Jukuri, open repository of the Natural Resources Institute Finland (Luke) All material supplied via Jukuri is protected by copyright and other intellectual property rights. Duplication or sale, in electronic or print form, of any part of the repository collections is prohibited. Making electronic or print copies of the material is permitted only for your own personal use or for educational purposes. For other purposes, this article may be used in accordance with the publisher’s terms. There may be differences between this version and the publisher’s version. You are advised to cite the publisher’s version. This is an electronic reprint of the original article. This reprint may differ from the original in pagination and typographic detail. Author(s): Gavin McNicol, Etienne Fluet-Chouinard, Zutao Ouyang, Sara Knox, Zhen Zhang, Tuula Aalto, Sheel Bansal, Kuang-Yu Chang, Min Chen, Kyle Delwiche, Sarah Feron, Mathias Goeckede, Jinxun Liu, Avni Malhotra, Joe R. Melton, William Riley, Rodrigo Vargas, Kunxiaojia Yuan, Qing Ying, Qing Zhu, Pavel Alekseychik, Mika Aurela, David P. Billesbach, David I. Campbell, Jiquan Chen, Housen Chu, Ankur R. Desai, Eugenie Euskirchen, Jordan Goodrich, Timothy Griffis, Manuel Helbig, Takashi Hirano, Hiroki Iwata, Gerald Jurasinski, John King, Franziska Koebsch, Randall Kolka, Ken Krauss, Annalea Lohila, Ivan Mammarella, Mats Nilson, Asko Noormets, Walter Oechel, Matthias Peichl, Torsten Sachs, Ayaka Sakabe, Christopher Schulze, Oliver Sonnentag, Ryan C. Sullivan, Eeva-Stiina Tuittila, Masahito Ueyama, Timo Vesala, Eric Ward, Christian Wille, Guan Xhuan Wong, Donatella Zona, Lisamarie Windham- Myers, Benjamin Poulter & Robert B. Jackson Title: Upscaling Wetland Methane Emissions From the FLUXNET-CH4 Eddy Covariance Network (UpCH4 v1.0): Model Development, Network Assessment, and Budget Comparison Year: 2023 Version: Published version Copyright: The Author(s) 2023 Rights: CC BY 4.0 Rights url: http://creativecommons.org/licenses/by/4.0/ Jukuri, open repository of the Natural Resources Institute Finland (Luke) All material supplied via Jukuri is protected by copyright and other intellectual property rights. Duplication or sale, in electronic or print form, of any part of the repository collections is prohibited. Making electronic or print copies of the material is permitted only for your own personal use or for educational purposes. For other purposes, this article may be used in accordance with the publisher’s terms. There may be differences between this version and the publisher’s version. You are advised to cite the publisher’s version. Please cite the original version: McNicol, G., Fluet-Chouinard, E., Ouyang, Z., Knox, S., Zhang, Z., Aalto, T., et al. (2023). Upscaling wetland methane emissions from the FLUXNET-CH4 eddy covariance network (UpCH4 v1.0): Model development, network assessment, and budget comparison. AGU Advances, 4, e2023AV000956. https://doi.org/10.1029/2023AV000956 MCNICOL ET AL. © 2023. The Authors. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Upscaling Wetland Methane Emissions From the FLUXNET-CH4 Eddy Covariance Network (UpCH4 v1.0): Model Development, Network Assessment, and Budget Comparison Gavin McNicol1,2  , Etienne Fluet-Chouinard1,3  , Zutao Ouyang1  , Sara Knox4  , Zhen Zhang5  , Tuula Aalto6  , Sheel Bansal7  , Kuang-Yu Chang8  , Min Chen9  , Kyle Delwiche10  , Sarah Feron11  , Mathias Goeckede12  , Jinxun Liu13  , Avni Malhotra14, Joe R. Melton15  , William Riley8  , Rodrigo Vargas16  , Kunxiaojia Yuan8  , Qing Ying17  , Qing Zhu8  , Pavel Alekseychik18  , Mika Aurela6  , David P. Billesbach19  , David I. Campbell20, Jiquan Chen21  , Housen Chu22  , Ankur R. Desai23  , Eugenie Euskirchen24  , Jordan Goodrich20  , Timothy Griffis25  , Manuel Helbig26  , Takashi Hirano27  , Hiroki Iwata28  , Gerald Jurasinski29,30  , John King31  , Franziska Koebsch32, Randall Kolka33  , Ken Krauss34  , Annalea Lohila6,35  , Ivan Mammarella35  , Mats Nilson36, Asko Noormets37  , Walter Oechel38  , Matthias Peichl36, Torsten Sachs39, Ayaka Sakabe40  , Christopher Schulze41  , Oliver Sonnentag42, Ryan C. Sullivan43  , Eeva-Stiina Tuittila44  , Masahito Ueyama45  , Timo Vesala35,46  , Eric Ward33  , Christian Wille39, Guan Xhuan Wong47, Donatella Zona38,48, Lisamarie Windham-Myers49  , Benjamin Poulter50  , and Robert B. Jackson1,51,52  1Department of Earth System Science, Stanford University, Stanford, CA, USA, 2Now at Department of Earth and Environmental Sciences, University of Illinois Chicago, Chicago, IL, USA, 3Department of Environmental Systems Science, Institute for Atmospheric and Climate Science, ETH Zurich, Zurich, Switzerland, 4Department of Geography, The University of British Columbia, Vancouver, BC, Canada, 5Department of Geographical Sciences, University of Maryland, College Park, MD, USA, 6Finnish Meteorological Institute, Climate Change, Helsinki, Finland, 7U.S. Geological Survey, Northern Prairie Wildlife Research Center, Jamestown, ND, USA, 8Earth and Environmental Sciences Area, Lawrence Berkeley National Laboratory, Berkeley, CA, USA, 9Department of Forest and Wildlife Ecology, University of Wisconsin-Madison, Madison, WI, USA, 10Department of Environmental Science, Policy, and Management, University of California, Berkeley, CA, USA, 11Knowledge Infrastructure, University of Groningen, Groningen, The Netherlands, 12Department for Biogeochemical Signals, Max Planck Institute for Biogeochemistry, Jena, Germany, 13U.S. Geological Survey, Western Geographic Science Center, Moffett Field, CA, USA, 14Department of Geography, University of Zurich, Zurich, Switzerland, 15Climate Research Division, Environment and Climate Change Canada, Victoria, BC, Canada, 16Department of Plant and Soil Sciences, University of Delaware, Newark, DE, USA, 17Earth System Science Interdisciplinary Center, University of Maryland, College Park, MD, USA, 18Natural Resources Institute Finland (LUKE), Helsinki, Finland, 19Department of Biological Systems Engineering, University of Nebraska-Lincoln, Lincoln, NE, USA, 20School of Science, University of Waikato, Hamilton, New Zealand, 21Center for Global Change and Earth Observations, Michigan State University, East Lansing, MI, USA, 22Climate and Ecosystem Sciences Division, Lawrence Berkeley National Lab, Berkeley, CA, USA, 23Department of Atmospheric and Oceanic Sciences, University of Wisconsin-Madison, Madison, WI, USA, 24University of Alaska Fairbanks, Institute of Arctic Biology, Fairbanks, AK 99775, USA, 25Department Soil, Water, and Climate, University of Minnesota Twin Cities, St. Paul, MN, USA, 26Department of Physics and Atmospheric Science, Dalhousie University, Halifax, NS, Canada, 27Research Faculty of Agriculture, Hokkaido University, Sapporo, Japan, 28Department of Environmental Science, Faculty of Science, Shinshu University, Matsumoto, Japan, 29Landcape Ecology, University of Rostock, Rostock, Germany, 30Peatland Science, University of Greifswald, Greifswald, Germany, 31Department of Forestry and Environmental Resources, North Carolina State University, Raleigh, NC, USA, 32Bioclimatology, University of Göttingen, Göttingen, Germany, 33USDA Forest Service Northern Research Station, Grand Rapids, MN, USA, 34USGS Wetland and Aquatic Research Center, Lafayette, LA, USA, 35Institute for Atmospheric and Earth System Research/Physics, Faculty of Science, University of Helsinki, Helsinki, Finland, 36Department of Forest Ecology and Management, Swedish University of Agricultural Sciences, Umeå, Sweden, 37Department of Ecology and Conservation Biology, Texas A&M University, College Station, TX, USA, 38Department Biology, San Diego State University, San Diego, CA, USA, 39GFZ German Research Centre for Geosciences, Potsdam, Germany, 40Hakubi center, Kyoto University, Kyoto, Japan, 41University of Alberta, Renewable Resources; Université de Montréal, Département de géographie, Université de Montréal, Montréal, QC, Canada, 42Département de géographie, Université de Montréal, Montréal, QC, Canada, 43Environmental Science Division, Argonne National Laboratory, Lemont, IL, USA, 44School of Forest Sciences, University of Eastern Finland, Joesnuu, Finland, 45Graduate School of Agriculture, Osaka Metropolitan University, Sakai, Japan, 46Faculty of Agriculture and Forestry, Institute for Atmospheric and Earth System Research/Forest Sciences, University of Helsinki, Helsinki, Finland, 47Sarawak Tropical Peat Research Institute, Kota Samarahan, Malaysia, 48Department of Animal and Plant Sciences, University of Sheffield, Sheffield, UK, 49U.S. Geological Survey, Water Key Points: • Random forest models trained on FLUXNET-CH4 methane fluxes reproduced spatiotemporal patterns in extra-tropical wetlands (R 2: 0.59–0.64) • Globally upscaled annual wetland methane emissions (146 TgCH4 y −1) overlapped with land surface and inversion model ensemble estimates • Humid/monsoon tropics dominate upscaled wetland methane emissions (∼68%) and uncertainties (∼78%) due to limited FLUXNET-CH4 site coverage Supporting Information: Supporting Information may be found in the online version of this article. Correspondence to: G. McNicol and E. Fluet-Chouinard, gmcnicol@uic.edu; efluet@stanford.edu Citation: McNicol, G., Fluet-Chouinard, E., Ouyang, Z., Knox, S., Zhang, Z., Aalto, T., et al. (2023). Upscaling wetland methane emissions from the FLUXNET-CH4 eddy covariance network (UpCH4 v1.0): Model development, network assessment, and budget comparison. AGU Advances, 4, e2023AV000956. https://doi. org/10.1029/2023AV000956 Received 12 JUL 2022 Accepted 15 MAY 2023 Author Contributions: Conceptualization: Gavin McNicol, Etienne Fluet-Chouinard, Zutao Ouyang, Sara Knox, Benjamin Poulter, Robert B. Jackson 10.1029/2023AV000956 Peer Review The peer review history for this article is available as a PDF in the Supporting Information. RESEARCH ARTICLE 1 of 24 AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 2 of 24 1. Introduction The post-industrial rise in atmospheric methane (CH4) concentrations has had a large climate warming effect, 60% the size of that for carbon dioxide (CO2) (IPCC,  2021). The short atmospheric lifetime of CH4 also promises relatively fast climate change mitigation effects following CH4 emissions reductions, rather than century-or-more timescales for CO2 reductions (Abernethy et al., 2021; Turner et al., 2019). However, current emissions trajectories more closely track high emissions scenarios (Zhang et  al.,  2023). Since 2014, there has been an accelerating increase in the CH4 growth rate that reached a record level in 2022, at 18.2 ppb y −1 (Lan et al., 2023), and these increases could continue as global temperatures rise (Bansal et al., 2023; Zhang et al., 2017). Large uncertainties around total (8%–39%) and individual CH4 sources (Table 1) prevent CH4 budget closure at regional-to-global scales. Better constrained CH4 budgets are needed to more accurately attribute and mitigate the sources causing the accelerating rise in CH4 emissions (Nisbet et al., 2022). Improving wetland CH4 emissions estimates will help constrain the global CH4 budget as wetlands comprise both the largest natural CH4 emissions source (20%–30% of total emissions) and the second largest uncertainty in the CH4 budget (Saunois et al., 2020). Mission Area, Menlo Park, CA, USA, 50Biospheric Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, MD, USA, 51Woods Institute for the Environment, Stanford University, Stanford, CA, USA, 52Precourt Institute for Energy, Stanford University, Stanford, CA, USA Abstract Wetlands are responsible for 20%–31% of global methane (CH4) emissions and account for a large source of uncertainty in the global CH4 budget. Data-driven upscaling of CH4 fluxes from eddy covariance measurements can provide new and independent bottom-up estimates of wetland CH4 emissions. Here, we develop a six-predictor random forest upscaling model (UpCH4), trained on 119 site-years of eddy covariance CH4 flux data from 43 freshwater wetland sites in the FLUXNET-CH4 Community Product. Network patterns in site-level annual means and mean seasonal cycles of CH4 fluxes were reproduced accurately in tundra, boreal, and temperate regions (Nash-Sutcliffe Efficiency ∼0.52–0.63 and 0.53). UpCH4 estimated annual global wetland CH4 emissions of 146 ± 43 TgCH4 y −1 for 2001–2018 which agrees closely with current bottom-up land surface models (102–181 TgCH4 y −1) and overlaps with top-down atmospheric inversion models (155–200 TgCH4 y −1). However, UpCH4 diverged from both types of models in the spatial pattern and seasonal dynamics of tropical wetland emissions. We conclude that upscaling of eddy covariance CH4 fluxes has the potential to produce realistic extra-tropical wetland CH4 emissions estimates which will improve with more flux data. To reduce uncertainty in upscaled estimates, researchers could prioritize new wetland flux sites along humid-to-arid tropical climate gradients, from major rainforest basins (Congo, Amazon, and SE Asia), into monsoon (Bangladesh and India) and savannah regions (African Sahel) and be paired with improved knowledge of wetland extent seasonal dynamics in these regions. The monthly wetland methane products gridded at 0.25° from UpCH4 are available via ORNL DAAC (https://doi.org/10.3334/ ORNLDAAC/2253). Plain Language Summary Wetlands account for a large share of global methane emissions to the atmosphere, but current estimates vary widely in magnitude (∼30% uncertainty on annual global emissions) and spatial distribution, with diverging predictions for tropical rice growing (e.g., Bengal basin), rainforest (e.g., Amazon basin), and floodplain savannah (e.g., Sudd) regions. Wetland methane model estimates could be improved by increased use of land surface methane flux data. Upscaling approaches use flux data collected across globally distributed measurement networks in a machine learning framework to extrapolate fluxes in space and time. Here, we train and evaluate a methane upscaling model (UpCH4) and use it to generate monthly, globally gridded wetland methane emissions estimates for 2001–2018. The UpCH4 model uses only six predictor variables among which temperature is dominant. Global annual methane emissions estimates and associated uncertainty ranges from upscaling fall within state-of-the-art model ensemble