Introduction

The permafrost region covers approximately 15% of the land area in the northern hemisphere (Obu et al., 2019). The broad-scale distribution of permafrost on Earth is controlled by climate conditions, with the largest areas occurring in the Arctic and boreal regions, which are the focus of this study (Figure 1). Extensive permafrost is also found on the Tibetan plateau (Yang et al., 2010). Permafrost affects many aspects of ecosystem function, including hydrology, vegetation, and carbon and nutrient cycling (Schuur et al., 2008). Permafrost soils are often carbon (C) rich because cold and wet conditions limit microbial decomposition of organic material, allowing for the accumulation of a globally significant soil C stock (Hugelius et al., 2014;Strauss et al., 2021). However, climate warming is increasing soil temperatures (Biskaborn et al., 2019) and thawing permafrost (Nitze et al., 2018), enabling microbial transformation of some portion of these long-protected soil C stocks, contributing to greenhouse gas emissions and climate change (Schaefer et al., 2014;Schuur et al., 2015Schuur et al., , 2022)). However, there is large uncertainty in future climate projections with implications for international greenhouse gas emissions policy decisions (Natali et al., 2022).

Over the last 20 years, research on permafrost region C cycling and climate feedbacks has seen tremendous progress and growth (Sjöberg et al., 2020) through the integration of traditionally separate disciplines including ecology, soil science, biogeochemistry, atmospheric science, hydrology, geophysics, remote sensing, and modeling. In this paper, we synthesize current knowledge of permafrost ecosystem characteristics controlling C cycling as well as the measured and modeled terrestrial carbon dioxide (CO 2 ) and methane (CH 4 ) exchange between permafrost ecosystems and the atmosphere to identify next steps in understanding permafrost region C cycling.

Permafrost Region Overview: Extent and Characteristics

Permafrost is defined as subsurface earth material with temperature at or below 0°C for at least two consecutive years (Harris et al., 1988). Located between the ground surface and the continuously frozen permafrost, the "active layer" thaws and refreezes annually. Here, the majority of soil biological processes occur, including the formation and decomposition of soil organic matter. Permafrost occurs throughout the boreal, sub-Arctic and tundra landscapes (Figure 1). Within the broader climatic constraints of the permafrost domain, permafrost occurrence at a given site is moderated by local factors, such as slope and aspect, hydrology and soil moisture conditions, winter snow depth, vegetation cover, as well as the soil properties and ground ice (Shur & Jorgenson, 2007). These factors can vary considerably over distances of meters to kilometers, so areas with and without permafrost can coexist under similar climate. Additional key variables characterizing the state of permafrost include ground temperature, active layer thickness, ground ice content, and permafrost formation history (Jorgenson & Osterkamp, 2005;Osterkamp & Romanovsky, 1999;Romanovsky & Osterkamp, 2000;Shur et al., 2005; S. L. Smith et al., 2022).

The circum-Arctic permafrost region is often mapped as four regions: a continuous zone (90%-100% of land surface covered by permafrost), a discontinuous zone (50%-90% permafrost), a sporadic zone (10%-50%) and isolated (0%-10%) zone (Brown et al., 1998(Brown et al., , revised 2001)). Multiple new spatial data products for permafrost characteristics in the northern high latitudes are now available (Table 1). These products suggest relatively similar aerial extents for permafrost in the exposed land area (14 and 15.7 × 10 6 km 2 ; Obu, 2021). If the entire permafrost region with its discontinuous zones without permafrost are considered, the permafrost region can cover up to 23 × 10 6 km 2 (Table 1); the Arctic-boreal permafrost domain, the focus of our review, covers 18.4 × 10 6 km 2 (Hugelius et al., 2023). Many permafrost maps largely build on the first permafrost map of the International Permafrost Association (IPA) (Brown et al., 1998(Brown et al., , revised 2001)). This was based on field mapping and manual digitizing of permafrost in different regions-a formidable effort that has not been repeated since. Most "modern" mapping approaches either rely on statistical relationships between climatic conditions and permafrost variables or on process-based models simulating ground thermal regimes (Obu et al., 2019;Ran et al., 2022). With such methods, gridded products of climate variables, such as air temperatures from climate re-analyses or remotely sensed land surface temperature, can be combined with geospatial data characterizing the landscape so that the effect of local factors on the ground thermal regime are better captured.

Permafrost maps are generally designed as "static" on timescales of several decades, and while useful to identify the spatial distribution of permafrost, the static concept is challenged by rapidly warming climate conditions in most permafrost areas (Rantanen et al., 2022). In-situ monitoring networks show increasing ground temperatures and a deepening of the active layer throughout most of the permafrost domain (Biskaborn et al., 2019;S. L. Smith et al., 2022). Furthermore, the formation of taliks, or the persistent unfrozen soil layer in a permafrost soil that forms when soils no longer freeze down to permafrost, is now widespread across Alaska (Farquharson et al., 2022). More abrupt disturbances such as retrogressive thaw slumps (mass movement and erosion on slopes), thermokarst lake and wetland formation, and thermokarst landscapes in general (i.e., land surface where the thawing of ice-rich permafrost terrain causes land subsidence) have been reported across all permafrost zones (Jorgenson et al., 2006;Nitze et al., 2018;Payette et al., 2004). Consequently, while the broad-scale extent and (Hugelius et al., 2020), the distribution of Yedoma (purple; Strauss et al., 2022), landscapes with very high potential thermokarst coverage (Olefeldt et al., 2016), and (b) the distribution of the boreal biome and the soil organic carbon stocks within the permafrost region (Hugelius et al., 2014), and (c) vegetation types across the permafrost region following Virkkala et al., 2021 (note that the wetland extent on this map is likely underestimated). All maps also show the extent of the northern permafrost region as defined in the previous RECAPP-2 permafrost synthesis (Hugelius et al., 2023). characteristics of permafrost under relatively stable conditions can be adequately quantified (i.e., static maps), dynamically mapping these under rapidly changing climate conditions remains a challenge, hindering our understanding of the large-scale extent and implications of permafrost thaw. Permafrost region soils: characteristics, extent, C stocks Yedoma domain extent (Strauss et al., 2021) Harmonized geological maps, remote sensing and Field mapping, including manual digitalization Pan-Arctic Polygons of variable size Polygon Peatland extent, depth, and C densities Hugelius et al. (2020) Harmonized soil maps and statistical modeling North of 23°N 10 km Raster Soil class, soil properties, C density Tarnocai et al. (2009) NCSCD Harmonized soil maps and statistical modeling Permafrost region Polygon Soil class, soil properties, C density Hugelius et al. (2013) NCSCDv2.0 Harmonized soil maps and statistical modeling Permafrost region polygon Soil class, soil properties, C density Mishra et al. (2021) Machine learning using harmonized soil profiles and remote-sensing data products Permafrost region 250 m Raster Soil class, soil properties, C density Hengl et al. (2017); Poggio et al. (2021) SoilGrids250 m/ 2.0 Machine learning using soil profiles and remote-sensing data products Global 250 m Raster

Permafrost Region Vegetation: A Key Control on C Cycling

There is considerable variation in northern permafrost region vegetation from the sparsely vegetated low-statured treeless tundra environments to the densely vegetated boreal forests in the south. High densities of lakes, ponds, and wetlands are found in these northern high latitudes, with wetlands alone covering between 5% and 25% of the permafrost region (Figure 1; Karesdotter et al., 2021;Olefeldt et al., 2021;Raynolds et al., 2019). Extensive lake and peatland formation is linked to the relatively flat landscapes created by glacial retreat, increases in available moisture, and thermokarst development (Alexandrov et al., 2016;Brosius et al., 2021;Gorham et al., 2007). Tundra vegetation is often distributed along soil moisture gradients, with graminoid vegetation found in areas with high soil moisture (e.g., topographical depressions or flat areas), whereas shrubs dominate in better drained, more elevated or sloping areas (Heijmans et al., 2022). Evergreen forests comprise the majority of boreal forests in the North American permafrost region followed by deciduous broadleaf forests (Wang et al., 2020); deciduous larch forests cover large areas in the Russian permafrost region (Shevtsova et al., 2020).

