The BAWLD domain includes all of the northern boreal and tundra ecoregions and also areas of rock and ice at latitudes >50∘ N (Olson et al., 2001). The BAWLD domain thus covers 25.5 × 106 km2, or 17 % of the global land surface. Although northern peat-forming wetlands can also be found in temperate ecoregions, our decision to define the southern limit of BAWLD by the transition from boreal to temperate ecoregions was based on the greater human footprint and the increased biogeographic diversity of temperate ecoregions, which would require additional land cover classes (Venter et al., 2016). A network of 0.5∘ grid cells, cropped along coasts and at the transition from boreal to temperate ecoregions, was created for the BAWLD domain.

Grid cells in BAWLD were populated with data from 15 publicly available spatial datasets, yielding 53 variables with spatial information (Table 1). Most datasets that were included have data at higher resolution than the 0.5∘ BAWLD grid cells; hence information was averaged for each grid cell. For datasets where the spatial resolution was coarser or where spatial data were not aligned with the 0.5∘ grid cells, data were first apportioned into BAWLD grid cells before area-weighted averages were calculated. Climate data from the WorldClim2 (WC2) dataset (Fick and Hijmans, 2017) were averaged for each grid cell, including “mean annual air temperature”, “mean annual precipitation”, and “climate moisture index”. Information on soils and permafrost conditions were summarized as fractional coverage within each grid cell and included “permafrost extent” from the Permafrost Zonation and Terrain Ruggedness Index (PZI) dataset (Gruber, 2012); permafrost zonation, ground ice content, and overburden thickness from the Circum-Arctic Map of Permafrost and Ground-Ice (CAPG) dataset (Brown et al., 2002); “yedoma ground” from the Ice-Rich Yedoma Permafrost (IRYP) dataset (Strauss et al., 2017); and non-permafrost peat “histosol”, permafrost peat “histel”, and “aqueous” wetland soils from the Northern Circumpolar Soil Carbon Database (NCSCD; hereafter NCS) (Hugelius et al., 2013). Four independent datasets provided information on wetland coverage, although without further differentiation between distinct wetland classes: the “regularly flooded shrub and/or herbaceous cover” area from the Global Land Cover Database 2000 (GLC2) (Bartholomé and Belward, 2005), the “wetlands” area in the GlobLand30 (GL30) dataset (Chen et al., 2015), and the “inundation and presence of wetlands” area from the Global Wetlands (GWET) dataset (Matthews and Fung, 1987) and the Circumpolar Arctic Vegetation Map (CAVM) dataset (Walker et al., 2005). Two datasets provided information of the extent of forested regions – the GLC2 and the Terrestrial Ecoregions of the World (TEW) dataset (Olson et al., 2001) – while three datasets provided information on the extents of tundra vegetation: the CAVM, the GL30, and the TEW. Three datasets provided information on extent of glaciers and permanent snow: the NCS, the GL30, and the TEW. The NCS dataset also provided information about the extents of “rocklands”, while the PZI dataset had extents of topographic ruggedness (“flat”, “undulating”, “hilly”, “mountainous”, and “rugged”). Information on river extents was found in two datasets: the “river area” in the BasinATLAS (BAS) dataset (Linke et al., 2019) and “rivers” in the Global Lakes and Wetland (GLW) dataset, which includes sixth-order rivers and greater (Lehner and Döll, 2004). Inundation dynamics was provided by two datasets, with “mean annual minimum” and “mean annual maximum” inundation in the Global Inundation Map (GIM) dataset (Fluet-Chouinard et al., 2015) and an analysis of temporal inundation from the Global Surface Water (GSW) dataset (Pekel et al., 2016), where we defined inundation of individual 30 m pixels as being inundated “rarely” (>0 % to 5 % of all available Landsat images), “occasionally” (5 % to 50 %), “regularly” (50 % to 95 %), or “permanently” (95 % to 100 %). Four datasets included information about static extents of open water, including “open water” in NCS; “water bodies” in GL30; “water bodies” in GLC2; and information about lakes in the Hydrolakes (HL) dataset (Messager et al., 2016), where we differentiated between the area of “large lakes” (lakes >10 km2) and “midsize lakes” (lakes between 0.1 and 10 km2). High-latitude data were not available for the GL30 (>82∘ N) and HL (>80∘ N) datasets and were coded as missing data. Regions outside the spatial extents of the CAVM, CAPG, and IRYP datasets were coded as 0 as it suggested absence of tundra vegetation, permafrost, and yedoma soils.

