Global wetland contribution to 2000–2012 atmospheric methane growth rate dynamics

To reduce the uncertainty for wetland area dynamics resulting from predictive modeling approaches such as TOPMODEL (Gedney and Cox 2003), we combined remote sensing and inventory data to develop a monthly global wetland area dataset. Current satellite remote sensing of wetlands uses coarse-spatial resolution passive and active microwave sensors, ~25 km², that observe surface water generally not obscured by vegetation (Bohn et al 2015, Schroeder et al 2015). This includes open water (e.g. lakes, rivers and ocean) as well as surface inundated wetlands comprising mainly of open plant canopies, and thus excludes exposed wetlands with no observable surface flooding as well as surface inundated wetlands beneath closed (forest) canopies. Consequently, whereas ground-based wetland inventories estimate between 8.2 and 10.1 Mkm² of wetlands globally (Lehner and Doll 2004), remote sensing surface water estimates of wetlands are far lower, i.e. ~6.5 Mkm² excluding coastal grid regions (Schroeder et al 2015). To develop a comprehensive wetland dynamics dataset, we integrated the Global Lakes and Wetlands Dataset, or GLWD (Lehner and Doll 2004), with the seasonal cycle of surface water inundation from the Surface WAter Microwave Product Series (SWAMPS; Schroeder et al 2015).

The SWAMPS dataset maps fractional surface water dynamics using remote sensing data from multiple passive and active microwave satellite missions using a 28 day screening procedure to mask snow, ice cover, and melting snow. In our analysis, SWAMPS surface inundation was derived from the Special Sensor Microwave Imager version 1 and 2 (SSMI v1/v2), SeaWinds-on-QuickSCAT (QSCAT) from January 2000 to October 2008, and from the European Space Agency Advanced Scatterometer (ASCAT) from November 2008 to December 2012. Land-cover data from MOD12Q1 V004 (Friedl et al 2010) was used to exclude permanent open water (water bodies, rivers, snow/ice) and thus avoid double counting of wetlands, with an additional global FAO land mask (Zobler 1986) applied to remove coastal grid cells where brackish and salt-water wetlands were not considered as a source of methane. The monthly SWAMPS dataset ('fw_28_swe') was re-projected from its native 0.25° EASE grid to a geographic 0.5° rectilinear grid (WGS84) using a conservative remapping interpolation to preserve the original wetland area.

The GLWD Level 3 dataset was first reclassified to remove Classes 1–3, lakes, reservoirs and rivers, and then aggregated by summing wetland area to 0.5° from 30 arc-second resolution. GLWD is commonly used as a benchmark for various remote sensing and wetland mapping activities because it incorporates the highest-quality country-level inventory coverage of wetlands (Peregon et al 2008, Fluet-Chouinard et al 2014, Bohn et al 2015). The integration of SWAMPS and GLWD took place in three phases; first the maximum annual surface water fraction at the per-pixel level (FwMax_x,y ) for the 2000–2012 period was compared with GLWD (GLWDmax_x,y ) to estimate the relative SWAMPS detection bias (FwMaxCor_x,y ).

$$\begin{equation} {\rm FwMaxCor}_{x,y} = \frac{{{\rm GLWD}_{x,y} }}{{{\rm FwMax}_{x,y} }} \end{equation} \tag{ 1 }$$

Second, FwMax_x,y from SWAMPS was adjusted using the FwMaxCor_x,y correction factor from equation (1) so that the maximum surface-water fraction from SWAMPS matched the GLWD estimate equation (2). For areas approximately northwest of the Hudson Bay Lowlands, the GLWD classifies the entire region as 'lakes', and so in cases where the merging SWAMPS-GLWD resulted in lower wetland area, the original SWAMPS surface-water values were used equation (3). Seasonal wetlands in desert systems, mapped in the GLWD, were retained in the SWAMPS-GLWD product.

$$\begin{equation} {\rm FwMaxGLWD}_{x,y} = {\rm FwMaxCor}_{x,y} {\rm FwMax}_{x,y} \end{equation} \tag{ 2 }$$

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\rm FwMaxGLWD}_{x,y} & = & {\rm if}\;({\rm FwMaxGLWD}_{x,y}\\ & < & {\rm FwMax}_{x,y} ,{\rm FwMax}_{x,y} ) \end{array} \end{equation} \tag{ 3 }$$

Third, the original monthly SWAMPS surface-inundation (Fw_x,y,m ) was rescaled equation (4) for each year as a fraction of that same year's (uncorrected) maximum inundation, resulting in a unique monthly scalar (0–1) for each year (FwScalar_x,y,m ). Lastly, FwMaxCor_x,y was multiplied by the annual fractional inundation cycle, FwScalar_x,y,m (equation 5).