Warming in the permafrost region is expected to enhance vegetation growth as well as shift species composition, which can affect C cycling both directly and indirectly. Vegetation changes have consequences for many additional ecosystem functions through effects on energy balance, hydrology, soil temperatures, C inputs to soil, and susceptibility to wildfire (Chapin et al., 1996;Mack et al., 2021;Sturm et al., 2005). Both greening (enhanced vegetation productivity; often associated with tree and shrub expansion) and browning (decreased productivity due to vegetation dieback or slower growth) are expected in permafrost regions under current warming trajectories, although the responses differ locally (Berner et al., 2020;C. X. Liu et al., 2021;Myers-Smith et al., 2020;Reid et al., 2022). Greening during the 1985-2016 has been more widespread, covering ca. 37% of the tundra, whereas browning occurs in only 5% of the tundra (Berner et al., 2020). Meta-analyses of direct warming effects on vegetation suggest that warming increases vascular plant abundance and height, especially shrubs, but again, results are spatially variable (Elmendorf et al., 2012;Sistla et al., 2013). Permafrost thaw can also increase nutrient availability and contribute to increased productivity (Hewitt et al., 2019;Salmon et al., 2016). However, enhanced vegetation growth may not translate into enhanced ecosystem C stocks due to feedbacks between snow conditions and soil temperatures, vegetation, litter, and decomposition (Hartley et al., 2012;Sistla et al., 2013).

For example, increased plant growth (both above-and belowground) could increase C inputs to soil, but enhanced root-derived C into soils could also increase soil C decomposition via microbial priming (Keuper et al., 2020).

Recent reviews discuss interactions between shrub expansion (shrubification), permafrost, and C cycling with the overall conclusion that it is not known whether shrubification results in increased or decreased soil carbon stocks (Heijmans et al., 2022;Mekonnen et al., 2021).

Many spatial data products are available to map ecosystem types in the permafrost region based on vegetation or land cover. These map products, ranging from global to regional coverage, are often used for spatial extrapolation of processes related to permafrost C cycling including soil mapping (Mishra et al., 2021;Palmtag et al., 2022) and for upscaling C fluxes (Virkkala et al., 2021a). The most widely-used vegetation map, the Circumpolar Arctic Vegetation Map, is pan-Arctic in extent but does not include the boreal or sub-Arctic parts of the permafrost region (Raynolds et al., 2019;D. A. Walker et al., 2005). Global products often fail to separate key land cover types for permafrost C cycling, such as different dominant tree species, shrub and wetland types (Chasmer et al., 2020). As image resolution improves, higher resolution vegetation classifications can be expected but will require additional approaches to overcome limitations in determining critical land cover types.

Permafrost Soils: A Globally Significant C Reservoir

Soils within the permafrost region have accumulated C over millennia, with different dynamics depending on the extent of glaciation during the last glacial maximum (LGM; Harden et al., 1992;Lindgren et al., 2018). Northern peatlands and soils are distributed across the permafrost region in areas that were glaciated at LGM (Figure 1a) and contain substantial C stocks (Frolking et al., 2011;Yu et al., 2010). Large C stocks in areas that were not glaciated at LGM (Figure 1a), such as the Yedoma region, generally accumulated during the Pleistocene and consist of perennially frozen, fine-grained, organic-bearing, and ice-rich sediments (Strauss et al., 2017). The accumulation and persistence of soil C in this region are driven by limitations on decomposition of soil organic matter by temperature and soil saturation as well as repeated frost heave (cryoturbation) or repeated sediment deposition, which incorporates soil C from the surface deeper in the soil profile (Harden et al., 2012;Strauss et al., 2017). These processes have resulted in large soil C stocks within the permafrost region, with best estimates ranging from 1,014 (95% CI: 839-1,208) to 1,035 ± 150 Pg C for 0-3 m depth (Hugelius et al., 2014;Mishra et al., 2021) and 1,307 Pg C including deep (>3 m depth) Yedoma deposits, deltaic alluvium, and peats (Strauss et al., 2021). The most carbon-rich reservoirs in the 0-3 m of the permafrost soils are in peatlands and some tundra regions primarily in Hudson Bay Lowland, West Siberian Lowlands, western parts of the Northwest Territories, Alberta and British Columbia in Canada, and parts of northern Alaska (Figure 1b; Hugelius et al., 2014;Tarnocai et al., 2009).

Deep soil C deposits have been the most challenging reservoirs to quantify, but new estimates have recently been published for peatlands and Yedoma deposits (Figure 1a; Hugelius et al., 2020;Strauss et al., 2021;Strauss et al., 2017). These estimates highlight the critical role of peat deposits in the overall C stock of the permafrost region, including areas with and without permafrost (Hugelius et al., 2020). The insulating properties of peat can protect permafrost from thawing, resulting in the presence of residual or relict patches of permafrost in landscapes otherwise free of permafrost (Shur & Jorgenson, 2007;Vitt et al., 2000). Northern peatlands store approximately 415 ± 147 Pg C in peat, of which 185 ± 66 Pg C is located in permafrost-affected peatlands (Hugelius et al., 2020); a synthesis dataset of permafrost peat properties showed that permafrost formation in peatlands can both enhance or decrease C accumulation rates depending on site characteristics and timing of formation (Treat et al., 2016).

Yedoma deposits can reach a thickness of up to tens of meters and often containing large syngenetic ice wedges. Today, these are found in areas that remained deglaciated during the last glaciation of Siberia, Alaska and the Yukon (Figure 1a), and contain 115 Pg C (95% CI: 83-129 Pg C; Strauss et al., 2021). Together with other deep deposits in the Yedoma domain such as Holocene thawed and refrozen sediment, the Yedoma domain contains 400 Pg C (95% CI: 327-466 Pg C; Strauss et al., 2017). Arctic delta deposits are also considered as deep (up to 60 m depth), heterogeneous deposits (H. J. Walker, 1998) and are estimated to store approximately 67 Pg organic carbon but this estimate is highly uncertain (Hugelius et al., 2014). Due to increasing river discharge, sea level rise and permafrost thaw, Arctic delta sediment deposits might degrade and thaw resulting in a release of bio-available C into the near-shore of the Arctic Ocean or as CO 2 into the atmosphere (Overeem et al., 2022).

The most recent terrestrial C stock estimates for the permafrost region have incorporated over 2,700 soil profiles, but northern regions are still under-sampled compared with temperate regions (Mishra et al., 2021). Overall, permafrost region C stock estimates have been improved by concerted efforts to compile, harmonize, synthesize, and create open datasets of existing soil profile characterizations (Malhotra et al., 2019;Palmtag et al., 2022;Tarnocai et al., 2009). Hugelius et al. (2014) discuss remaining sources of uncertainty in the soil C dataset for the permafrost region, which include extensive spatial gaps over Russia, Scandinavia, Greenland, Svalbard and eastern Canada. Areas with thin soils and low C stocks in the High Arctic and mountainous regions also remain under-sampled, contributing to high uncertainty in spatially explicit C density mapping (Mishra et al., 2021).