The land cover classification in BAWLD was constructed with the goal to enable upscaling of CH4 fluxes for large spatial extents. As such, we aimed to include as few classes as possible to facilitate large-scale mapping while still including classes that allow for separation among ecosystems with distinct hydrology, ecology, biogeochemistry, and thus net CH4 fluxes. The BAWLD land cover classification is hierarchical, with five wetland classes, seven lake classes, and three river classes, along with four other classes: glaciers, dry tundra, boreal forest, and rocklands. The class descriptions (see Kuhn et al., 2021, for further details) were provided to all experts for their land cover assessments and thus effectively serve as the BAWLD class definitions.

Expert assessments can be used to inform various environmental assessments and are particularly useful to assess levels of uncertainty and to provide data that cannot be obtained through other means (Olefeldt et al., 2016; Loisel et al., 2021; Abbott et al., 2016; Sayedi et al., 2020). We solicited an expert assessment to aid in the modelling of fractional coverage of the 19 land cover classes within each BAWLD grid cell. Researchers associated with the Permafrost Carbon Network (http://www.permafrostcarbon.org, last access: 30 October 2021) with expertise from wetland, lake, and/or river ecosystems within the BAWLD domain were invited to participate. We also included a few additional referrals to suitable experts outside the Permafrost Carbon Network. A total of 29 researchers completed the expert assessment and are included as co-authors of the BAWLD dataset. Each expert was asked to identify a region within the BAWLD domain for which they considered themselves familiar. Experts were then assigned 10 random cells from their region of familiarity and 10 cells distributed across the BAWLD domain that allowed for an overall balanced distribution of training cells (Fig. S1 in the Supplement). No cell was assessed more than once, and in total ∼ 3 % of the area of the BAWLD domain was included in the expert assessment. Each expert was asked to assess the percent coverage of each of the 19 land cover classes within their 20 training cells. To guide their assessment, each expert was provided step-by-step instructions, plus information on the definitions of each land cover class and a KML file with the data extracted from available spatial datasets for each grid cell (Table 1). Experts were asked to use their knowledge of typical wetland and lake classes within specific high-latitude regions, their ability to interpret satellite imagery as provided by Google Earth, and their judgement of the quality and relevance of available spatial datasets to make their assessments of fractional cover. The information provided to experts to carry out the assessment is provided in the Supplement.

Random forest regression models were created to predict the percent coverage of all 19 individual BAWLD land cover classes, along with three additional models for total wetland, lake, and river coverage. The regression models used the expert assessment of land cover fractional extent as the response variables. Each land cover class was at first modelled separately, which was followed by minor adjustments, described below, that ensured that the total land cover within each cell added up to 100 %. All statistical analysis and modelling were done using R 4.0.2 (R Core Team, 2020) and the packages Boruta (v7.0.0; Kursa and Rudnicki, 2010), caret (v6.0-86; Kuhn, 2020), randomForest (v4.6-14; Liaw and Wiener, 2002), and factoextra (v1.0.7; Kassambara and Mundt, 2020).

Prior to running the random forest analyses, we performed an automatic feature selection using a Boruta algorithm (Kursa and Rudnicki, 2010). The Boruta algorithm completed 150 runs for each land cover class, after which subsets of the 53 possible data variables (Table 1) were deemed important and selected for inclusion in subsequent random forest models (Table 2). The random forest models (Kuhn, 2020; Liaw and Wiener, 2002) then used boot-strapped samples (i.e., the expert assessments of land cover fractional grid cell coverages) to grow 500 decision trees (ntree), with a subset of randomized data variables as predictors at each tree node (mtry). We used a 10-fold cross-validation with five repetitions providing mtry as a tuneable parameter for model training. The random forest model output included the root mean square error (RMSE), the percent of the expert assessment variability that was explained (%Var), and relative variable importance (Table 2). Relative variable importance assigns a 100 % importance to the variable with the most influence on the model and then ranks all other variables relative to the influence of that variable. A bias correction (Song, 2015) was applied to the predicted data of land cover class coverages as the models were found to overestimate low coverages and underestimate high coverages. After the bias correction, all bias-adjusted predictions <0 % were set to 0 %, while those >100 % were set to 100 % (for examples of the bias correction, see Fig. S2). Next, we ensured that the combined coverage of all 19 land cover classes within each grid cell added up to 100 % by applying a proportional adjustment. In order to estimate the 5th and 95th percentile confidence bounds of the land cover predictions, we repeated the random forest analysis, as outlined above, an additional 20 times for each class. Each new run completely excluded 20 % of the expert assessments, and the data were reshuffled four times. Each grid cell thus had 21 predictions of coverage for each of the 19 land cover classes and for the cumulative wetland, lake, and river coverages, and the variability in these predictions were used to define the 5th and 95th percentile confidence bounds.