$$\begin{equation} {\rm FwScalar}_{x,y,m} = \frac{{{\rm Fw}_{x,y,m} }}{{{\rm FwMax}_{x,y,m} }}\;{\rm where}\;{\rm m} = 1..12 \end{equation} \tag{ 4 }$$

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\rm FwCor}_{x,y,t} & = & {\rm FwMaxGLWD}_{x,y} {\rm FwScalar}_{x,y,t}\\ & & \;{\rm where}\;t = 1..{\rm all}\;{\rm months} \end{array} \end{equation} \tag{ 5 }$$

The adjusted SWAMPS-GLWD product results in a maximum wetland area of 10.5 Mkm², and in agreement with the GLWD and other studies (Fluet-Chouinard et al 2014), but the product also maintains the seasonal cycle and inter-annual trends of inundation mapped by SWAMPS. Key wetland areas are retained in the SWAMPS-GLWD in areas such as Amazonia, the Congo Basin, and the Western Siberian Lowlands, which in previous studies have been poorly represented (Bohn et al 2015).

We also conducted a sensitivity test to account for how differences in view angle between the QSCAT and ASCAT instruments might influence trends in surface inundation. SWAMPS accounted for changes in the angle-of-incidence between sensors by applying a time-averaged normalization approach to the backscatter retrievals (Schroeder et al 2015), however sensor-based offsets in grid cells with low surface inundation may affect the trends. We removed low surface-inundation grid cells (defined by their maximum annual value) using a per-pixel threshold of 0.5%, 1%, 2.5% and 5%, and compared the change in methane emissions for each scenario with no filter applied.

A common simulation protocol was followed by each of the wetland modeling teams (listed in table S1 available at http://stacks.iop.org/ERL/12/094013/mmedia) using standardized climate, atmospheric CO₂ and dynamic wetland area data (used to map CH₄ producing regions), and also to specify boundary conditions for model spin-up and transient runs. CRU-NCEP v4.0 was used as the meteorology data, which includes long and shortwave radiation, air pressure, specific humidity, total precipitation, air temperature, and wind speed and direction. CRU-NCEP v4.0 combines the higher spatial resolution of CRU TS3.22 (Harris et al 2013) with the higher temporal resolution from the NCEP Reanalysis product (Kanamitsu et al 2002), to produce a meteorological forcing dataset that covers years 1901–2012 at 6 hourly temporal and 0.5 degree spatial resolution, and is used as the climate driver for biogeochemical models included in the annual Global Carbon Project CO₂ budget (Le Quéré et al 2015). Global atmospheric CO₂ concentrations were provided at an annual resolution for 1860–2012, with data prior to 1958 from ice cores (Joos and Spahni 2008) and after 1958 from the average of NOAA measurements at Mauna Loa (MLO) and the South Pole (SPO) stations.

Models were run to equilibrium during a spin-up phase where the first thirty years of climate data, 1901–1930, were recycled with pre-industrial CO₂ concentrations ~276 ppm). Soil texture data was prescribed using model-specific global soil databases such as the Harmonized World Soils Database (FAO/IIASA/ISRIC/ISSCAS/JRC 2012) and with pedo-transfer functions (i.e. Cosby et al 1984) to determine water-holding capacity. Land use (i.e. agriculture or pasture) and land-cover change were not simulated, and CH₄ emissions by fire excluded from our analysis. Modeling groups used their default vegetation distributions determined by either a dynamic vegetation model or by prescribed satellite vegetation products (Poulter et al 2015).

Atmospheric observations of CH₄ were accessed from the NOAA ESRL cooperative air sampling network (Dlugokencky et al 1994). We carried out a comparison of wetland CH₄ emissions with atmospheric growth rate data from surface flask measurements at MLO and with the globally averaged marine surface annual mean dataset, which uses selected sites representative of a well-mixed marine boundary layer. In addition to anthropogenic contributions, the growth rate at MLO integrates terrestrial flux processes and has been demonstrated to be useful as a representative station for diagnosing CO₂ and CH₄ exchange between the biosphere and atmosphere (Fung et al 1991, Dlugokencky et al 1995, Bousquet et al 2006, Wang et al 2014, Meng et al 2015). Annual mean CH₄ concentrations from 2000–2012 were first detrended, removing the long-term increase in CH₄ concentrations following 2006, because we were interested in evaluating the role of interannual climate variability on CH₄ emissions (assuming minimal variability in OH at interannual timescales) and then using a conversion of 2.78 Tg CH₄ per ppb to estimate changes in the 'atmospheric burden' (Fung et al 1991). The interannual variability of CH₄ concentrations and emissions was then calculated as Y_i +Y_(i+1) (where Y = year). As could be expected, the variability in the MLO observations was more highly correlated with wetland CH₄ emission variability than with the globally averaged observations, where the averaging across multiple, mainly marine, stations across latitudes partly dampens the contribution from land to inter-annual variability. Thus, in the following, only the MLO observations are used to discuss the contribution of the modeled fluxes to atmospheric variability.

