Soil respiration strongly offsets carbon uptake in Alaska and Northwest Canada

Soil respiration (i.e. from soils and roots) provides one of the largest global fluxes of carbon dioxide (CO2) to the atmosphere and is likely to increase with warming, yet the magnitude of soil respiration from rapidly thawing Arctic-boreal regions is not well understood. To address this knowledge gap, we first compiled a new CO2 flux database for permafrost-affected tundra and boreal ecosystems in Alaska and Northwest Canada. We then used the CO2 database, multi-sensor satellite imagery, and random forest models to assess the regional magnitude of soil respiration. The flux database includes a new Soil Respiration Station network of chamber-based fluxes, and fluxes from eddy covariance towers. Our site-level data, spanning September 2016 to August 2017, revealed that the largest soil respiration emissions occurred during the summer (June–August) and that summer fluxes were higher in boreal sites (1.87 ± 0.67 g CO2–C m−2 d−1) relative to tundra (0.94 ± 0.4 g CO2–C m−2 d−1). We also observed considerable emissions (boreal: 0.24 ± 0.2 g CO2–C m−2 d−1; tundra: 0.18 ± 0.16 g CO2–C m−2 d−1) from soils during the winter (November–March) despite frozen surface conditions. Our model estimates indicated an annual region-wide loss from soil respiration of 591 ± 120 Tg CO2–C during the 2016–2017 period. Summer months contributed to 58% of the regional soil respiration, winter months contributed to 15%, and the shoulder months contributed to 27%. In total, soil respiration offset 54% of annual gross primary productivity (GPP) across the study domain. We also found that in tundra environments, transitional tundra/boreal ecotones, and in landscapes recently affected by fire, soil respiration often exceeded GPP, resulting in a net annual source of CO2 to the atmosphere. As this region continues to warm, soil respiration may increasingly offset GPP, further amplifying global climate change.


Introduction
The northern permafrost region holds over 50% of the global soil organic carbon (SOC) pool and approximately one trillion tonnes of carbon in the top 3 m of soil alone (Hugelius et al 2014, Meredith et al 2019. Historically, SOC in permafrost-affected ground and seasonally thawed active layers was largely protected from microbial decomposition by lowtemperatures (Faucherre et al 2018). However, arctic air temperatures have increased rapidly (Box et al 2019), rising 2.7 • C (annual average) and 3.1 • C (October-May) between 1971 and 2017. This warming has increased the length of the non-frozen season (Kim et al 2012) and has deepened soil thaw (Luo et al 2016) in Alaska and Canada. Soil warming can increase microbial activity (Natali et al 2014) and may result in large amounts of soil carbon being released into the atmosphere, predominantly as carbon dioxide (CO 2 ; Schuur et al 2015, Turetsky et al 2020.
Soil root and microbial respiration (herein referred to as soil respiration) are dominant components of an ecosystem's annual CO 2 emission (Mahecha et al 2010). Soil respiration in boreal forests is estimated to account for 48%-68% of total ecosystem respiration (ER; soil + aboveground components; Hermle et al 2010, Parker et al 2020). In tundra, soil respiration is the primary source of CO 2 efflux and summer emissions alone may account for 60%-90% of annual ER (Sommerkorn et al 1999, Gagnon et al 2018, Strimbeck et al 2018. Generally, the seasonality and magnitude of soil respiration are influenced by soil temperature, soil water content, root activity, and microbial-community access to SOC (Bond-Lamberty et al 2004, Schuur et al 2009, Nagano et al 2018. As northern landscapes continue to warm, CO 2 emissions resulting from soil respiration may increasingly offset carbon uptake by plants (i.e. gross primary productivity, GPP). Moreover, the fastest rate of warming in the Arctic-boreal region is occurring in autumn, winter, and spring (Box et al 2019), a period when microbial respiration continues but plant productivity is limited. Recent tundra and boreal carbon budgets in northern Alaska and Canada using eddy covariance (EC) flux observations show that enhanced soil respiration during an anomalously warm winter (2015)(2016) offset any carbon gains provided by GPP (Liu et al 2020). Similarly, annual soil respiration offset 75% of the total forest GPP in a boreal Finland study (Pumpanen et al 2015). In northern Sweden, a steady increase in soil respiration, and no change in forest GPP, resulted in a transition from net annual ecosystem CO 2 sink to source (Hadden and Grelle 2016). An atmospheric study of North Slope, Alaska tundra reported late autumn and early winter CO 2 emissions had increased by 73% since 1975 (Commane et al 2017). These observed increases in soil respiration have been attributed to increased ground thaw (Kim et al 2006) and residual unfrozen water in soil pore space (Faucherre et al 2018). Further, a recent synthesis of soil flux indicated soil respiration from Arctic-boreal permafrost regions may already outweigh ecosystem CO 2 uptake under contemporary climate conditions .
Little is known about the spatiotemporal patterns of soil CO 2 emissions from tundra and boreal biomes at the regional level, in part due to the lack of spatial representation by in situ observations. Existing in situ (e.g. EC) and satellite-based CO 2 monitoring networks are unlikely to detect changes in soil respiration across the permafrost domain (Parazoo et al 2016), especially in winter months, or identify local changes in net ecosystem exchange (NEE) or component (i.e. GPP and respiration) CO 2 fluxes (Schimel et al 2015).
Process-based terrestrial models can be useful tools to diagnose how components of the carbon cycle might change in response to shifts in ecosystem properties and climate but are hampered in representing seasonal and spatial patterns by the lack of integrated observations (Fisher et al 2018. In many regions, including Northern Eurasia and Alaska, process-models have failed to agree on flux magnitudes and even the sink vs source status of ecosystem carbon budgets (Fisher et al 2014, Rawlins et al 2015. Improving process-level understanding of soil respiration requires integrating in situ flux data, observations of ecosystem properties (e.g. vegetation characteristics, thermal and moisture state) from satellite remote sensing, and data-informed modeling (Jeong et al 2018, Schimel et al 2019. This study addresses knowledge gaps in our understanding of soil respiration from permafrost ecosystems. We seek to improve understanding of the spatiotemporal patterns of soil respiration in boreal and tundra landscapes, the magnitudes of seasonal and annual soil CO 2 loss, and how soil respiration impacts ecosystem carbon budgets. Here we apply information gained from a new network of Soil Respiration Stations (SRSs) within the NASA Arctic Boreal Vulnerability Experiment (ABoVE) domain. We also incorporate a complementary suite of flux records from EC towers within the region. We used random forest (RF) models and remote sensing to extrapolate soil fluxes to the ABoVE domain for the 2016-2017 period, obtaining spatially and seasonally disaggregated regional estimates of soil emissions. Last, we determine the seasonal and annual offset of GPP by respiration (soil, and ecosystem) to identify landscape net annual carbon source, or sink, status under contemporary climate conditions.