While each cell in BAWLD has a distinct land cover combination, we were also interested in identifying cells with similarities in their land cover compositions to distinguish between regions of the boreal and arctic domain that represent characteristic landscapes. We carried out a k-means clustering (Kassambara and Mundt, 2020) to group grid cells with similarities in their predicted land cover compositions. The k-means clustering was based on within-cluster sum of squares, and we evaluated resulting maps with between 10 and 20 distinct classes. Using 15 clusters was deemed to balance the within-cluster sum of squares and interpretability of the resulting map. We henceforth refer to these clusters as “wetscapes” as each cluster was defined largely by the relative dominance (or absence) of different wetland, lake, and river classes.

We evaluated the predictions of wetland coverage in BAWLD against four independent, high-resolution regional land cover datasets. These four datasets were chosen as they included more than one wetland class, thus enabling evaluation against both total wetland coverage and subsets of wetland classes. Two of these datasets were specifically aimed at mapping of wetlands, including Ducks Unlimited Canada's wetland inventories for western Canada as part of the Canadian Wetland Inventory (CWI; Canadian Wetland Inventory Technical Committee, 2016) and wetland mapping of the West Siberian Lowlands (WSL) (Terentieva et al., 2016). The other two datasets, the 2016 National Land Cover Database (NLCD) of Alaska (Homer et al., 2020) and the 2018 CORINE Land Cover (CLC) (Büttner, 2014) of northern Europe, represent more general land cover datasets. Data from these four datasets were summarized for each BAWLD grid cell where there was complete coverage. Data filtration was done for the CWI to remove cells if >10 % of the cell was classified as burned, cloud, or shadow. There were few cases where there were equivalent wetland classes in BAWLD and these four regional datasets, and as such comparisons were generally made between groups of wetland classes that were considered generally comparable. Similar evaluations were not possible for the lake classes as there are no regional or circumpolar spatial datasets with information on lake genesis or sediment type.

Wetlands were predicted to cover a total of 3.2 × 106 km2, or 12.5 % of the BAWLD domain. The wetland area was dominated by Fens (29 % of total wetland area), Bogs (28 %), and Permafrost Bogs (27 %), while Marshes and Tundra Wetlands, which have relatively higher CH4 emissions, covered 5 % and 12 % of the wetland area, respectively (Table 3). This estimate of total wetland area was greater than previously mapped within the BAWLD domain in GLC2 at 0.9 × 106 km2 (Bartholomé and Belward, 2005), GL30 at 1.4 × 106 km2 (Chen et al., 2015), and GWET at 2.3 × 106 km2 (Matthews and Fung, 1987) but similar to the area of wetland soils in NCS (sum of “histosols”, “histels”, and “aqueous” soil coverage) at 3.0 × 106 km2 (Hugelius et al., 2014). Differences between BAWLD and other estimates of wetland area likely stem partially from differences in wetland definitions, where, for example, definitions of wetlands in GLC2 and GL30 likely do not include wooded bogs, fens, and permafrost bogs. While estimates of total wetland area GWET and in the NCS were closer to BAWLD, there were differences in the spatial distribution. Wetland cover in BAWLD was generally greater than in GWET and NCS in regions with low wetland cover. This likely reflects the ability of experts to infer the presence of small or transitional wetlands that may otherwise be underestimated when mapped using other methodologies. Conversely, wetland cover in BAWLD was generally lower than in GWET and NCS in regions with high wetland cover. This was likely due to differences in definitions, especially the exclusion of all open-water ecosystems from wetlands in BAWLD. For example, it was common in the West Siberian Lowlands for the summed coverage of wetland soils in NCS and the “open-water” coverage in GL30 to be substantially greater than 100 %, suggesting that NCS included peatland pools and small ponds within its wetland soil coverage. Overall, the predictive random forest model of total wetland coverage was able to explain 86 % of the variability in the expert assessments, and it was primarily influenced by the area of “wetlands” in GLC30 and the wetland soil categories in NCS, followed by the coverage of “flat topography” in PZI (Gruber, 2012) (Table 2).