For the stabilization period (2000–2006), global wetland CH₄ emissions were estimated at 184 ± 21 Tg CH₄ yr⁻¹, where the uncertainty is estimated as one standard deviation of the model ensemble mean. Wetland emissions remained statistically similar during the period of renewed growth (2007–2012) at 183 ± 23 Tg CH₄ yr⁻¹, with a slightly larger value in the last year of analysis, 2012, of 186 ± 23 Tg CH₄ yr⁻¹ (table 1). For both time periods, tropical biomes, defined in figure 1(a) as regions 7 to 9, dominated the global flux with representative 2012 emissions, for example, of 69 ± 12 Tg CH₄ yr⁻¹, followed by boreal (44 ± 19 Tg CH₄ yr⁻¹), temperate (44 ± 10 Tg CH₄ yr⁻¹), and sub-tropical biomes (20 ± 5 Tg CH₄ yr⁻¹). The global fluxes are consistent with a range of previously published estimates using satellite based approaches, i.e. 170 Tg CH₄ yr⁻¹ (Bloom et al 2010), atmospheric inversions, i.e. 149–159 Tg CH₄ yr⁻¹ (Ghosh et al 2015) and 165 ± 9 Tg CH₄ yr⁻¹ (Bousquet et al 2011), process-based models, i.e. 190 ± 39 Tg CH₄ yr⁻¹ (Melton et al 2013) and a combination of inversion and process-model, i.e. 172 Tg CH₄ yr⁻¹ (Spahni et al 2011).

Zoom In Zoom Out Reset image size

Between the two time periods (the 2000–2006 'stabilization' and the 2007–2012 'renewed growth' periods), no statistically significant change in the average model ensemble emissions (two-sided Student's t-test; α = 0.1) was found at the global scale or regionally (figure 2(a) and figure 3(a)). Among individual models, the change in global emissions between 2000–2006 and 2007–2012 ranged from a 5.4 Tg CH₄ yr⁻¹ decrease for ORCHIDEE to an increase of 4.8 Tg CH₄ yr⁻¹ for LPJ-MPI, with an ensemble average change of −0.5 ± 0.9 Tg CH₄ yr⁻¹ (figure 1(b)). At the regional scale (figure 1(b)), an increase in boreal wetland CH₄ emissions of 1.2 ± 0.3 Tg CH₄ yr⁻¹ was found for the ensemble, with only CLM4.5 estimating reduced emissions of 0.2 Tg CH₄ yr⁻¹ and with LPJ-MPI providing the largest increase of 3.5 Tg CH₄ yr⁻¹. Six of the individual models had statistically significant increasing trends for boreal CH₄ emissions (linear regression, p < 0.1) and of these six, all models agreed with an increase in emissions occurring for June−August (JJA) and September−November (SON). A decrease in tropical emissions of 0.9 ± 0.3 Tg CH₄ yr⁻¹ between 2000–2006 and 2007–2012 was observed across the model ensemble, with just two models estimating an increase (CLM4.5 and LPJ-MPI, 0.1 and 1.1 Tg CH₄ yr⁻¹, respectively) and ORCHIDEE estimating the largest decrease of 3.1 Tg CH₄ yr⁻¹. Four of the individual models had a statistically significant (p < 0.1) decrease in tropical CH₄ emissions between JJA and SON. Changes in tropical wetland emissions were sensitive to the filtering of the low-surface inundated wetlands, carried out to detect inter-sensor bias, but no statistically significant change was detected (table S3). A decrease in temperate regional emissions of 1.4 ± 0.4 Tg CH₄ yr⁻¹ and almost no change in semi-arid emissions (0.4 ± 0.2 Tg CH₄ yr⁻¹) was also obtained for the ensemble, but with a larger spread across models than for the boreal and tropical regions.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

During the 2000–2012 period, a linear regression analysis with climate forcing based on the Climate Research Unit, CRU TS3.22 (Harris et al 2013), masked to match wetland containing grid cells only (as an average over the 2000–2012 time period), revealed variable spatial and seasonal trends in precipitation and air temperature (for annual trends, see figure 3(c) and (d). Total global December–February (DJF) precipitation increased by 2.5 mm yr⁻¹ (p < 0.05) but did not change significantly in other seasons (figure 3(c)). Increasing boreal winter (DJF) precipitation contributed to about half of the annual global increase, 1.4 mm yr⁻¹ (p = 0.1), with other boreal seasons showing no change, tropical DJF precipitation increased by 8.2 mm yr⁻¹ (p = 0.01), and semi-arid DJF precipitation increased by 1.8 mm yr⁻¹ (p = 0.04). Global annual air temperature over wetlands was nearly constant (figure 3(d)), and increased at a rate of 0.02 °C yr⁻¹ in JJA (p < 0.05) and 0.04 °C yr⁻¹ in SON (p < 0.05). The change in global air temperature was mainly due to increasing air temperature in boreal regions with a significant (p < 0.1) rate of increase of 0.06 °C yr⁻¹ from 2000–2012 in SON, and in tropical biomes where air temperature also increased slightly in JJA and SON at a rate of 0.02 °C yr⁻¹ (p < 0.1). Cloud cover (not shown) increased by 0.1% yr⁻¹ (p = 0.05) between March–May (MAM) in the tropics and decreased during boreal MAM by −0.2% yr⁻¹ (p = 0.03).