Other key data gaps include Arctic delta deposits and peat deposits buried under mineral soils that glaciation and permafrost have preserved (Treat et al., 2019). Understanding how soil C stocks will change with disturbance continues to be an important topic, including the response to gradual and abrupt permafrost thaw and resulting hydrologic changes (e.g., M. C. Jones et al., 2017;Plaza et al., 2019).

Terrestrial Carbon Fluxes in the Permafrost Region

CO 2 and CH 4 Flux Magnitudes and Underlying Mechanisms

Northern permafrost regions have been a net sink of atmospheric CO 2 and smaller source of CH 4 since the beginning of the Holocene (Frolking & Roulet, 2007;Harden et al., 1992;Lindgren et al., 2018;Shi et al., 2020). Overall, carbon uptake has exceeded carbon emissions, as evidenced by the large soil carbon stocks of the region.

For recent decades (primarily 1990-2015), estimates of mean annual terrestrial net ecosystem exchange (NEE, i. e., the balance between gross primary productivity (GPP) and ecosystem respiration, ER) range from -1,800 (net sink) to 600 Tg C yr -1 (net source) (Bruhwiler et al., 2021;McGuire et al., 2016;Virkkala et al., 2021a;Watts et al., 2023), with most of the recent estimates averaging at -300 Tg C yr -1 (Watts et al., 2023). Wetlands and lakes in the permafrost region emit between 5.3 and 37.5 Tg CH 4 -C yr -1 (net source), with the majority of estimates being close to 22.5 Tg CH 4 -C yr -1 (Bruhwiler et al., 2021;Christensen et al., 2017;McGuire et al., 2012;McNicol et al., 2023;Peltola et al., 2019a). However, the spatial domains included in these reviews were variable and were sometimes based on latitudinal limits (e.g. >60°N) or the entire Arctic-boreal or permafrost regions. In addition to ecosystem-mediated C exchange, direct emissions from Arctic-boreal fires are between 100 and 400 Tg C yr -1 (on average 142 Tg C yr -1 ) (McGuire et al., 2016;van Wees et al., 2022;Veraverbeke et al., 2021). Lateral fluxes of CO 2 , CH 4 , and dissolved organic matter from terrestrial ecosystems to riverine and lacustrine systems can comprise a key part of the C budgets, ranging from 2% to 16% of NEE in areas with intact permafrost or up to 60% of NEE in upland areas experiencing thaw slumping (McGuire et al., 2009;Olefeldt et al., 2012;Zolkos et al., 2022). Earlier reviews have discussed lateral fluxes and controls on aquatic system C cycling in the permafrost region (Ramage et al., 2023;Tank et al., 2020;Vonk et al., 2015). Here, we focus on terrestrial ecosystem C exchange with the atmosphere.

The annual CO 2 sink is primarily driven by intense plant activity during the relatively short growing seasons (typically lasting 2-5 months; Lund et al., 2010;Virkkala et al., 2021a). However, the net ecosystem C accumulation is driven by belowground dynamics in soils and biomass rather than accumulation in above-ground vegetation C stocks (Bradshaw & Warkentin, 2015;Hartley et al., 2012;Shaver et al., 1992). The growing season sink strength has been relatively well synthesized across different moisture gradients and continents (McGuire et al., 2012), biomes (Virkkala et al., 2021), and vegetation types (Ramage et al., 2023). Net growing season C uptake is highest in the boreal permafrost region, particularly in warm evergreen and larch forests and can range between -150 and -240 g C m -2 month -1 during the June-August period (Hiyama et al., 2021); moist to wet graminoid-dominated tundra ecosystems also show strong growing season C uptake between -90 and -150 g C m -2 month -1 (Celis et al., 2017;Kittler et al., 2017;Pirk et al., 2017). Peatlands have low rates of net CO 2 uptake both from low plant productivity and even lower rates of decomposition due to anoxic soil conditions (Euskirchen et al., 2014;Frolking et al., 2011); mean long-term apparent C accumulation rates range from 20 to 35 g C m -2 y -1 , but are higher in recently accumulated peat and lower in boreal permafrost peatlands (14 g C m -2 y -1 ; Treat et al., 2016).

Arctic and permafrost regions are a net source of CH 4 to the atmosphere (McGuire et al., 2012;Saunois et al., 2020). Methane emissions are the net of production in anoxic soils and oxidation in the overlying aerobic soils, which can be bypassed by plant-mediated transport and ebullition (Christensen et al., 2003;Whalen, 2005).

Methane fluxes from permafrost regions can show different patterns than permafrost-free regions. Unlike upland areas in temperate regions that are net sinks of atmospheric CH 4 (Le Mer & Roger, 2001), upland (i.e., nonwetland) areas in tundra and boreal forest can be net CH 4 sources to the atmosphere due to periodically saturated conditions and cold-season emissions (Hashemi et al., 2021;Hiyama et al., 2021;Kuhn et al., 2021b;Treat et al., 2018b;Zona et al., 2016). However, upland tundra can also oxidize more CH 4 than previously thought (Jorgensen et al., 2015;Oh et al., 2020;Voigt et al., 2023); understanding the controls on these differences and net effect remains to be explored. For permafrost wetlands, CH 4 emissions are generally smaller than in permafrostfree wetlands due to the lower temperatures (Kuhn et al., 2021b;Olefeldt et al., 2013;Treat et al., 2018b). Moreover, airborne data have helped detect unexpectedly high CH 4 emissions from tundra (Miller et al., 2016), hotspots at lake margins (Elder et al., 2021), and strong geologic emissions in the Mackenzie River Delta (Kohnert et al., 2017). Some emissions hotspots are known to be thermogenic CH 4 (Kleber et al., 2023;Kohnert et al., 2017;Walter Anthony et al., 2012). Several previous efforts have extensively reviewed aspects of CH 4 fluxes in northern regions including key abiotic drivers such as temperature, water table position, and vegetation (Bridgham et al., 2013;Kuhn et al., 2021b;Olefeldt et al., 2013;Segers, 1998;Whalen, 2005), interactions with vegetation (Bastviken et al., 2022), feedbacks to climate (Dean et al., 2018), in peatlands (Blodau, 2002;Lai, 2009), production rates (Schädel et al., 2016;Treat et al., 2015), and generally for the permafrost region (Miner et al., 2022).

In-situ terrestrial CO 2 and CH 4 fluxes in the permafrost region have been synthesized in nearly 20 studies over the past decades with varying spatial extents (Figure 2). Virkkala et al. (2022) summarized the existing CO 2 flux syntheses for the permafrost region (Table 1 in Virkkala et al., 2022;Figure 2b here), showing an increase in CO 2 flux measurements over time in the permafrost region from ∼30 sites to over 200 sites in just one and a half decades. However, these 200 sites are not all currently active; the number of active eddy covariance sites measuring CO 2 and CH 4 fluxes in 2022 was 119 and 45 sites, respectively (Pallandt et al., 2022). Methane fluxes have been synthesized in 10 studies for both the permafrost region as well as smaller regions (Figure 2, Table 2); recent syntheses include between 18 (eddy covariance) and 96 (eddy covariance + flux chambers) unique sites in the permafrost region.

A key motivation for these syntheses has been to quantify CO 2 and CH 4 flux magnitudes and their controls across the permafrost region. Early estimates established that Arctic and boreal regions are a significant source of CH 4 to the atmosphere (Bartlett & Harriss, 1993;Matthews & Fung, 1987) but the CO 2 balance in the region has remained less certain (Chapin et al., 2000;Hayes et al., 2022). Recent in-situ estimates indicate that the boreal biome within the permafrost region has acted as an annual CO 2 sink over the past two decades, while the tundra biome appears to be either CO 2 neutral or a small CO 2 source, although there is considerable uncertainty associated with these findings (Bradshaw & Warkentin, 2015; Z.-L. Li et al., 2021;Natali et al., 2019;Virkkala et al., 2021a). Some parts of the permafrost region, such as Alaska, might be annual net CO 2 sources in both biomes (Commane et al., 2017).