estimates from the Global Carbon Project (GCP) methane budget. In some tropical regions, the spatial pattern of UpCH4 emissions diverged from GCP predictions, however, inclusion of flux measurements from additional ground-based sites, together with refined maps of tropical wetlands extent, could reduce these prediction uncertainties. Data curation: Gavin McNicol, Etienne Fluet-Chouinard, Zutao Ouyang, Sara Knox Formal analysis: Gavin McNicol Funding acquisition: Sara Knox, Lisamarie Windham-Myers, Benjamin Poulter, Robert B. Jackson Methodology: Zhen Zhang Project Administration: Sara Knox, Robert B. Jackson Supervision: Benjamin Poulter, Robert B. Jackson Validation: Zhen Zhang Visualization: Etienne Fluet-Chouinard Writing – original draft: Gavin McNicol, Etienne Fluet-Chouinard Writing – review & editing: Gavin McNicol, Etienne Fluet-Chouinard, Zutao Ouyang, Sara Knox, Zhen Zhang, Lisamarie Windham-Myers, Benjamin Poulter, Robert B. Jackson 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 3 of 24 Wetlands, as defined here, include both seasonally and permanently inundated soils that are vegetated includ- ing riparian and floodplain forests, inland marsh systems, and peat-forming wetlands (peatlands) but exclude rice agriculture, tidal and non-vegetated waterbodies such as ponds, lakes, streams, rivers, and estuaries (Zhang et al., 2021). Current methods to estimate global wetland CH4 emissions generally fall into one of two approaches: top-down (TD) atmospheric observation-based inversions and bottom-up (BU) land surface models. Although both approaches involve state-of-the-art methods, emissions estimates vary significantly within and between the two approaches (Saunois et al., 2020). For the decade 2008–2017, a comprehensive TD model ensemble, compiled as part of the Global Carbon Project (GCP) methane budget activity, estimated wetland emissions of 159–200 (mean 181) TgCH4 yr −1. TD inversions use approaches such as Bayesian inference or ensemble Kalman filters to esti- mate the (posterior) wetland emissions required to reproduce in-situ and satellite atmospheric CH4 concentration retrievals. Since observations are limited, wetland source attribution strongly depends on prior assumptions as additional sources of information, constraining sectoral contributions (i.e., natural, industrial, agricultural and waste emissions) in space and time (Jacob et  al.,  2022). Considerable variability exists between TD models, primarily due to differences in assumptions and whether satellite retrievals are included as constraints (Kirschke et al., 2013; Saunois et al., 2020). In contrast to TD models, BU model estimates are not constrained by atmospheric CH4 concentration data and attempt to directly represent wetland CH4 fluxes and underlying flux processes with varying complexity (Riley et al., 2011; Ueyama et al., 2023). For the same decade (2008–2017), the GCP's 13-member BU model ensemble estimated emissions of 102–182 (mean 149) TgCH4 yr −1, ∼20% lower than the TD ensemble mean (Saunois et al., 2020). The widespread observed among BU models arises from differences in model param- eterization, which is informed by process knowledge and literature estimates of parameter values and some- times by calibration to observed wetland CH4 fluxes at a limited number of sites (e.g., 3 northern wetlands for Source sector Ensemble Average emissions [ensemble range] (TgCH4 y −1) Absolute range (TgCH4 y −1) Normalized range Total BU 737 [594–881] 287 39% TD 576 [550–594] 44 8% BU—TD 161 Natural wetlands BU 149 [102–182] 80 54% TD 181 [159–200] 41 23% BU - TD −32 Other natural BU 222 [143–306] 163 73% TD 37 [21–50] 29 78% BU—TD 185 Agriculture and waste BU 206 [191–223] 32 16% TD 217 [207–240] 33 15% BU—TD −11 Fossil fuels BU 128 [113–154] 41 32% TD 111 [81–131] 50 45% BU—TD 17 Biomass and biofuel burning BU 30 [26–40] 14 47% TD 30 [22–36] 14 47% BU—TD 0 Note. “Other natural” combines open (non-vegetated) freshwaters, geological (on and offshore), wild animal, termite, wildfire, permafrost, and biological oceanic sources. Normalized ranges are reported as absolute ranges divided by ensemble average emissions. Table 1 Natural and Anthropogenic Methane Emissions by Source Sector for the Decade 2008–2017 From Global Carbon Project Top-Down (TD) and Bottom-Up (BU) Model Ensembles (Saunois et al., 2020) 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 4 of 24 Wetland-DNDC in Zhang et al., 2002). However, Chang et al.  (2021) showed a large variability in CH4 flux temperature dependency across a network of eddy covariance tower sites (FLUXNET-CH4; Delwiche et al., 2021; Knox et al., 2019), indicating that upscaling from few sites is likely to introduce errors during global extrapola- tion. To date, process-based models have yet to use networked data, for instance FLUXNET-CH4, for multi-site calibration, in part because this approach is more technically challenging than for machine learning models. Equally large uncertainties (∼50% of total uncertainty) are introduced when BU models simulate independent wetland extents in prognostic runs versus using prescribed global wetland extents in diagnostic runs (Melton et al., 2013). Notably, the substantial GCP BU spread (±30%-50% of ensemble mean emissions) was observed even in diagnostic model runs where all models were prescribed a common Wetland Area and Dynamics for Methane Modeling (WAD2M) wetland extent (Zhang et al., 2021), underscoring the need to reduce wetland CH4 flux uncertainties as well as wetland extent uncertainties. No global benchmark data set exists to favor or falsify, with a strong degree of confidence, any BU or TD model (Saunois et  al.,  2020). Insights about model parameterization and sources of uncertainty can only be gained at present via model intercomparisons, such as the BU Wetland CH4 Inter-Comparison of Models Project (WETCHIMP) activity (Melton et al., 2013) and the TD Wetland CH4 emission and uncertainty ensemble data set for Atmospheric Chemistry and Transport modeling (WetCHARTS) activity (Bloom et al., 2017), the GCP wetland CH4 synthesis (Poulter et al., 2017), by regional scale evaluation of converging or diverging TD and BU estimates (Stavert et al., 2021; example in Figure 1), or by comparison to satellite retrievals (Parker et al., 2018). Independent estimates of global wetland CH4 emissions, incorporating new data for calibration and model constraints, and implementing new modeling approaches, such as machine learning algorithms, are emerging alternatives for refining models and reducing uncertainties around wetland CH4 sources (Saunois et al., 2020). One data stream that can improve wetland CH4 models and global emission estimates is the growing availa- bility of CH4 fluxes measured near the land surface. In situ eddy covariance flux towers provide long-term, semi-automated, and quasi-continuous fluxes at ecosystem scales (<1 km 2) with minimal disturbance to soils or canopy structure/function (Baldocchi, 2014; Chu et al., 2021). Although BU models have been parameterized using CH4 flux data at individual sites, network CH4 data has not been fully utilized. Since the late-1990s, FLUX- NET has provided standardized CO2 flux data measured using eddy covariance across hundreds of locations around the world, enabling independent benchmarking of satellite measurements and Earth system models (Jung et al., 2020; Pastorello et al., 2020). Upscaling is a workflow combining statistical models and data to transfer information across scales, often using machine learning, and has been used by projects such as FLUXCOM to extrapolate FLUXNET data from 224 sites (∼850 site years) and predict global terrestrial ecosystem carbon and energy fluxes (Bodesheim et al., 2018; Jung et al., 2020; Tramontana et al., 2016). The FLUXNET-CH4 data set now provides similar opportunities to refine model parameterization and generate independent, data-driven esti- mates of regional-to-global CH4 emissions (Chang et al., 2021; Delwiche et al., 2021; Knox et al., 2019). Peltola et al. (2019) independently acquired eddy covariance data from 25 high-latitude sites and developed a wetland CH4 flux upscaling workflow to predict monthly, regional (>45°N) wetland CH4 emissions for 2012–2013. Annual emissions of 31–38 TgCH4 y −1 agreed well with previous bottom-up estimates (e.g., Chen et al., 2015; Treat et al., 2018; Zhang et al., 2017) but were higher than those of top-down estimates (23–28 TgCH4 y −1) for the region (e.g., Bruhwiler et al., 2014; Spahni et al., 2011). Peltola et al. (2019) thus demonstrated that upscaling estimates from eddy covariance data could produce plausible CH4 fluxes at regional scales. To date, however, no upscaling project has taken advantage of the full FLUXNET-CH4 site network to make and evaluate global wetland CH4 emissions predictions. Here, we develop a wetland CH4 upscaling workflow (UpCH4) that combines FLUXNET-CH4 and globally gridded predictor data to train random forest model ensembles, including validation and test routines optimized for spatial prediction applications. Given that surface flux measurement may networks take decades to grow, our first goal is to robustly evaluate the ability of machine learning models to extrapolate beyond training conditions in space and time. We then use this CH4 flux upscaling workflow (UpCH4) to predict wetland CH4 emissions globally and compare them to current top-down and bottom-up model estimates. Given knowledge of model structure and network coverage, our second goal is to identify regions of convergence and diagnose regions of divergence between UpCH4 and existing estimates. Finally, we use knowledge of the model structure to conduct a FLUXNET-CH4 network dissimilarity analysis with the goal of informing strategic improvements in eddy covariance site coverage. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 5 of 24 2. Methods 2.1. Predictive Modeling 2.1.1. Eddy Covariance CH4 Flux Data UpCH4 was designed to predict CH4 fluxes globally at inundated and non-inundated (i.e., shallow water table) freshwater vegetated wetlands. Forty-five natural and restored freshwater wetland sites from the FLUXNET-CH4 database qualified for model training (Delwiche et al., 2021). Two sites (RU-VrK and SE-St1) were excluded after quality control filtering and 1 year of data was excluded from a restored wetland site (US-Sne; Table 2) that had not yet developed vegetation cover. One site (DE-Hte) is coastal but was freshwater dominated during the observation period. The final eddy covariance tower data set consisted of 43 freshwater wetland sites covering Figure 1. (a) Large regional discrepancies exist between (b) bottom-up (BU) and (c) top-down (TD) model estimates of wetland CH4 emissions, in addition to different global totals. Mean daily natural wetland CH4 emissions for 2010–2017 estimated from a 17-member TD ensemble are subtracted from the mean from a 13-member BU process model ensemble (Saunois et al., 2020). High northern latitude bounding boxes correspond to locations of Hudson Bay Lowland (left) and West Siberian Lowland (right) wetland complexes. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 6 of 24 Cl us ter W ee ks Pe rc en t ID Si te na m e Co un try La tit ud e Lo ng itu de Cl as s M AT (C ) M AP (m m ) DO I r efe re nc es 1 14 0 2.2 BR -N pw No rth er n P an tan al  W etl an d Br az il −1 6.4 98 −5 6.4 12 Sw am p 25 .2 1,3 18 Vo ur lit is et  al.  (2 02 0) 2 34 0.5 BW -G um Gu m a Bo tsw an a −1 8.9 64 72 2 22 .37 11 11 1 Sw am p 23 .1 45 9 He lft er  (2 02 0a ) 2 33 0.5 BW -N xr Nx ar ag a Bo tsw an a −1 9.5 48 05 6 23 .17 91 66 7 Sw am p 23 .5 43 3 He lft er  (2 02 0b ) 3 11 4 1.8 CA -S CB Sc ot ty  C re ek  B og Ca na da 61 .30 89 −1 21 .29 84 Bo g −2 .8 41 4 So nn en tag an d He lb ig  (2 02 0a ) 3 11 6 1.8 CA -S CC Sc ot ty  C re ek  L an ds ca pe Ca na da 61 .30 79 −1 21 .29 92 Bo g −2 .9 41 4 So nn en tag an d He lb ig  (2 02 0b ) 4 37 9 6.0 DE -H te Hu ete lm oo r Ge rm an y 54 .21 02 78 12 .17 61 11 Fe n 8.5 58 4 Ko eb sc h a nd Ju ra sin sk i ( 20 20 ) 4 24 8 4.0 DE -Z rk Za rn ek ow Ge rm an y 53 .87 59 4 12 .88 90 1 Fe n 8.3 58 0 Sa ch s a nd W ill e ( 20 20 ) 5 11 5 1.8 DE -S fN Sc he ch en fil z N or d Ge rm an y 47 .80 63 9 11 .32 75 Bo g 8.3 1,1 23 Sc hm id an d K lat t ( 20 20 ) 6 25 7 4.1 FI -L om Lo m po lo jan kk a Fi nl an d 67 .99 72 4 24 .20 91 8 Fe n −1 .0 51 2 Lo hi la et  al.  (2 02 0) 7 12 5 2.0 FI -S i2 Si ik an ev a- 2 B og Fi nl an d 61 .83 72 24 .19 67 Bo g 3.2 66 4 Ve sa la, T ui tti la, M am m ar ell a, an d Al ek se yc hi k ( 20 20 ) 7 25 7 4.1 FI -S ii Si ik an ev a Fi nl an d 61 .83 26 5 24 .19 28 5 Fe n 3.2 66 6 Ve sa la, T ui tti la, M am m ar ell a, an d Ri nn e ( 20 20 ) 8 37 0.6 FR -L Gt La  G ue tte Fr an ce 47 .32 29 17 2.2 84 10 2 Fe n 11 .0 70 7 Ja co to t e t a l.  (2 02 0) 9 62 1.0 ID -P ag Pa lan gk ar ay a u nd ra in ed  fo re st In do ne sia −2 .32 11 3.9 Sw am p 27 .4 23 86 Sa ka be et  al . ( 20 20 ) 10 11 4 1.8 JP -B BY Bi ba i b og Ja pa n 43 .32 30 05 6 14 1.8 10 69 7 Bo g 6.7 1,1 53 Ue ya m a e t a l.  (2 02 0) 11 81 1.3 M Y- M LM M alu da m  N ati on al  Pa rk M ala ys ia 1.4 53 57 5 11 1.1 49 49 2 Sw am p 26 .9 3,4 01 W on g e t a l.  (2 02 0) 12 20 6 3.3 NZ -K op Ko pu ata i Ne w Z ea lan d −3 7.3 87 9 17 5.5 53 9 Bo g 13 .9 1,3 43 Ca m pb ell an d Go od ric h ( 20 20 ) 13 11 7 1.9 RU -C h2 Ch er sk y r efe re nc e Ru ss ia 68 .61 68 9 16 1.3 50 89 W et tu nd ra −1 2.3 17 2 Go ec ke de  (2 02 0) 14 78 1.2 RU -C ok Ch ok ur da kh Ru ss ia 70 .82 91 4 14 7.4 94 28 W et tu nd ra −1 4.1 21 0 Do lm an et  al . ( 20 20 ) 15 24 6 3.9 SE -D eg De ge ro Sw ed en 64 .18 20 29 19 .55 65 39 Fe n 1.7 62 0 Ni lss on an d P eic hl  (2 02 0) 16 58 0.9 US -A 03 AR M -A M F3 -O lik to k Un ite d St ate s 70 .49 53 28 −1 49 .88 23 W et tu nd ra −1 1.9 14 4 Bi lle sb ac h a nd Su lli va n ( 20 20 a) 16 72 1.1 US -IC s Im na va it  Cr ee k W ate rsh ed  W et  Se dg e T un dr a Un ite d St ate s 68 .60 58 −1 49 .31 1 W et tu nd ra −8 .9 24 2 Eu sk irc he n e t a l.  (2 02 0c ) 17 36 0.6 US -A 10 AR M -N SA -B ar ro w Un ite d St ate s 71 .32 42 −1 56 .61 49 W et tu nd ra −1 2.0 10 7 Bi lle sb ac h a nd Su lli va n ( 20 20 b) 17 84 1.3 US -A tq At qa su k Un ite d St ate s 70 .46 96 −1 57 .40 89 W et tu nd ra −1 0.3 13 3 Zo na an d O ec he l ( 20 20 a) 17 74 1.2 US -B eo Ba rro w  En vi ro nm en tal  O bs er va to ry  (B EO ) t ow er Un ite d St ate s 71 .28 1 −1 56 .61 23 W et tu nd ra −1 1.9 10 9 Zo na an d O ec he l ( 20 20 b) Ta bl e 2 C lu ste r A ss ig nm en ts , N um be r o f W ee ks a nd P er ce nt o f T ot al D at a Se t, an d W et la nd S ite ID , L oc at io n, C la ss , C lim at e, a nd D ig ita l O bj ec t I de nt ifi er s ( D O I) fo r t he 4 3 W et la nd s I nc lu de d in U ps ca lin g 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 7 of 24 Cl us ter W ee ks Pe rc en t ID Si te na m e Co un try La tit ud e Lo ng itu de Cl as s M AT (C ) M AP (m m ) DO I r efe re nc es 17 12 8 2.0 US -B es Ba rro w- Be s ( Bi oc om pl ex ity  E xp er im en t S ou th  to we r) Un ite d St ate s 71 .28 09 −1 56 .59 65 W et tu nd ra −1 2.0 10 9 Zo na an d O ec he l ( 20 20 c) 17 13 1 2.1 US -N GB NG EE  A rc tic  B ar ro w Un ite d St ate s 71 .28 00 44 −1 56 .60 91 8 W et tu nd ra −1 1.9 10 9 To rn an d D en ge l ( 20 20 a) 18 97 1.5 US -B ZB Bo na nz a C re ek  T he rm ok ar st  Bo g Un ite d St ate s 64 .69 55 47 −1 48 .32 08 4 Bo g −2 .4 29 2 Eu sk irc he n a nd Ed ga r ( 20 20 a) 18 91 1.5 US -B ZF Bo na nz a C re ek  R ich  F en Un ite d St ate s 64 .70 37 33 −1 48 .31 33 3 Fe n −2 .5 29 4 Eu sk irc he n a nd Ed ga r ( 20 20 b) 18 20 6 3.3 US -U af Un ive rsi ty  of  A las ka , F air ba nk s Un ite d St ate s 64 .86 62 7 −1 47 .85 55 3 Bo g −2 .8 29 8 Iw ata et  al . ( 20 20 ) 19 17 7 2.8 US -D PW Di sn ey  W ild er ne ss  P re se rv e W etl an d Un ite d St ate s 28 .05 20 6 −8 1.4 36 11 M ar sh 22 .1 1,2 23 Hi nk le an d Br ac ho  (2 02 0) 20 16 5 2.6 US -Iv o Iv ot uk Un ite d St ate s 68 .48 65 −1 55 .75 03 W et tu nd ra −8 .5 24 7 Zo na an d O ec he l ( 20 20 d) 21 92 1.5 US -L A2 Sa lva do r W M A  Fr es hw ate r M ar sh Un ite d St ate s 29 .85 87 −9 0.2 86 9 M ar sh 20 .0 1,6 16 Ho lm et  al . ( 20 20 ) 22 27 0 4.3 US -L os Lo st  Cr ee k Un ite d St ate s 46 .08 27 −8 9.9 79 2 Fe n 4.1 83 3 De sa i ( 20 20 ) 23 47 8 7.6 US -M yb M ay be rry  W etl an d Un ite d St ate s 38 .04 98 61 −1 21 .76 49 8 M ar sh 15 .4 34 6 M att he s e t a l.  (2 02 0) 23 11 4 1.8 US -S ne Sh er m an  Is lan d R es to re d W etl an d Un ite d St ate s 38 .03 69 −1 21 .75 47 M ar sh 15 .5 34 0 Sh or t e t a l.  (2 02 0) 23 28 5 4.5 US -T w1 Tw itc he ll  W etl an d W es t P on d Un ite d St ate s 38 .10 74 −1 21 .64 69 M ar sh 15 .4 37 1 Va lac h, Sz ut u, et  al.  (2 02 0) 23 30 4 4.8 US -T w4 Tw itc he ll  Ea st  En d W etl an d Un ite d St ate s 38 .10 27 43 6 −1 21 .64 13 3 M ar sh 15 .4 37 0 Ei ch elm an n e t a l.  (2 02 0) 23 34 0.5 US -T w5 Ea st  Po nd  W etl an d Un ite d St ate s 38 .10 71 55 −1 21 .64 25 7 M ar sh 15 .4 37 1 Va lac h, Ka sa k, et  al.  (2 02 0) 24 22 4 3.6 US -N C4 NC _A lli ga to rR ive r Un ite d St ate s 35 .78 79 −7 5.9 03 8 Sw am p 16 .5 1,3 22 No or m ets et  al . ( 20 20 ) 25 33 0.5 US -N GC NG EE  A rc tic  C ou nc il Un ite d St ate s 64 .86 14 −1 63 .70 08 W et tu nd ra −3 .1 41 3 To rn an d D en ge l ( 20 20 b) 26 19 1 3.0 US -O Rv Ol en tan gy  R ive r W etl an d R es ea rc h P ar k Un ite d St ate s 40 .02 01 −8 3.0 18 3 M ar sh 11 .0 95 4 Bo hr er an d M or in  (2 02 0) 26 31 0.5 US -O W C Ol d W om an  C re ek Un ite d St ate s 41 .37 95 16 7 −8 2.5 12 46 7 M ar sh 9.9 89 8 Bo hr er et  al . ( 20 20 ) 26 13 8 2.2 US -W PT W in ou s P oi nt  N or th  M ar sh Un ite d St ate s 41 .46 46 39 −8 2.9 96 15 7 M ar sh 9.9 88 1 Ch en an d C hu  (2 02 0) No te . M ea n a nn ua l t em pe ra tu re (M AT ; ° C) an d m ea n a nn ua l p re cip ita tio n ( M AP ; m m ) w er e e xt ra cte d f ro m W or ld Cl im 2. 0 ( Fi ck & H ijm an s,  20 17 ). Ta bl e 2 C on tin ue d 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 8 of 24 bog (8), fen (8), marsh (10), swamp (6), and wet tundra (11) wetland classes and distributed across Arctic-boreal (20), temperate (16), and (sub)tropical (7) climate zones. Weekly mean CH4 fluxes were computed from half-hourly FLUXNET-CH4 Version 1.0 fluxes, which are availa- ble as a standardized gap-filled product (Delwiche et al., 2021). Although current wetland CH4 emission products are resolved at monthly timesteps, a finer-resolution (here, weekly) increases training data size and helps capture sub-monthly functional dependencies between predictors and flux (Jung et al., 2020; Tramontana et al., 2016). Weekly fluxes were only retained when there was a minimum of 1 day (48 half-hours or ∼14%) of CH4 obser- vations. This gap-filling threshold was chosen to retain as much training data as possible while minimizing the errors introduced by filling long gaps (Dengel et al., 2013; Peltola et al., 2019). Most gaps were less than 5 hr in length, and the maximum possible gap-length was 12  days (Figure S1 in Supporting Information  S1). A detailed study of machine learning-based gap-filling of eddy covariance CH4 fluxes found that bias introduced by gap-filling remains small and consistent across gap-lengths of 12 days or less (Irvin et al., 2021). After applying the gap-filling threshold, the final flux data set consisted of 6,210 weekly observations spanning 2006–2018, with 96% of the data recorded after 2010, and 38% recorded after 2015. Sites within 300  km of each other were grouped together, resulting in 26 clusters that were used for spatial leave-one-out cross-validation (LOOCV) of the machine learning model, where each training/validation fold consisted of all data except one hold-out cluster (Meyer et al., 2019) (Figure 2). Spatial LOOCV has previously been applied to evaluate models used in global upscaling of CO2 and energy fluxes (Tramontana et al., 2016) and is suitable for making spatio-temporal predictions from spatially sparse time series data (Roberts et al., 2017). Further information on the wetland site class, geolocation, climate, site investigators, and data source is provided in Table 2. Additional information about the FLUXNET-CH4 sites considered for this study, including data digi- tal object identifiers, site references, and source locations, are detailed in (Table S1), on the FLUXNET website and in Delwiche et al. (2021). 2.1.2. Predictor Data A total of 140 candidate predictors were considered for data-driven upscaling. These candidate predictors were organized into five broad classes: climatic (e.g., temperature, precipitation), biometeorological (i.e., flux tower-measured air temperature, and ecosystem carbon and energy fluxes), land cover class and properties (e.g., vegetation class) (Tuanmu & Jetz, 2014, 2015), soil physical and chemical properties (e.g., clay content) (Hengl Figure 2. Location, class, and size of 26 globally distributed freshwater wetland clusters. Symbol sizes reflect the number of weekly CH4 fluxes in each cluster expressed as a percent of the total data set considered in this study. Clusters combine data from sites that occur within 300 km of each other. At least one site was available from each major climate zone (Arctic-boreal, temperate, and tropical) and all major wetland classes were represented. Mean annual maximum wetland area fraction over 2000–2017 is shown from the Wetland Area Dynamics for Methane Modeling (WAD2M) product (Zhang et al., 2021). 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 9 of 24 et al., 2017; Lamarque et al., 2013), topography (e.g., slope), and vegetation indices (e.g., vegetation greenness) (Gao, 1996; Hall & Riggs, 2016; Huete et al., 2002; Jensen & Mcdonald, 2019; Myneni et al., 2015; Vermote, 2015; Wan et al., 2015). Each predictor class contains data from different sources (e.g., models, tower-observations, and/or remote sensing) and information content (e.g., spatial only, temporal only, or spatio-temporal). Gridded data products were extracted at pixels corresponding to tower locations and used directly in model development whereas tower-measured data, when available, were used preferentially in model training and then substituted by a globally gridded product to force the model during upscaling runs (e.g., tower-measured air temperature was mapped to the MERRA-2 reanalysis air temperature gridded product). Predictor data sources for each class are described in detail in (Text S1 and Table S2 in Supporting Information S1; Table S3). Lead and lag times of one, two, or 3 weeks or months (for weekly and monthly data, respectively) were imposed on all temporally-resolved predictors, corresponding to multi-day and seasonal lead and lag timescales identified between wetland CH4 fluxes, and temperature, eddy covariance-derived gross primary production (GPP), and soil moisture-related drivers (Delwiche et al., 2021; Knox et al., 2021). For MODIS data, monthly mean seasonal cycles (MSC) and other annual metrics (e.g., site-year mean, minimum, maximum, amplitude) were also derived. Quality control figures (example shown in Figure S2 in Supporting Information S1) were generated for all predic- tors and used primarily to identify and replace outliers with values from a proximate site within a cluster, an adjacent pixel (when sites were isolated within a cluster), or the site-year median (when varying in time). After deriving predictors, a total of 273 predictors were available for model training. 2.1.3. Predictor Selection We used forward feature selection (FFS) to identify the optimal predictor subset from across all possible predic- tors to use in the final model (Gregorutti et al., 2017; Meyer et al., 2018). For each FFS step, we trained a random forest model algorithm (Breiman, 2001) with all possible predictors and computed the cost function (here, mean absolute error (MAE)) on validation data to identify the predictor(s) associated with the smallest MAE. In the first FFS step, we evaluated all possible predictor pairs (33,670 possible combinations) and selected the pair that resulted in the smallest MAE. In the second FFS step, we evaluated all possible single predictors (of 273), and selected the predictor that, when combined with the first pair, resulted in the smallest MAE. The second step was repeated 10 times to ensure the identification of a global MAE minima. More details are provided in Text S2 in Supporting Information S1 and predictors identified via FFS are visualized in Figure S4 in Supporting Informa- tion S1. FFS is suitable for spatial prediction tasks because it adds predictors when they reduce the cost function computed on held-out data. In contrast, recursive selection relies on importance rankings generated from the training data itself, which increases chances of overfitting (Meyer et al., 2018). 2.1.4. Cross Validation After FFS, random forest models were re-trained using the final predictor set and their predictive performance was evaluated using leave one (cluster) out cross validation (cross validation, hereafter) (Meyer et al., 2018). During model training, a full hyperparameter grid-search was performed, which allowed for deeper, more complex trees (with a small minimum leaf node size). Model performance was evaluated by comparing predicted and observed CH4 fluxes using the coefficient of determination (R 2), MAE normalized by flux standard deviation (nMAE), and bias, computed as the mean of residuals. Nash-Sutcliffe Efficiency (NSE) was also computed as an integrative measure of model performance. NSE is equal to R 2 when model bias is zero, and a NSE > 0 corresponds to a model performance better than simply taking the average of the data. Performance was evaluated with respect to three data set components: site mean flux, mean seasonal cycle (MSC) calculated as the average monthly anom- aly from site mean, and interannual monthly anomalies from the MSC (Jung et al., 2020; Peltola et al., 2019; Tramontana et al., 2016). These components distinguish spatial prediction performance (annual site means) from monthly mean seasonal cycles (MSC), and interannual variability (weekly or monthly anomalies). 2.2. Upscaling Global CH4 Flux 2.2.1. Final Model Ensemble A final random forest model ensemble was trained to propagate uncertainties in training data to a gridded product. First, Monte Carlo simulations were used to create 1,000 simulated training datasets (each composed of CH4 flux plus the final predictor variables) where each weekly observation was drawn from a normal (Gaussian) distribution, 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 10 of 24 except CH4 flux for which the measurement uncertainty is defined as the variance of a double-exponential (Laplace) distribution (Irvin et al., 2021; Knox et al., 2019). For gridded products (e.g., static WorldClim data), dispersion around the true observations (distribution mean) was parameterized as a standard deviation of a 0.25° bounding box around each extracted pixel, for unitless MODIS enhanced vegetation index (EVI) products with the overall measurement uncertainty (0.015), for CH4 flux as the weekly mean uncertainty (incorporating both random and gap-filling uncertainties (Irvin et al., 2021; Knox et al., 2019), and for tower-measured air temper- ature as a conservative estimate of standard temperature sensor precision (0.5°C) (Campbell Scientific, Utah). Second, 500 datasets were bootstrap sampled with replacement from the 1,000 simulations, and in each sample one site cluster was dropped at random. The site US-OWC was also always excluded due to its exceptionally high fluxes which could not be reproduced accurately by the best model (Figure 3a) and could introduce unnecessary bias error at other sites. Training on each bootstrap data set resulted in a final 500-member random forest model ensemble. Full details on Monte Carlo simulations are provided in Table S4 in Supporting Information S1. 