Average annual maximum global wetland area for the merged SWAMPS-GLWD was 10.5 million km² (see Methods for comparison with original SWAMPS) and was in agreement with the GLWD inventoried global wetland area (Lehner and Doll 2004) used as the basis for several benchmarking activities (Fluet-Chouinard et al 2014, Bohn et al 2015). Globally and regionally, the SWAMPS-GLWD dataset had a similar seasonal phase (figure S1) for wetland area as GIEMS (R² > 0.85 for all except the semi-arid region) yet SWAMPS-GLWD had larger seasonal amplitude because of the addition of permanent wetlands from GLWD. The overlapping time period for GIEMS and SWAMPS, years 2000–2007, showed no significant trends globally or regionally for both datasets. Between the CH₄ stabilization and renewed growth periods, global mean annual wetland area decreased by 93 000 km² (2% of average annual wetland area, figure 1(c) and figure 3(b)). At the seasonal scale, a large part of the decrease was explained by negative JJA trends in wetland area, where a statistically significant decrease of 27 000 km² yr⁻¹ was observed (p < 0.01). Much of the seasonal decrease was explained by statistically significant changes in tropical, temperate, and semi-arid JJA wetland area of −4600, −1400, and −6000 km² yr⁻¹, respectively, with additional changes in DJF tropical (−3500 km² yr⁻¹) and semi-arid (−7300 km² yr⁻¹) wetland area observed. In the boreal regions, wetland area increased by 3000 and 16 400 km² yr⁻¹ in DJF and SON, respectively (p < 0.01). Overall, a complex pattern of regional and seasonal contributions in declining global wetland area was observed, consistent with decadal and multi-decadal observations of land-water storage and open-water bodies (Dieng et al 2015, Donchyts et al 2016), and with tropical wetland area decreasing (−3.4%) and boreal wetland increasing (1.8%) in area (figures 2(b) and (c)) between the stabilization and the renewed growth periods.

A partial correlation analysis was carried out to determine the effect of wetland area and climate on CH₄ emissions, and to determine the interaction between local climate and large-scale climatic teleconnections, including the Multivariate El Niño Index (MEI) and the North Atlantic Oscillation (NAO), on regional wetland area dynamics. Partial correlation analysis is an appropriate statistic to provide estimates on the correlation coefficient for a set of variables while simultaneously controlling for their interactions. The resulting partial correlations, r, range from −1 to 1 with absolute values closer to unity reflecting higher explanatory power, either with a negative or positive relationship between the independent and dependent variables. Monthly time series for each variable were correlated for the period 2000–2012 and the data were not detrended beforehand because there were no significant trends detected.

The MEI and NAO represent two major global climatic teleconnections, with the MEI linking Pacific sea surface temperature anomalies (lagged by one month) with a warming and drying in tropical regions in its positive El Niño phase and a wetting of mid-latitude arid regions in its negative La Niña phase (Wolter and Timlin 1993). The MEI is similar in its temporal dynamics to the Oceanic Niño Index that uses sea surface temperature anomalies from the Niño 3.4 region. In contrast, the NAO measures the difference in air pressure between the Icelandic low and Azores high (Barnston and Livezey 1987), reflecting mid-to-high latitude climates, with a positive NAO characterized by above average annual temperature and wet winters in Eastern North America and northern Europe and below-average temperatures in the arctic. In contrast, during the negative NAO phase, cooler and drier than average conditions persist in eastern North America and northern Europe, with warmer than average conditions in the Arctic.