The existing CH 4 flux syntheses have established the magnitude of CH 4 fluxes during the growing season and annual emissions for a wide range of sites and ecosystems across the northern permafrost region (Figure 2; Table 2). Multiple syntheses show significant differences in CH 4 emissions observed among wetland classes and compared to uplands (Figure 3; Knox et al., 2019;Kuhn et al., 2021b;Treat et al., 2018b). Specifically, marshes and fens have significantly larger CH 4 fluxes than permafrost bogs (including palsas, peat plateaus) and upland tundra, ranging from 5.5x-7.5x larger to 18x-23x larger, respectively, as demonstrated by our quantitative summary of these syntheses shown in Figure 3. However, CH 4 fluxes from other permafrost wetlands do not differ significantly from the other wetland categories (marshes, fens, bogs), and differences between permafrost and non-permafrost bogs were not significant, implying that it is important to capture both permafrost (temperature/substrate) effects on CH 4 fluxes and vegetation differences, likely related to the presence of aerenchymous plants facilitating CH 4 transport versus Sphagnum mosses and shrubs (Bastviken et al., 2022).

Emerging evidence highlights the key role of non-growing seasons in understanding the annual CO 2 and CH 4 balances (Commane et al., 2017;Natali et al., 2019;Treat et al., 2018b;Zona et al., 2016). Shoulder seasons, the transition periods close to the growing season (i.e., spring and fall), may be particularly important. For example, in fall and early winter, deeper soils are often thawed despite soils at the surface being frozen, boosting decomposition of deeper (and potentially older) soil organic matter while plant activity remains limited (Euskirchen et al., 2017;Pedron et al., 2022;Schuur et al., 2009); increased connectivity with groundwater pathways may enhance export (Hirst et al., 2023). As the soils freeze and thaw during the "zero-curtain" window (Outcalt et al., 1990), microbial activity can persist at low rates even when average soil temperatures are at or below zero (Clein & Schimel, 1995;Öquist et al., 2009). Emissions occurring during this extended period can add up to a substantial annual flux, up to 50% of annual ER and CH 4 emissions (Celis et al., 2017;Hashemi et al., 2021;Treat et al., 2018b;Zona et al., 2016). At some sites, the non-growing season CO 2 emissions currently offset or exceed growing season uptake and ultimately determine the annual C balance (Hashemi et al., 2021;Z. Liu et al., 2022;Watts et al., 2021). However, only ca. 20% of current eddy covariance sites measuring both CO 2 and CH 4 fluxes year-round; these sites are representative for only 10%-20% of the pan-Arctic (Pallandt et al., 2022). Most of these sites are in warmer areas that are in general easier to access and maintain (northern Scandinavia, Alaska, Journal of Geophysical Research: Biogeosciences 10.1029/2023JG007638 southern parts of Canada), while areas that are more remote remain under sampled. Continued research on the evolving seasonal freeze-thaw and soil moisture dynamics and effects on C emissions following permafrost thaw is critical for gaining a deeper understanding of the permafrost C feedback.

Regional Variability in CO 2 and CH 4 fluxes

In addition to regional differences in climate warming, differences across the permafrost region may affect the vulnerability of permafrost C to decomposition and release to the atmosphere (Gulev et al., 2021;Jorgenson & Osterkamp, 2005). The permafrost region varies in characteristics such as temperature, permafrost extent, ice content, and the degree of ecosystem protection of permafrost (e.g., insulating organic layers) (e.g., Shur & Jorgenson, 2007). Together with variability in observed and projected degree of warming, this makes some areas more likely to experience widespread permafrost degradation than others (Fewster et al., 2022;Olefeldt et al., 2016). The abundance of lakes and wetlands, vegetation composition, permafrost growth and formation history, soil C stocks and geomorphology also differ across the permafrost domain (e.g., Sections 1.2, 1.3), influencing the controls on CO 2 and CH 4 fluxes over broad spatial scales.

As the number and distribution of measurement sites across the permafrost domain has grown, we can compare the different datasets and approaches across policy-relevant domains (Figure 2) to see how flux magnitude and direction differ (Supporting Information S1). We analyze regional variability of CO 2 and CH 4 fluxes using recently published datasets and models to study the general spatial patterns in C fluxes and convergence across datasets and models. Terrestrial ecosystem NEE fluxes are derived from various recent model inter-comparisons and outputs and in-situ synthesis datasets (Table S1 in Supporting Information S1); annual CH 4 fluxes are from two in-situ syntheses (Kuhn et al., 2021b;Treat et al., 2018b) and one statistical upscaling-based on eddycovariance (Peltola et al., 2019a). For North America, the regions included Alaska, Canadian tundra, boreal Western Canada, and Eastern Canada. For Eurasia, these included Western Eurasia, Siberian tundra, Eastern Siberia, and Western Siberia. This regional approach can help to target new areas for measurements based on key differences indicative of a lack of understanding of the underlying processes. We limited these datasets to the permafrost region within the northern tundra and boreal biomes, similar to the Regional Carbon Cycle Assessment and Processes Project 2 (RECAPP-2) permafrost effort (Ciais et al., 2022;Hugelius et al., 2023).

The results from our comparison among datasets and models show stronger regional CO 2 sinks in the southern permafrost region, while lower net CO 2 uptake or net CO 2 emissions occur toward the north (Figures 4 and 5). for different ecosystem and wetland classes found in the permafrost region using two different synthesis datasets (BAWLD: Kuhn et al., 2021a;Treat et al., 2018a). Significant differences were found in CH 4 emissions between ecosystem classes (F 6,202 = 6.0, p < 0.0001) but not between datasets. Ecosystem classes were categorized as marsh, fen, bog, permafrost wetland (PermWet), permafrost bog (PermBog, including peat plateaus and palsas), boreal forest (Boreal), and upland tundra (UpTundra).

Journal of Geophysical Research: Biogeosciences

10.1029/2023JG007638

This regional pattern in CO 2 fluxes is likely related to temperature, radiation regime, and growing season length, in agreement with earlier syntheses (McGuire et al., 2012;Virkkala et al., 2021a). The highest median annual CO 2 sinks were located in western Canada (-52 g C m -2 yr -1 ) and western Siberia (-41 g C m -2 yr -1 ), and smallest CO 2 sinks in Alaska (-6 g C m -2 yr -1 ) and Siberian tundra (-5 g C m -2 yr -1 ; Table 3). Some statistically significant differences occurred between regions that were strong sinks and small sinks to net sources (Figure 4; F 7,31 = 4.29, p < 0.01). The CH 4 syntheses show highest annual fluxes from the Siberian tundra and Western Eurasian regions (Figure 5, median = 15.5-17.9 g CH 4 m -2 y -1 ) but no statistically significant differences between regions were found.

Regional differences in wetland CH 4 fluxes were highly variable among chamber-based synthesis studies (Figure 5a), with regional medians ranging from 1.6 to 18 g CH 4 m -2 y -1 . The variability was smaller for the eddy-covariance based upscaling (5.7-13 g CH 4 m -2 y -1 ). Colder regions with thinner sediments in Canadian tundra and Eastern Canada tended to have lower CH 4 fluxes (Figures 5a and 1a) while highest annual CH 4 fluxes were found in Eurasia. Relatively few annual measurements have been reported for Hudson Bay Lowlands and Taiga Plains (Canada) and Western Siberia (Figures 2b and 5b), home to the largest peatland complexes in the world (Hugelius et al., 2020). Comparing the coefficient of variation among the datasets showed a mean of 0.29 across the regions with the best agreement in western Canada (0.05) and worst in eastern Canada (0.53), despite having a similar number of observations. Given that CH 4 emissions vary strongly among wetland classes (Figure 3a), some variability among the methods may be due to differences among the wetland types measured and synthesized within the regions (Treat et al., 2018b), which may or may not reflect the distribution of wetland types across the landscape (Kuhn et al., 2021b;Olefeldt et al., 2021).