Study region
The spatial domain of this study, which includes permafrost-affected landscapes of Alaska and Northwest Canada, represents the core region of the NASA ABoVE Field Campaign (Kasischke et al 2014, Loboda et al 2019) and spans gradients of climate, permafrost distribution (or prevalence), vegetation, and ecosystem disturbance from fires (figure 1). Approximately 24% of the region has been recently burned (between 2000 and2017;Loboda et al 2017a, 2017b, Pastick et al 2018. Because our flux sampling locations only represent permafrost-affected ecosystems, our analyses excluded landscapes where permafrost was absent (Gruber 2012); we also excluded barren lands (<10% vegetation) and open water.

SRS chamber data
We used CO 2 flux data from ten SRS (Minions et al 2019) installed along a north-south gradient in Alaska, spanning the North Slope to Eight Mile Lake near Denali National Park (figure 1; supporting information, SI table 1 (available online at stacks.iop.org/ERL/16/084051/mmedia)). Each SRS is a fully automated system that measures soil surface CO 2 flux using three forced diffusion (FD) chambers. The SRS technique was designed to provide yearround measurements of soil emissions (live aboveground vegetation was removed during chamber installations to ensure that flux measurements do not reflect net CO 2 exchange), even during periods of snow cover. Detailed information about the SRS system and FD processing is provided in the supplement (SI section 1). In addition to the SRS records, chamber-based fluxes collected using an Eosense eosFD portable sensor near Council, Alaska were obtained from project partners. Six of the 11 FD stations (SRS and the eosFD site) are in tundra and five in the boreal region. Six of the SRS sites represent paired burned and unburned ecosystems (SI table 1).

EC tower data
We used AmeriFlux (ameriflux.lbl.gov) and ECinvestigator provided quality-controlled CO 2 flux records primarily from September 2016 through August 2017 (matching the period of highest data availability from the SRS sites) from 15 EC towers (figure 1, see SI table 1, SI section 2.1); eight tower sites were in tundra and seven in boreal. The half hourly EC records included NEE, GPP, and ER. NEE was obtained directly from the EC records and indicates the net of ecosystem CO 2 respiration and CO 2 uptake; GPP and ER were obtained using standard EC flux partitioning algorithms (Reichstein et al 2005, Lasslop et al 2010. Quality data were available yearround for at least ten sites (SI table 1; SI figure 2).