For the model ensemble, variability in global CH₄ emissions was most highly correlated with wetland area (r = 0.64), followed by temperature (r = 0.37) and with negligible correlations for precipitation (r = 0.09) and cloud cover (r = 0.11). A two to three month lag between the CH₄ emissions response and climate increased the precipitation correlation by a small amount, from 0.09 (with no time lag) to 0.11 with a one month lag. Monthly to seasonal scale lags have also been observed in atmospheric inversion and hydrologic studies (Papa et al 2015, Ribeiro et al 2016, Wilson et al 2016) where transit time of water within a basin and other hydrologic processes, such as evapotranspiration, decouple the more immediate interactions between precipitation and emissions. At the regional scale, wetland area was also the most important variable for CH₄ emissions, with a correlation of 0.89 and 0.72 in tropical and temperate regions, respectively. In contrast, for the individual models, global CH₄ emission for JULES and LPJ-MPI was more highly correlated with air temperature than with wetland area due to their greater temperature sensitivity than other models, whereas the remaining models were correlated first with wetland area, and then with air temperature followed by smaller precipitation or cloud cover correlations. The ranking of global wetland area as the main driver of CH₄ production, followed by temperature and then precipitation was similar for the boreal, tropical, temperate, and arid regions, and consistent with results from a multi-model CH₄ sensitivity experiment carried out by Melton et al (2013). Overall, the higher air temperature sensitivity of CH₄ emissions was responsible for moderate correlations, with varying time lags (t minus number months, n), with the MEI for boreal (r_t−0 = −0.16), tropical (r_t−6 = 0.33), and temperate emissions (r_t−3 = 0.24), with the NAO also only weakly correlated with the model ensemble for boreal regions (r_t−3 = −0.13). At the global scale, the MEI and NAO were most highly correlated with CH₄ emissions with a five-month lag, r_t−5 = 0.26 and r_t−5 = 0.08, respectively, slightly lower than in previously published studies (Bousquet et al 2006, Hodson et al 2011).

Wetland area dynamics from the SWAMPS-GLWD dataset at global and regional scales were moderately correlated with air temperature (masked for wetland grid cells), r_t−0 = 0.24 globally, and r_t−0 = 0.33 for boreal regions, suggesting surface-flooding increased following seasonal permafrost thaw under warmer temperatures (Schuur et al 2015). Precipitation (also masked for wetland grid cells) was weakly correlated with global wetland area (r_t−0 = −0.11), and with wetland area in temperate (r_t−0 = 0.18), tropical (r_t−0 = −0.12), and arid regions (r_t−0 = −0.15). The introduction of time lags (up to +6 months) in the climate variables did not significantly improve the correlations with wetland area except in semi-arid regions where a one month lag increased the precipitation correlation with wetland area (r_t−1 = 0.30). These results highlight the importance of incorporating sub-grid cell topographic variation, as well as cell-to-cell interactions, when modeling feedbacks between hydrologic flow paths and surface inundation dynamics. At the global scale, the MEI was positively, albeit weakly, correlated with wetland area (r_t−5 = 0.33) mainly because of the relationship with tropical (r_t−5 = 0.29) and temperate wetland dynamics (r_t−5 = 0.25). Boreal regions were negatively correlated with MEI with a one-month lag (r_t−1 = −0.14) and both the MEI and NAO positively correlated with wetland area in temperate regions (r_t−0 = 0.27 and r_t−0 = 0.15, respectively). The moderate correlations, compared with earlier studies, were partly due to the short time series, where between 2000–2012, the NAO was mainly in negative phase (meaning below-average precipitation in mid-high latitudes, cooler eastern North America and northern European temperatures, and warmer arctic conditions), and no large El Niño events, whereas a record magnitude La Niña lasted from late 2009 to 2011 (Evans and Boyer-Souchet 2012). In addition, our regional definitions may also interfere with the strength of the teleconnection correlations by introducing a mix of biome types with varying climatic responses (Zhang et al 2015).

From 2000–2012, global wetland emissions appear to have remained stable and with regional increasing and decreasing trends closely compensating for one another, with no net contribution to the observed renewed atmospheric growth rate. By shifting the period of comparison to 2003–2005 and 2010–2012 to evaluate the sensitivity of our definition for the stabilization and renewed growth periods, we find only a slightly larger increase in emissions, from 185 ± 22 Tg CH₄ yr⁻¹ to 186 ± 24 Tg CH₄ yr⁻¹, an average 1.23 ± 1.1 Tg CH₄ increase and also not large enough to explain the 2007 renewed atmospheric CH₄ growth rate of ~17 Tg CH₄ yr⁻¹. Additionally, the increase between the 2003–2005 and 2010–2012 periods is not robust and almost entirely driven by just one model, which also has the highest temperature sensitivity, LPJ-MPI (12.6 Tg CH₄ yr⁻¹ increase). The increase in boreal emissions from 2000–2012 appears to be closely linked to both increasing air temperature and wetland area, with an anomalously warm event in 2007 (Bruhwiler et al 2014). In high latitude regions, evidence for warming air temperature is well documented and the feedbacks between increasing air temperature, sea-ice cover loss, and terrestrial CH₄ emissions is becoming increasingly clear (Karl et al 2015, Parmentier et al 2015). Finer-temporal and spatial remote-sensing based analyses are also consistent with the evidence presented here for a net increase in boreal wetland area and CH₄ emissions from 2003 to 2011 (Watts et al 2014). Overall, these changes are consistent with field observations (Sweeney et al 2016) and with what could be expected from projected climate change and warming impacts on high latitude systems that link temperature sensitivity as a dominant control on arctic wetland CH₄ emissions (Schaefer et al 2011, Chen et al 2015, Schuur et al 2015).

