Increasing Methane Emissions and Widespread Cold‐Season Release From High‐Arctic Regions Detected Through Atmospheric Measurements

Rising Arctic temperatures pose a threat to the large carbon stores trapped in Arctic permafrost. To assess methane emissions in high‐Arctic regions, we analyzed atmospheric data from Alaska and Siberia using two methods: (a) a wind sector approach to calculate emission changes based on concentration enhancements using wind direction, and (b) an inversion method utilizing a high‐resolution atmospheric transport model. Incorporating data after 2015, we observed a significant rise in methane emissions (0.018 ± 0.005 Tg yr−2 from 2000 to 2021) from Alaska's North Slope, indicating a shift from previous analyses. We find 34%–50% of yearly emissions occurred in the late season (September–December) consistently across multiple years and regions, which is historically underestimated in models and inventories. Our findings reveal significant changes occurring in the Arctic, highlighting the crucial role of long‐term atmospheric measurements in monitoring the region, especially during the cold season.


Introduction
The Arctic is experiencing temperature increase four times higher than the global average (Rantanen et al., 2022), yet its role in the global carbon budget is poorly constrained (Meredith et al., 2019).With 1,100-1,500 Pg of carbon stored within the permafrost (Hugelius et al., 2014), rising temperatures could result in this carbon being released into the atmosphere as carbon dioxide and methane.In particular, release as methane could have a major impact on the global climate, due to the potency of methane as a greenhouse gas (Myhre et al., 2013).It is therefore crucial to understand and quantify terrestrial sources of methane in the Arctic, and to advance methods that offer long-term monitoring in the region.
Recent studies have shown that significant tundra emissions of methane may be occurring during the cold season (September-May), once temperatures have lowered and snow cover has returned (Arndt et al., 2019;Bao et al., 2021;Mastepanov et al., 2008Mastepanov et al., , 2013;;Treat et al., 2018;Zona et al., 2016).These cold season emissions are the result of a prolonged "zero-curtain", allowing the permafrost active layer to remain thawed whilst the ground above and below is frozen (Hinkel et al., 2001).Over 50% of the Arctic's annual methane emissions could be originating from the cold season (Zona et al., 2016), a finding often overlooked in global emission assessments (Saunois et al., 2020).
The majority of high-Arctic zero-curtain studies have been conducted in Alaska (Arndt et al., 2019;Bao et al., 2021;Tao et al., 2021;Zona et al., 2016).The few studies made over Siberia also suggest the potential significance of the cold season on the annual methane budget (Kittler et al., 2017;Rößger et al., 2022).Most zerocurtain studies rely on flux measurements, which capture small spatial scales (hundreds of meters) and may not accurately reflect pan-Arctic scales.
A small number of atmospheric measurement stations situated across the Arctic coastline continuously measure atmospheric methane mole fraction.In contrast to flux measurements, well-calibrated atmospheric mole fraction measurements detect small changes in atmospheric composition that represent emissions over a large spatial scale (hundreds of kilometers).These measurements are made using instruments that often are in operation for years to decades, allowing the detection of gradual changes in emissions over time.Moreover, measuring fluxes during the cold season in the Arctic has historically posed challenges due to adverse weather conditions.Continuous and automated atmospheric measurements offer a means to monitor these periods with fewer logistical constraints.Despite significant increases in surface air temperature over the past decades, Sweeney et al. (2016), showed that changes in July-September methane emissions from the North Slope of Alaska between 1986 and 2015 were not statistically different from zero.However, a significant trend in November-December emissions was observed.This study employed mole fraction measurements from the National Oceanic and Atmospheric Administration's Barrow Observatory in Utqiagvik, Alaska.Wind direction was used to identify the times that CH 4 mole fraction reflected a clean "background" sector (air coming from the ocean before arriving at Barrow), versus times that mole fractions represented land emissions (air passing over land).The difference between land and background mole fractions reflected the strength of emissions.
Emissions can also be inferred from atmospheric mole fractions through inverse modeling, which has the benefit over directly using mole fraction enhancements (as in Sweeney et al. (2016)), in that it takes into account more complex meteorological information such as changes in planetary boundary layer height and wind speed.Arctic methane emissions derived from inverse methods (Berchet et al., 2015;Ishizawa et al., 2019;Tenkanen et al., 2021;Thompson et al., 2017;Tsuruta et al., 2023;Wittig et al., 2023) have primarily focused on large-scale regions, such as entire countries or highly inundated regions such as the West Siberian Lowlands or Hudson Bay Lowlands.Additionally, these studies have not focused on late-season emissions from the high Arctic, a region that contains a large proportion of dry tundra (less than 20% inundation), which has the potential to emit significantly during the cold season (Treat et al., 2018;Zona et al., 2016).Zona et al. (2016) found that tundra with less than 5% inundation was strongly emitting during the cold season.
Our study uses multiple methods to quantify emissions over three high-latitude Arctic tundra regions-the North Slope of Alaska, the Taymyr Peninsula and the East Siberian Lowlands.These regions were selected because they are in the rapidly warming high Arctic (Rantanen et al., 2022;Turetsky et al., 2019), and contain large stores of carbon (Hugelius et al., 2014).The defined regions also contain a high proportion of dry tundra (Olefeldt et al., 2021), which are much less inundated than more monitored regions such as the Hudson Bay Lowlands and West Siberian Lowlands.We determine whether late season emissions are prevalent across the three regions and are persistent over time, and whether emissions are exhibiting significant changes over decadal timescales.