The predictive random forest models for individual wetland classes differed both in terms of how much of the variability in the expert assessment data was explained and in terms of which spatial data were most influential (Table 2). The model for Permafrost Bog coverage explained 84 % of the variability in the expert assessments and was very strongly influenced by “histel” distribution in the NCS (Hugelius et al., 2014). Predictive models explained ∼ 75 % of the variability in the expert assessments for Bogs and Fens separately (Table 2) but 87 % when considered jointly. This shows that the available predictor variables were less suitable for modelling Bogs and Fens separately than jointly, which could partly be due to lower agreement among experts in assessments of Bog and Fen coverages compared to their sum. This would not be surprising as bogs and fens (and swamps) occur along hydrological and nutrient gradients and can have vegetation characteristics that make them difficult to distinguish. Models for Bogs and Fens were both strongly influenced by the “histosol” distribution in NCS, with secondary influences from the area of “wetlands” in GL30, “permafrost extent” in PZI, and “mean annual air temperature” in WC2. Predictive models for Marsh and Tundra Wetlands explained less of the variability in expert assessments, at 54 % and 47 %, respectively. The predictive models for Marsh and Tundra Wetlands were influenced by variables that indicate a transition between terrestrial and aquatic ecosystems, e.g., area of “occasional inundation” in GSW, “rivers” in BAS, and “midsize lakes” in HL, but then differed in the influence of climate and permafrost conditions.

Each wetland class had a distinct spatial distribution (Fig. 2). Bogs and Fens were the dominant wetland classes in relatively warmer climates, with high densities in the West Siberian Lowlands, Hudson Bay Lowlands, and the Mackenzie River Basin. While Bogs and Fens had similarities in their spatial distributions, there was also a relative shift in dominance from Bogs to Fens in relatively colder and drier climates (Fig. S3). These trends are supported by bog-to-fen transitions observed both within and between regions (Packalen et al., 2016; Vitt et al., 2000a; Väliranta et al., 2017) but may not be universal (Kremenetski et al., 2003). Marshes were also found in warmer climates and largely associated with Bogs and Fens but with a more evenly spread distribution. The highest abundance of Marsh coverage was predicted for the Ob River floodplains, a region with very few field studies of CH4 emissions (Terentieva et al., 2019; Glagolev et al., 2011). Bogs, Fens, and Marshes all decreased in abundance in colder climates, with Permafrost Bogs becoming more abundant than Bogs when mean annual temperatures were below -2.5 ∘C, corresponding to findings from western Canada, Fennoscandia, and the West Siberian Lowlands (Vitt et al., 2000b; Seppälä, 2011; Terentieva et al., 2016). Tundra Wetlands became dominant over Fens and Marshes when mean annual air temperatures were below -5.5 ∘C (Fig. 3). Tundra Wetlands were predicted to be most abundant in the lowland regions across the Arctic Ocean coast, with especially high abundance in northern Alaska, eastern Siberia, and on the Yamal and Gydan peninsulas in western Siberia.

We found good agreement between the distribution of wetlands in BAWLD and that of four independent regional spatial datasets (Figs. 4, S4). The best agreements for total wetland cover were between BAWLD and the two datasets dedicated specifically to wetland mapping, with R2 of 0.76 with the WSL dataset and 0.72 with the CWI dataset. There were also strong relationships between BAWLD and the WSL dataset for the distribution of specific wetland classes, for both drier wetland classes (“ridge” + “ryam” + “palsa” vs. Permafrost Bog + Bog) and wetter classes (“fen” + “hollow” vs. Fen). When comparing “wet hollow” of the WSL dataset and Marsh in BAWLD there were discrepancies, but they were primarily attributed to the explicit exclusion of the Ob River floodplains in the WSL dataset (Fig. S4). For the wettest classes, we had only a weak relationship (R2=0.19) between the CWI “marsh” class and the sum of the BAWLD Marsh and Tundra Wetland classes, but the overall average abundance for comparable grid cells was similar, at 1.4 % and 2.2 %, respectively. Agreements between BAWLD and the NLCD and CLC datasets were lower, especially for the relatively drier wetland classes (Fig. S4). Lower agreement between BAWLD and some classes of regional wetland datasets should not be interpreted to demonstrate poor accuracy of BAWLD as differences can be due to class definitions, large mapping units, and relatively low accuracy of the non-wetland-specific regional datasets.