In tropical regions, high interannual variability in precipitation makes detecting decadal scale carbon-cycle trends challenging (Jung et al 2010, Zhang et al 2015). In terms of wetland area dynamics, Papa et al (2010) reported a decrease of 19 600 km² yr⁻¹ in tropical surface inundation between 1993 and 2005 based on the GIEMS data, and losses of tropical surface inundation appear to have continued through 2012 at a rate of 4000 km² yr⁻¹ (Schroeder et al 2015). As a consequence, declining tropical wetland CH₄ emissions have been found in a range of studies using GIEMS, for example, Meng et al (2015), who found a decline of 1.68 Tg CH₄ yr⁻¹ from 1993–2004. However, even with models that used fixed or static, i.e. the GLWD, rather than dynamic areal extent of wetlands, declining tropical wetland CH₄ emissions were simulated (Zhu et al 2015), suggesting that trends in climatic drivers that force changes in wetland area may be an equally important constraint on tropical CH₄ production. Over Amazonia and the Congo Basins, large consecutive droughts, in 2005 and 2010, combined with regional warming, have resulted in widespread declines in tropical forest canopy greenness (Hilker et al 2014, Zhou et al 2014). These Amazonian droughts are superimposed on an intensification of the hydrologic cycle in the wet season (Gloor et al 2013) rather than an increase in the duration of the wet season throughout the year. Declining precipitation trends between 2010–2012 were observed in Western Amazonia and Eastern Congo, but over Southeast Asia, increases in precipitation were observed (figure 3(c)). Degradation of global wetlands due to human activities is also a large component in declining wetland function (Petrescu et al 2015, Donchyts et al 2016), and losses of tropical wetland area due to drainage are the highest globally, ranging up to 2% per year (Davidson 2014, Papa et al 2015).

Outside of tropical regions, declining wetland area in temperate regions also coincided with long-term drying of soils from 1950–2005 (Mueller and Zhang 2015), however the precipitation trends from 2000–2012 of relevance in this study suggest soil moisture actually increased in this time period (figure 3(c)). Semi-arid regions, i.e. eastern Australia and South America, showed decreasing annual precipitation trends, despite large swings in seasonal precipitation related to a series of strong La Niña events (Boening et al 2012).

Zoom In Zoom Out Reset image size

ENSO has previously been highlighted as a key driver of interannual variability in global wetland CH₄ production (Bousquet et al 2006, Hodson et al 2011, Dalsøren et al 2015). Here, we find a possibly lower role for ENSO in driving global wetland CH₄ production that is due in part to i) the duration of the brief time series where no strong El Niño was observed and ii) the use of a new integrated wetland–surface–water dataset. Previous analyses have used longer time series, such as Hodson et al (2011), who scaled modeled soil-moisture to wetland area based on a GIEMS calibrated hydrologic model. For the time period 1950–2005 they found a slightly higher global correlation of global CH₄ emissions with ENSO, R² = 0.39 (with a three-month lag) and R² = 0.56 for the tropics. In addition, using an atmospheric inversion, Bousquet et al (2006) partitioned CH₄ emissions to anthropogenic and natural sources for the period 1984–2003. The study concluded that the dominant role of CH₄ surface sources was from high interannual variability of wetland area that was synchronized with ENSO. The longer 1950–2005 time period includes a wider range of both positive and negative phase ENSO events, whereas the 2000–2012 period evaluated here includes two major La Niña and only moderate El Niño events.

In addition, many studies have relied on GIEMS surface inundation data to constrain wetland areal dynamics, and have found GIEMS to be highly correlated with ENSO (Prigent et al 2007). While the new SWAMPS-GLWD dataset used here was found to enhance seasonal variation in wetland area, the dataset also partially decoupled the interannual surface-water variability from climate because of the integration of permanently inundated wetlands (with no surface flooding) from GLWD, to address known limitations in microwave remote sensing of wetlands, particularly in forested areas (Bohn et al 2015). We evaluated how the development of the SWAMPS-GLWD wetland dynamics dataset affected interannual variability (IAV) of wetland CH₄ emissions by comparing with the detrended observations of atmospheric CH₄ variability from MLO. We found the observed IAV at MLO to range from −13 to 22 Tg CH₄ yr⁻¹ from 2000 to 2012 (figure 4). In comparison, the IAV of the SWAMPS-GLWD driven wetland model ensemble ranged from −13 to 19 Tg CH₄ yr⁻¹, and across models, the range varied from small IAV (−8 to 1 Tg CH₄ yr⁻¹ for CTEM) to large IAV (3 to 19 Tg CH₄ yr⁻¹ for ORCHIDEE). Compared to observations, the sensitivity of the model ensemble results provide confidence in the use of SWAMPS-GLWD for partially driving a CH₄ IAV consistent with previous top-down and isotopic studies, e.g. Bousquet et al (2006), that demonstrate wetland CH₄ emissions explain a large portion of the IAV in atmospheric growth (r = 0.46 for the model ensemble, with the individual models ranging from r = 0.2 (TRIPLEX-GHG) to r = 0.6 (SDGVM)). Notably, the contribution of boreal wetlands to global CH₄ IAV appears to decline from 2000–2012 relative to an increase from tropical contributions (figure 4), however, overall we found a trend toward decreasing IAV in the observed CH₄ growth rate. Wildfires, not considered in this study, can contribute between 10–20 Tg CH₄ yr⁻¹ of emission IAV, however no significant trend over time has been observed to date (van der Werf et al 2006, van der Werf et al 2010, Worden et al 2013). Additionally, year-to-year variability in the atmospheric oxidative sink for methane may also affect variability in growth rate anomalies (Rigby et al 2008).

