). Southern Alaska has relatively low permafrost probability
(
Regional mean of model-simulated carbon fluxes (a), Rh
fluxes from different soil depths
Across Alaska, annual GPP from 2001 to 2017 shows overall positive
productivity trends mostly in western Alaska and the interior of Alaska (Fig. 8a), with 66.8 % of areas showing positive trends and 32.9 % of areas showing negative trends. However, only a very small portion of the areas show significant (
Temporal trends of model-estimated annual carbon fluxes from 2001 to 2017. For NEE, positive trends indicate decreasing net carbon uptake activity, while negative trends indicate enhanced net ecosystem carbon uptake.
The attribution analysis results using the GBR method confirmed that NDVI
and the annual thawing index are the two most important variables affecting the estimated annual carbon fluxes, which was generally consistent across
different vegetation types (Fig. 9). For annual GPP flux, NDVI was the most
important variable, followed by the annual thawing index and downward solar
radiation, while, for annual Rh fluxes, the annual thawing index was the most
important variable, followed by NDVI, with other variables playing a very
minor role. Despite the importance of the annual thawing index in controlling
annual GPP and Rh fluxes, the snow offset showed little importance to both
fluxes. This was likely due to the low temporal resolution of the MODIS snow
cover data (i.e., 8 d composite) used to calculate the snow offset, which
was calculated as the center date of the 8 d composite period. The low
temporal resolution of the snow offset and a strong correlation (
Mean relative importance values of selected environmental
variables in controlling model-estimated annual carbon fluxes in Alaska (in which
The model-simulated growing-season Rh shows overall positive trends during
the study period, while the contribution of surface (
Temporal trends of model-estimated total Rh, GPP, and
surface soil contribution to Rh (Rh fraction) during the early and peak
growing season from 2001 to 2017. In
The timing of snow offset or surface thaw onset shows the highest
correlation with the surface soil Rh fraction during the growing season but
with opposing respiration responses during the early (April–May;
Regional mean correlation coefficient between the environmental variables and estimated Rh fraction of surface (0–13 cm) soils during the growing season from 2001 to 2017. Unless indicated, the variables were calculated during the same period as the Rh fraction. The thaw onset was derived from MODIS LST data, and the snow offset was derived from Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA-2), downscaled snow depth data.
Total Rh during the early cold season, from September to November, shows
overall positive trends from 2001 to 2017, except for portions of interior
and southwestern Alaska, while the Rh contribution from surface (
Regional trends of total Rh
Regional mean correlation coefficient between the environmental variables and estimated surface (0–13 cm) soil contribution to total Rh during the early cold season (September to November). Unless indicated, the variables were calculated during the same period as the Rh fraction.
The spatial pattern in the soil respiration trends during the early cold
season can be explained by the temporal lag (days) between the onset of
surface freezing and freezing in deeper (23 cm) soil layers, i.e., the soil-freezing delay or the duration of the zero-curtain period in areas with
permafrost occurrence (Fig. 12). The model simulations show an advance of
Sensitivity of model-simulated soil-freezing process to
snow cover changes across Alaska. The mean
Comparisons of the statistical distribution of model-simulated carbon fluxes
at the 1 and 10 km resolutions show an enhanced NEE seasonal amplitude
from the coarser-scale model simulations (Fig. 13). A larger difference in
the distributions is seen in the model-simulated Rh fluxes, with slightly
reduced Rh flux during summer and enhanced Rh flux from October to November
at a 10 km resolution. The largest differences in the Rh fluxes occur in
October and November, with daily mean differences of
Comparisons of the statistical distribution of model inputs and
simulations at the 1 and 10 km resolutions across Alaska.
Based on the simulations of a newly developed 1 km permafrost carbon model,
we highlight the important role of snow cover variability in controlling
soil heterotrophic respiration and the CO
Our results show that earlier snow melting, associated with spring warming, enhances soil heterotrophic respiration throughout the growing season, leading to a reduction in net carbon uptake later in the growing season in Alaska (Fig. S12). Previous studies reported that earlier snow melting generally results in enhanced vegetation productivity and carbon uptake during the early growing season, consistent with our simulations, while its impact on net ecosystem exchange during the later growing season may vary with local climate and site conditions (Aurela et al., 2004; Humphreys and Lafleur, 2011; Pulliainen et al., 2017). The variable impact of snow on the seasonal carbon cycle can be explained by the divergent responses of vegetation productivity and Rh to soil moisture and soil temperature during the later growing season. Earlier snow melting in spring can lead to depleted soil water conditions during the later growing season, resulting in a decrease in vegetation productivity and weaker net ecosystem carbon sink activity, especially in the boreal region (Buermann et al., 2013; Sulla-Menashe et al., 2018). However, our simulations indicate that deeper soil warming associated with early snow melting is mainly responsible for the enhanced ecosystem carbon loss later in the growing season. Surface warming and earlier disappearance in spring snow cover are associated with a deeper thaw depth in the permafrost region (Park et al., 2016; Yi et al., 2018). Field studies have shown that deeper permafrost thawing is associated with enhanced ecosystem respiration and, thus, reduced carbon sink activity during the later summer (Natali et al., 2011; Lund et al., 2012; Webb et al., 2016). Other studies also indicate that ecosystem respiration may dominate the NEE response to spring snow cover conditions and warming in the Arctic tundra; however, divergent responses have been observed in different tundra ecosystems (Humphreys and Lafleur, 2011; Parmentier et al., 2011; Lund et al., 2012; Darrouzet-Nardi et al., 2019).