Methods
We employ two approaches to determine trends and to quantify emissions.The first method extends the results of Sweeney et al. (2016), including data from 2016 to 2021 in order to assess any changes North Slope emission trends.As this method is based on calculating enhancements in methane mole fraction from the land versus background sectors, it is only able to determine relative changes in emissions over time.The second method uses a hierarchical Bayesian inversion framework, with an atmospheric transport model (Numerical Atmosphericdispersion Modeling Environment (NAME) (Jones et al., 2007)).The atmospheric transport model provides a pathway for directly estimating emissions from mole fractions by simulating the dilution of emissions in the atmosphere, and thus offers absolute quantification of emissions.Model meteorological fields used to run NAME are available from the year 2000 and emissions are quantified from 2000 to 2021 for the North Slope and from 2014 to 2021 for the two Siberian regions.Emissions derived through the inversion are masked for anthropogenic emissions using filters and sector attribution methods.
Throughout this work, we define the early (April-June), summertime (July-August) and late (September-December) seasons to correspond to typical spring thaw, growing season and fall zero curtain periods (Zona et al., 2016) for the high Arctic.

Measurements
We used high frequency in situ atmospheric mole fraction measurements of CH 4 from 11 sites across the Arctic (described in Table S1 in Supporting Information S1, locations shown in Figure 1).In North America, we used the continuous record from the NOAA Observatory in Barrow (BRW) (Lan et al., 2022) from 1986 to 2021.We used measurements three coastal Siberian sites, Ambarchik (AMB) (Reum et al., 2019), Cape Baranova (CBR) and Tiksi (TIK) (Laurila et al., 2017).Measurements at AMB were made by the Max Planck Institute for Biogeochemistry (MPI-BGC).Measurements at CBR (2015-2020) and TIK (2013-2020) are both part of the Finnish Meteorological Institute (FMI) network.In the West Siberian Lowlands we used 5 Japan-Russia Siberian Tall Tower Inland Observation Network (JR-STATION) sites from the National Institute for Environmental Studies (NIES)-Azovo (AZV), Berezorechka (BRZ), Karasevoe (KRS), Noyabrsk (NOY) and Vaganovo (VGN) (Sasakawa et al., 2010).These sites were reported on the NIES-94 scale and converted to the WMOX2004A scale for this work.There are multiple inlets used for each site, in each case we use the measurements from the highest inlet.In Europe we used the FMI site Pallas, Finland (PAL) (Laurila et al., 2017), and the National Institute of Polar Research (NIPR) site Ny-Ålesund, Svalbard (NYA) (Morimoto et al., 2017).