Flux modeling
We used published values from field and laboratory studies to separate aboveground respiration components from the EC-based ER records (SI section 2.2). We acknowledge that the literature-based ratio approach does not account for seasonal variability in aboveground respiration, and variability from other factors including temperature, species type, total biomass, and ecosystem stress. However, this approach was used because more detailed information was not available. We then applied the combined SRS FD and EC ER dataset, information from remote sensing, and ancillary geospatial layers (SI section 3) to obtain data-driven RF models (SI section 4) developed separately for summer (June-August), autumn (September, October), winter (November-March) and spring (April and May). These seasons were based on observed seasonality in the tundra and boreal SRS and EC flux records (SI figure 3). Candidate variables used in the models are described in the supplement (SI section 3) and included information about vegetation greenness and productivity, leaf area, topography, soil characteristics (e.g. permafrost status, soil texture, SOC content), and other environmental conditions (e.g. albedo, radiation, temperature, snow cover, soil moisture (SM) status).

RF models and spatial prediction
RF is a machine learning method that uses an ensemble approach to regression by means of multiple decision trees and bootstrap sampling (Liaw andWiener 2002, Cutler et al 2012). RFs have been widely used in ecological studies (Pearson et al 2013, Clewley et al 2017 and carbon budget assessments (Tramontana et al 2015, Jung et al 2020. Strengths of RF include the ability to handle high-dimensional problems, noise, and non-linearity, and its ability to provide robust internal estimates of error and variable importance (Cutler et al 2012).
We developed RF models in the R computing environment (R Core Team 2019) using the random-Forest package (Liaw 2018). Each tree was constructed using a random selection (i.e. bagging) of approximately 2/3 of the samples (42 site-flux observations in the autumn model, 110 in the winter model, 48 in the spring model and 65 in the summer model; see SI section 4.1). The remaining 1/3 of the observations was used to validate each forest (1000 trees per trained RF model). Predictor variable (SI table 4) selection was achieved using the variable selection (Genuer et al 2010, 2019) R package which was designed to reduce high (>70%) cross-correlations between the selected inputs. The tuneRF algorithm (Liaw 2018) was applied to optimize the Mtry parameter (the number of variables available for splitting at each tree node). Variable importance was assessed using ran-domForest varImpPlot (Liaw 2018) and the rfPermute (Archer 2020) R package was used to provide corresponding estimates of parameter significance.
This process was applied to obtain optimal RF models for each season (SI section 4.2). The final models were applied to the raster predictor datasets (raster package in R; Hijmans et al 2020) to obtain 300 m resolution maps of monthly average soil respiration.

ABoVE region carbon budgets
We used the monthly average soil CO 2 emission maps (g CO 2 -C m −2 d −1 ) from the RF models to obtain regional flux budgets. The emission estimates were scaled to the terrestrial spatial domain within each 300 × 300 m grid cell by removing fractions of identified open water within each grid cell. Fractional water was derived using the 30 m Wang et al (2019) land cover map for 2014. We then obtained monthly and annual soil respiration totals for the ABoVE domain (Tg CO 2 -C period −1 ). To determine the extent that soil respiration offset the annual ecosystem uptake of CO 2 (i.e. GPP), we obtained estimates from an ensemble of satellite observation based GPP  (Li and Xiao 2019). Lastly, to gauge the potential impact of regional NEE on annual GPP, we used literature-based flux ratios (SI section 2) to provide estimates of emissions from aboveground respiration, in addition to our RF-estimates of soil respiration.

