FLUXNET-CH4: A global, multi-ecosystem dataset and analysis 1 of methane seasonality from freshwater wetlands 2

. Methane (CH 4 ) emissions from natural landscapes constitute roughly half of global CH 4 contributions to 137 the atmosphere, yet large uncertainties remain in the absolute magnitude and the seasonality of emission quantities 138 and drivers. Eddy covariance (EC) measurements of CH 4 flux are ideal for constraining ecosystem-scale CH 4 139 emissions due to quasi-continuous and high temporal resolution of CH 4 flux measurements, coincident carbon dioxide, 140 water, and energy flux measurements, lack of ecosystem disturbance, and increased availability of datasets over the 141 last decade. Here, we 1) describe the newly published dataset, FLUXNET-CH4 Version 1.0, the first, open source 142 global dataset of CH 4 EC measurements (available at https://fluxnet.org/data/fluxnet-ch4-community-product/). 143 FLUXNET-CH4 includes half-hourly and daily gap-filled and non gap-filled aggregated CH 4 fluxes and meteorological data from 79 sites globally: 42 freshwater wetlands, 6 brackish and saline wetlands, 7 formerly drained ecosystems, 7 rice paddy sites, 2 lakes, and 15 uplands. Then, we 2) evaluate FLUXNET-CH4 representativeness for 146 freshwater wetland coverage globally, because the majority of sites in FLUXNET-CH4 Version 1.0 are freshwater 147 wetlands which are a substantial source of total atmospheric CH 4 emissions; and 3) provide the first global estimates 148 of the seasonal variability and seasonality predictors of freshwater wetland CH 4 fluxes. Our representativeness analysis suggests that the freshwater wetland sites in the dataset cover global wetland bioclimatic attributes (encompassing energy, moisture, and vegetation-related parameters) in arctic, boreal, and temperate regions, but only sparsely cover humid tropical regions. Seasonality metrics of wetland CH 4 emissions vary considerably across latitudinal bands. In freshwater wetlands (except those between 20° S to 20° N) the spring onset of elevated CH 4 emissions starts three days earlier, and the CH 4 emission season lasts 4 days longer, for each degree C increase in mean annual air temperature. On average, the spring onset of increasing CH 4 emissions lags soil warming by one month, with very few sites experiencing increased CH 4 emissions prior to the onset of soil warming. In contrast, roughly half of these sites experience the spring onset of rising CH 4 emissions prior to the spring increase in gross primary productivity 157 (GPP). The timing of peak summer CH 4 emissions does not correlate with the timing for either peak summer 158 temperature or peak GPP. Our results provide seasonality parameters for CH 4 modeling, and highlight seasonality 159 metrics that cannot be predicted by temperature or GPP (i.e., seasonality of CH 4 peak). FLUXNET-CH4 is a powerful new resource for diagnosing and understanding the role of terrestrial ecosystems and climate drivers in the global CH 4 161 cycle; and future additions of sites in tropical ecosystems and site-years of data collection will provide added value to 162 this database. All seasonality parameters are available at https://doi.org/10.5281/zenodo.4672601. Additionally, raw 163 FLUXNET-CH4 data used to extract seasonality parameters can be downloaded from https://fluxnet.org/data/fluxnet- ch4-community-product/, and a complete list of the 79 individual site data DOIs is provided in Table 2 in the Data 165 Availability section of this document. or GPP individual latitudinal

days earlier, and the CH4 emission season lasts 4 days longer, for each degree C increase in mean annual air 154 temperature. On average, the spring onset of increasing CH4 emissions lags soil warming by one month, with very 155 few sites experiencing increased CH4 emissions prior to the onset of soil warming. In contrast, roughly half of these 156 sites experience the spring onset of rising CH4 emissions prior to the spring increase in gross primary productivity 157 (GPP). The timing of peak summer CH4 emissions does not correlate with the timing for either peak summer 158 temperature or peak GPP. Our results provide seasonality parameters for CH4 modeling, and highlight seasonality 159 metrics that cannot be predicted by temperature or GPP (i.e., seasonality of CH4 peak). FLUXNET-CH4 is a powerful 160 new resource for diagnosing and understanding the role of terrestrial ecosystems and climate drivers in the global CH4 161 cycle; and future additions of sites in tropical ecosystems and site-years of data collection will provide added value to 162 this database. All seasonality parameters are available at https://doi.org/10.5281/zenodo.4672601. Additionally, raw 163 FLUXNET-CH4 data used to extract seasonality parameters can be downloaded from https://fluxnet.org/data/fluxnet-164 ch4-community-product/, and a complete list of the 79 individual site data DOIs is provided in Table 2   year time scale (Myhre et al., 2013), and its atmospheric concentration has increased by >1000 ppb since 1800 174 (Etheridge et al., 1998). While atmospheric CH4 concentrations are substantially lower than those of CO2, CH4 has 175 contributed 20-25% as much radiative forcing as CO2 since 1750 (Etminan et al., 2016). Despite its importance to