Wind Sector Analysis at Barrow
This analysis was closely modeled on the framework established by Sweeney et al. (2016), adhering as closely as possible to their original design.To do this we developed new analysis code, incorporating guidance directly from the authors, to replicate the procedures outlined in their manuscript.Each hourly averaged CH 4 measurement in the data set was categorized by wind direction into sectors.The background sector was 0-90°, considered to be representative of well-mixed background air.The "land" sector at 150°-210°, is a region of predominately Arctic soils and lakes.The remaining regions were included anthropogenic emissions, with Utqiagvik (210°-360°) and Prudhoe Bay fossil fuel extraction field (90°-150°) in close proximity.The measurements were further filtered for times when the standard deviation within the hour was <10 ppb, and the wind speed was <3.0 ms 1 , to avoid analysis of stagnant air.Background measurements were averaged monthly and smoothed using a harmonic fitting routine described in Thoning et al. (1989) (Figure S1 in Supporting Information S1).Enhancements in CH 4 mole fraction were calculated as mole fractions from the land sector minus mole fractions from the background sector.Trends in enhancements were studied for different three time periods across the year, to quantify seasonal land sector emissions and investigate the presence of zero curtain emissions.Calculation of uncertainties in trends are described in Text S1 in Supporting Information S1.

Inversion Method
To calculate emissions estimates from the atmospheric mole fraction data, we first calculate a prior modeled mole fraction estimate using sensitivity maps derived from the UK Met Office NAME model and a priori flux fields.Arctic wide a priori flux fields were produced using bottom-up emissions inventories of terrestrial sources, anthropogenic sources and fires.Further details of the prior emissions are shown in Tables S2 and S3 in Supporting Information S1, along with prior (Figure S2 in Supporting Information S1).We calculate a posterior distribution of emissions using a hierarchical Bayesian Markov Chain Monte Carlo inversion framework.
The model representation of y, a vector containing CH 4 mole fractions, is defined in Equation 1 as: where H is a sensitivity matrix, describing the relationship between atmospheric mole fractions and emissions from the surface, x contains the emissions and ϵ contains the model-measurement errors.H is the product of sensitivity maps that are produced using an atmospheric transport model, and a priori flux fields.
We solved for x using a hierarchical Bayesian inversion framework (Equation 2).This method is described in Ganesan et al. (2014) and Say et al. (2020).To avoid using subjective uncertainties chosen by judgment, we included additional "hyper-parameters", represented by θ.