The 95 % confidence intervals for predictions of abundance varied both between wetland classes and among regions (Table 3, Figs. S5, S6). The confidence interval for total wetland area was between 2.8 and 3.8 × 106 km2, i.e., a range that represented 31 % of the central estimate. The range of the confidence interval depends both on how much consensus there is among experts in their assessments and how well the available spatial datasets used in the random forest modelling can explain the expert assessments. The considerable range of the confidence interval for wetlands likely stems from a combination of these two components. The confidence interval for total area of individual wetland classes varied between representing 42 % (Fens) and 71 % (Marshes) of respective central estimates. The absolute range of confidence intervals for individual cells generally increased with higher central estimates of abundances, but the range of confidence intervals decreased if expressed as a percent of the central estimate (Fig. S7).

Lakes were predicted to cover a total of 1.44  × 106 km2, or 5.6 % of the BAWLD domain. Large Lakes had the greatest lake area (44 % of total lake area), followed by Midsize Glacial Lakes (26 %) and Midsize Peatland Lakes (10 %) (Table 3). The lake classes with the highest CH4 emissions, Small Yedoma Lakes and Small Peatland Lakes, jointly covered 10 % of the total lake area. The total predicted lake area in BAWLD was higher than the area of lakes in HL (1.20 × 106 km2), which only includes lakes >0.1 km2 and was similar to the area of “open water” in GL30 (1.43 × 106 km2). The “open-water” class in the GL30 dataset is, however, based on Landsat 30 m resolution data and thus excludes very small open-water areas, while it includes both lentic and lotic open water. The 95 % confidence interval for the total lake area in BAWLD was 0.24 × 106 km2, or 17 % of the central estimate.

The predictive models for the three midsize lake classes each explained between 69 % and 75 % of the variability in expert assessments, while a model for the sum of the three midsize lake classes explained 99.1 %. The predictive model for the sum of the three midsize lake classes was almost exclusively influenced by the area of “midsize lakes” in HL, while the three midsize lake classes were differentiated through further influences by the area of “yedoma ground” (Midsize Yedoma Lakes), by the area of “histosols” and “histels” in NCS, and “wetlands” in GL30 (Midsize Peatland Lakes) and by “shoreline length” in HL (Midsize Glacial Lakes). The influence of “shoreline length” for Midsize Glacial Lakes shows that experts associated glacial lakes with high shoreline development and peatland and yedoma lakes with low shoreline development. Despite similarities in how much of the expert assessments were explained by the predictive models (69 %–75 %), the extrapolation to the BAWLD domain led to large differences in the 95 % confidence interval, which represented only 26 % of the central estimate for Midsize Glacial Lakes while representing 69 % and 140 % for Midsize Peatland and Midsize Yedoma Lakes, respectively (Table 3, Fig. S8). Midsize Glacial Lakes were predominately predicted to have high abundances on the Canadian Shield and in Fennoscandia, while Midsize Yedoma Lakes were associated with the lowland, coastal tundra regions of northeastern Siberia and Alaska, and Midsize Peatland Lakes were especially common in the West Siberian Lowlands but also common in the peatland regions of the Hudson Bay Lowlands, the Mackenzie River Basin, and in coastal lowland regions (Fig. 5).