Soil emission characteristics
Site-level fluxes showed strong seasonal emission patterns (figure 2) closely tied to changes in air and soil temperature (figure 3). Soil respiration (regional mean ± standard deviation) was largest in summer (boreal: 1.87 ± 0.67 g CO 2 -C m −2 d −1 ; tundra: 0.94 ± 4 g CO 2 -C m −2 d −1 ) and peak daily-averaged respiration was often observed in July (SI figure 3), the warmest month (air temperatures >10 • C, at EC and SRS flux sites). This was followed by a steady decline in autumn (boreal: 0.8 ± 0.4 g CO 2 -C m −2 d −1 ; tundra: 0.42 ± 0.2 g CO 2 -C m −2 d −1 ). Winter respiration persisted even under snow cover, and cold air and soil (10-15 cm depth) temperatures averaging −18 • C ± 6 • C and −3.5 • C ± 2.7 • C, respectively. In winter, boreal soil respiration averaged 0.24 ± 0.2 g CO 2 -C m −2 d −1 and tundra averaged 0.18 ± 0.16 g CO 2 -C m −2 d −1 . Soil respiration began to increase again in spring as ecosystems warmed (average boreal/tundra soil temperatures of −1.98 • C in April, −0.07 • C in May, and 1.82 • C in June). Soil respiration from boreal sites was systematically higher than those from tundra in all seasons, excluding winter (t-test; p = 0.03 in autumn, p = 0.22 in winter, p = 0.002 in spring, p < 0.001 in summer). T-test significance for monthly flux averages is shown in SI figure 3 and seasonal flux patterns according to biome (i.e. tundra or boreal) and flux location are shown in SI figure 4.
Air temperature (p = 0.009) and soil temperature at 10-15 cm depth (p = 0.01) explained 65% of the observed variability in monthly soil respiration at the site level, in a linear regression analysis that included fluxes from all seasons. During the 2016-2017 period, soil respiration was observed even at air temperatures approaching −30 • C and at soil temperatures (∼15 cm depth) below −10 • C (figures 3(a) and (b); SI figure 5). Soil respiration increased steadily after ground thaw. Soil respiration for the 14 tundra and boreal sites where in situ SM was available indicated that higher fluxes in summer most often occurred where soils (⩽15 cm depth) were relatively wet but not saturated (SI figure 6). Observed relationships between the seasonal site-level soil respiration fluxes and important remote-sensing based indicators of permafrost status, temperature, SM, and GPP is provided in SI figure 7.

RF model performance and variable importance
The RF models explained much of the variance in soil respiration, with moderate-to-low root mean squared error (RMSE) and mean absolute error (MAE; SI table 5, SI figure 8). The R 2 values were 0.68 for the summer model, 0.57 for autumn, 0.65 for winter, and 0.76 for spring. The respective RMSE (g CO 2 -C m −2 d −1 ) values were 0.35 (summer), 0.24 (autumn), 0.10 (winter), and 0.25 (spring). The positive MAE (averaging 0.2 ± 0.09 g C m −2 d −1 ) indicated a slight underestimation of soil respiration by the models. In the summer RF model, MODIS (MOD) GPP was the most important variable, followed by soil sand content (an indicator of water and nutrient retention), MODIS leaf area index (LAI), tree cover, and normalized difference vegetation index (NDVI; an indicator of greenness). In the autumn model, SMAP root Observed relationships between (a) air temperature or (b) in situ soil temperature (∼10-15 cm depth) and average monthly soil respiration (g CO2-C m −2 d −1 ) from EC (triangles) and SRSs (circles) in Alaska and Northwest Canada for boreal (green) and tundra (blue) sites. Fitted curves (black lines) were obtained using locally weighted loess smoothing; gray shading represents confidence intervals (±/−standard error).

Annual carbon flux estimates for ABoVE domain
Annual soil respiration emission for the study domain was 591.2 Tg C-CO 2 ± 120 Tg C-CO 2 during the 2016-2017 period (figure 4; SI table 6). Monthly soil respiration maps are provided in SI figure 9 and seasonal respiration budgets are shown in SI figure 10 (SI figure 11 shows associated emission uncertainty maps). Summer (June-August) contributed to 58% of annual soil respiration, the longer winter (November-March) period generated 15%, with comparable proportions occurring in autumn (15%, September, October) and spring (12%, April, May). Across the ABoVE region, the largest soil respiration budgets occurred in the boreal zones and the warmer, more southern, forest-tundra ecotone. Over half of regional soil respiration emissions (54% of annual total) were from colder landscapes having a widespread occurrence of near-surface permafrost (i.e. where the PZI was >75%; spanning 70% of the domain) and the remaining 46% of emissions were from warmer permafrost (0% < PZI > 75%; 30% of the domain; table 2; SI figure 12). The area covered by Shrubland/Herbaceous vegetation produced the majority (46%) of soil respiration, followed by Sparse Vegetation and Evergreen Forest. Annual GPP for the whole domain, obtained from MOD17, GOSIF, and SMAP L4_C products (section 2.4.2, SI figure 13), was 1046-1256 Tg CO 2 -C in 2016 and 1025-1134 Tg CO 2 -C in 2017 (SI table 7) with an estimated uncertainty of 310 Tg CO 2 -C yr −1 (SI section 3). Annual GPP was considerably higher (>600 g CO 2 -C m −2 yr −1 ) in the boreal regions relative to tundra (<300 g CO 2 -C m −2 yr −1 ; SI figure  13). Our extrapolations indicate soil respiration offset approximately 54% of GPP across the domain (averaging 1101 Tg CO 2 -C). The offset of GPP by soil respiration varied considerably across the region ( figure 5; SI figure 14). Offsets of ⩾100% (i.e. annual net carbon source areas) were identified in far northern tundra and mountainous landscapes, along transitional tundra-boreal ecotones, and in landscapes recently disturbed by fire (e.g. west of Hudson Bay and south of the Selwym Mountains in Canada). We estimate that approximately 8% of the ABoVE region was a net carbon source (100% offset of GPP) in 2016-2017, based on soil respiration alone and not accounting for aboveground respiration and nonterrestrial carbon emissions (i.e. aquatic bodies).