p(x, θ| y) ∝ p( y|x,θ)p(x|θ)p(θ)
(2) These hyper-parameters describe the unknown model-measurement uncertainties as additional probability density functions (PDFs) that are then propagated in the inversion.Thus, the derived emissions and their uncertainties account for this additional level of model uncertainty.
There is no analytical solution to solve for a posterior optimization of x, so a sampling technique is required.We used Markov Chain Monte Carlo (MCMC) with No U-Turn Sampler described in Say et al. (2020).MCMC also allows for the use of non-Gaussian PDFs to describe each parameter.
Hourly model sensitivity maps or footprints were produced to provide the quantitative relationship between the hourly averaged measured mole fractions at each site and the surface emissions.Footprints were calculated using the Met Office's NAME, a Lagrangian Particle Dispersion Model (Jones et al., 2007).This model tracks particles released from the measurement site backwards in time using meteorological data from a numerical weather prediction model, the Met Office's Unified Model (UM).Particles were tracked for up to 30 days, or when they have left the regional domain.Recording the exit times and locations provides the sensitivity of measured mole fractions to the boundary conditions of the regional domain.Our footprints were spatially resolved on a 0.234°× 0.352°grid across the whole Arctic domain.A model time step of 1 min was chosen to prevent grid cells from being skipped during a time step due to the small size of grid cells near the poles.Southward of 80°N, it is not expected that model grid cells will be skipped within the 1 min time step.The Arctic domain used here spans 36°t o 90°N and all longitudes.This zonal coverage was chosen to include the impacts of sources that could traverse the Arctic within the model simulation.Average model-measurement sensitivity maps for each inversion are shown in Figure S3 in Supporting Information S1.
A quadtree algorithm was used to aggregate grid cells so that 100 emission parameters were solved for in the inversion.Higher resolution regions are estimated near the sites and coarser resolution regions farther away, where the measurements have lower sensitivity.The model solved for the scaling of the prior in each of these regions.These basis functions were produced for the BRW inversion and the AMB-CBR-TIK inversion separately as shown in Figure S4 in Supporting Information S1.The emissions scaling factors had a prior PDF that is Gaussian, truncated at zero to prevent non-physical negative emissions.These scaling factor PDFs were given a mean of 1 and a standard deviation of 2.
Due to the footprints extending across all latitudes, there is only one boundary at the southern boundary.For the Siberian inversions, mole fraction curtains at the boundary of the domain were determined by the global methane model CAMS v19r1 (Inness et al., 2019).This product runs until 2019, so for subsequent years the background was assumed to be the same as that of 2019.A scaling parameter for the Southern boundary condition is included in the inversion with a PDF that is Gaussian, truncated at 0, with a mean of 1 and a standard deviation of 0.02 (corresponding approximately to 40 ppb), thus the background is optimized during the inversion alongside the emissions and model-measurement error.
For the BRW inversions, we subtracted a smoothed background (as described in the wind sector method, see main paper methods) from the raw data, so our results from the wind sector and inversions used consistent approaches.This smoothed background is shown in Figure S1 in Supporting Information S1.This background only fits the data well for the April-December months, so January-March months were not included in the inversion.We also carried out a sensitivity test for the BRW inversion using the CAMS background (same method as the Siberian inversion), which can be seen in Figure S5 in Supporting Information S1.
The uncertainty in observations is a combination of model and measurement uncertainty.Measurement uncertainty was included as the uncertainty directly taken from each hourly measurement.The model uncertainty was a hyper-parameter with a uniform PDF between 0.5 and 30 ppb.The model and measurement uncertainties were combined in quadrature to result in the overall uncertainty applied to each data point.
We filtered out data that is influenced by local emissions, which cannot be not resolved in the inversion.At BRW, the first filter was wind direction.With the town of Utqiagvik located to the West, times when the measured wind direction is 210°-360°are removed because of the potential influence of anthropogenic emissions from the town.
The second filter was to remove times when measured wind speeds were less than 3 m/s to avoid stagnant conditions.For the other sites, we did not use measured wind direction and speeds because there was no clear pattern of wind directions being heavily influenced by anthropogenic emissions.We instead used a "local influence filter".The local influence filter uses a similar method to Lunt et al., 2021;Lunt et al., 2021), whereby if the 25 grid cells surrounding the site contributed to greater than 10% of the total NAME footprint for that time, the measurement is filtered.The same local influence filter at Barrow was also used as a sensitivity study (see Figure S6 in Supporting Information S1), which showed no difference to the wind speed/direction filter.
We partitioned our posterior emissions using the fraction of the natural, anthropogenic and fire emissions in each grid cell of our prior, similarly to Tunnicliffe et al. (2020).In the study regions, the types of source are largely separated spatially as shown in Figure S7 in Supporting Information S1.
The performance of each inversion performance is characterized in Figure S8 in Supporting Information S1 for the North Slope inversion and Figure S9 in Supporting Information S1 for the Siberia inversion.