183
Tower-based eddy covariance (EC) measurements provide ecosystem-scale CH4 fluxes at high temporal 184 resolution across years, are coupled with measurements of key CH4 drivers such as temperature, water and recent 185 substrate input (inferred from CO2 flux), and thus help constrain bottom-up CH4 budgets and improve CH4 predictions.

186
Although EC towers began measuring CO2 fluxes in the late 1970s (Desjardins 1974; Anderson et al., 1984), and 187 some towers began measuring CH4 in the 1990s (Verma et al., 1992), most CH4 flux EC measurements began within 188 the last decade (2010s). Given that many EC CH4 sites are relatively new, the flux community has only recently 189 compiled them for global synthesis efforts (e.g., Chang et al., 2021in press) and is still working to standardize CH4  Broad-scale wetland CH4 seasonality estimates, such as when fluxes increase, peak, and decrease and the 198 predictors of seasonality, remain relatively unconstrained across wetlands globally. These key seasonality metrics 199 vary considerably across high-emitting systems such as wetlands and other aquatic systems (Desjardins, 1974;Dise,

208
CH4 flux is also driven by inundation depth since anoxic conditions are typically necessary for methanogenesis (Lai,  wetland CH4 seasonality globally remain a knowledge gap that high-frequency and long-term EC data can help fill.

219
Here, we first describe Version 1.0 of the FLUXNET-CH4 dataset (available at 220 https://fluxnet.org/data/fluxnet-ch4-community-product/). Version 1.0 of the dataset expands and formalizes the 221 publication of data scattered among regional flux networks as described previously in Knox et al. (2019). FLUXNET-

222
CH4 includes half-hourly and daily gap-filled and non gap-filled aggregated CH4 fluxes and meteorological data from 223 79 sites globally: 42 freshwater wetlands, 6 brackish and saline wetlands, 7 formerly drained ecosystems, 7 rice paddy 224 sites, 2 lakes, and 15 upland ecosystems. FLUXNET-CH4 includes an additional 2 wetland sites (RU-Vrk and SE-225 St1), but they are not available under the CC BY 4.0 data policy and thus are excluded from this analysis. Since the 226 majority of sites in FLUXNET-CH4 Version 1.0 (hereafter referred to solely as "FLUXNET-CH4") are freshwater 227 wetlands, which are a substantial source of total atmospheric CH4 emissions, we use the subset of data from freshwater 228 wetlands to evaluate the representativeness of freshwater wetland coverage in the FLUXNET-CH4 dataset relative to 229 wetlands globally, and provide the first assessment of global variability and predictors of freshwater wetland CH4 flux 230 seasonality. We quantify a suite of CH4 seasonality metrics and evaluate temperature and GPP (a proxy for recent 231 substrate input) as predictors of seasonality across four latitudinal bands (northern, temperate, subtropical, and 232 tropical). Due to a lack of high-temporal resolution water

243
ETC) to avoid duplication of efforts, as most sites are part of different regional networks (albeit with different data 244 products). We collected and standardized data for FLUXNET-CH4 with assistance from the regional flux networks,

245
AmeriFlux's "Year of Methane", FLUXNET, the EU's Readiness of ICOS for Necessities of Integrated Global 246 Observations (RINGO) project, and a U.S. Geological Survey Powell Center working group. FLUXNET-CH4 is a 247 community-led project, so while we developed it with assistance from FLUXNET, we do not necessarily use standard 248 FLUXNET data variables, formats, or methods.