Our simulations also indicate that the arrival of seasonal snow cover and
the number of snow-free days after land surface freeze play a major role in
controlling subsurface soil freeze-up and soil respiration during the early
cold season. Earlier snow onset relative to surface freeze onset (i.e., a
short snow-free period after surface freezing) keeps the soil warm and
results in a longer soil-freezing delay and zero-curtain period in
permafrost areas, with enhanced soil respiration during the early cold
season (Fig. 11). Due to strong snow insulation effects, underlying soils
can remain unfrozen for a substantial period long after the surface soil
freezes, i.e., the zero-curtain period. Field studies have shown persistent
carbon emissions during this zero-curtain period and also throughout the
winter season, while the resulting cold-season soil carbon emissions may
partially offset or even exceed the growing season net carbon uptake
(Elberling and Brandt, 2003; Luers et al., 2014; Webb et al., 2016;
Euskirchen et al., 2017). A recent study showed that Alaskan ecosystems were
either a carbon source or carbon neutral during the recent observational
period (2012–2014) due to a large contribution of cold-season carbon
emissions, with larger emissions in the early cold season based on CO
However, large uncertainties are associated with cold-season carbon
emissions in our estimates and other studies based on either in situ data or
atmospheric inversions. An analysis using satellite and airborne CO
An important feature of boreal–arctic landscapes is strong surface
heterogeneity, which may not be well represented in current global-scale
models operating at the order of tens of kilometers or more (Koven et al.,
2013b; Yi et al., 2015; Tao et al., 2019). Our comparisons between the 1
and 10 km model simulations showed a nonnegligible influence of landscape
heterogeneity on the model-simulated CO
Our current and previous assessment of the permafrost soil model also identified several areas in which improvements should be made to enhance model capabilities, especially in boreal forest. Comparisons with in situ measurements indicate larger discrepancies between model ALT simulations and in situ data in the boreal interior of Alaska characterized by a greater density of woody vegetation overlain with discontinuous or sporadic permafrost (Yi et al., 2018). Model-simulated soil temperatures also showed a larger bias at the boreal forest sites in relation to the in situ winter flux synthesis data (Sect. 3.1.2). The larger apparent uncertainty may reflect poor model representation of the vegetation canopy influence on thermal energy loading at the soil surface. Previous studies have shown that the MODIS vegetation index, leaf area index, and tree cover data are sensitive to boreal forest structure and postfire disturbance recovery (Mastepanov et al., 2013). These data sets can be used to account for the temperature difference between the soil surface and canopy skin temperature indicated by the MODIS LST data for different vegetation categories, either through simple empirical models or more sophisticated approaches derived from canopy radiative transfer models (Paul et al., 2004; Verhoef et al., 2007; Dolschak et al., 2015).
In addition, better understanding of the scaling behavior of environmental
controls on soil moisture is needed to improve model representation of
active layer conditions and carbon emissions (Mishra and Riley, 2015).
Previous studies indicate that topography and soil conditions are the
dominant factors affecting soil moisture variability at finer scales (Crow
et al., 2012), which are not sufficiently represented by the
coarse resolution (
Other notable uncertainties in the model-estimated carbon fluxes include
insufficient representation of the soil moisture migration with permafrost
thaw and winter processes. Earlier spring thaw and snowmelt have been linked
with active layer deepening and permafrost degradation, exacerbating the
soil water deficit during the later growing season, especially in the
southern boreal forest areas (Buermann et al., 2013; Park et al., 2016;
Zhang et al., 2019). Using external soil moisture inputs, the current
permafrost model was not able to fully represent this phenomenon, which
requires a more complete depiction of soil water, energy and carbon
processes, and linkages (Walvoord and Kuryly, 2016). On the other hand,
insufficient winter process representation in our model may partly explain
the inconsistency between the model-simulated and observation-based
temperature response curve of the winter flux indicated by the EC
tower-based measurements (Fig. 6). For example, field studies have shown
that the soil CO
We developed a remote-sensing-driven permafrost carbon model at an intermediate scale (
The regional model simulations will be archived and distributed for public access through the NASA ABoVE archive at the NASA ORNL DAAC. All data used in this study were obtained from
free and open data repositories. The model code used in this study is
available from GitHub at
The supplement related to this article is available online at:
YY, JSK, and CEM initiated the study. YY did the calculations and wrote the paper. All coauthors contributed to the data and provided feedback on the final version.
The authors declare that they have no conflict of interest.
This study was funded by the NASA Terrestrial Ecology Program as part of the Arctic Boreal Vulnerability Experiment (ABoVE). A portion of the research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA.
This research has been supported by the NASA Terrestrial Ecology Program (grant no. NNH18ZDA001N).
This paper was edited by Michael Weintraub and reviewed by two anonymous referees.