Discussion
This study provides new estimates of soil respiration for the ABoVE domain and insights into how soil respiration is offsetting net annual GPP across permafrost-affected tundra and boreal landscapes. Our analysis of in situ observations and RF-model results indicate that soil respiration was generally highest under warmer (above freezing) soil temperatures and deeper seasonal soil thaw, in moderate-to-moist soils (0.5-0.8 m 3 m −3 ), and in areas with higher vegetation productivity. Accordingly, the largest annual soil respiration rates occurred in boreal ecosystems where trees and shrubs were present, especially along the more southern portions of the domain having substantial permafrost thaw.

The temperature-soil respiration relationship
Consistent with earlier studies (e.g. Wickland et al 2006, Natali et al 2014, Loranty et al 2018, we found temperature to be an important driver of soil respiration at the site level. Our regional flux assessments showed highest soil respiration rates in summer (contributing to 58% of annual soil respiration) when soil temperatures were warmer and soil thaw was deepest. Higher emissions in warmer soils are not only from increased microbial decomposition of SOC, but likely also from increased root activity (i.e. belowground autotrophic respiration), a strong source of CO 2 in thawing permafrost systems (Hicks Pries et al 2015, 2016. Although gridded estimates of belowground root density are not available for this region, LAI, % tree cover, and vegetation indices (important predictors in the RF models) provided proxies of vegetation productivity (e.g. Street et al 2006), and indirect information about root respiration.
Within the site-level soil respiration database, larger, and sometimes episodic, CO 2 emissions (>0.5 g CO 2 -C m −2 d −1 ) were observed as soil temperatures approached 0 • C, especially as soil layers began to freeze in the autumn. Like our site-level findings, an atmospheric study of Alaska's North Slope also identified high CO 2 emissions in autumn and early winter (October-December; Commane et al 2017) during the landscape freeze. Although our RF model approach represented regional flux characteristics relatively well, the autumn RF model had the lowest performance of the four seasonal models, resulting from its inability to capture spatiotemporally episodic releases of CO 2 observed in situ. As a result, the model underestimated regional CO 2 emissions (by ⩾0.2 g C m −2 d −1 , based on MAE estimates) during the autumn period.

Regional predictors of soil respiration
Our regional assessments show that carbon source/ sink status is highly heterogeneous. Annual carbon status of an ecosystem is influenced by many factors, including GPP and plant community type (e.g. Rouse  Kim et al 2021). SM is also an extremely influential factor that is very heterogeneous across the landscape and affects both vegetation productivity and soil respiration (Grogan 1999), yet this environmental variable can be radically altered by permafrost thaw (Jorgenson et al 2013) and is especially difficult to monitor regionally at finer landscape-level scales (Du et al 2019).
Burn status (i.e. burned or unburned) was not a significant predictor of the regional monthlyaveraged soil respiration emissions examined in this analysis, which could be in part due to our database containing information from only three burn sites (representing tundra and forest landscapes 11-15 years after fire), or because of rapid postfire recovery. Following a fire event, the combination of warmer and drier soils can substantially increase CO 2 flux from soils (O'Neill et al 2002, 2003, Ueyama et al 2019. However, a review of fire disturbance at high latitudes reported that soil and root respiration in forests may stabilize after a decade (Ribeiro-Kumara et al 2020). As a result, our