Small Glacial, Yedoma, and Peatland Lakes were jointly estimated to cover 0.9 % of the BAWLD domain. The predictive models explained 16 %, 39 %, and 66 % of the variability in the expert assessments, respectively (Table 2). The relatively lower predictive power for small lakes was not unexpected given the lack of information on the smallest open-water systems in the available spatial data, the variable abundance of very small open-water systems among landscapes (Muster et al., 2019), and a lower relative consensus among experts when assessing classes with generally small fractional coverages. Models for all three small lake classes were influenced by the area of “occasional inundation” in GSW but were then differentiated by variables largely similar to those that were characteristic of the corresponding midsize lake classes (Table 2). The predicted distributions of the small lake classes were also largely similar to that of the corresponding midsize lake type classes (Fig. 5). The overall predicted area of small lakes was 0.24 × 106 km2, representing 17 % of the total lake area. The combined 95 % uncertainty for the three classes ranged between 0.15 and 0.38 × 106 km2 (Table 3, Fig. S8), suggesting that small lakes represent between 11 and 26 % of the total lake area. Previous assessments have estimated that open-water ecosystems < 0.1 km2 represent between 21 and 31 % of global lake area (Holgerson and Raymond, 2016) but relied on assumptions in the statistical modelling which may lead to bias for boreal and arctic regions (Cael and Seekell, 2016; Muster et al., 2019).

Rivers were predicted to cover a total of 0.12 × 106 km2, or 0.47 % of the BAWLD domain. Large Rivers accounted for 65 % of the total river area in BAWLD. These estimates were similar to global assessments, where streams and rivers have been estimated to cover between 0.30 % and 0.56 % of the land area, with 65 % of the river area consisting of large rivers of sixth or greater stream order (Downing et al., 2012). The predictive model for Large Rivers was strongly influenced by the area of “large rivers” in GLWD, but experts consistently made lower assessments which led to an overall 15 % lower area of Large Rivers compared to the area of rivers in GLWD within the BAWLD domain.

Small Organic-Poor and Small Organic-Rich Rivers were estimated to represent 27 % and 8 %, respectively, of the total river area. The predictive models for the Small Organic-Poor and Organic-Rich Rivers explained 19 % and 59 % of the expert assessments and were distinctly influenced by the area of “occasional inundation” in GSW and “wetlands” in GLC30, respectively. The estimated area of small rivers varied among experts, reflecting difficulties in consistent assessments among experts for land cover classes with low extents (<1 % in most grid cells). The distributions of expert assessments for small river areas were non-normal, leading to a long upper tail for the 95 % confidence interval (Fig. S9). For example, the low, central, and high estimates for the area of Small Organic-Rich Rivers were 0.005, 0.10, and 0.54 × 106 km2, respectively. The predicted distributions showed that the Small Organic-Rich Rivers class was closely associated with the distribution of the BAWLD wetland classes, while Small Organic-Poor Rivers dominated elsewhere, with especially high abundances in regions with higher mean annual precipitation (Fig. 5).

Boreal Forest, Dry Tundra, Rocklands, and Glaciers were predicted to cover 10.7, 5.3, 2.7, and 2.1 × 106 km2, respectively, within the BAWLD domain (Fig. S10). The predictive models explained between 96 % (Glaciers) and 67 % (Rocklands) of the variability in expert assessments. While the predictive models for Glaciers were almost exclusively influenced by the area of “permanent snow and ice” in GL30, several variables influenced predictions of Rocklands – including area of “rocklands” in NCS, “mountainous” and “rugged” terrain in PZI, and “barrens” in CAVM. The predictive models for Boreal Forest and Tundra suggested that the transition between these classes was strongly influenced by the area “forest” in GLC2 and by the distinction between “tundra” and “boreal” terrestrial ecoregions in TEW.

We defined “wetscapes” as regions with characteristic composition of specific wetland, lake, and river classes. Our clustering analysis distinguished 15 typical wetscapes within the BAWLD domain (Fig. 6), each defined by the relative presence or absence of the 19 BAWLD classes (Table S1). Visualizing the distribution of wetscapes provides information on regions that are likely to have similarities in the magnitude, seasonality, and climatic controls over CH4 emissions.

Three wetscapes common in boreal regions were differentiated based on the abundance of non-permafrost wetlands. The Sparse, Common, and Dominant Boreal Wetlands wetscapes all had limited lake coverage (<6 % on average) but had 15 %, 35 %, and 60 % combined coverages of Bogs, Fens, and Marshes, respectively. The Dominant Boreal Wetlands wetscape was almost exclusive to the non-permafrost regions of the Hudson Bay Lowlands and the West Siberian Lowlands. The Common Boreal Wetlands wetscape was more widespread, found adjacent to the core areas of the Hudson Bay Lowlands and the West Siberian Lowlands but also in the Mackenzie River Basin, northern Finland, European Russia, and in the Kamchatka Lowlands. The Sparse Boreal Wetlands wetscape was widespread in Sweden, Finland, European Russia, and the southern boreal regions of Canada outside of Yukon. Emissions of CH4 from these regions are likely dominated by wetlands rather than lakes, with main sensitivity to climate change being altered water balance (Tarnocai, 2006; Olefeldt et al., 2017; Olson et al., 2013).