2.2.2. Model Forcing We applied the 500-member final model ensemble to an 18-year (216-month) time series of global grids of final predictor variables covering 2001–2018 from which the mean and standard deviation of predictions at each pixel globally was used as mean CH4 flux and data-driven uncertainty. The reconstruction period (2001–2018) aligns with the current Global Carbon Project CH4 modeling protocol and the monthly timestep of the WAD2M wetland extent product (Saunois et al., 2020; Zhang et al., 2021), though weekly data were used to train the machine learning model. For MODIS data, Google Earth Engine (Gorelick et al., 2017) was first used to prepare global monthly grids at 10-km resolution, excluding low quality observations, aggregating from 8-day values to monthly averages using the average of all good 8-day observations within a month, and aggregating from 500-m resolution to 10-km resolution using the average of all good quality 500-m or 1000-m observations within the 10-km pixel. Gaps of time-series of MODIS images were filled using the same methods as for site-based MODIS time series. All global grids were then resampled to a common 0.25° resolution and cropped to exclude Antarctica. Data were reprojected to WGS-84 geographic coordinates. Monthly positive and negative lags were imposed on grid stacks by shifting the stack by whole-month time steps, and linear interpolation between these shifts was used to create weekly time shifts. For each temporal predictor, a mean seasonal cycle stack was created by averaging rasters for each month across all available years for use in the global dissimilarity and tower constituency analyses (see Sections 2.2.3 and 2.2.4). 2.2.3. Wetland Area Products We weighted each grid cell CH4 flux prediction by fractional grid cell wetland extent to estimate CH4 emissions using the primary WAD2M (Zhang et al., 2021) product and an alternate Global Inundation Estimate from Multi- ple Satellites GIEMS version 2; Prigent et al., 2020) global wetland map. We used WAD2M as the primary map because it was also used in the GCP bottom-up model ensemble allowing for direct flux prediction comparisons (Saunois et al., 2020). However, global wetland area is the largest source of uncertainty in wetland CH4 emissions along with flux rates (Bloom et al., 2017; Melton et al., 2013), and the two maps enable a preliminary illustration of the sensitivity of our predicted emissions to different wetland area products. A robust evaluation of wetland extent uncertainties on upscaled emission estimates would require a detailed intercomparison such as that of Melton et al. (2013). Both WAD2M and GIEMS-2 maps were modified with several correction data layers to represent the monthly area covered by vegetated wetlands, excluding open water and coastal wetlands (Text S3 in Supporting Information S1) (Pekel et al., 2016). The maps were generated based on distinct multi-sensor meth- odologies estimating monthly inundated wetland area, to which a set of tailored correction layers and steps were applied to isolate only vegetated wetland area, following the methodology of Zhang et al. (2021). 2.2.4. Global Applicability and Tower Constituency Similar to the challenges faced in global wetland CH4 prediction using BU process models upscaled model predictions were extrapolated across a much larger spatial (Stell et al., 2021) and temporal (Chu et al., 2017) domain than that captured in the training data. Model extrapolation is likely to reduce the accuracy of flux predic- tions and can distort uncertainty estimates (Stell et al., 2021). To measure extrapolation during CH4 upscaling, we first computed point-based dissimilarity (Hoffman et al., 2013) globally, defined as the minimum Euclidean distance between each grid cell-to-flux tower pair in predictor space, normalized by the mean distance among flux towers (Meyer & Buchta, 2020; Meyer et al., 2018, 2019). We then used the dissimilarity map to define a 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 11 of 24 Figure 3. Random forest model predicted versus observed values for(a–d) the mean seasonal cycle (MSC) of methane (CH4) flux for sites in (a) tundra, (b) boreal, (c) temperate, and (d) tropical climate regions and (e) the overall sites mean CH4 flux, during cross validation. Although the US-OWC site is plotted in (a), it is excluded from calculation of Nash-Sutcliffe Efficiency (NSE), Coefficient of Determination (R 2), and mean absolute error (MAE; nmol m −2 s −1) performance metrics, and sample count (n). The 1:1 fit is shown as a dashed black line. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 12 of 24 monthly, global area of model applicability (AOA) and corresponding area of extrapolation, using a dissimilarity threshold (Meyer & Pebesma, 2022). Finally, we defined the global constituency of each site cluster to identify which training conditions dominate global predictions and further evaluate the plausibility of global extrapola- tions. Each pixel was assigned as a constituent of the site cluster that was closest in predictor space (Hargrove & Hoffman, 2004). We propose that combining constituencies with AOA provides a semi-quantitative and inform- ative approach to evaluating global representativeness and extrapolation confidence in upscaling models. We demonstrate this approach by identifying the regions that are most like training conditions and which training conditions each region is most similar to. This approach allows us to diagnose CH4 flux prediction patterns in extrapolations. 3. Results 3.1. Predictive Modeling 3.1.1. Model Predictors Only the first six predictors from the FFS were used in the final upscaling model, as they accounted for ∼85% of the MAE reduction. These predictors were tower-measured air temperature, with and without a 2-week lag, MODIS EVI with a 3-week lag, mean temperature of the driest quarter, precipitation of the wettest month, and vegetation canopy height (Figure S4 in Supporting Information S1). It is notable that three of the final six predic- tors were temperature related. An additional five predictors not included in the final model extended the FFS to the MAE minimum (Table S5 in Supporting Information S1) and included MODIS snow cover, tower-measured GPP with a 2-week lead, and the annual minimum in MODIS EVI. The random forest variable importance rankings deviated slightly in order from FFS, with air temperature and the two static climatological predic- tors ranked as the most important final predictors, while canopy height and MODIS EVI were less important (Figure S5 in Supporting Information S1). A strong exponential dependency was observed between CH4 flux and air temperature, likely explaining its dominance in the variable importance, while more complex and/or less tightly correlated dependencies were observed between CH4 flux and the other predictors (Figure S6 in Support- ing Information S1). Temperature hysteresis was not reproduced in the model in many sites where it has been observed in the CH4 flux observations (Chang et al., 2021) (Figure S7 in Supporting Information S1). 3.1.2. Cross Validation Performance Model residuals (errors) were normally distributed around zero at bogs, fens, swamps, and wet tundra sites (Figure S8 in Supporting Information S1). At freshwater marshes, residuals displayed more negative outliers due to one site (US-OWC) that displayed exceptionally high CH4 fluxes (>10× higher than overall median) that the model did not reproduce. The fluxes at US-OWC are plausible because the site is situated in a eutrophic estuarine marsh on the southern shore of Lake Erie, USA, (Rey-Sanchez et al., 2018), which displays very high rates of sediment methanogenesis (Angle et al., 2017). However, evaluating the global scale emissions from eutrophic wetland is beyond the scope of this first wetland upscaling effort and therefore, hereafter, cross validation metrics are reported with the exclusion of the exceptionally high fluxes at site US-OWC. When predicting site mean CH4 fluxes during cross validation (Figure 3a), the model achieved an NSE of 0.54 and nMAE of 0.42, and low model bias (2.6 nmol m −2 s −1) relative to the overall site mean CH4 flux (61.5 nmol m −2 s −1). Site mean CH4 flux errors were not spread evenly among wetland classes. Mean absolute errors (MAE) increased from wet tundra (11.7 nmol m −2 s −1) to bogs (13.2 nmol m −2 s −1), fens (25.7 nmol m −2 s −1), marshes (35.2 nmol m −2 s −1), and swamps (62.2 nmol m −2 s −1). However, after normalizing by flux standard deviation, nMAE increased from marshes (0.53) to fens (0.54), bogs (0.64), swamps (0.78), and wet tundra (1.48), reflecting low CH4 flux variability at wet tundra and high flux variability at marshes (Table S6 in Supporting Information S1). Mean seasonal cycles were, overall, predicted by the model with comparable accuracy to the site means (NSE = 0.53 and nMAE = 0.41). However, model prediction performance on MSC differed greatly by climate region (Figures 3b–3e), and decreased from higher at temperate (R 2 = 0.64; NSE = 0.63; nMAE = 0.37), boreal (R 2 = 0.62; NSE = 0.54; nMAE = 0.37), and tundra sites (R 2 = 0.59; NSE = 0.52; nMAE = 0.53), to lower at tropical sites (R 2 = 0.08; NSE = −0.12; nMAE = 0.81). Although all tropical sites were swamps, the pattern was less clear when sites were grouped by wetland class rather than by climate region, because the model achieved high MSC performance at a temperate swamp, US-NC4 (R 2 = 0.78; NSE = 0.50; nMAE = 0.58) (Table S6 in Supporting Information S1). Finally, the model was unable to predict interannual monthly anomalies from the MSC (NSE = −5.11; R 2 = 2e −3; nMAE = 37.2). 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 13 of 24 3.2. Global Upscaling 3.2.1. Unweighted Global Methane Flux Predictions Gridded freshwater wetland CH4 flux predictions, before being weighted by wetland extent (Figure S9 in Supporting Information S1), were compared pixel-wise to the original 43 training and four additional test flux tower sites (Figure 4). Globally, the model achieved an R 2 of 0.53, NSE of 0.51, nMAE of 0.35, and Bias of −6.0 n mol m −2 s −1. The slight improvement compared to cross-validation performance is expected as all the training data appears many times across the 500 bootstrap-sampled data sets models. Examples of gridded prod- uct predictions at training sites and their uncertainties are shown in Figure S10 in Supporting Information S1. Notably, the model also performed well at three of four additional test sites (Figure 4), reproducing the site mean with nMAE ranging from 0.39 to 0.49 at the boreal and temperate sites. However, as was observed in training evaluations, predictions at the tropical forest test site (PE-QFR) did not reproduce the seasonal signal and exhib- ited the largest nMAE (1.35). 3.2.2. Global Model Applicability and Tower Constituency Confidence in gridded model predictions was also evaluated semi-quantitatively using global dissimilarity and tower constituency analyses. Global dissimilarity was low, even in areas geographically distant from existing towers (Figure S11 in Supporting Information  S1), suggesting that the model was not forced to extrapolate far from training conditions, even when making global predictions. The most dissimilar regions were Eastern Figure 4. The mean seasonal cycle of model-predicted CH4 flux (solid black line) and uncertainty range (gray ribbon) compared against monthly observed mean fluxes (open circles) and (a–c) standard deviation or (d) 25th–75th percentiles (vertical bars) for four AmeriFlux test sites. The model reproduced mean fluxes and the seasonal cycle best at (a) a boreal fen (CA-CF2; Tenuta, 2020), (b) a temperate fen (US-ALQ; Olson, 2018), and (c) a boreal bog site (US-MBP; Roman et al., 2021), whereas seasonal cycle performance was not reproduced at (d) one humid tropical forest site (PE-QFR; Roman et al., 2020), which is an upland site that experiences very wet soil conditions and supports substantial but highly variable CH4 fluxes. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 14 of 24 Siberia, South Asia (e.g., India), the Sahel and Congo regions of sub-Saharan Africa, and the Amazon basin. The most dissimilar sites were in tropical and temperate regions, with lower dissimilarity in boreal and tundra regions (Figure S11 in Supporting Information  S1). The tower constituency map (Figure S12 in Supporting Information S1) visualized how the model is likely to extrapolate from training data at sites to geographically distant regions, given their similar predictor conditions. The Amazon and Congo basins, which lack flux towers, fall within the constituencies of two SE Asian towers (ID-Pag and MY-MLM). The constituency of the Brazilian Pantanal tower, BR-Npw, encompasses other wet savanna and monsoon regions of South America, Africa, and Asia. Much of central and western Australia and semi-arid tropics of South America and the Sahel are assigned to the subtropical Botswana site constituencies (BW-Nxr and BW-Gum) or Mediterranean California (USA) site constituencies (US-Myb, US-Tw1). These semi-arid or arid tropical constituency assignments are reasonable on the basis of climate; however, it is important to note that the wetland sites in these environments where the afore- mentioned towers are located are associated with large inland deltas (i.e., the Okavango Delta, Botswana, or the Sacramento Delta, California), which provide water to support marsh and swamp wetlands. The final model did not include surface hydrological variables and therefore extrapolation to these regions may not be easily evaluated by the dissimilarity and constituency analysis employed here. No relationship was observed between model error or variance and site-month dissimilarity (Figure S13 in Supporting Information S1) that could be used to scale errors in regions of extrapolation (Jung et al., 2020). 