256
Data gaps of CH4 flux were filled using artificial neural network (ANN) methods first described in Knox et al. (2015) 257 and in Knox et al. (2019), and summarized here in Sect. 2.1.2. Gap-filled data for gaps exceeding two months are 258 provided and flagged for quality. Please see Table B1 for variable description and units, as well as quality flag 259 information. For the seasonality analysis in this paper we excluded data from gaps exceeding two months, and we 260 encourage future users of FLUXNET-CH4 to critically evaluate gap-filled values from long data gaps before including 261 them in analyses (Dengel et al., 2013;Kim et al., 2020).

262
In addition to half-hourly data, the FLUXNET-CH4 Version 1.0 release also contains a full set of daily mean 263 values for all parameters except wind direction and precipitation. Daily precipitation is included as the daily sum of 264 the half-hourly data, and daily average wind direction is not included. 265 2.1.2 Gap-filling methods and uncertainty estimates 266 As described in Knox et al. (2015) and in Knox et al. (2019), the ANN routine used to gap-fill the CH4 data 267 was optimized for generalizability and representativeness. To avoid biasing the ANN toward environmental conditions 268 with typically better data coverage (e.g., summer-time and daytime measurements), the explanatory data were divided 269 into a maximum of 15 clusters using a k-means clustering algorithm. Data used to train, test, and validate the ANN 270 were proportionally sampled from these clusters. For generalizability, the simplest ANN architecture with good 271 performance (<5% gain in model accuracy for additional increases in architecture complexity) was selected for 20 272 extractions of the training, test, and validation data. Within each extraction, each tested ANN architecture was 273 reinitialized 10 times, and the initialization with the lowest root-mean-square-error was selected to avoid local minima.

274
The median of the 20 predictions was used to fill each gap. A standard set of variables available across all sites was 275 used to gap-fill CH4 fluxes (Dengel et al., 2013), which included the previously mentioned meteorological variables 276 TA, SW_IN, WS, PA, and sine and cosine functions to represent seasonality. These meteorological variables were 277 selected for their relevance to CH4 exchange and were gap-filled using the ERA-I reanalysis data. Other variables 278 related to CH4 flux (e.g., water table depth [WTD] and soil temperature [TS]) were not included as explanatory 279 variables as they were not available across all sites or had large gaps that could not be filled using the ERA-I reanalysis

281
While the median of the 20 predictions was used to fill each gap, the spread of the predictions was used to 282 provide a measure of uncertainty resulting from the ANN gap-filling procedure. Specifically, the combined annual   Table B2. Throughout this paper, we include 296 uncertainties on individual site years when discussing single years of data. In sites with multiple years of data, we 297 report the standard deviation of the multiple years.

323
In addition to the variable description table (Table B1) and the site metadata (Table B3), we provide several 324 more tables to complement our analysis. Table B4 includes the climatic data used in the representativeness analysis.

325
Table 5 provides seasonality parameters for CH4 flux, air temperature, soil temperature (from the probe closest to the 326 ground surface), and GPP. For sites with multiple soil temperature probes, the full set of soil temperature parameters 327 are in Table B6.    Table B4). We use 357 EVI because it is a more direct measurement than GPP from global gridded products and is considered a reasonable 358 proxy for GPP (Sims et al., 2006). Together, these environmental variables account for, or are, proxies for key controls 359 of CH4 production, oxidation at the surface, and transport (Bridgham et al., 2013). We use a principal components   To further identify geographical gaps in the coverage of the FLUXNET-CH4 Version 1.0 network, we 366 quantified the dissimilarity of global wetlands from the tower network, using a similar approach to that taken for CO2 367 flux towers (Meyer and Pebesma 2020). We calculated the 4-dimensional Euclidean distance from the four bioclimatic 368 variables between every point at the land surface to every tower location at the FLUXNET-CH4 network. We then 369 divided these distances by the average distance between towers to produce a dissimilarity index. Dissimilarity scores 370 <1 represent areas whose nearest tower is closer than the average distance among towers, while areas with scores >1 371 are more distant. Lastly, we identified the importance of an individual tower in the network by estimating the 372 geographical area to which it is most analogous in bioclimate space. We divided the world's land surface according 373 to closest towers in bioclimatic space. The area to which each tower is nearest is defined as the tower's constituency.  (Fig. 1). We also calculate parameters such as amplitude (peak flux -380 baseline, which is the average of spring and fall baselines; ("e" -(("a" + "b")/2) in Fig. 1), and relative peak timing ( 381 ( "g" -"f" ) / ("h" -"f") in Fig. 1). Timesat uses a double-logistic fitting function to create a series of localized fits 382 centered on data minima and maxima. Localized fits are determined by minimizing minimized using a merit 383 function and with the Levenberg-Marquardt method (Madsen et al., 2004;Nielsen, 1999). These localized fits are 384 then merged using a global function to create a smooth fit over the full time interval. To fit CH4 time-series in 385 Timesat, we used gap-filled data after removing gaps exceeding two months. We do not report Timesat parameters 386 when large gaps occur during CH4 emissions spring increase, peak, or fall decrease.