The Lake-Rich Peatlands and the Permafrost Peatlands wetscapes were both found in lowland regions with discontinuous permafrost, near the boreal-to-tundra transition. The Lake-Rich Peatlands wetscape was almost exclusively found in the West Siberian Lowlands, north of the Ob River. This wetscape was characterized by roughly equal abundances of Bogs, Fens, and Permafrost Bogs (each 14 %–16 %), along with 8 % Marshes, 9 % Small Peatland Lakes, and 5 % Midsize Peatland Lakes. It is notable that this wetscape, with the highest coverages of high-CH4-emitting marshes and peatland lakes, has no presence in North America. The Permafrost Peatlands wetscape was conversely primarily found in the Hudson Bay Lowlands and the Mackenzie River Basin, with additional coverage along the Arctic Ocean coast in European Russia, in interior Alaska, and in the Anadyr Lowlands of far-eastern Russia. This wetscape had the greatest abundance of Permafrost Bogs (27 %), with less contribution from other wetland classes (16 %) and relatively low abundance of lakes (7 %). The Lake-Rich Peatlands wetscape likely has the highest regional CH4 emissions, while the Permafrost Peatlands wetscape likely has low to moderate emissions. However, CH4 emissions from both these wetscapes are likely highly sensitive to climate change due to the rapid ongoing and future permafrost thaw that causes expansion of thermokarst lakes and non-permafrost wetlands at the expense of Permafrost Bogs (Bäckstrand et al., 2008; Turetsky et al., 2002).

Three wetscapes were found in lowland tundra regions and varied in relative dominance of different wetland and lake classes. Wetland-Rich Tundra had 23 % wetlands but only 7 % lakes and was found on the Gydan and Taymyr peninsulas in northern Siberia, with minor extents in far-eastern Siberia and in Alaska. Wetland- and Lake-Rich Tundra had similar wetland cover (24 %) but twice the coverage of lakes (15 %), split equally between glacial and peatland lakes. It was found on the Alaska North Slope along with minor extents on the Yamal Peninsula, the Mackenzie River Delta, and on sections of Baffin Island. Lastly, the Wetland- and Lake-Rich Yedoma Tundra was characterized by the highest abundance of yedoma lakes (8 %) and a total wetland and lake coverage of 46 % and was primarily found in the Kolyma Lowlands, with minor extents in the Yukon–Kuskokwim Delta and on the Alaska North Slope. These regions may have sensitive CH4 emissions, particularly associated with thermokarst lake expansion where highly labile yedoma sediments fuel high CH4 production (Walter Anthony et al., 2016).

The remaining seven wetscapes are likely to have overall low CH4 emissions, or even net uptake, resulting from either the dominance of low-CH4-emitting classes or the relative absence of wetland and lake classes. The Dry Tundra wetscape was common in regions of undulating topography of northernmost Siberia, the Alaska North Slope, and the western Canadian arctic and was characterized by relatively low abundances of wetlands (9 %) and lakes (3 %). The Lake-Rich Shield wetscape was exclusive to the Canadian Shield, and although it had a high abundance of lakes (18 %), these were almost completely dominated by low-CH4-emitting large lakes and glacial lakes. The Upland Boreal wetscape dominates boreal regions of Siberia but is also found in the Yukon, Alaska, and Quebec and was defined by having <5 % wetlands and 0.5 % lakes. The Alpine and Tundra Barrens wetscape had <2 % wetlands and ∼ 1.5 % lakes and dominates the Greenland coast; the high-latitude polar deserts of the Canadian Arctic Archipelago; and the mountain ranges in Fennoscandia, Alaska, Yukon, and eastern Siberia. Lastly, the Glaciers, Large Lakes, and Large Rivers wetscapes were defined by the dominance of the namesake BAWLD classes.