These synthesis datasets also show some biases toward C hotspots: most sites measuring CO 2 and CH 4 fluxes are in wetlands or moist-wet ecosystems with high CH 4 emissions and high growing season CO 2 sinks (Figure 3b; Virkkala et al., 2022). Drier ecosystems including boreal forests, sparsely vegetated regions, and mountainous areas remain less studied (Figure 3b; Pallandt et al., 2022;Virkkala et al., 2022) despite covering ca. 80% of the permafrost region (Karesdotter et al., 2021;Olefeldt et al., 2021). This limits our ability to detect changes in C fluxes because even small changes in the site distribution (e.g., new sites being set up in new environments), methodology (e.g., chambers or towers synthesized), and data coverage can impact the average sign of fluxes or direction in trends when data are aggregated over larger domains (Belshe et al., 2013;McGuire et al., 2012). Note. The in-situ column also includes the number of sites from the entire permafrost domain which is relatively similar to the proportion of measurement years in total in the dataset. Standard deviations were calculated for each year and model separately and averaged across all models, and thus represents average standard deviation around the mean and describes the spatial flux variability within the region. Note that inversion estimates include lake CO 2 fluxes as well, but fossil fuel emissions, cement carbonation sink, lateral fluxes and fire emissions have been masked away.

Manual flux chamber measurements are distributed more broadly across the permafrost region than eddy covariance measurements and could help to offset some spatial biases and data gaps, particularly for CH 4 fluxes (Figure 2). However, barriers remain to using these manual chamber data for modeling because of the limited spatial and temporal scales of measurements; statistical upscaling may offer some possibilities to further use these data (Natali et al., 2019;Virkkala et al., 2021a). Semi-permanent mobile towers or automated chambers could be utilized to enhance spatial coverage and complement the existing flux network of long-term monitoring sites (Varner et al., 2022;Voigt et al., 2023). Further improvements in flux estimates can be expected as new sites are Journal of Geophysical Research: Biogeosciences 10.1029/2023JG007638

added, more recent data are integrated to repositories, and newer methods are developed to leverage the sparse and disparate existing datasets.

Long-Term Trends in CO 2 and CH 4 Fluxes

How CO 2 and CH 4 exchange have changed over time in the permafrost region remains unknown. Circumpolar CO 2 trend analyses show an increasing growing season sink in the tundra (Belshe et al., 2013), a small and relatively negligible trend in non-growing season NEE in the permafrost region (Natali et al., 2019), but no clear changes in annual NEE despite increases in GPP and ER in the tundra (Belshe et al., 2013;Z.-L. Li et al., 2021).

Long-term (>15-year) of measurements of CO 2 in sub-Arctic tundra sites show diverging trends: one shows an increasing net loss of CO 2 (Schuur et al., 2021), while the other shows enhanced CO 2 uptake following changes in vegetation with permafrost thaw (Varner et al., 2022).

Long-term measurements of CH 4 fluxes are rare (Christensen et al., 2017;Pallandt et al., 2022) but flux magnitudes have been shown to be increasing at the site-level for two permafrost sites in Eurasia over the past decades (Rößger et al., 2022;Varner et al., 2022). However, in North America, an analysis of concentration enhancements on the Alaska Slope found no change in CH 4 flux magnitude over time (Sweeney et al., 2016). Similarly, there was no trend in 10 years of CH 4 flux measurement at a fen in interior Alaska (Olefeldt et al., 2017). Unfortunately, the data density in the CH 4 synthesis datasets included here was not sufficient to detect trends in emissions (e.g., Basu et al., 2022) or response to regionally warm and wet conditions that might enhance wetland CH 4 emissions to the extent that they affect global atmospheric CH 4 concentrations (Peng et al., 2022). Additional long-term measurements are needed to establish whether trends are occurring against a background of interannual variability and local processes (Hiyama et al., 2021). A synthesis of the limited long-term records of CO 2 and CH 4 exchange across multiple sites within the permafrost domain would be valuable.

CO 2 and CH 4 fluxes in Changing and Disturbed Environments

Understanding trends in C fluxes is challenging, because climate warming is affecting the timing and characteristics of seasonality in permafrost ecosystems, which has complex interactions with the environmental controls on C cycling. Warmer air temperatures in the winter and shoulder seasons result in longer duration of soil thaw (Farquharson et al., 2022;Y. Kim et al., 2012), lengthening the duration of microbial activity in the soil and affecting cold season fluxes as discussed above. The timing of snowmelt and the onset of the growing season are key controls of growing season NEE (Bellisario et al., 1998;Groendahl et al., 2007); the timing of these events has shifted earlier in the last decades (Xu et al., 2018). There is some evidence that the lengthening of the growing season increases the growing season C sink due to enhanced plant C uptake and increased vegetation biomass (Belshe et al., 2013;Bruhwiler et al., 2021). However, interactions with moisture seem to be a key determinant of the net growing season C uptake. For example, warmer peak growing season temperature can increase net summer C uptake through enhanced photosynthesis but warming also increases evapotranspiration, reducing available soil moisture and potentially increasing ER (J. Kim et al., 2021). Further, while earlier snowmelt might enhance net C uptake at the beginning of the growing season, the dry and warm conditions resulting from earlier snowmelt might increase ecosystem CO 2 losses during the late growing season (Belshe et al., 2013;Helbig et al., 2022). Further observations and enhanced linkages between biophysical processes, vegetation, and C cycles are needed.

Permafrost thaw and the associated carbon feedbacks have been increasingly well-studied (Schuur et al., 2022;Sjöberg et al., 2020;Virkkala et al., 2018), both as gradual thaw and abrupt thaw. Site-level studies indicate that CH 4 and CO 2 emissions can be strongly positively correlated with active layer depth due to the effects of increasing soil temperature on microbial activity, so gradual thaw of permafrost that deepens the soil active layer results in larger C emissions (Celis et al., 2017;Galera et al., 2023). Estimates of C loss from abrupt thaw may exceed those from active layer deepening but are highly uncertain (Estop-Aragonés et al., 2020;Zolkos et al., 2022). For example, less than ten site-level studies were available to use for a recent in-situ-based greenhouse gas budget estimate that showed that areas affected by abrupt thaw were net emitters of 31 (21, 42) Tg CO 2 -C yr -1 and 31 (20, 42) Tg CH 4 -C yr -1 (Ramage et al., 2023;Turetsky et al., 2020); the large uncertainties represent the potential spatial distribution of abrupt thaw areas that have only been quantified in limited regions (Nitze et al., 2018). To our knowledge, terrestrial sites experiencing abrupt thaw that have measured multi-year CO 2 or CH 4 fluxes are limited to wet graminoid ecosystems in Alaska (Schuur et al., 2021), boreal black spruce lowlands in Canada and Alaska (Euskirchen et al., 2017;Helbig et al., 2017), and collapsing palsas from Fennoscandia (Varner et al., 2022). However, the current site network misses thaw slumps, gullies, and active layer detachments (Cassidy et al., 2016) that cover <1% of the areas affected by abrupt thaw; overall abrupt thaw is estimated to affect ∼7% of the permafrost region in total (Ramage et al., 2023). Gradual and abrupt permafrost thaw cause changes in hydrology, often increasing soil moisture and/or lake extent, thus often increasing CH 4 emissions (Helbig et al., 2017;Miner et al., 2022;Varner et al., 2022). Many sites that have been observed to experience gradual or abrupt permafrost thaw are currently net C sources to the atmosphere (Euskirchen et al., 2017;Schuur et al., 2021); historically, some sites have shifted back to sequestering C centuries to millennia after permafrost thaw (M. C. Jones et al., 2017;Walter Anthony et al., 2014) but it is unclear whether this can be expected in the next centuries if temperatures continue to rise (M. C. Jones et al., 2023).