The depletion of atmospheric δ¹³C of CH₄ since 2007 presents three scenarios, (i) a change in average biogenic wetland δ¹³C source signature, (ii) an overall increasing biogenic source, (iii) a decreasing thermogenic or pyrogenic source, or some combination of all (Kirschke et al 2013). Thermogenic, or fossil-fuel related emissions, are unlikely to have decreased in recent years (Bergamaschi et al 2013, Nisbet et al 2014), and recent studies based on isotopic δ¹³C confirm a large biogenic source (Dlugokencky et al 2011, Rice et al 2016, Schaefer et al 2016). In addition to the wetland types considered in this study, there are several additional sources of biogenic emissions that could contribute to the depletion of atmospheric δ¹³C. These include river systems (Bastviken et al 2011, Borges et al 2015), lakes (Verpoorter et al 2014, Tan and Zhuang 2015), and agriculture (Leff et al 2004, Chen et al 2013). For example, taken together, river and lake system CH₄ emissions are highly uncertain and are estimated to emit as much as 100 ± 50 Tg CH₄ yr⁻¹ (Bastviken et al 2011), or the equivalent of ~30%–50% of global wetland emissions, and would require a reassessment of other source terms to close the global methane budget (Saunois et al 2016). These emission hotspots are also geographically distributed across arctic (Walter Anthony et al 2014), temperate (Chen et al 2013) and tropical systems (Borges et al 2015). The temporal response of agricultural CH₄ emissions (excluding biomass burning) is poorly understood, yet agriculture accounts for about 30% of total wetland CH₄ emissions (Kirschke et al 2013) and is produced from rice cultivation and enteric fermentation of livestock ruminants. These agricultural emissions can change on annual to decadal time scales in response to climate (Li et al 2002), but also in response to farming practices where land management can rapidly respond to socio-economic drivers (Chen et al 2013) and contribute to atmospheric IAV and to long-term trends (Tian et al 2016).

By using a multi-model approach to investigate the temporal trends and spatial patterns in global CH₄ emissions, the model uncertainty can be quantified more robustly. Here, the sources of uncertainty can be partitioned to (i) driver data, (ii) model structure, and (iii) parameter uncertainty. By providing a consistent set of climate, atmospheric CO₂, and wetland area data, the model spread was reduced from 123 Tg CH₄ yr⁻¹ (WETCHIMP, which used a similar model ensemble) to 80 Tg CH₄ yr⁻¹ (this study). This reduction highlights that uncertainties in wetland area are almost equally important to our mechanistic understanding of in situ CH₄ production and consumption processes. Combining SWAMPS and GLWD led to wetland area estimates consistent with more detailed regional estimates for Amazonia (Wilson et al 2007, Draper et al 2014, Hess et al 2015), southeast Asia (Hooijer et al 2010), and high-latitude systems, such as the Western Siberia Lowlands (Bohn et al 2015, Zhang et al 2016b).

Model structure is another key source of uncertainty (Wania et al 2013, Xu et al 2016a), as illustrated by the range of temperature-emission sensitivities for the current model ensemble. About half of the models used here (JULES, LPJ-wsl, ORCHIDEE, SDGVM, CTEM) were based on the semi-empirical model approach of Christensen et al (1996), whereas the other models (LPX-Bern, LPJ-MPI, CLM4.5, TRIPLEX-GHG) were based on more mechanistic first order approaches based on the framework developed by Walter et al (2001), see table S2 for a summary. One topic of large uncertainty are the oxidative processes that consume CH₄ (Ridgwell et al 1999), and that may change over time and alter the CO₂:CH₄ production ratios used in the semi-empirical approaches (Curry 2007). However, there was no clustering of model structure in terms of global or regional emission trends.