3.2.3. Wetland Area-Weighted CH4 Fluxes Time series of mean freshwater wetland CH4 fluxes from UpCH4 (2001–2018), weighted by WAD2M pixel wetland area, displayed regional patterns that reflected the interaction of wetland area and the model's flux predictions (Figure 5a). The highest wetland area-weighted fluxes (>30 mg CH4 m −2 d −1) were predicted in both high- and low-latitude regions with extensive wetland area (e.g., Hudson Bay Lowlands (HBL), Congo Basin) and in semi-arid regions where wetland cover is low but model flux predictions were very high (Figure S9 in Supporting Information S1). Relative uncertainties (Figure 5b) were the smallest for high-latitude high emission (>20 mgCH4 m −2 d −1) hotspots associated with the HBL and West Siberian Lowlands (WSL) wetland complexes and were largest in monsoon regions of relatively low flux (<5 mgCH4 m −2 d −1) sandwiched between the semi- arid and humid tropics. Upscaled (UpCH4) fluxes were compared to three alternative global wetland CH4 emission datasets using either the same WAD2M wetland area (i.e., GCP BU ensemble; Figure 5c), or variable wetland products (i.e., GCP TD ensemble; Figure  5e). At high-latitude wetland complexes (HBL and WSL), UpCH4 predicted slightly lower emissions than the mean of the GCP BU ensemble, similar to the GCP TD ensemble, but higher than WetCHARTS. Given the same WAD2M wetland area was applied in upscaling as in the GCP BU ensemble, this difference can be attributed to lower predicted CH4 fluxes by UpCH4. At mid-to-low latitudes, UpCH4 predicted 10–30 mgCH4 m −2 d −1 higher fluxes than the other products from the semi-arid tropics, including the Sahel and Horn of Africa, central and western Australia, and western Asia, while also predicting 20–30 mgCH4 m −2 d −1 lower fluxes from the humid tropics, including the large wetland complexes of the Amazon and Congo Basins, and SE Asia, especially in Indonesia and Malaysia. Again, as the GCP BU ensemble used the same wetland area, these product differences can be attributed to wetland CH4 flux rates rather than extent. The regional pattern of tropical emissions in WetCHARTS was more similar to the GCP TD ensemble pattern than the GCP BU ensem- ble pattern, which differ as described in Figure 1. 3.2.4. Temporal Trends and Spatial Patterns in UpCH4 Emissions Upscaling model (UpCH4) emissions using both WAD2M and GIEMS-2 wetland area products are seasonal, with a JJA peak that corresponds with the expansion of wetland area, the warmest soil temperatures, and peak productivity at northern high-latitudes, (Figure 6). As indicated by globally integrated fluxes, UpCH4-WAD2M emissions correspond well with the average of the GCP BU ensemble range, whereas UpCH4-GIEMS-2 emis- sions are significantly lower, although the amplitude of the seasonal cycle is larger. No clear long term annual trend is predicted by UpCH4 (or other models), though interannual variability is apparent, driven by wetland extent changes. The latitudinal pattern of UpCH4 emissions using either wetland extent lacks the year-round elevated equato- rial band found in the GCP products (Figure 7). The temporal pattern of this band varies slightly between the GCP BU and TD ensemble, which show a clear seasonal peak in AMJ. The UpCH4-WAD2M upscaling still 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 15 of 24 produces a similar global total to the GCP BU ensemble from a much larger latitudinal range for moderate fluxes. Increased UpCH4 emissions are observed for narrow tropical N hemisphere (10–20°N) and subtropical S hemi- sphere (20–30°S) bands, likely related to high flux predictions in the semi-arid climate whereas GCP BU and TD ensembles show enhanced fluxes around 30°N and 10°N. 4. Discussion 4.1. Model Performance UpCH4 achieved comparable metrics based on our spatial cross-validation (global R 2  =  0.50; Northern site R 2 = 0.48) to a recent northern high-latitude upscaling of CH4 fluxes from 25 eddy covariance towers (Peltola et al., 2019), despite a larger and more variable global flux data set. UpCH4 also achieved better performance than global upscaling models for net ecosystem exchange of CO2 (R 2 < 0.5) (FLUXCOM; Tramontana et al., 2016). As noted by Peltola et al. (2019), net CH4 fluxes measured with eddy covariance may be difficult for machine learning models to reproduce with limited datasets as they display complex and non-linear behavior, and are subject to storage effects and lags (Knox et al., 2021; Sturtevant et al., 2016) due to underlying CH4 production, oxidation, and transport processes (Bridgham et al., 2013; Chang et al., 2021) identified substantial hysteresis in the seasonal temperature dependency of wetland CH4 flux and Delwiche et al. (2021) identified lags and leads of various lengths between peak growing season air temperature and peak CH4 flux. Temperature lags may be due to the substrate control of CH4 production and would agree with coherence with ecosystem production (Knox Figure 5. Global maps of: (a) Upscaled (UpCH4) mean 2001–2018 CH4 flux using WAD2M wetland area (b) CH4 flux uncertainty computed as 1 standard deviation of the random forest ensemble expressed as a percent of the mean (i.e., coefficient of variation) 2001–2018 flux; (c, e) the mean 2001–2018 CH4 flux from the Global Carbon Project (GCP) bottom-up/top-down process model ensemble (also using WAD2M) subtracted from the UpCH4 mean (a); (d, f) the correlation, expressed as the correlation coefficient of determination (R 2), between the mean seasonal cycle (MSC) of UpCH4 and the GCP bottom-up/top-down ensemble. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 16 of 24 et al., 2021; Mitra et al., 2020). In UpCH4, leads and lags of various lengths were imposed on all temporal predic- tors to address part of this complexity and two lagged predictors were selected in the final model (i.e., lagged temperature and EVI), indicating their utility when using non-temporal machine learning algorithms such as random forests. However, observed hysteresis between temperature and CH4 flux was not reproduced (Figure S7 in Supporting Information S1), suggesting that simply imposing lags on predictors cannot capture the complex biogeochemical processes that drive intra-seasonal variability (Chang et al., 2021). Deep learning models able to learn temporal dependencies, such as Long Short Term Memory (LSTM) neural networks, and/or able to incor- porate process knowledge as constraints, could be considered as the core algorithms for CH4 upscaling; however, this would require adapting model architecture for spatio-temporal predictions (Reichstein et al., 2019). Performance improvements may also be expected if wetland CH4 fluxes measured using eddy covariance can be partitioned into ebullition, diffusion, and plant-mediated transport pathways as has been done to partition diffusion and ebullition in lakes and bogs (Iwata et  al.,  2018; Ueyama et  al.,  2022). Each of these transport Figure 6. Monthly time series of upscaled wetland CH4 emissions (UpCH4) using WAD2M (2001–2018) and GIEMS-2 (2001–2015) compared with the 13-member GCP BU ensemble (2001–2017) and 17-member subset of GCP TD ensemble (2010–2017). The UpCH4 mean and spread are difficult to distinguish from those of the GCP BU ensemble due to small differences between the two estimates. Figure 7. Average monthly freshwater wetland CH4 emissions (TgCH4 month −1) for ∼1°latitude bands from UpCH4 for the two wetland maps (WAD2M and GIEMS-2) over 2001–2017; three alternative global datasets (Bottom-up GCP ensemble over 2001–2017, Top-down ensemble over 2010–2017, and WetCHARTS v1.0 over 2000–2010) are also shown. Average CH4 flux estimates from five data sources showing differences in total emissions. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 17 of 24 processes is regulated by distinct drivers and produces flux signals that may be more directly attributed to these processes when separated from each other. Improved models may also be possible if we can reconcile current wetland classifications with real wetland differences in the mean and/or vari- ance of methane fluxes. 4.2. Budget Comparisons Global freshwater wetland CH4 emissions and overall uncertainties during the period 2001–2018 from UpCH4 were 146  ±  42.7 TgCH4  y −1 using WAD2M wetland area (Figure 8a), which closely matches emissions from the GCP bottom-up model ensemble for 2007–2018 (mean 149 TgCH4 y −1; range 102–182 TgCH4 y −1) but is substantially lower (31–44 TgCH4 y −1) than GCP top-down emissions (mean 181 TgCH4 y −1; range 159–200 TgCH4 y −1) (Saunois et  al.,  2020) and WETCHIMP emissions (190 ± 39 TgCH4 y −1) (Melton et  al.,  2013). Notably, UpCH4-WAD2M global emissions were also similar to that of Nzotungicimpaye et  al.  (2021) (158.6 TgCH4  y −1) who implemented WETMETH—a CH4 process model, within the UVic Earth System Climate Model, and Ma et al.  (2021) (148 TgCH4 y −1) who used satellite-based observations to refine estimates from 42 BU process models. Using GIEMS-2 corrected for only vegetated wetland area, rather than WAD2M, cut UpCH4 global emissions by more than half (71.8 ± 22.6 TgCH4 y −1), highlighting the sensitivity of total emissions to the wetland map used (Figures S14–S16 in Supporting Information S1). At the global scale, UpCH4 mean CH4 emissions agreed more closely with GCP BU emissions for tropical (60S–30N), temperate (30–60 N), northern (>45 N), and Arctic (60–90 N) latitudinal ranges compared to GCP TD esti- mates, which were consistently higher, most notably in the tropics. However, UpCH4 uncertainty ranges overlapped with both TD and BU estimates and supported the previously reported observation that ∼68% of wetland CH4 emissions originate from tropical wetlands (Saunois et  al.,  2020). Within major wetland complexes at northern latitudes (Figure 8b), UpCH4 agreed more closely with lower TD estimates for the WSL, and UpCH4 emissions for the HBL and Prairie Pothole Region were also lower than both BU and TD estimates. For tropical wetland regions, estimates diverged more significantly between UpCH4 and the GCP ensembles (Figure 8c) with larger uncertainties likely due to lack of data in tropics. Although UpCH4 agreed closely with BU estimates for the tropical latitude total, UpCH4 emissions were much (∼3x) higher for the semi-arid monsoon Sahel compared to either GCP ensemble product, and much lower emissions estimated for the humid tropical forested wetlands of the Amazon, Congo, and the Indonesian archipelago. 4.3. Interpreting Product Differences Interpreting similarities and differences in spatial patterns between global estimates of wetland CH4 emissions is challenging because they arise from differences in both wetland fluxes and wetland area (akin to Melton et al., 2013). However, some broad conclu- sions can be drawn. GCP BU ensemble comparisons are informative at subtropical and tropical latitudes because in this study the shared use of the WAD2M wetland area enables model differences to be fully attributed to flux processes. As UpCH4 global upscaling emissions agreed most closely with GCP BU ensemble, it is nota- ble that tropical and subtropical regional patterns diverged substantially (Figure 5c). The semi-arid to humid tropics gradient is inverted in UpCH4 when compared with GCP and thus these models produce similar global totals but different regional distributions. Given that inversion and observational studies in the Amazon Basin indicate a very large CH4 source (Devol et al., 1988; Gauci et al., 2022; Pangala et al., 2017), it is plausible that Figure 8. (a) Global and (b, c) regional comparisons of annual wetland CH4 emissions from upscaling (UpCH4; blue circles), GCP BU ensemble (bottom-up; black square), and TD inversion ensemble (top-down; green triangle). Regional wetland complex initialisms for West Siberian Lowlands (WSL; 52–74°N, 60–95°W), Hudson Bay Lowlands (HBL; 50–60°N, 75–96°W); Prairie Pothole Region (PPR; 42–55°N, 92–115°W). 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 18 of 24 the upscaling pattern is biased, and that semi-arid and humid tropical wetland emissions are over- and under- estimated, respectively. However, observations of exceptional CH4 emissions from seasonal wetlands in the hot semi-arid tropics (e.g., BR-Nxr and BR-Gum, the Okavango Delta wetlands in this study) underscore that more measurements could support more robust validation of these patterns. If semi-arid/monsoon wetland CH4 sources are currently being underestimated in BU models, this missing source would significantly close the gap between current BU and TD global wetland CH4 emissions estimates. 4.4. Data Limitations The training data contained a geographic bias with many more temperate and boreal northern hemisphere sites than tropical and southern hemisphere sites. Only 9% of the training data (5 sites) were acquired from south of 23°N, despite the tropics accounting for the vast majority of the non-frozen global wetland area and an estimated two-thirds of global CH4 emissions (Melton et al., 2013; Saunois et al., 2020). As a result, tropical CH4 flux predictions were based on flux data from a few towers that are unlikely to be representative of the entire tropical region and, correspondingly, tropical prediction uncertainties from UpCH4 were high (Figure 5b). For instance, training data for semi-arid subtropical regions were dominated by site clusters in the Sacramento Delta, United States (US-Myb, US-Sne, US-Tw1, US-Tw4, US-Tw5), and the Okavango Delta, Botswana (BW-Gum, BW-Nxr), which together accounted for ∼20% of all training data. In both regions, minerotrophic deltaic wetlands are dependent on large seasonal or permanent allochthonous riverine inputs from regional-scale drainage basins, and thus sustain very productive marsh or swamp conditions conducive to high CH4 fluxes that contrast sharply from the surrounding dryland environments (Hemes et al., 2019; Knox et al., 2015). The lack of wetland data from more varied hydrologic classes of wetlands (e.g., riverine, isolated or rainfed, i.e., ombrotrophic) under similar climate conditions, combined with the biases described above, may have led to the high predicted CH4 fluxes for the semi-arid subtropics. The CH4 flux predictions and uncertainties (Figure 5), combined with the tower constituency and model appli- cability maps (Figures S10 and S11 in Supporting Information S1), provide the first global survey for situating new eddy covariance measurement sites based on both CH4 flux and environmental information. Moreover, an expanded site constituency representativeness analysis confirmed that global wetland CH4 upscaling will benefit most from additional tropical wetland flux data (Figure 9). The large humid tropical site constituencies (e.g., ID-Pag—a SE Asian peat swamp forest) imply very wide model extrapolation across the Amazon and Congo Basins, where there is currently no available CH4 flux data and where fluxes could be quite different. Simi- larly, the largest CH4 flux uncertainties were observed for transitional climate regions between the semi-arid and humid tropics (Figure 5) and establishing towers to bracket or traverse these regions could help capture important gradients in tropical wetland conditions relevant to CH4 flux variability. Overall, the lack of observations for Figure 9. (a) Global wetland CH4 flux site-cluster constituencies for which pixels are assigned to one of 26 site-clusters (colors) based on similarity to UpCH4 predictor conditions at the sites within that cluster. (b) The ranked difference in predicted CH4 emissions (TgCH4 y −1) between UpCH4 predictions for a given constituency (colors) and a simple extrapolation of the monthly mean flux for the site cluster to the entire constituency (ignoring flux data from other sites). Constituencies and their upscaling-based emission estimates are likely to be insensitive to adding additional sites (more transparent colors in map (a)) when emissions differences are close to zero (short segments in (b)). In contrast, constituencies and emission estimates are likely to be more sensitive to additional sites (opaque colors in map (a)) where absolute emissions differences are large (long segments in (b)). The size of the segment end points is proportional to the annual constituency CH4 flux. A full constituency map without variable color transparency is provided in Figure S12 in Supporting Information S1. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 19 of 24 large regions of the tropics combined with the distinctive hydrological regimes characterizing different tropical wetland regions (Apers et al., 2022; Dalmagro et al., 2018), likely account for the upscaling discrepancy with past and typically higher tropical emission estimates, such as for the Amazon Basin (Figure 6). Refined methods to evaluate CH4 flux tower network representativeness along different dimensions of variability could result in improved estimates, as has been undertaken at regional scales (Malone et al., 2022; Villarreal & Vargas, 2021). Similarly, use of more finely resolved spatial forcing data can more accurately represent wetland conditions and may improve model functional responses (e.g., 30-m resolution in Bansal et al., 2023). In addition to geographic bias at global scales, wetland flux towers are also likely biased with respect to average grid cell conditions. Wetlands are often a minority cover type at landscape scales meaning that training data based on grid cell averages alone, for example, MODIS vegetation products may not be representative of tower conditions (Chu et al., 2021). Applied to the case described above of minerotrophic swamp or marsh surrounded by drylands, scale-mismatch will likely result in erroneous wetland vegetation indices, such as greenness and phenology metrics, as well as derived products such as GPP, which many studies, including this study, indicate as important for predicting CH4 fluxes(e.g., Bridgham et al., 2013; Chang et al., 2021; Delwiche et al., 2021; Knox et al., 2019; Whiting & Chanton, 1993). Spatial biases at global and landscape scales could be addressed in future work by improving the geographic and grid cell representativeness of CH4 flux and predictor data, respectively. Developing methods to reconcile and integrate chamber flux data with tower flux data could be prioritized to gain information from the large amount of existing chamber data, such as in Bansal et al. (2023), Kuhn et al. (2021), and Turetsky et al. (2014). Chamber methods offer a relatively inexpensive and accessible means to gather data in underrepresented regions (Harriss & Matson, 2009), may provide insights into patch-scale effects when paired with tower data via environmental response functions methods (Xu et al., 2017), and can extend the temporal representativeness of flux data (Chu et al., 2017). 5. Conclusions We develop a wetland CH4 flux upscaling workflow (UpCH4) for eddy covariance flux data and evaluate global CH4 emission predictions. Extratropical estimates from UpCH4 can provide insights from comparisons to exist- ing CH4 model predictions. The use of UpCH4 tropical wetland emissions estimates should include consid- eration of uncertainties and joint use of additional regional data constraints is strongly encouraged. UpCH4 estimates average annual freshwater wetland CH4 emissions of 146 ± 42.7 TgCH4 y −1 for 2001–2018 which aligns closely with the most recent GCP BU estimates (149 TgCH4 y −1) (Saunois et al., 2020) and a hybrid study that constrained a large ensemble of process models with satellite data (148 TgCH4 y −1) (Ma et al., 2021). UpCH4 emission uncertainties were larger than, and overlapped with, both GCP BU and TD estimates, and the sensitivity to wetland extent products is illustrated by the halving of the global emissions total (71.8 ± 22.6 TgCH4 y −1) when using GIEMS-2 corrected for only vegetated wetland area. All gridded emissions products are available via ORNL DAAC (https://doi.org/10.3334/ORNLDAAC/2253). UpCH4 is most suitable for comparison to other bottom-up and top-down models within temperate, boreal, and arctic climate zones from 2010 onwards, and will improve in tropical regions as EC data coverage is expanded over time. Conflict of Interest The authors declare no conflicts of interest relevant to this study. Data Availability Statement Different gridded CH4 flux outputs were generated with and without applying the WAD2M or GIEMS-2 wetland extent masks (Text S1 and Figure S3 in Supporting Information S1). Thus, final gridded products are: (a) grid- ded unweighted wetland CH4 fluxes in nmol CH4 m −2  s −1 and g C-CH4 m −2  d −1; (b) wetland area-weighted fluxes in mg CH4 m −2 d −1; and (c) wetland area-weighted fluxes in TgCH4 grid cell −1 month −1. Main data prod- ucts are available via DOE ORNL DAAC and Zenodo (UpCH4: https://doi.org/10.3334/ORNLDAAC/2253; WAD2M: https://doi.org/10.5281/zenodo.5553187). The GIEMS-2 wetland extent product is available at request from Catherine Prigent (catherine.prigent@obspm.fr). Code notebooks for random forest model development 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 20 of 24 and validation are available via Zenodo (https://doi.org/10.5281/zenodo.7978099). The Global Carbon Project methane budget is available at https://doi.org/10.18160/GCP-CH4-2019. References Abernethy, S., O'Connor, F. M., Jones, C. D., & Jackson, R. B. (2021). Methane removal and the proportional reductions in surface tempera- ture and ozone. Philosophical Transactions. Series A, Mathematical, Physical, and Engineering Sciences, 379(2210), 20210104. https://doi. org/10.1098/rsta.2021.0104 Angle, J. C., Morin, T. H., Solden, L. M., Narrowe, A. B., Smith, G. J., Borton, M. A., et al. (2017). Methanogenesis in oxygenated soils is a substantial fraction of wetland methane emissions. Nature Communications, 8(1), 1567. https://doi.org/10.1038/s41467-017-01753-4 Apers, S., De Lannoy, G. J. M., Baird, A. J., Cobb, A. R., Dargie, G. C., Pasquel, J., et al. (2022). Tropical peatland hydrology simulated with a global land surface model. Journal of Advances in Modeling Earth Systems, 14(3), e2021MS002784. https://doi.org/10.1029/2021ms002784 Baldocchi, D. (2014). Measuring fluxes of trace gases and energy between ecosystems and the atmosphere—The state and future of the eddy covariance method. Global Change Biology, 20(12), 3600–3609. https://doi.org/10.1111/gcb.12649 Bansal, S., Post van der Burg, M., Fern, R. R., Jones, J. W., Lo, R., McKenna, O. P., et al. (2023). Large increases in methane emissions expected from North America's largest wetland complex. Science Advances, 9(9), eade1112. https://doi.org/10.1126/sciadv.ade1112 Billesbach, D., & Sullivan, R. (2020a). FLUXNET-CH4 US-A03 ARM-AMF3-Oliktok [Dataset]. FluxNet; Argonne National Laboratory. https:// doi.org/10.18140/FLX/1669661 Billesbach, D., & Sullivan, R. (2020b). FLUXNET-CH4 US-A10 ARM-NSA-Barrow [Dataset]. FluxNet; Argonne National Laboratory. https:// doi.org/10.18140/FLX/1669662 Bloom, A., Bowman, W. K., Lee, M., Turner, J. A., Schroeder, R., Worden, R. J., et al. (2017). A global wetland methane emissions and uncer- tainty dataset for atmospheric chemical transport models (WetCHARTs version 1.0). Geoscientific Model Development, 10(6), 2141–2156. https://doi.org/10.5194/gmd-10-2141-2017 Bodesheim, P., Jung, M., Gans, F., Mahecha, M. D., & Reichstein, M. (2018). Upscaled diurnal cycles of land--atmosphere fluxes: A new global half-hourly data product. Earth System Science Data, 10(3), 1327–1365. https://doi.org/10.5194/essd-10-1327-2018 Bohrer, G., Kerns, J., Morin, T., Rey-Sanchez, A., Villa, J., & Ju, Y. (2020). FLUXNET-CH4 US-OWC old woman creek [Dataset]. FluxNet; Old Woman Creek National Estuarine Research Reserve; The Ohio State University. https://doi.org/10.18140/FLX/1669690 Bohrer, G., & Morin, T. (2020). FLUXNET-CH4 US-ORv Olentangy River wetland research park [Dataset]. FluxNet; The Ohio State University. https://doi.org/10.18140/FLX/1669689 Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324 Bridgham, S. D., Cadillo-Quiroz, H., Keller, J. K., & Zhuang, Q. (2013). Methane emissions from wetlands: Biogeochemical, microbial, and modeling perspectives from local to global scales. Global Change Biology, 19(5), 1325–1346. https://doi.org/10.1111/gcb.12131 Bruhwiler, L., Dlugokencky, E., Masarie, K., Ishizawa, M., Andrews, A., Miller, J., et  al. (2014). CarbonTracker-CH_4: An assimilation system for estimating emissions of atmospheric methane. Atmospheric Chemistry and Physics, 14(16), 8269–8293. https://doi.org/10.5194/ acp-14-8269-2014 Campbell, D., & Goodrich, J. (2020). FLUXNET-CH4 NZ-Kop Kopuatai [Dataset]. FluxNet; University of Waikato. https://doi.org/10.18140/ FLX/1669652 Chang, K.-Y., Riley, W. J., Knox, S. H., Jackson, R. B., McNicol, G., Poulter, B., et al. (2021). Substantial hysteresis in emergent temperature sensitivity of global wetland CH4 emissions. Nature Communications, 12(1), 2266. https://doi.org/10.1038/s41467-021-22452-1 Chen, J., & Chu, H. (2020). FLUXNET-CH4 US-WPT winous point North Marsh [Dataset]. FluxNet; University of Toledo / Michigan State University. https://doi.org/10.18140/FLX/1669702 Chen, X., Bohn, T. J., & Lettenmaier, D. P. (2015). Model estimates of climate controls on pan-Arctic wetland methane emissions. Biogeo- sciences, 12(21), 6259–6277. https://doi.org/10.5194/bg-12-6259-2015 Chu, H., Baldocchi, D. D., John, R., Wolf, S., & Reichstein, M. (2017). Fluxes all of the time? A primer on the temporal representativeness of FLUXNET. Journal of Geophysical Research: Biogeosciences, 122(2), 289–307. https://doi.org/10.1002/2016JG003576 Chu, H., Luo, X., Ouyang, Z., Chan, W. S., Dengel, S., Biraud, S. C., et al. (2021). Representativeness of Eddy-Covariance flux footprints for areas surrounding AmeriFlux sites. Agricultural and Forest Meteorology, 301–302, 108350. https://doi.org/10.1016/j.agrformet.2021.108350 Dalmagro, H. J., Lathuillière, M. J., Hawthorne, I., Morais, D. D., Pinto, O. B., Jr., Couto, E. G., & Johnson, M. S. (2018). Carbon biogeochem- istry of a flooded Pantanal forest over three annual flood cycles. Biogeochemistry, 139(1), 1–18. https://doi.org/10.1007/s10533-018-0450-1 Delwiche, K. B., Knox, S. H., Malhotra, A., Fluet-Chouinard, E., McNicol, G., Feron, S., et al. (2021). FLUXNET-CH4: A global, multi-ecosystem dataset and analysis of methane seasonality from freshwater wetlands. Earth System Science Data, 13(7), 3607–3689. https://doi.org/10.5194/ essd-13-3607-2021 Dengel, S., Zona, D., Sachs, T., Aurela, M., Jammet, M., Parmentier, F. J. W., et al. (2013). Testing the applicability of neural networks as a gap-filling method using CH_4 flux data from high latitude wetlands. Biogeosciences Discussions, 10(12), 8185–8200. https://doi.org/10.5194/ bg-10-8185-2013 Desai, A. (2020). FLUXNET-CH4 US-Los lost creek [Dataset]. FluxNet; University of Wisconsin. https://doi.org/10.18140/FLX/1669682 Devol, A. H., Richey, J. E., Clark, W. A., King, S. L., & Martinelli, L. A. (1988). Methane emissions to the troposphere from the Amazon flood- plain. Journal of Geophysical Research, D: Atmospheres, 93(D2), 1583–1592. https://doi.org/10.1029/JD093iD02p01583 Dolman, H., Maximox, T., Parmentier, F., & Budishev, A. (2020). FLUXNET-CH4 RU-Cok Chokurdakh [Dataset]. FluxNet; Vrije Universiteit Amsterdam. https://doi.org/10.18140/FLX/1669656 Eichelmann, E., Knox, S., Sanchez, C., Valach, A., Sturtevant, C., Szutu, D., et al. (2020). FLUXNET-CH4 US-Tw4 Twitchell East end wetland [Dataset]. FluxNet; University of California. https://doi.org/10.18140/FLX/1669698 Euskirchen, E., Bret-Harte, M., & Edgar, C. (2020). FLUXNET-CH4 US-ICs imnavait creek watershed wet sedge tundra [Dataset]. FluxNet; Marine Biological Laboratory; University of Alaska Fairbanks. https://doi.org/10.18140/FLX/1669678 Euskirchen, E., & Edgar, C. (2020a). FLUXNET-CH4 US-BZB bonanza creek thermokarst bog [Dataset]. FluxNet. https://doi.org/10.18140/ FLX/1669668 Euskirchen, E., & Edgar, C. (2020b). FLUXNET-CH4 US-BZF bonanza creek rich fen [Dataset]. FluxNet; University of Alaska Fairbanks, Institute of Arctic Biology. https://doi.org/10.18140/FLX/1669669 Fick, S. E., & Hijmans, R. J. (2017). WorldClim 2: New 1-km spatial resolution climate surfaces for global land areas. International Journal of Climatology, 37(12), 4302–4315. https://doi.org/10.1002/joc.5086 Acknowledgments Primary support for the project came from the Gordon and Betty Moore Foun- dation (Grant GBMF5439, “Advancing Understanding of the Global Methane Cycle”) to Stanford University, with additional support from the John Wesley Powell Center for Analysis and Synthesis of the U.S. Geological Survey (USGS “Wetland FLUXNET Synthesis for Meth- ane” working group). Further support came from the National Aeronautics and Space Administration Carbon Monitoring System award to Lawrence Berkeley National Laboratory and the University of Illinois Chicago (Zhu Prototyping a monitoring system of global wetland CH4 emissions with machine learning and satellite remote sensing) (20-CMS20- 0039; NNH20ZDA001N). Observa- tions from the Atmospheric Radiation Measurement (ARM) user facility are supported by the U.S. Department of Energy (DOE) Office of Science user facility managed by the Biological and Environmental Research Program. Work at Argonne National Laboratory was supported by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, under contract DEAC0206CH11357. Sarah Feron acknowledges the support of ANID (ANILLO ACT210046) and CORFO (19BP-117358). Work at USGS was additionally supported by the Biological Carbon Sequestration program. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. ICOS Finland was funded by the Academy of Finland, the Ministry of Transport and Communications, and University of Helsinki. We thank Dario Papale, Markus Reichstein, Martin Jung, Martijn Pallandt, Hanna Meyer, and Olli Peltola for consultations on upscaling methodology. We thank Catherine Prigent for the use of the GIEMS-2 wetland data set (Prigent et al., 2020). Selected data of this study are available through the USGS ScienceBase repository (Krauss et al., 2018). We thank Tamara Wilson and Geneva Chong (USGS) for signed reviews and two further anonymous reviewers, and the research teams contrib- uting to the FLUXNET-CH4 Synthesis Activity. 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 21 of 24 Gao, B.-C. (1996). NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 58(3), 257–266. https://doi.org/10.1016/S0034-4257(96)00067-3 Gauci, V., Figueiredo, V., Gedney, N., Pangala, S. R., Stauffer, T., Weedon, G. P., & Enrich-Prast, A. (2022). Non-flooded riparian Amazon trees are a regionally significant methane source. Philosophical Transactions. Series A, Mathematical, Physical, and Engineering Sciences, 380(2215), 20200446. https://doi.org/10.1098/rsta.2020.0446 Goeckede, M. (2020). FLUXNET-CH4 RU-Ch2 Chersky reference [Dataset]. FluxNet; Max Planck Institute for Biogeochemistry. https://doi. org/10.18140/FLX/1669654 Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth engine: Planetary-scale geospatial analysis for everyone. Gregorutti, B., Michel, B., & Saint-Pierre, P. (2017). Correlation and variable importance in random forests. Statistics and Computing, 27(3), 659–678. https://doi.org/10.1007/s11222-016-9646-1 Hall, D. K., & Riggs, G. A. (2016). MODIS/Terra snow cover daily L3 global 500m SIN grid, version 6 [Dataset]. National Snow and Ice Data Center. https://doi.org/10.5067/MODIS/MOD10A1.006 Hargrove, W. W., & Hoffman, F. M. (2004). Potential of multivariate quantitative methods for delineation and visualization of ecoregions. Envi- ronmental Management, 34(1), S39–S60. https://doi.org/10.1007/s00267-003-1084-0 Harriss, R. C., & Matson, P. A. (2009). Biogenic trace gases: Measuring emissions from soil and water. John Wiley & Sons. Helfter, C. (2020a). FLUXNET-CH4 BW-Gum Guma [Dataset]. FluxNet; UK Centre for Ecology and Hydrology. https://doi.org/10.18140/ FLX/1669370 Helfter, C. (2020b). FLUXNET-CH4 BW-Nxr Nxaraga [Dataset]. FluxNet; UK Centre for Ecology and Hydrology. https://doi.org/10.18140/ FLX/1669518 Hemes, K. S., Chamberlain, S. D., Eichelmann, E., Anthony, T., Valach, A., Kasak, K., et al. (2019). Assessing the carbon and climate benefit of restoring degraded agricultural peat soils to managed wetlands. Agricultural and Forest Meteorology, 268(January), 202–214. https://doi. org/10.1016/j.agrformet.2019.01.017 Hengl, T., Mendes de Jesus, J., Heuvelink, G. B. M., Ruiperez Gonzalez, M., Kilibarda, M., Blagotić, A., et al. (2017). SoilGrids250m: Global gridded soil information based on machine learning. PloS One, 12(2), e0169748. https://doi.org/10.1371/journal.pone.0169748 Hinkle, C., & Bracho, R. (2020). FLUXNET-CH4 US-DPW Disney wilderness preserve wetland [Dataset]. FluxNet; University of Central Flor- ida; University of Central Florida (UCF). https://doi.org/10.18140/FLX/1669672 Hoffman, F. M., Kumar, J., Mills, R. T., & Hargrove, W. W. (2013). Representativeness-based sampling network design for the State of Alaska. Landscape Ecology, 28(8), 1567–1586. https://doi.org/10.1007/s10980-013-9902-0 Holm, G., Perez, B., McWhorter, D., Krauss, K., Raynie, R., & Killebrew, C. (2020). FLUXNET-CH4 US-LA2 Salvador WMA freshwater marsh [Dataset]. FluxNet; US Geological Survey; USGS-Wetland and Aquatic Research Center. https://doi.org/10.18140/FLX/1669681 Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002). Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment, 83(1), 195–213. https://doi.org/10.1016/S0034-4257(02)00096-2 IPCC. (2021). Summary for policymakers. In V. Masson-Delmotte, P. Zhai, A. Pirani, S. L. Connors, C. Péan, et al. (Eds.), Climate change 2021: The physical science basis. Contribution of working group I to the sixth assessment report of the intergovernmental panel on climate change (pp. 3–32). Cambridge University Press. https://doi.org/10.1017/9781009157896.001 Irvin, J., Zhou, S., McNicol, G., Lu, F., Liu, V., Fluet-Chouinard, E., et al. (2021). Gap-filling eddy covariance methane fluxes: Comparison of machine learning model predictions and uncertainties at FLUXNET-CH4 wetlands. Agricultural and Forest Meteorology, 308–309, 108528. https://doi.org/10.1016/j.agrformet.2021.108528 Iwata, H., Hirata, R., Takahashi, Y., Miyabara, Y., Itoh, M., & Iizuka, K. (2018). Partitioning eddy-covariance methane fluxes from a shallow lake into diffusive and ebullitive fluxes. Bound. -Layer Meteorol., 169(3), 413–428. https://doi.org/10.1007/s10546-018-0383-1 Iwata, H., Ueyama, M., & Harazono, Y. (2020). FLUXNET-CH4 US-Uaf University of Alaska, Fairbanks [Dataset]. FluxNet; Osaka Prefecture University; Shinshu University. https://doi.org/10.18140/FLX/1669701 Jacob, D. J., Varon, D. J., Cusworth, D. H., Dennison, P. E., Frankenberg, C., Gautam, R., et al. (2022). Quantifying methane emissions from the global scale down to point sources using satellite observations of atmospheric methane. Atmospheric Chemistry and Physics, 22(14), 9617–9646. https://doi.org/10.5194/acp-22-9617-2022 Jacotot, A., Gogo, S., & Laggoun-Défarge, F. (2020). FLUXNET-CH4 FR-LGt La Guette [Dataset]. FluxNet; Observatoire des Sciences de l'Univers en région Centre. https://doi.org/10.18140/FLX/1669641 Jensen, K., & Mcdonald, K. (2019). Surface water microwave product series version 3: A near-real time and 25-year historical global inundated area fraction time series from active and passive microwave remote sensing. IEEE Geoscience and Remote Sensing Letters, 16(9), 1402–1406. https://doi.org/10.1109/LGRS.2019.2898779 Jung, M., Schwalm, C., Migliavacca, M., Walther, S., Camps-Valls, G., Koirala, S., et al. (2020). Scaling carbon fluxes from eddy covariance sites to globe: Synthesis and evaluation of the FLUXCOM approach. Biogeosciences, 17(5), 1343–1365. https://doi.org/10.5194/bg-17-1343-2020 Kirschke, S., Bousquet, P., Ciais, P., Saunois, M., Canadell, J. G., Dlugokencky, E. J., et al. (2013). Three decades of global methane sources and sinks. Nature Geoscience, 6(10), 813–823. https://doi.org/10.1038/ngeo1955 Knox, S. H., Bansal, S., McNicol, G., Schafer, K., Sturtevant, C., Ueyama, M., et al. (2021). Identifying dominant environmental predictors of freshwater wetland methane fluxes across diurnal to seasonal time scales. Global Change Biology, 27(15), 3582–3604. https://doi.org/10.1111/ gcb.15661 Knox, S. H., Jackson, R. B., Poulter, B., McNicol, G., Fluet-Chouinard, E., Zhang, Z., et al. (2019). FLUXNET-CH4 synthesis activity: Objec- tives, observations, and future directions. Bulletin of the American Meteorological Society, 100(12), 2607–2632. https://doi.org/10.1175/ BAMS-D-18-0268.1 Knox, S. H., Sturtevant, C., Matthes, J. H., Koteen, L., Verfaillie, J., & Baldocchi, D. (2015). Agricultural peatland restoration: Effects of land- use change on greenhouse gas (CO2 and CH4) fluxes in the Sacramento-San Joaquin delta. Global Change Biology, 21(2), 750–765. https:// doi.org/10.1111/gcb.12745 Koebsch, F., & Jurasinski, G. (2020). FLUXNET-CH4 DE-Hte Huetelmoor [Dataset]. FluxNet; Landscape Ecology, University of Rostock. https://doi.org/10.18140/FLX/1669634 Krauss, K. W., Raynie, R., Killebrew, C., McWhorter, D. E., Holm, G. O., Jr., & Perez, B. C. (2018). Net ecosystem exchange of CO2 and CH4 from two Louisiana coastal marshes [Dataset]. U.S. Geological Survey. https://doi.org/10.5066/F7MG7NSV Kuhn, M. A., Varner, R. K., Bastviken, D., Crill, P., MacIntyre, S., Turetsky, M., et al. (2021). BAWLD-CH 4: A comprehensive dataset of methane fluxes from boreal and arctic ecosystems. Earth System Science Data, 13(11), 5151–5189. https://doi.org/10.5194/essd-13-5151-2021 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 22 of 24 Lamarque, J.-F., Dentener, F., McConnell, J., Ro, C.-U., Shaw, M., Vet, R., et al. (2013). Multi-model mean nitrogen and sulfur deposition from the Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP): Evaluation of historical and projected future changes [Dataset]. EGU, 13, 7997–8018. https://doi.org/10.5194/acp-13-7997-2013 Lan, X., Thoning, K. W., & Dlugokencky, E. J. (2023). Trends in globally-averaged CH4, N2O, and SF6 determined from NOAA global monitor- ing laboratory measurements. Version 2023-08. https://doi.org/10.15138/P8XG-AA10 Lohila, A., Aurela, M., Tuovinen, J.-P., Laurila, T., Hatakka, J., Rainne, J., & Mäkelä, T. (2020). FLUXNET-CH4 FI-Lom Lompolojankka [Data- set]. FluxNet; Finnish Meteorological Institute. https://doi.org/10.18140/FLX/1669638 Ma, S., Worden, J. R., Bloom, A. A., Zhang, Y., Poulter, B., Cusworth, D. H., et al. (2021). Satellite constraints on the latitudinal distribution and temperature sensitivity of wetland methane emissions. AGU Advances, 2(3). https://doi.org/10.1029/2021av000408 Malone, S. L., Oh, Y., Arndt, K. A., Burba, G., Commane, R., Contosta, A. R., et al. (2022). Gaps in network infrastructure limit our under- standing of biogenic methane emissions for the United States. Biogeosciences, 19(9), 2507–2522. https://doi.org/10.5194/bg-19-2507-2022 Matthes, J., Sturtevant, C., Oikawa, P., Chamberlain, S., Szutu, D., Ortiz, A., et al. (2020). FLUXNET-CH4 US-Myb Mayberry wetland [Dataset]. FluxNet; University of California. https://doi.org/10.18140/FLX/1669685 Melton, J. R., Wania, R., Hodson, E. L., Poulter, B., Ringeval, B., Spahni, R., et al. (2013). Present state of global wetland extent and wetland methane modelling: Conclusions from a model inter-comparison project (WETCHIMP). Biogeosciences, 10(2), 753–788. https://doi. org/10.5194/bg-10-753-2013 Meyer, D., & Buchta, C. (2020). Distance and similarity measures [R package proxy version 0.4-24]. Retrieved from https://CRAN.R-project. org/package=proxy Meyer, H., & Pebesma, E. (2022). Machine learning-based global maps of ecological variables and the challenge of assessing them. Nature Communications, 13(1), 2208. https://doi.org/10.1038/s41467-022-29838-9 Meyer, H., Reudenbach, C., Hengl, T., Katurji, M., & Nauss, T. (2018). Improving performance of spatio-temporal machine learning models using forward feature selection and target-oriented validation. Environmental Modelling & Software, 101, 1–9. https://doi.org/10.1016/j. envsoft.2017.12.001 Meyer, H., Reudenbach, C., Wöllauer, S., & Nauss, T. (2019). Importance of spatial predictor variable selection in machine learning applications -- Moving from data reproduction to spatial prediction. Ecological Modelling, 411, 108815. https://doi.org/10.1016/j.ecolmodel.2019.108815 Mitra, B., Minick, K., Miao, G., Domec, J.-C., Prajapati, P., McNulty, S. G., et al. (2020). Spectral evidence for substrate availability rather than environmental control of methane emissions from a coastal forested wetland. Agricultural and Forest Meteorology, 291, 108062. https://doi. org/10.1016/j.agrformet.2020.108062 Myneni, R., Knyazikhin, Y., & Park, T. (2015). MCD15A2H MODIS/Terra+Aqua leaf area index/FPAR 8-day L4 global 500m SIN grid V006 [Dataset]. NASA EOSDIS Land Processes DAAC. https://doi.org/10.5067/MODIS/MCD15A2H.006 Nilsson, M., & Peichl, M. (2020). FLUXNET-CH4 SE-Deg Degero [Dataset]. FluxNet; Department of Forest Ecology and Management; Swedish University of Agricultural Sciences. https://doi.org/10.18140/FLX/1669659 Nisbet, E. G., Jones, A. E., Pyle, J. A., & Skiba, U. (2022). Rising methane: Is there a methane emergency? Philosophical Transactions. Series A, Mathematical, Physical, and Engineering Sciences, 380(2215), 20210334. https://doi.org/10.1098/rsta.2021.0334 Noormets, A., King, J., Mitra, B., Miao, G., Aguilos, M., Minick, K., et al. (2020). FLUXNET-CH4 US-NC4 NC_AlligatorRiver [Dataset]. FluxNet; Texas A&M University. https://doi.org/10.18140/FLX/1669686 Nzotungicimpaye, C.-M., Zickfeld, K., MacDougall, A. H., Melton, J. R., Treat, C. C., Eby, M., & Lesack, L. F. W. (2021). WETMETH 1.0: A new wetland methane model for implementation in Earth system models. Geoscientific Model Development, 14(10), 6215–6240. https://doi. org/10.5194/gmd-14-6215-2021 Olson, B. (2018). AmeriFlux AmeriFlux US-ALQ Allequash creek site [Dataset]. AmeriFlux; USGS. https://doi.org/10.17190/AMF/1480323 Pangala, S. R., Enrich-Prast, A., Basso, L. S., Peixoto, R. B., Bastviken, D., Hornibrook, E. R. C., et al. (2017). Large emissions from floodplain trees close the Amazon methane budget. Nature, 552(7684), 230–234. https://doi.org/10.1038/nature24639 Parker, R. J., Boesch, H., McNorton, J., Comyn-Platt, E., Gloor, M., Wilson, C., et  al. (2018). Evaluating year-to-year anomalies in tropi- cal wetland methane emissions using satellite CH4 observations. Remote Sensing of Environment, 211, 261–275. https://doi.org/10.1016/j. rse.2018.02.011 Pastorello, G., Trotta, C., Canfora, E., Chu, H., Christianson, D., Cheah, Y.-W., et al. (2020). The FLUXNET2015 dataset and the ONEFlux processing pipeline for eddy covariance data. Scientific Data, 7(1), 225. https://doi.org/10.1038/s41597-020-0534-3 Pekel, J.-F., Cottam, A., Gorelick, N., & Belward, A. S. (2016). High-resolution mapping of global surface water and its long-term changes. Nature, 540(7633), 418–422. https://doi.org/10.1038/nature20584 Peltola, O., Vesala, T., Gao, Y., Räty, O., Alekseychik, P., Aurela, M., et al. (2019). Monthly gridded data product of northern wetland meth- ane emissions based on upscaling eddy covariance observations. Earth System Science Data, 11(3), 1263–1289. https://doi.org/10.5194/ essd-11-1263-2019 Poulter, B., Bousquet, P., Canadell, J. G., Ciais, P., Peregon, A., Saunois, M., et al. (2017). Global wetland contribution to 2000–2012 atmospheric Methane growth rate dynamics. Environmental Research Letters: ERL [Web Site], 12(9), 094013. https://doi.org/10.1088/1748-9326/aa8391 Prigent, C., Jimenez, C., & Bousquet, P. (2020). Satellite-derived global surface water extent and dynamics over the last 25 years (GIEMS-2). Journal of Geophysical Research, 125(3). https://doi.org/10.1029/2019jd030711 Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., & Prabhat (2019). Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743), 195–204. https://doi.org/10.1038/s41586-019-0912-1 Rey-Sanchez, A. C., Morin, T. H., Stefanik, K. C., Wrighton, K., & Bohrer, G. (2018). Determining total emissions and environmental drivers of methane flux in a Lake Erie estuarine marsh. Ecological Engineering, 114, 7–15. https://doi.org/10.1016/j.ecoleng.2017.06.042 Riley, W. J., Subin, Z. M., Lawrence, D. M., Swenson, S. C., Torn, M. S., Meng, L., et al. (2011). Barriers to predicting changes in global terres- trial methane fluxes: Analyses using CLM4Me, a methane biogeochemistry model integrated in CESM. Biogeosciences, 8(7), 1925–1953. https://doi.org/10.5194/bg-8-1925-2011 Roberts, D. R., Bahn, V., Ciuti, S., Boyce, M. S., Elith, J., Guillera-Arroita, G., et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8), 913–929. https://doi.org/10.1111/ecog.02881 Roman, T., Griffis, T., Kolka, R., Wayson, C., Lilleskov, E., Torres, D., et al. (2020). AmeriFlux AmeriFlux PE-QFR Quistococha forest reserve [Dataset]. AmeriFlux; University of Minnesota; USDA-Forest Service-International Programs. https://doi.org/10.17190/AMF/1671889 Roman, T., Kolka, R., Griffis, T., & Deventer, J. (2021). AmeriFlux AmeriFlux US-MBP Marcell bog lake peatland [Dataset]. AmeriFlux; University of Minnesota; USDA- Forest Service. https://doi.org/10.17190/AMF/1767835 Sachs, T., & Wille, C. (2020). FLUXNET-CH4 DE-Zrk Zarnekow [Dataset]. FluxNet; GFZ German Research Centre for Geosciences. https://doi.org/10.18140/FLX/1669636 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 23 of 24 Sakabe, A., Itoh, M., Hirano, T., & Kusin, K. (2020). FLUXNET-CH4 ID-Pag Palangkaraya undrained forest [Dataset]. FluxNet; Hokkaido University; Kyoto University; University of Hyogo; University of Palangkaraya. https://doi.org/10.18140/FLX/1669643 Saunois, M., Stavert, A. R., Poulter, B., Bousquet, P., Canadell, J. G., Jackson, R. B., et al. (2020). The global methane budget 2000--2017. Earth Syst. Sci. Data, 12(3), 1561–1623. https://doi.org/10.5194/essd-12-1561-2020 Schmid, H., & Klatt, J. (2020). FLUXNET-CH4 DE-SfN Schechenfilz Nord [Dataset]. FluxNet; Karlsruhe Institute of Technology, Institute of Meteorology and Climate Research (IMK-IFU). https://doi.org/10.18140/FLX/1669635 Shortt, R., Hemes, K., Szutu, D., Verfaillie, J., & Baldocchi, D. (2020). FLUXNET-CH4 US-Sne Sherman Island restored wetland [Dataset]. FluxNet; University of California. https://doi.org/10.18140/FLX/1669693 Sonnentag, O., & Helbig, M. (2020a). FLUXNET-CH4 CA-SCB Scotty creek bog [Dataset]. FluxNet; Université de Montréal; Wilfrid Laurier University. https://doi.org/10.18140/FLX/1669613 Sonnentag, O., & Helbig, M. (2020b). FLUXNET-CH4 CA-SCC Scotty creek landscape [Dataset]. FluxNet; Université de Montréal; Wilfrid Laurier University. https://doi.org/10.18140/FLX/1669628 Spahni, R., Wania, R., Neef, L., van Weele, M., Pison, I., Bousquet, P., et al. (2011). Constraining global methane emissions and uptake by ecosystems. Biogeosciences, 8(6), 1643–1665. https://doi.org/10.5194/bg-8-1643-2011 Stavert, A. R., Saunois, M., Canadell, J. G., Poulter, B., Jackson, R. B., Regnier, P., et al. (2021). Regional trends and drivers of the global methane budget. Global Change Biology, 28(1), 182–200. https://doi.org/10.1111/gcb.15901 Stell, E., Warner, D., Jian, J., Bond-Lamberty, B., & Vargas, R. (2021). Spatial biases of information influence global estimates of soil respiration: How can we improve global predictions? Global Change Biology, 27(16), 3923–3938. https://doi.org/10.1111/gcb.15666 Sturtevant, C., Ruddell, B. L., Knox, S. H., Verfaillie, J., Matthes, J. H., Oikawa, P. Y., & Baldocchi, D. (2016). Identifying scale-emergent, nonlinear, asynchronous processes of wetland methane exchange. Journal of Geophysical Research: Biogeosciences, 121(1), 188–204. https:// doi.org/10.1002/2015jg003054 Tenuta, M. (2020). AmeriFlux AmeriFlux CA-CF2 Churchill fen site 2 [Dataset]. AmeriFlux; University of Manitoba. https://doi.org/10.17190/ AMF/1634879 Torn, M., & Dengel, S. (2020a). FLUXNET-CH4 US-NGB NGEE Arctic Barrow [Dataset]. FluxNet; Lawrence Berkeley National Laboratory. https://doi.org/10.18140/FLX/1669687 Torn, M., & Dengel, S. (2020b). FLUXNET-CH4 US-NGC NGEE Arctic Council [Dataset]. FluxNet; Berkeley Lab; Lawrence Berkeley National Laboratory. https://doi.org/10.18140/FLX/1669688 Tramontana, G., Jung, M., Schwalm, C. R., Ichii, K., Camps-Valls, G., Ráduly, B., et al. (2016). Predicting carbon dioxide and energy fluxes across global FLUXNET sites with regression algorithms. Biogeosciences, 13(14), 4291–4313. https://doi.org/10.5194/bg-13-4291-2016 Treat, C. C., Bloom, A. A., & Marushchak, M. E. (2018). Nongrowing season methane emissions-a significant component of annual emissions across northern ecosystems. Global Change Biology, 24(8), 3331–3343. https://doi.org/10.1111/gcb.14137 Tuanmu, M.-N., & Jetz, W. (2014). A global 1-km consensus land-cover product for biodiversity and ecosystem modelling: Consensus land cover. Global Ecology and Biogeography: A Journal of Macroecology, 23(9), 1031–1045. https://doi.org/10.1111/geb.12182 Tuanmu, M.-N., & Jetz, W. (2015). A global, remote sensing-based characterization of terrestrial habitat heterogeneity for biodiversity and ecosystem modelling: Global habitat heterogeneity. Global Ecology and Biogeography: A Journal of Macroecology, 24(11), 1329–1339. https://doi.org/10.1111/geb.12365 Turetsky, M. R., Kotowska, A., Bubier, J., Dise, N. B., Crill, P., Hornibrook, E. R. C., et al. (2014). A synthesis of methane emissions from 71 northern, temperate, and subtropical wetlands. Global Change Biology, 20(7), 2183–2197. https://doi.org/10.1111/gcb.12580 Turner, A. J., Frankenberg, C., & Kort, E. A. (2019). Interpreting contemporary trends in atmospheric methane. Proceedings of the National Academy of Sciences of the United States of America, 116(8), 2805–2813. https://doi.org/10.1073/pnas.1814297116 Ueyama, M., Hirano, T., & Kominami, Y. (2020). FLUXNET-CH4 JP-BBY Bibai bog [Dataset]. https://doi.org/10.18140/FLX/1669646. Flux- Net; Osaka Prefecture Univeristy Ueyama, M., Knox, S. H., Delwiche, K. B., Bansal, S., Riley, W. J., Baldocchi, D., et al. (2023). Modeled production, oxidation, and transport processes of wetland methane emissions in temperate, boreal, and Arctic regions. Global Change Biology, 29(8), 2313–2334. https://doi. org/10.1111/gcb.16594 Ueyama, M., Yazaki, T., Hirano, T., & Endo, R. (2022). Partitioning methane flux by the eddy covariance method in a cool temperate bog based on a Bayesian framework. Agricultural and Forest Meteorology, 316, 108852. https://doi.org/10.1016/j.agrformet.2022.108852 Valach, A., Kasak, K., Szutu, D., Verfaillie, J., & Baldocchi, D. (2020). FLUXNET-CH4 US-Tw5 East Pond wetland [Dataset]. FluxNet; Univer- sity of California. https://doi.org/10.18140/FLX/1669699 Valach, A., Szutu, D., Eichelmann, E., Knox, S., Verfaillie, J., & Baldocchi, D. (2020). FLUXNET-CH4 US-Tw1 Twitchell wetland West Pond [Dataset]. FluxNet; University of California. https://doi.org/10.18140/FLX/1669696 Vermote, E. (2015). MOD09A1 MODIS/Terra surface reflectance 8-day L3 global 500m SIN grid V006 [Dataset]. NASA EOSDIS Land Processes DAAC. https://doi.org/10.5067/MODIS/MOD09A1.006 Vesala, T., Tuittila, E.-S., Mammarella, I., & Alekseychik, P. (2020). FLUXNET-CH4 FI-Si2 Siikaneva-2 bog [Dataset]. FluxNet; University of Eastern Finland; University of Helsinki. https://doi.org/10.18140/FLX/1669639 Vesala, T., Tuittila, E.-S., Mammarella, I., & Rinne, J. (2020). FLUXNET-CH4 FI-Sii Siikaneva [Dataset]. FluxNet; University of Eastern Finland; University of Helsinki. https://doi.org/10.18140/FLX/1669640 Villarreal, S., & Vargas, R. (2021). Representativeness of FLUXNET sites across Latin America. Journal of Geophysical Research: Biogeo- sciences, 126(3), e2020JG006090. https://doi.org/10.1029/2020jg006090 Vourlitis, G., Dalmagro, H., de Nogueira, J., Johnson, M., & Arruda, P. (2020). FLUXNET-CH4 BR-Npw northern Pantanal wetland [Dataset]. FluxNet; California State University, San Marcos; Universidade de Cuiabá; Universidade Federal de Mato Grosso; University of British Columbia. https://doi.org/10.18140/FLX/1669368 Wan, Z., Hook, S., & Hulley, G. (2015). MOD11A2 MODIS/Terra land surface temperature/Emissivity 8-day L3 global 1km SIN grid V006 [Dataset]. USGS. https://doi.org/10.5067/MODIS/MOD11A2.006 Whiting, G. J., & Chanton, J. P. (1993). Primary production control of methane emission from wetlands. Nature, 364(6440), 794–795. https:// doi.org/10.1038/364794a0 Wong, G., Melling, L., Tang, A., Aeries, E., Waili, J., Musin, K., et al. (2020). FLUXNET-CH4 MY-MLM Maludam national park [Dataset]. FluxNet; Sarawak Tropical Peat Research Institute. https://doi.org/10.18140/FLX/1669650 Xu, K., Metzger, S., & Desai, A. R. (2017). Upscaling tower-observed turbulent exchange at fine spatio-temporal resolution using environmental response functions. Agricultural and Forest Meteorology, 232, 10–22. https://doi.org/10.1016/j.agrformet.2016.07.019 Zhang, Y., Li, C., Trettin, C. C., Li, H., & Sun, G. (2002). An integrated model of soil, hydrology, and vegetation for carbon dynamics in wetland ecosystems. Global Biogeochemical Cycles, 16(4), 9-1–9-17. https://doi.org/10.1029/2001gb001838 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License AGU Advances MCNICOL ET AL. 10.1029/2023AV000956 24 of 24 Zhang, Z., Fluet-Chouinard, E., Jensen, K., McDonald, K., Hugelius, G., Gumbricht, T., et al. (2021). Development of the global dataset of wetland area and dynamics for methane modeling (WAD2M). Earth System Science Data, 13(5), 2001–2023. https://doi.org/10.5194/ essd-13-2001-2021 Zhang, Z., Poulter, B., Feldman, A. F., Ying, Q., Ciais, P., Peng, S., & Li, X. (2023). Recent intensification of wetland methane feedback. Nature Climate Change, 13(5), 430–433. https://doi.org/10.1038/s41558-023-01629-0 Zhang, Z., Zimmermann, N. E., Stenke, A., Li, X., Hodson, E. L., Zhu, G., et al. (2017). Emerging role of wetland methane emissions in driving 21st century climate change. Proceedings of the National Academy of Sciences of the United States of America, 114(36), 9647–9652. https:// doi.org/10.1073/pnas.1618765114 Zona, D., & Oechel, W. (2020a). FLUXNET-CH4 US-Atq Atqasuk [Dataset]. FluxNet; San Diego State University. https://doi.org/10.18140/ FLX/1669663 Zona, D., & Oechel, W. (2020b). FLUXNET-CH4 US-Beo Barrow environmental observatory (BEO) tower [Dataset]. FluxNet; San Diego State University. https://doi.org/10.18140/FLX/1669664 Zona, D., & Oechel, W. (2020c). FLUXNET-CH4 US-Bes Barrow-Bes (Biocomplexity Experiment South tower) [Dataset]. FluxNet; San Diego State University. https://doi.org/10.18140/FLX/1669665 Zona, D., & Oechel, W. (2020d). FLUXNET-CH4 US-Ivo Ivotuk [Dataset]. FluxNet; San Diego State University. https://doi.org/10.18140/ FLX/1669679 References From the Supporting Information Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A., & Hegewisch, K. C. (2018). TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958-2015. Scientific Data, 5(1), 170191. https://doi.org/10.1038/sdata.2017.191 Amatulli, G., Domisch, S., Tuanmu, M.-N., Parmentier, B., Ranipeta, A., Malczyk, J., & Jetz, W. (2018). A suite of global, cross-scale topo- graphic variables for environmental and biodiversity modeling. Scientific Data, 5(1), 180040. https://doi.org/10.1038/sdata.2018.40 Amatulli, G., McInerney, D., Sethi, T., Strobl, P., & Domisch, S. (2019). Geomorpho90m - global high-resolution geomorphometry layers: Empirical evaluation and accuracy assessment. PeerJ Preprints. https://doi.org/10.7287/peerj.preprints.27595 Behrens, T., Schmidt, K., Viscarra Rossel, R. A., Gries, P., Scholten, T., & MacMillan, R. A. (2018). Spatial modelling with Euclidean distance fields and machine learning. European Journal of Soil Science, 69(5), 757–770. https://doi.org/10.1111/ejss.12687 Kuhn, M. (2020). caret: Classification and regression training. Retrieved from https://CRAN.R-project.org/package=caret Morin, T. H. (2019). Advances in the eddy covariance approach to CH4 monitoring over two and a half decades. Journal of Geophysical Research: Biogeosciences, 124(3), 453–460. https://doi.org/10.1029/2018JG004796 R Core Team. (2022). R: A language and environment for statistical computing. Schroeder, R., McDonald, K. C., Chapman, B. D., Jensen, K., Podest, E., Tessler, Z. D., et al. (2015). Development and evaluation of a multi-year fractional surface water data set derived from active/passive microwave remote sensing data. Remote Sensing, 7(12), 16688–16732. https:// doi.org/10.3390/rs71215843 Sexton, J. O., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., et al. (2013). Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. International Journal of Digital Earth, 6(5), 427–448. https://doi.org/10.1080/17538947.2013.786146 Simard, M., Pinto, N., Fisher, J. B., & Baccini, A. (2011). Mapping forest canopy height globally with spaceborne lidar. Journal of Geophysical Research, 116(4), 1–12. https://doi.org/10.1029/2011JG001708 Wright, M. N., & Ziegler, A. (2017). ranger: A fast implementation of random forests for high dimensional data in C++ and R. Journal of Statis- tical Software, 77(1). https://doi.org/10.18637/jss.v077.i01 Xiao, X., Boles, S., Frolking, S., Salas, W., Moore, B., Li, C., et al. (2002). Observation of flooding and rice transplanting of paddy rice fields at the site to landscape scales in China using VEGETATION sensor data. International Journal of Remote Sensing, 23(15), 3009–3022. https:// doi.org/10.1080/01431160110107734 Zarco-Tejada, P. J., & Ustin, S. L. (2001). Modeling canopy water content for carbon estimates from MODIS data at land EOS validation sites. In IGARSS 2001. Scanning the present and resolving the future. Proceedings. IEEE 2001 international geoscience and remote sensing symposium (Cat. No.01CH37217) (Vol. 1, pp. 342–344). https://doi.org/10.1109/IGARSS.2001.976152 2576604x, 2023, 5, D ow nloaded from https://agupubs.onlinelibrary.w iley.com /doi/10.1029/2023A V 000956 by Luonnonvarakeskus, W iley O nline Library on [02/01/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on W iley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License