Warming is increasing the magnitude, extent, and severity of other disturbances in the permafrost region including wildfire, insect outbreaks, flooding, and drought (Foster et al., 2022;Meredith et al., 2019). These disturbances can impact C cycling directly through, for example, C emissions from fire combustion, and indirectly, by altering environmental conditions that control C fluxes, such as soil moisture, temperature, light availability, and species composition. Wildfire extent and severity has been increasing in the past decades (M. W. Jones et al., 2022); wildfire-induced changes to vegetation and soils can affect permafrost stability (Holloway et al., 2020), likely driving compounded effects on ecosystem C cycling (Harden et al., 2006;X.-Y. Li et al., 2021;Mack et al., 2021). The time required for C accumulation post-fire to offset wildfire C emissions takes decades and remains an open question (Mack et al., 2021;Ueyama et al., 2019;X. J. Walker et al., 2019). Additionally, overwintering fires are fundamentally changing fire dynamics and accelerating the fire season (Scholten et al., 2021). The effects of insect outbreaks might be severe during the outbreak but increased C uptake during the following years can compensate for the earlier losses (Lund et al., 2017;Ruess et al., 2021). Similar dynamics might occur with extreme meteorological events such as drought, flooding, and lack of snow but impacts are unclear (Olefeldt et al., 2017;Treharne et al., 2019). Interactions between permafrost, large herbivores, and soil C are an interesting area of research, however, the introduction of large herbivores is unlikely to stop the increasing carbon emissions from permafrost thaw at a circumpolar scale (Zimov et al., 2009). Increasing human presence is also impacting Arctic lands (Friedrich et al., 2022), but little is understood about effects on emissions such as increased fugitive CH 4 emissions (e.g., leaky infrastructure; Klotz et al., 2023), land use change emissions (Strack et al., 2019), or effects of the interactions between land use change and permafrost thaw (Ward Jones et al., 2022).

Overall, an improved understanding requires new cross-disciplinary approaches to understand the magnitude of these processes across the entire permafrost domain.

Modeling the Carbon Fluxes in the Terrestrial Permafrost Region

Main Modeling Approaches for C Exchange

Bottom-up C cycle models, that is, mechanistic process models, statistical and machine learning-based upscaling approaches, and top-down models (atmospheric inversions) are critical tools for estimating permafrost region C budgets. Process models are widely used to extrapolate and predict C fluxes both into the past and future (Koven et al., 2015;Lawrence et al., 2012;McGuire et al., 2016McGuire et al., , 2018b) ) because they represent mechanistic understanding of processes at various scales. In the context of Arctic-boreal C budgets, land surface models (LSMs) of varying complexity can be used to represent relevant processes, such as dynamic vegetation and permafrost carbon. These can either be included within an earth system model (ESM) or driven in standalone mode by meteorological data. ESMs simulate coupled and dynamic interactions between Earth's climate system of oceans, atmosphere, cryosphere, and land surface and can include feedbacks from the land surface onto the atmosphere (Fisher et al., 2014). In addition to individual process-based models, coordinated research collaborations facilitating large model intercomparisons and ensembles (MIPs) have been key in exploring C budgets and several process model intercomparison studies exist for the permafrost region in addition to individual process models (McGuire et al., 2012(McGuire et al., , 2016(McGuire et al., , 2018b)).

A few pan-Arctic studies have used statistical and machine learning models to upscale recent or current C fluxes at high spatial resolutions across larger domains or higher temporal resolutions (Jung et al., 2020;McNicol et al., 2023;Natali et al., 2019;Peltola et al., 2019a;Virkkala et al., 2021a). Earlier approaches often used simpler empirical upscaling of flux measurements (e.g., Bartlett & Harriss, 1993). These model types can be flexible with driver data and new datasets can thus easily be integrated but they have limited predictive capability; here, data Journal of Geophysical Research: Biogeosciences 10.1029/2023JG007638 assimilation systems such as the CARbon DAta MOdel (CARDAMOM) that integrates various data sources with less complex process models might be a solution for better predictions (López-Blanco et al., 2019;Y. Q. Luo et al., 2012). Additionally, top-down atmospheric inversion models are constrained by atmospheric data where concentration changes are linked to flux and atmospheric transport and are often spatially coarser than the bottomup approaches (Bruhwiler et al., 2021;Byrne et al., 2023;Z. Liu et al., 2022).

Bottom-up and top-down models have different main uses as well as strengths and limitations. Flux upscaling using statistical and machine learning approaches is still a relatively new field and has only been used in a few pan-Arctic studies; model intercomparisons may not yet be possible and may be limited by the number of pan-Arctic sites. Inversions have been used in permafrost region flux studies for over a decade already, but the number of inversion intercomparisons is still relatively low, and atmospheric observations in this area are scarce (Bruhwiler et al., 2021;Z. Liu et al., 2022;McGuire et al., 2012). In summary, bottom-up and top-down approaches complement each other and are important for predicting C emission and uptake patterns across the permafrost region.

Modeling Insights Into CO 2 Cycling in the Permafrost Region

Here we compared magnitudes of NEE among process-based modeling, inversion modeling, and statistical upscaling of in-situ data approaches for the regions used in earlier analysis (Supporting Information S1, Table 1).

The models include results from the Coupled Model Intercomparison Phase 6 (CMIP6) assessed for the IPCC AR6 report (Canadell et al., 2021; IPCC, and the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP), which provides historical runs and projections across the 21st century using various different driving data (Lange, 2019); other intercomparison projects not addressed here include the Coupled Climate Carbon Cycle MIP (C4MIP; Canadell et al., 2021), the TRENDY project (Friedlingstein et al., 2022;Sitch et al., 2015), and the Multi-scale Synthesis and Terrestrial Model Intercomparison Project (MsTMIP; Huntzinger et al., 2020).

In general, models and in-situ data had some agreement in regional NEE estimates with many of the approaches in each region agreeing on the sign of NEE (i.e., net sink or source). However, differences in NEE among approaches were still relatively high, with the average range of annual NEE estimates of 41 g C m -2 yr -1 (Figures 4 and 6). The best agreement in average NEE was found in the Siberian tundra and Eastern Canada which were small to moderate CO 2 sinks, respectively (Figure 4). This was unexpected, because these are also areas that have low flux data coverage (Table 3). The largest variability in mean NEE was found in western Siberia where the ISIMIP and inversion models showed a much stronger (>25 g C m -2 yr -1 ) average sink than the other approaches; recent remote sensing analyses show a decreasing sink strength in Siberia driven by disturbance (Fan et al., 2023). While part of this disagreement is simply due to the high overall fluxes in this forest-dominated region, new measurements and process-level understanding of disturbance effects in this domain are critical to resolving this issue.

The largest differences among approaches were found between ISIMIP models and in-situ and/or upscaled estimates (e.g., in Alaska and Siberian tundra; Figures 4 and 6). This might suggest that the ISIMIP LSMs underestimate CO 2 emissions in this region, assuming that in-situ based estimates are reliable and representative of each region (Figure 4). The CMIP6 ESMs show weaker sink strength than both the ISIMIP LSMs and the inversions (both on average ca. 20 g C m -2 yr -1 weaker), which might be related to CMIP6 models underestimating the C sink strength in the permafrost region (see Section 3.3). While one could assume that the in-situ based averages and upscaling provide the most accurate estimates as they integrate recent data, they also suffer from severe data gaps and thus extrapolation uncertainties in some regions (see Section 2.2). Overall, the variability among approaches highlights the need for both additional data and development of predictive models as discussed in key challenges below.