In addition to meteorological and wetland area interannual variability, atmospheric CO₂ rose by 24 ppm to 393 ppm from 2000 to 2012 (based on observations from Mauna Loa, MLO). Net primary production in carbon cycle models tends to respond positively to trends in elevated CO₂ (Hickler et al 2008), and would be expected to provide a sustained increase in substrate in the form of soil organic carbon for anaerobic processes to produce CH₄. A strong CO₂ driven response in CH₄ emissions was not observed by the ensemble mean because of the high IAV of climate and wetland area that appear to be more limiting for CH₄ emissions than substrate. Over longer timescales, i.e. multi-decadal to centennial, a strong CO₂ feedback on CH₄ emissions is expected, with simulated increases in global emissions of up to 73% ± 49% at 857 ppm CO₂ (Melton et al 2013).

Lastly, model parameters are difficult to robustly estimate because CH₄ production occurs in complex landscapes where anaerobic soil conditions can be very heterogeneous. To estimate emissions at scales of 50 km² or larger, where CH₄ production may be occurring in small topographic depressions, remains a large challenge (Lara et al 2015, Shi et al 2015). While the area-weighted monthly-average flux estimates for the model ensemble ranged within observations, i.e. from 2.7 to 3.9 g CH₄ m⁻² month⁻¹ globally, 0.7 to 3.1 g CH₄ m⁻² month⁻¹ for boreal wetlands (observations 2–4 g CH₄ m⁻² month⁻¹ in boreal systems (Turetsky et al 2014), and 5.2 to 8.2 g CH₄ m⁻² month⁻¹ for tropical wetlands (observations 0.1 to 29 g CH₄ m⁻² month⁻¹ (Sjogersten et al 2014). Benchmarking of process models with flux tower measurements or airborne campaigns remains critical for improving model structure and parameters (Miller et al 2016) and addressing scaling artifacts that may obscure non-linear methane production and consumption processes.

Interpreting the interannual and decadal dynamics of the CH₄ atmospheric growth rate has presented significant challenges over the past three decades, with the sources and sinks remaining poorly understood (Kirschke et al 2013). Using an ensemble of global wetland models constrained with satellite and inventory based surface inundation and wetland area seasonality and trends, we now provide a comprehensive and updated estimate of the role of wetlands in the recent increase of the atmospheric growth rate that began in 2007. We show that the role of wetlands in the renewed period of atmospheric CH₄ growth appears minimal to non-existent, and that:

• At the global scale, wetland CH₄ emissions have remained constant from 2000–2012 at 184 ± 22 Tg g CH₄ yr⁻¹ but that significant spatial variability in trends are masked by the global perspective (figure S2 and S3).
• In boreal regions, increasing CH₄ emissions corresponds to increasing wetland area and air temperature, whereas in the tropics, decreasing wetland area and large variability in precipitation has led to decreased emissions.
• At global and the regional scales defined in our study, the role of climatic teleconnections such as ENSO and the NAO are smaller than what has been reported in previous work; however, we confirm that the IAV of the atmospheric growth rate is largely explained by wetlands.
• The interannual variability in global wetland emissions is dominated by boreal regions from 2000–2006 and then with increasing contribution from tropical regions possibly coinciding with larger droughts over Amazonia and the Congo Basin (figure 4). However, there has been no consistent shift in the IAV of wetland CH₄ emissions over the 2000–2012 time period (figure 4).
• The range of the modelled interannual variability in global wetland emissions in 2007–2012 is similar to the IAV observed at the MLO station, while it is less than observed for 2000–2006. Therefore, the period 2000–2006 is anomalous not only due to the absent trend in the growth rate of atmospheric CH₄ concentrations, but also due to anomalously high IAV not fully explained by natural wetland emissions.
• Our results, interpreted in the context of a depletion in atmospheric δ¹³C observed since 2007, suggests that either a shift in δ¹³C biogenic source signature occurred or other agricultural biogenic sources are required to explain the recent and sustained atmospheric increase in CH₄, or that, less likely, a decrease in thermogenic and pyrogenic emission has occurred. This is consistent with recent work of Schaefer et al (2016) who present isotopic evidence suggesting an increasing role of livestock and agriculture in the growth rate of atmospheric CH₄.
• The pattern of increasing high-latitude emissions and possibly decreasing to stable tropical emissions are consistent with climate change projections that forecast a general increase in boreal air temperatures and a decrease in tropical precipitation (Scholze et al 2006). Thus the past decade presents an observational test case for climate and socio economic impact studies on CH₄ production (Lawrence et al 2015, Petrescu et al 2015).
• To reduce uncertainties in wetland dynamics mapping we recommend that (1) multi-platform remote sensing using both radar and optical observations are integrated at higher spatial resolution to resolve issues associated with low-detection probabilities in closed-forest canopy regions, (2) that inter-sensor calibration and effects on inter-annual and seasonal trends are clearly accounted for, and (3) that ground-based wetland inventories are continually updated and made available to benchmark and calibrate remote sensing algorithms, and with clear terminology to avoid double counting.