Regional carbon budgets
Our 2016-2017 assessment shows an annual soil respiration loss of 591 Tg CO 2 -C for the permafrostaffected ABoVE domain. A comparison of our RFbased results with the Natali et al (2019) pan-Arctic estimates (referred to as NCC 2019 and subset to the ABoVE permafrost-affected study area) showed that soil respiration estimates in the NCC 2019 record was substantially higher (∼1.6×) than our RF budgets during the winter and early spring (SI table 6 and SI figure 16). A corresponding model analysis by Schiferl et al (2021) used a stochastic time-inverted Lagrangian transport (Lin 2003) model and atmospheric CO 2 observations influenced by Alaska North Slope tundra (obtained from the Utqiaġvik tall tower) to verify the NCC 2019 and RF-model results. The study determined that our RF-model approach underestimated atmospheric enhancements in October-December by 2-3× but the RF-estimates were much better aligned with atmospheric observations, relative to NCC 2019, during the January-April period (SI section 5, SI figure 17). While episodic bursts of CO 2 from freezing soils may contribute to the larger atmospheric CO 2 levels observed October-December across the North Slope, our assessments also indicate that very large emissions of CO 2 to the atmosphere could result from the turnover and freeze of lakes and ponds which are widespread throughout the region (SI section 5; Preskienis et al 2021). If this assessment is correct, then the Natali et al (2019) results also overestimate soil CO 2 emissions for the North Slope during the autumn season.
For the ABoVE study domain in 2016-2017, soil respiration only partially offset GPP, by approximately 54%-60%. However, for many grid cells in northern tundra, mountainous regions, or where boreal forest GPP was reduced by recent fire (SI figures 14 and 15) soil respiration alone (not accounting for aboveground autotrophic respiration) equaled or exceeded annual GPP, indicating that some sites are net CO 2 sources. The Belshe et al (2013) meta-analysis of EC fluxes from high-latitude tundra sites concluded that tundra systems are currently CO 2 sources. Similarly, Natali et al (2019) determined the permafrost-affected Arctic-boreal zone to likely be a net CO 2 source when considering winter contributions from soils. Using published ratio estimates of aboveground vs belowground (soil) contributions to ER for boreal and tundra biomes we estimate an annual ER between 820 and 1171 Tg CO 2 -C, respectively offsetting 74%-106% of annual GPP (SI figure 18). This estimate suggests that tundra is currently a CO 2 source, while the boreal is a CO 2 sink.

Conclusion
Soil respiration can strongly impact the carbon sink or source status of high latitude permafrost regions. When considering the permafrost-affected tundra and boreal biomes of Alaska and Northwest Canada as a whole, soil respiration offset annual GPP in 2016-2017 by 54%-60%. However, in sparsely vegetated tundra regions and recently burned landscapes, soil respiration exceeded GPP. Although a majority (58%) of annual soil respiration emissions occurred in the summer months, we found considerable contributions of soil CO 2 in the shoulder and winter seasons. Our soil emission estimate of ∼591 ± 120 Tg CO 2 -C for the domain is likely conservative due to the inability of our statistical model approach to capture episodic bursts of CO 2 during soil freeze and thaw, and a lack of soil respiration data from very recent fire scars. We also acknowledge uncertainties introduced by using a simple literature-based flux correction ratio method to remove aboveground components from tower-based ER observations, which does not account for variability in aboveground respiration by species, temperature, stand age and other factors. We also note that the 2016-2017 period was characterized by record breaking high air temperatures across much of the region relative to previous years and the longer-term 1981-2019 normal (ACRS 2016(ACRS , 2017. Warming records have been repeatedly broken in more recent years and we estimate that post-2017 soil respiration budgets will exceed those reported here. Our data-driven gridded soil respiration budgets provide new, valuable records that will be useful for the future benchmarking of process-based models. Although our assessment is limited to a one-year period, efforts to ensure the continued operation of SRS and EC sites will allow future regional studies to better understand interannual variability and spatiotemporal trends in soil respiration across the rapidly changing Arctic-boreal environment. As current spaceborne observations of CO 2 are not yet able to track changing emission contributions in winter, nor able to identify finer landscape-level patterns of soil emissions (Parazoo et al 2016), the continuation if not expansion of existing in situ monitoring networks is urgently needed to document changes in soil respiration and ecosystem carbon sink/source status across the thawing permafrost region in North America and elsewhere, including Siberia and the Tibetan Plateau.

Data availability statement
Data from this study are included within the article and supplementary information and are available through the ORNL DAAC.
The data that support the findings of this study are openly available at the following URL/DOI: https://doi.org/10.3334/ORNLDAAC/1935.