CH 4 Emissions and Trends From the North Slope of Alaska
While Sweeney et al. (2016) previously showed no significant increase in summertime CH 4 enhancements, we show by extending the observational record through 2021, that a statistically significant trend can now be detected across July-August.We report a wind sector trend of 0.87 ± 0.40 ppb/yr in summertime CH 4 enhancements from the North Slope of Alaska over the period 1986-2021 (Figure 2a)).The wind sector trend over the period 2000-2021 is 2.32 ± 0.76 ppb/yr.Using the inversion method, this 2000-2021 trend corresponds to a growth of 0.018 ± 0.005 Tg yr 2 .Average summertime inversion emissions increased from 0.50 ± 0.09 Tg/yr during 2000-2010 to 0.90 ± 0.14 Tg/yr during 2010-2021.Both methods suggest that summertime emissions from the North Slope have increased between 90% and 100% over the last 20 years.
Our results also confirm that the late season (September-December) trend first shown by Sweeney et al. (2016) has persisted through to 2021.We find a wind-sector trend of 0.36 ± 0.34 ppb/yr from 1986 to 2021, and a windsector trend of 0.87 ± 0.71 ppb/yr from 2000 to 2021, which corresponds an inversion trend of 0.017 ± 0.003 Tg yr 2 .We find a small positive trend in wind-sector enhancements in the early season for the 1986-2021 and 2000-2021, however this is not reflected with our inversion method.We suspect that this uncertain trend is largely driven by emissions from June, a month representing mixed conditions.
Comparisons with previous inversion studies over the North Slope are challenging due to their reliance on shortterm campaigns or quantification over all of North America.However, Miller et al. (2016) found North Slope emissions of 1.74 ± 0.44 Tg/yr May-October using measurements from the CARVE missions across 2012-2014.We find lower emissions with our inversion method of 0.70 ± 0.09 Tg/yr for the 2012-2014 mean.
Given the Arctic's rapid warming over the last few decades, we investigated temperature trends over the North Slope.We find significant positive trends in air temperature anomaly, across the early and late seasons (Figure 3a).Anomalies (temperatures relative to the 1986-2021 mean) were calculated by taking the temperatures when winds were coming from the land sector and subtracting the mean monthly background temperature as per the wind sector method.Every 10 years at Barrow, the air temperature increases by 1.6°in the September-December time frame and 1.0°in the April-June time frame.We extend the method of Sweeney et al. (2016) to examine the temperature sensitivity of tundra methane emissions.
We find an air temperature sensitivity for the July-August time period of 5.3 ± 2.6 ppb/degree C (Figure 3b).This corresponds to a Q 10 value of 2.1, consistent with the findings of Sweeney et al. (2016).Q 10 is a parameter used in wetland models to estimate methane fluxes, indicating how methane production would change with a 10°temperature rise (Bloom et al., 2017;Xu et al., 2016).Our empirically-derived Q 10 value (Q 10 = [ΔCH 4 ] T2 [ΔCH 4 ] T1 10 T2 T1 .)sits within the middle of values used in the WetCHARTs full ensemble, which employs Q 10 values between 1 and 3 (Bloom et al., 2017).
Outside of summer months, air temperature is not a reliable predictor of emissions due to the insulating effect of snow (Arndt et al., 2019).Unfortunately, long-term soil temperature data sets are not available to evaluate the temperature sensitivity during the late season.
This new summertime trend challenges the idea that methane trends in the Arctic may be too small for detection by atmospheric measurements (Euskirchen et al., 2022).It not only marks a significant shift in this specific region but also may have broader implications for the pan-Arctic region given increasing Arctic temperatures (Rantanen et al., 2022), and the temperature sensitivity of emissions from the region over this time period (Figure 3b).The North Slope of Alaska is one of many biomes of the Arctic.Given the Arctic's diverse land and vegetation types (Olefeldt et al., 2021), trends the North Slope of Alaska might not represent the whole of the Arctic's complexity.Therefore, it highlights the importance of maintaining and expanding longterm measurement campaigns throughout the Arctic.