Key Advancements and Challenges in Modeling Carbon Cycling in the Permafrost Region

LSMs have improved their representation of permafrost over the years, for example, by realistically simulating the thermal and hydraulic properties of soil, including phase change of soil water, and by accounting for the insulating effects of moss and snow cover (Chadburn et al., 2015;Ekici et al., 2014;Nicolsky et al., 2007). Despite these important advances to their land surface schemes, the CMIP6 ESMs included in the latest IPCC report still have a limited representation of C cycle processes in high-latitude regions. In the CMIP6 model ensemble, soil C stocks across the permafrost region were severely underestimated (Varney et al., 2022), likely

Journal of Geophysical Research: Biogeosciences 10.1029/2023JG007638

leading to an underestimation of the potential for C-climate feedbacks from these frozen soils. Only two of the CMIP6 models included a representation of permafrost C in soils (CESM and NorESM), which improved C stocks estimates in the permafrost region. The relatively short spin-up time of some models (on the order of centuries) compared to the slow build-up time of permafrost C over many millennia-especially for C-rich Pleistocene Yedoma deposits (Lindgren et al., 2018) and Holocene peatlands (Yu et al., 2010)-may be one reason for this underestimation (Huntzinger et al., 2020;Schwalm et al., 2019). Alternatively, inaccurate representations of vegetation cover and plant-derived C and nutrient inputs to the soil may also be responsible for low soil C stocks (Varney et al., 2022). Given the important role of soil C stocks in the permafrost C feedback, as well as the potential for C accumulation in soils with permafrost thaw (Treat et al., 2021), it is crucial to both simulate soil C stocks as well as demonstrate the potential for both soil C accumulation and loss.

Capturing vegetation dynamics is also critical to modeling permafrost dynamics but many dynamic global vegetation models (DGVMs; a type of LSMs that addresses the behavior and changes in vegetation) were originally developed to represent the biomes of lower latitudes where extreme winter conditions are absent (Bruhwiler et al., 2021;Lambert et al., 2022). The high degree of disagreement among models predicting future C balance in the permafrost region is attributed to uncertainty about whether plant productivity and subsequent ecosystem C uptake will compensate for permafrost C release (McGuire et al., 2018b). One limitation in the CMIP6 models was that only a few included vegetation dynamics (Canadell et al., 2021); those that did simulated Arctic grasses rather than dwarf shrubs and struggled to correctly simulate the seasonal trends of leaf area index (LAI; Song et al., 2021). In addition, accounting for nutrient limitations is essential to avoid an unrealistically strong vegetation response to CO 2 fertilization (Zaehle et al., 2015), but of the 11 land carbon cycle models used in CMIP6 ESMs, only six included a nitrogen cycle (Canadell et al., 2021).

Future model projections remain highly uncertain whether the permafrost region will act as a C source or sink (Braghiere et al., 2023). In addition to challenges with soils and vegetation, current LSMs miss the capability to ). The "Across all models" map was produced so that each modeling approach (inversions, process-based, and upscaling models) received equal weight. Note that inversion estimates include lake CO 2 fluxes as well, but fossil fuel emissions, cement carbonation sink, lateral fluxes and fire emissions have been removed; and the upscaling only includes one model and agreement cannot be calculated; thus values are either 0 (not a sink/neutral/source) or 100 (is a sink/neutral/ source).

simulate abrupt changes following disturbances. While five of 11 models included in the land carbon cycle models used in CMIP6 ESMs simulated fire, none of them included fire-permafrost-carbon interactions (Canadell et al., 2021). Thermokarst processes are also absent although they can to a certain extent be represented in LSMs (N. D. Smith et al., 2022). Vegetation-specific disturbances such as insect outbreaks, frost damage, and droughts can affect the C balance (Reichstein et al., 2013), but improvements to vegetation dynamics should be priority. Furthermore, the contribution of peatland, inland aquatic ecosystems, and the lateral carbon fluxes between terrestrial and aquatic systems are not included in CMIP6 models but are included in regional modeling studies of C fluxes in the permafrost region (Chaudhary et al., 2020;Kicklighter et al., 2013;Lyu et al., 2018;McGuire et al., 2018a). The limited representation of processes is due to their complexity as well as the lack of observations integrating interactions between terrestrial and aquatic systems (Vonk et al., 2019). Overall, the potential for C sequestration in peatland and other soils (Treat et al., 2021), and other region-specific disturbances such as abrupt permafrost thaw (Turetsky et al., 2020) should be a major focus of future model development to achieve a more accurate quantification of the permafrost C feedback.

Progress in modeling wetland CH 4 fluxes in high-latitude regions has been made over the past decades (Xiaofeng Xu, Yuan, et al., 2016). Site-scale validation of process-based LSMs suggest that models generally capture wetland CH 4 variability well at seasonal and longer time scales but perform poorly at shorter time scales (<15 days; Zhen Zhang et al., 2023). Model-data comparisons show some issues with seasonality, including a strong underestimation of non-growing season (October-April) CH 4 emissions by as much as two-thirds (Ito et al., 2023;Miller et al., 2016;Treat et al., 2018b;Xiyan Xu, Yuan, et al., 2016). Nevertheless, these data-model integration efforts do highlight that Arctic-boreal wetland CH 4 processes are better captured than those in tropical wetlands (Delwiche et al., 2021;McNicol et al., 2023;Zhen Zhang et al., 2023).

Methane flux models still face challenges and uncertainties, particularly in defining the past and present extent of wetlands (Bloom et al., 2017;Peltola et al., 2019a;Saunois et al., 2020), capturing the spatial and temporal heterogeneity of wetland ecosystems in terms of soil moisture, inundation variability, including the vegetation communities, and predicting the effects of permafrost thaw on CH 4 dynamics (Koven et al., 2011(Koven et al., , 2015)). These factors add uncertainty to data-driven flux upscaling and atmospheric inversions through a priori flux assumptions (Bruhwiler et al., 2021;Peltola et al., 2019a;Saunois et al., 2020). However, improvements in the recent wetland maps in Boreal-Arctic Wetland Lake Database (BAWLD) and Wetland Area and Dynamics for Methane Modeling (WAD2M) are promising (Olefeldt et al., 2021;Z. Zhang et al., 2021). Model intercomparisons have generated important maps and budget estimates of CO 2 fluxes but are relatively uncommon for CH 4 (Bloom et al., 2017;Collier et al., 2018;Ito et al., 2023;Melton et al., 2013), and should be undertaken as more models are developed. Challenges also remain for modeling CH 4 cycling beyond the borders of wetlands, particularly in uplands and lakes. Uplands cover close to 80% of the permafrost region and can be both annual CH 4 sources (Zona et al., 2016) and sinks (Oh et al., 2020;Voigt et al., 2023). Wetlands and lakes have differing CH 4 emissions and processes (Kuhn et al., 2021b;Wik et al., 2016), but distinguishing these landforms in observations and remote sensing images can be difficult, leading to possible double counting of emissions sources (Thornton et al., 2016). Hybrid process modeling together with remote sensing and eddy covariance data have been used to estimate wetland CH 4 fluxes relatively accurately (Watts et al., 2023), which incorporates important factors such as soil moisture, temperature, vegetation characteristics, and hydrological dynamics to estimate wetland CH 4 fluxes.