387
We estimate 'start of elevated emissions season' when CH4 emissions begin to increase in the spring ( "f"

398
We also used Timesat to extract seasonality metrics for GPP, partitioned using the daytime-based approach

421
We regressed the CH4 seasonality parameters from Timesat against annual temperature, annual water table 422 depth, and Timesat seasonality parameters for air temperature, soil temperature, and GPP (proxy for recent carbon 423 input available as substrate) using linear mixed-effect modeling with the lmer command (with site as a random

436
We also compared quarterly CH4 flux sums by dividing data into quarterly periods:   crop sites (excluding rice), two alpine meadows, one grassland, one mixed forest, one tundra, and one urban site. The 459 drained sites represent former wetlands that have been artificially drained for use as grasslands (n = 3) or croplands 460 (n = 3). FLUXNET-CH4 sites span the globe, though are concentrated in North America and Europe (Fig. 3). Table   461 B3 includes characteristics of all sites in the dataset.

467
Sites represent a range of ecosystem types, latitudes, median fluxes, and seasonality patterns (Table 1).

468
Across all FLUXNET-CH4 sites (including non-wetland sites), mean average annual CH4 flux is positively skewed 469 with a median flux of 9.5 g C m -2 yr -1 , a mean flux of 16.9 g C m -2 yr -1 , and numerous annual CH4 fluxes exceeding 470 60 g C m -2 yr -1 . Marshes and swamps have the highest median flux, and upland, salt marsh, and tundra sites have 471 the lowest (Fig. 4). Lake emissions are highly variable due to one high-flux lake site (JP-SWL). Flux data at many 472 sites show strong seasonality in CH4 emissions, but data coverage is also lower outside the growing season (Table   473 1). Data coverage is lowest during the JFM quarter (on average 20% of half-hourly time periods contain flux data) 474 reflecting the predominance of Northern hemisphere sites and the practical difficulties in maintaining EC tower sites 475 during colder winter months (Table 1). Bogs, fens, and marshes have pronounced seasonality, with fluxes being 476 highest in the AMJ and JAS quarters. In contrast, CH4 fluxes from uplands, drained sites, and salt marshes are more 477 uniform and low year-round.   bogs, fens, wet tundra, marshes, and swamps, and a range of annual CH4 emission rates (Fig. 4). The majority of 487 freshwater wetlands in our dataset emit 0-20 g C m -2 yr -1 , with 10 emitting 20-60 g C m -2 yr -1 , and one more than 60 g 488 C m -2 yr -1 . Differences in annual CH4 flux among wetland types is partially driven by temperature (which is often 489 linked to site type), with mean annual air temperature explaining 51% of the variance between sites (Fig. 5,   annual CH4 flux of -0.1 ± 0.1 g C m -2 yr -1 (see Table B3 for site acronyms and metadata). The average agricultural site 513 emissions are 1.3 ± 0.8 g C m -2 yr -1 and non-agricultural site emissions are 1.6 ± 1.2 g C m -2 yr -1 across sites.