CH 4 Emissions From Siberia
We also quantify tundra emissions from two Siberian regions-the Taymyr Peninsula and East Siberian Lowlands-from late 2014 to 2021 (Figures 4a and 4b respectively).For the Siberian locations we used data from all 11 measurement sites shown in 1.The three coastal Siberian sites, Ambarchik (AMB), Tiksi (TIK) and Cape Baranova (CBR), have high sensitivity to our Siberian regions.Unlike Barrow Observatory measurements, which began in 1986, measurements from these stations only began in 2014 for AMB, 2015 for CBR and 2010 for TIK.As emissions are only quantified from late 2014, no trend analysis was performed.
We find average annual emissions from Arctic tundra in each region to be 0.47 ± 0.07 and 2.39 ± 0.19 Tg/yr for the Taymyr Peninsula and East Siberian Lowlands, respectively.Within these annual emissions we find summertime emissions of 0.12 ± 0.03 and 5.65 ± 0.4 Tg/yr and late season emissions of 0.63 ± 0.09 and 2.51 ± 0.19 Tg/yr for the Taymyr Peninsula and East Siberian Lowlands, respectively.Total emissions and emissions normalized by area for all three regions can be found in Table S5 in Supporting Information S1, and posterior emissions maps can be seen in Figure S10 in Supporting Information S1.
Our study was designed to investigate the three high-latitude Arctic regions at high spatial resolution and to perform in depth analyses on these smaller regions.We thus use observations from 11 stations that are sensitive to the target regions.We do not focus on the entire pan-Arctic region or North America, Canada, and Siberia to save computational cost.
Previous inversion studies over Siberia have mainly focused on larger regions (Thompson et al., 2017;Tsuruta et al., 2023;Wittig et al., 2023), for the purpose of pan-Arctic quantification, and therefore do not focus specifically on high latitude dry tundra regions.Other previous studies have also focused on boreal wetland regions such as the Hudson Bay Lowlands or West Siberian Lowlands (Thompson et al., 2017), which exhibit very different soil methane dynamics to our high-Arctic dry tundra regions.Consequently, these studies do not provide directly comparable results for our East Siberian Lowlands and Taymyr Peninsula regions.While Berchet et al. (2015) reported wetland emissions of 1-13 Tg/yr from a smaller Siberian lowlands region that overlapped a portion of our Taymyr Peninsula region, it also included emissions from the West Siberian Lowlands, an area with a variety of complex natural and anthropogenic emission sources.