Atmospheric inversion model ensembles are an integral part of determining global CO 2 and CH 4 budgets as they aggregate natural terrestrial and aquatic as well as anthropogenic sources over large domains (Friedlingstein et al., 2022;Saunois et al., 2020). Full ensembles have been less frequently used in the permafrost region where atmospheric inversions have a large model spread in CO 2 and CH 4 fluxes due to differing transport models, priors, and observations (Bruhwiler et al., 2021;Z. Liu et al., 2022). However, models are rapidly evolving. For example, airborne and satellite data are being more extensively used to define the prior estimates for inversions (Byrne et al., 2023;Tsuruta et al., 2023). While promising, satellite observations based on optical remote sensing still have some limitations for application during polar winter and with persistent cloud cover. Improvements should still be made toward better maps of surface conditions to better delineate flux surface fields (e.g., wetland distribution), an expanded tall tower network for better mixing ratio and isotopic data (Basu et al., 2022), and comprehensive sensitivity tests regarding transport modeling to understand Arctic-specific conditions (e.g., influence of polar vortex and shallow and stable boundary layers). Further iterations between top-down and bottomup modeling informed and constrained by observational data have strong potential to resolve discrepancies in Journal of Geophysical Research: Biogeosciences 10.1029/2023JG007638 permafrost C budgets (Commane et al., 2017;Elder et al., 2021;Miller et al., 2016); developments in model benchmarking systems and data assimilation will also help with furthering understanding and refining estimates (Collier et al., 2018;Y. Q. Luo et al., 2012;Stofferahn et al., 2019).

Summary of the Next Steps

This review highlights significant progress in permafrost C cycle science since early permafrost maps and C flux syntheses (Tables 1 and 2). Major recent methodological advances include new geospatial data products describing permafrost conditions and soil C, nearly continuous records of CO 2 and CH 4 fluxes from eddy covariance towers across the permafrost domain, and the incorporation of permafrost-relevant characteristics into multiple process and machine-learning based models that can be used to simulate CO 2 and CH 4 fluxes. Several new key research topics have also emerged. Non-growing season emissions have a larger role in the annual C balance than previously thought, and even more so in a warmer climate. Vegetation shifts and enhanced productivity are key processes potentially mitigating positive permafrost climate feedbacks but might not always lead to increasing net annual C uptake because they can also alter soil microclimate and chemistry in a way that accelerates C emissions. Permafrost thaw is known to impact C cycling not only gradually but also abruptly, and in interaction with other disturbances, such as wildfires, will likely increase terrestrial C emissions to the atmosphere. For CH 4 , new hotspots such as thermogenic vents and craters as well as coldspots (areas with high uptake rates) are still being investigated. With the Arctic warming potentially up to four times faster than the global average (Rantanen et al., 2022), and permafrost thaw already happening faster than predicted in some parts of the region (Fewster et al., 2022), new processes and potentially novel ecosystems will likely emerge.

The integration of new process understanding from individual sites to cross-site data syntheses, and from individual models to model intercomparisons has been critical to estimating permafrost region C budgets and their trends. These data-model integration efforts have shown that while permafrost regions are cold and processes are slow, they still play a substantial role in the global C cycle. The permafrost region CH 4 budget ranges between 10 and 50 Tg CH 4 yr -1 ; trends over time remain uncertain due to the sparsity of data. The terrestrial CO 2 budget (a balance between GPP and ER) represents a relatively strong CO 2 sink (-700 to -100 Tg C yr -1) , and there is evidence of both increasing growing season plant uptake and non-growing season C emissions. However, the partial disagreement across modeling approaches and syntheses, large spread of the estimated budgets, and unclear regional patterns and temporal trends shows fact that large uncertainties remain (Figures 456). The increased intensity and number of wildfires adds uncertainty to the evaluation of annual C balance in the permafrost region since a large fire year may offset multiple years of regional C uptake (M. W. Jones et al., 2022;Mack et al., 2021;X. J. Walker et al., 2019). Considering these challenges, we outline several research priorities below.

1. Process-based knowledge: Weather extremes and disturbances cause large inter-annual variability in C fluxes and change the contributions of the two key C fluxes-CO 2 and CH 4 -to the total C budget. At the same time, hydrological changes associated with permafrost thaw make understanding moisture gradients and terrestrial-aquatic interfaces more important to understand the controls of C cycling. As such, CO 2 and CH 4 exchange between ecosystems and the atmosphere do not capture the full response of permafrost C losses; lateral C fluxes also need to be quantified. New knowledge about extreme event impacts such as winter and summer droughts, fires, and insect outbreaks and their compound effects on C cycling derived from long-term field sites or controlled experiments targeting these extremes, and measurements in currently under-sampled drier upland landscapes and areas experiencing rapid disturbances, such as abrupt permafrost thaw, are crucial. 2. Observations and syntheses: While the network of sites with continuous observations is steadily increasing and subsequent data syntheses grow in scope (from 30 to 200 sites), detecting hotspots, hot moments, and longterm trends of in-situ CO 2 and CH 4 fluxes remains a challenge. Therefore, the observational network capacity must be increased to support the continuity of long-term eddy covariance CO data gaps (spatially and across ecosystem types). Further improvements to environmental data such as soil C, dominant plant species and their traits, and permafrost thaw status would help contextualize and upscale flux data. 3. Modeling: The three broad types of modeling approaches -statistical or machine learning-based upscaling, process modeling, and inversion approaches -are all needed to predict C fluxes in the permafrost domain.

Process models are the most widely used technique to predict C fluxes but there are limitations related to coldseason emissions, belowground plant-soil feedbacks, permafrost thaw, disturbance history, as well as capturing temporal lags, tipping points, and non-linear responses. In addition, dynamic and spatially higher resolution wetland, soil moisture, and disturbance maps are needed to capture the rapidly changing permafrost landscapes, for example, the distribution of gradual and abrupt permafrost thaw. Using monitoring data to inform process-based and inversion models through data assimilation techniques could allow substantial decrease in model uncertainties (Y. Luo & Schuur, 2020). As more geospatial permafrost-related data products become available and new study sites are measured, better simulations and analyses of the dynamic processes that drive change in the permafrost region are possible. 4. Model and data intercomparisons: Regularly benchmarking and exploring the model-based magnitudes, trends, and drivers of C fluxes is necessary to identify areas of convergence and divergence between models and in-situ measurements (Collier et al., 2018). Determining whether key processes for the permafrost region identified by observations are included or adequately represented can significantly improve process-based model performance (Koven et al., 2011), as is identifying benchmarking metrics to constrain predictions (Schwalm et al., 2019). In particular, new CH 4 model intercomparisons are needed, especially as CH 4 models become more numerous and incorporate additional attributes. This ongoing evaluation will help improve our understanding and predictions of the permafrost region C fluxes.

While knowledge gaps remain, we anticipate the next decades to bring significant improvements in our processlevel understanding and C budget estimates in the permafrost region. Continued coordinated efforts among the field, remote sensing, and modeling communities is required to integrate new knowledge throughout the knowledge chain from observations to modeling and predictions and finally to policy, and to most effectively constrain the permafrost region C budget (Fisher et al., 2018;Natali et al., 2022). Open data policies, reduced latency between observations and reporting, as well as improved methodological protocols, instrumentation and model intercomparisons need to be adopted moving forward. International networks addressing the permafrost region remain important, like the Permafrost Carbon Network and synthesis projects (Schuur et al., 2022), Arctic Monitoring and Assessment Program (AMAP) (Christensen et al., 2017), and RECCAPs (Ciais et al., 2022;McGuire et al., 2012) to understand and inform policy makers on ways to best protect and preserve these rapidly changing, sensitive permafrost ecosystems.