514
Rice sites (n = 7) have average annual emissions across all sites of 16.7 ± 7.7 g C m -2 yr -1 and are characterized    Table B2 for individual site-year CH4 flux sums and associated uncertainty) and minimal 532 seasonality. Two other FLUXNET-CH4 saltwater sites (US-La1 and US-StJ) have significantly higher fluxes, with 533 annual sums of 12.6 ± 0.6 and 9.6 ± 1.0 g C m -2 yr -1 , respectively, while the mangrove site HK-MPM has annual mean 534 fluxes of 11.1 ± 0.5 g C m -2 yr -1 . This range of CH4 fluxes across different saltwater ecosystems could be valuable for  Table B4). We exclude wetlands classified as

554
When considering the four bioclimatic variables, MAT, LE, EVI and SRWI in a PCA, we found that our 555 tower network generally samples the bioclimatic conditions of global wetland cover, but some noticeable gaps remain 556 (Fig. 6). Three clusters of the world's wetland-dense regions are identified, but are not equally sampled by the network. tropical) (Fig. 8). Annual CH4 fluxes for temperate, and subtropical sites were significantly higher than for northern 634 sites (8.7 ± 5.0, 29.7 ± 25.2, 40.1 ± 14.6, and 24.5 ± 20.7 g C m -2 yr -1 for northern, temperate, subtropical, and tropical,

667
We found that latitudinal groups showed strong differences in absolute CH4 flux across quarters, and narrower 668 differences in percentage of annual CH4 flux (Fig. 9a versus 9b). Thus, the AMJ quarter had a similar relative The start of the elevated CH4 flux season, and how long the elevated flux season lasts, correlated strongly 692 with mean annual air temperature ( Fig. 10; p<0.0001 for each). Methane flux began to increase roughly two months 693 earlier in the warmest systems (mean annual temperature > 20 °C) compared to the coldest (mean annual 694 temperature near -10 °C), though several of the warmer sites had high variability. Our data suggest that the CH4 695 season started 2.8 ± 0.5 days earlier for every degree Celsius increase in mean annual temperature (Fig. 10a). In 696 contrast, the end of the CH4 emission season was not correlated with mean annual temperature, but a positive trend 697 existed despite high variability in warmest and coldest sites (Fig. 10b). The high variability seen in the end of CH4 704 emissions than the coldest sites (Fig. 10c). CH4 season length increased 3.6 ± 0.6 days for every degree Celsius 705 increase in mean annual temperature (note that these relationships are correlations, and we cannot disentangle 706 causality with this analysis). Temperature is highly correlated with other parameters (i.e., radiation, days of snow 707 cover, etc.), so CH4 flux is also likely to correlate with other environmental parameters.

713
Although the spring onset of increasing CH4 emissions correlated with mean annual air temperature, on 714 average it lagged the spring increase in the shallowest soil temperatures by 31 ± 40 days (Fig. 11, lag is significantly 715 different than zero, p < 0.001), with very few instances of CH4 emissions beginning before seasonal soil 716 temperatures increase (and by 20 ± 50 days for the deepest temperature probes). In contrast, for roughly half of the 717 sites, CH4 emission increased prior to seasonal GPP (a proxy for fresh substrate availability) increases. This

718
suggests that the initiation of increased CH4 fluxes at the beginning of the season was not limited by availability of 719 substrate derived from recent photosynthate. Additionally, the onset of CH4 fluxes tended to occur closer to the 720 onset of soil temperature increase for cooler temperature sites (sites with later start dates tend to be cooler; Fig. 11a).

721
This result is likely attributable to the direct influence of increased temperature on microbial processes (Chadburn et

733
In contrast with the CH4 season-start timing, the timing of the CH4 peak did not correlate with either the 734 timing of the soil temperature peak or the GPP peak (Fig. A1). For 63% of the sites, the average timing of peak CH4 735 emissions lagged the soil temperature peak, and at 83% of the sites average peak CH4 lagged peak GPP (Fig. A1).

736
Although there was no simple relationship between absolute CH4 peak timing and the environmental drivers we 737 investigated, there was a correlation (p = 0.0005) between the relative timing of peak CH4 compared to season onset 738 (calculated as described in Section 2.3) and mean annual air temperature (Fig. 12a). For cooler sites, the peak of

840
Eleonora Canfora, and Carlo Trotta did the data collection, curation, and pre-processing for all of the sites outside

841
North and South America. Remaining co-authors contributed eddy-covariance data to FLUXNET-CH4 dataset 842 and/or participated in editing the manuscript.

844
The authors declare that they have no conflict of interest.