Seasonal Cycles Across Three Arctic Regions
For each of the regions studied in the inversion, we infer strong emissions from the late season (Figure 5).We tested the impact of prior emissions on these results by deriving emissions using alternate priors with varying seasonal profiles.
All priors used the Emissions Database for Global Atmospheric Research (EDGAR) (Crippa et al., 2020) for anthropogenic emissions, and the Global Fire Emissions Database (GFED) (Randerson et al., 2017) for emissions from fires.The magnitude of the annual wetland emissions in all cases was the Global Methane Budget top-down 30°-90°N emissions of 46 Tg yr 1 (Saunois et al., 2020).We carried out an inversion that included oceanic emissions using (Weber et al., 2019), and found that using these emissions had a minimal impact on the estimation of wetland emissions, thus we opted not to include ocean emissions into our priors.The difference between the three priors lies in the seasonal and spatial distributions, which are shown in Figure 5a and Figure S3 in Supporting Information S1.
The "WetCHARTs" prior has a seasonal profile derived from the mean of the WetCHARTs extended ensemble (Bloom et al., 2017).Emissions peak in July, are low in the late shoulder season and are almost zero for the winter and early Spring.This profile is typical of most Arctic wetland models.The spatial distribution of this prior is the monthly inter-annually varying fractional wetland satellite product SWAMPs (Surface WAter Microwave Product Series) (Jensen & Mcdonald, 2019) over the years 2000-2019 (for 2020-2021 the spatial distribution is assumed to be that of 2019).
The "Late-season Zona" prior has a seasonal profile that extends further into the late shoulder season, based on Zona et al. (2016), and the spatial distribution of emissions was set in all months to the July value of SWAMPs.As shown in Zona et al. (2016), late season emissions from the North Slope of Alaska come from less inundated areas, and by setting the spatial distribution to that of a summertime month, we increase the potentially emitting area for our inversions in the late and early seasons.This profile is not represented in most wetland emission products, which base spatial distribution on inundation only.The "Uniform" prior distributes emissions evenly across each month and uses the same spatial distribution as the "Late-season Zona" prior.
Emissions derived using the three priors are shown in Figure 5a.The posterior seasonal cycle from the inversions using the "WetCHARTs" prior are restricted by the near zero emissions values of the prior in the late season.We see that with the "Late-season Zona" prior, the atmospheric data derives larger emissions much later in the season, and even with the completely "Uniform" prior, this emissions seasonal profile is preserved.The profile shapes are similar across all Arctic regions, indicating the presence of late season emissions when inversions are allowed to attribute emissions to these months.For the "Late-season Zona" prior, we find that on average, around 50% of annual emissions from the North Slope come from the late season, in agreement with Zona et al. (2016).Tenkanen et al. (2021), conducted an inversion study using the remote sensing product Soil Moisture and Ocean Salinity Freeze/Thaw (SMOS F/T) (Rautiainen et al., 2016) to capture seasonal emissions from the cold season.They modified fluxes from the LPX-Bern DYPTOP model and assigned a low emission rate to the 1°× 1°grid cells that are at least 90% frozen.Over the high latitude regions, including the North Slope of Alaska, the Taymyr Peninsula, and the East Siberian Lowlands, this model results in an overall decrease in emissions between Octobar-December from the standard model, resulting in a profile very similar to the WetCHARTS prior.Methane can still be released after the surface is frozen due to the zero curtain.As the SMOS F/T product is a surface-only product and not sensitive to the sub-surface conditions, this inversion may be underestimating key emissions from the zero curtain period.Rößger et al. (2022) attribute 39% of emissions from the North Siberian Lena River Delta to be from the cold season.We find around 34% of emissions from East Siberian Lowland to be from the late season alone, and around 40% of emissions from the Taymyr Peninsula.These results show consistency in the prevalence of these late season emissions across three high Arctic regions.Our defined region for the East Siberian Lowland contains  S6 in Supporting Information S1. (Olefeldt et al., 2021), which could potentially explain the lower percentage of late season emissions we see from this region, given that late season emissions have been shown to be most significant in lowly inundated tundra (Zona et al., 2016).
The early season of April-June does not show strong emissions with any of the priors.Bao et al. (2021) found a much stronger methane emissions in the late shoulder season compared to the early shoulder season across three sites on the North Slope of Alaska.For the late season, a gradual freezing from both below and above the active layer prevents gas diffusion and therefore reduces CH 4 oxidation, and forces resultant CH 4 through cracks in the surface or plants.In the early season, surface-down thawing releases the remaining CH 4 gradually, which allows gas diffusion and ice melt to better oxidize the CH 4 .

Comparison to Wetland Models
Whilst recent studies have found that wetland emissions could be overestimated due to the double counting of wetland extent resulting from the conflation of wetlands and small lakes (Thornton et al., 2016), it has also been suggested that emissions from drier areas, especially during the late season, may be overlooked (Zona et al., 2016).This land type is not detected by satellites as inundated, which is typically used as the basis for mapping CH 4 emissions in flux models (Bohn et al., 2015).Additionally, some areas may appear "dry" on an inundation map because the surface is frozen, even though emissions are still occurring from the sub-surface due to the zero curtain effect.
The WetCHARTs extended ensemble is an aggregation of models with two different inventory based spatial extents, either the Global Wetlands and Lakes Data set (GLWD) (Lehner & Döll, 2004) or ESA's GLOBCOVER wetland and freshwater land cover map (Bontemps et al., 2011), both combined with ERA Interim precipitation to define the temporal variability.The WetCHARTS models produce very low emissions in the late season (Figure 6b) and the resulting seasonal profile restricts late-season emissions from being derived in inversions, as shown in Figure 5.
Recently, the Wetland Area and Dynamics for Methane Modeling (WAD2M) (Zhang et al., 2020) inundation data set was developed, which uses the satellite data set Surface WAter Microwave Product Series (SWAMPs), but includes the Northern Circumpolar Soil Carbon Database (NCSCD) (Hugelius et al., 2013) and the CircumArctic Wetlands based on Advanced Aperture Radar (CAWASAR) (Wilkman et al., 2018) inventories to capture underrepresented wetland areas.While WAD2M extends the emitting area in the high-latitude regions in the summer months, Figure 6c shows that WAD2M inundation does not capture emitting areas in the late season.
From the work in this study, and in studies such as Zona et al. (2016), we have seen higher emissions from October to December compared to existing models.If future fluxes are based on inundation maps, emissions from areas in the high latitude Arctic in the late season will not be included.We suggest that future flux models make spatial extent adjustments based on low inundation areas in the late season and frozen ground that is still emitting due to the zero curtain effect.

Conclusions
We present estimates of methane emissions from three high-Arctic regions in Alaska and Siberia.We show that growing summertime methane emissions from the North Slope of Alaska can now be detected for the first time since the observational record at Barrow Observatory began in 1986.We show that summertime and late season emissions from the North Slope have increased by approximately 50% and 30%, respectively, over the past three decades.This study also shows that strong late season emissions (34%-50% of annual emissions) are persistent across the three high-latitude Arctic regions.These late season emissions indicate a widespread "zero curtain" effect on a pan-Arctic scale, particularly in dry tundra environments, emissions from which are not included in wetland assessments of emissions due to these land classes not appearing as inundated from satellite observations.These results suggest that wetland models should be modified to increase the emitting areas of high Arctic tundra in the late season and included in global assessments of methane emissions.

Figure 1 .
Figure 1.Location of the atmospheric measurement stations and three high-latitude regions in this study-the Taymyr Peninsula, the East Siberian Lowlands and the North Slope of Alaska.

Figure 2 .
Figure 2. Trends in methane from the North Slope of Alaska.(a) CH 4 enhancements in ppb over the period 1986-2021.(b) Same as A but over 2000-2021.(c) Emissions in Tg/yr over 2000-2021.Uncertainties in the wind sector and inversion trends are found using a sampling method described in Text S1 in Supporting Information S1.

Figure 3 .
Figure 3. (a) Trends in surface temperature anomalies from Barrow (BRW) across April-June, July-August and September-December.(b) Surface temperature anomalies versus CH 4 enhancements for July-August at Barrow (BRW).

Figure 4 .
Figure 4. Methane emissions in Tg/yr from three high-Arctic regions.(a) The Taymyr Peninsula, (b) the East Siberian Lowlands and (c) the North Slope of Alaska.

Figure 5 .
Figure 5. Average seasonal profiles (April-Deceber) as fraction of total annual emissions for the inverse estimates derived using different prior emissions.(a) The average prior over the three regions.(b-d) Posterior profiles for, (b) the North Slope.(c) East Siberian Lowlands and, (d) Taymyr Peninsula.

Figure 6 .
Figure 6.Flux and inundation averages for October-December 2015 for (a) the Posterior Late-season Zona flux derived in this study, (b) the WetCHARTs extended ensemble flux (Bloom et al., 2017) and (c) Wetland Area and Dynamics for Methane Modeling (WAD2M) inundation (Zhang et al., 2020).Regions shown in gray on panel (a) from left to right the Taymyr Peninsula, East Siberian Lowlands and North Slope of Alaska.Note the order of magnitude smaller scale between the Late-season Zona posterior flux and the WetCHARTs extended ensemble flux.