A global wetland methane emissions and uncertainty dataset for atmospheric chemical transport models (WetCHARTs version 1.0)

Abstract

Wetland emissions remain one of the principal sources of uncertainty in the global atmospheric methane (CH${}_{\mathrm{4}})$ budget, largely due to poorly constrained process controls on CH${}_{\mathrm{4}}$ production in waterlogged soils. Process-based estimates of global wetland CH${}_{\mathrm{4}}$ emissions and their associated uncertainties can provide crucial prior information for model-based top-down CH${}_{\mathrm{4}}$ emission estimates. Here we construct a global wetland CH${}_{\mathrm{4}}$ emission model ensemble for use in atmospheric chemical transport models (WetCHARTs version 1.0). Our 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ resolution model ensemble is based on satellite-derived surface water extent and precipitation reanalyses, nine heterotrophic respiration simulations (eight carbon cycle models and a data-constrained terrestrial carbon cycle analysis) and three temperature dependence parameterizations for the period 2009–2010; an extended ensemble subset based solely on precipitation and the data-constrained terrestrial carbon cycle analysis is derived for the period 2001–2015. We incorporate the mean of the full and extended model ensembles into GEOS-Chem and compare the model against surface measurements of atmospheric CH${}_{\mathrm{4}}$; the model performance (site-level and zonal mean anomaly residuals) compares favourably against published wetland CH${}_{\mathrm{4}}$ emissions scenarios. We find that uncertainties in carbon decomposition rates and the wetland extent together account for more than 80 % of the dominant uncertainty in the timing, magnitude and seasonal variability in wetland CH${}_{\mathrm{4}}$ emissions, although uncertainty in the temperature CH${}_{\mathrm{4}}$ : C dependence is a significant contributor to seasonal variations in mid-latitude wetland CH${}_{\mathrm{4}}$ emissions. The combination of satellite, carbon cycle models and temperature dependence parameterizations provides a physically informed structural a priori uncertainty that is critical for top-down estimates of wetland CH${}_{\mathrm{4}}$ fluxes. Specifically, our ensemble can provide enhanced information on the prior CH${}_{\mathrm{4}}$ emission uncertainty and the error covariance structure, as well as a means for using posterior flux estimates and their uncertainties to quantitatively constrain the biogeochemical process controls of global wetland CH${}_{\mathrm{4}}$ emissions.

Introduction

Methane (CH${}_{\mathrm{4}})$ is a potent greenhouse gas with a global warming potential more than 25 times that of CO${}_{\mathrm{2}}$ on a 100-year time horizon (Myhre et al., 2013). The global CH${}_{\mathrm{4}}$ budget and growth rate remain poorly understood, largely due to the poorly resolved evolution of atmospheric CH${}_{\mathrm{4}}$ sources and sinks (Nisbet et al., 2014). Wetland CH${}_{\mathrm{4}}$ emissions are the largest natural source of atmospheric CH${}_{\mathrm{4}}$, amounting to roughly 20–40 % of global CH${}_{\mathrm{4}}$ emissions (Ciais et al., 2013). The large disparities between a range of top-down and bottom-up wetland CH${}_{\mathrm{4}}$ estimates (Kirschke et al., 2013; Melton et al., 2013) arise from large uncertainties in the timing, distribution and the underlying processes controlling net wetland CH${}_{\mathrm{4}}$ production.

In wetland soils, CH${}_{\mathrm{4}}$ is produced through the decomposition of organic matter in anaerobic (oxygen-depleted) environments. The dominant processes controlling the seasonal and inter-annual variations include the carbon availability (soil C substrate) and decomposition rate, wetland inundation extent and temperature (Yvon-Durocher et al., 2014). Other important controls on wetland CH${}_{\mathrm{4}}$ emissions include the presence of macrophytes (Laanbroek, 2010), organic C decomposition rates (Miyajima et al., 1997) and soil pH (Singh et al., 2000), amongst other factors. The link between terrestrial carbon–water cycling and wetland CH${}_{\mathrm{4}}$ emissions is of particular interest from a terrestrial greenhouse gas emissions standpoint: inter-annual variations in terrestrial carbon cycling (Le Quéré et al., 2013) can affect wetland CH${}_{\mathrm{4}}$ emissions on seasonal to century timescales (Hodson et al., 2011). The role of carbon cycle dynamics in global wetland CH${}_{\mathrm{4}}$ emissions is increasingly recognized. Temporal variations in gross primary production influence the short-term carbon supply (such as carbon inputs from root exudates and fine litter) as well as long-lived carbon stores (such as wood litter turnover or soil organic C) in wetland soils (Riley et al., 2011; Bloom et al., 2012; Melton et al., 2013). The combined response of CO${}_{\mathrm{2}}$ and CH${}_{\mathrm{4}}$ fluxes to climatic variability remains poorly characterized. For example, increasing temperatures in boreal ecosystems could lead to higher carbon uptake, increased respiration and drier soils (Watts et al., 2014), and it is currently unclear whether these processes amount to an amplifying or a dampening effect on boreal CH${}_{\mathrm{4}}$ emissions. From a greenhouse gas balance standpoint, quantifying the global-scale process links between terrestrial carbon cycling and wetland CH${}_{\mathrm{4}}$ emissions is crucial to characterizing the combined terrestrial biosphere CO${}_{\mathrm{2}}$ and CH${}_{\mathrm{4}}$ flux response to climatic variability.

The quantification of regional wetland CH${}_{\mathrm{4}}$ emissions remains challenging. While wetland CH${}_{\mathrm{4}}$ emissions are relatively well constrained on a global scale (Kirschke et al., 2013; Saunois et al., 2016), regional CH${}_{\mathrm{4}}$ fluxes are difficult to detect due to their comparatively diffuse nature relative to anthropogenic point sources and the scarcity of direct measurements of wetland CH${}_{\mathrm{4}}$ emissions. From a bottom-up perspective, challenges in wetland CH${}_{\mathrm{4}}$ modelling stem from order-of-magnitude uncertainties in wetland CH${}_{\mathrm{4}}$ emission factors and their spatio-temporal dependence on biogeochemical process controls. Nonetheless, for top-down CH${}_{\mathrm{4}}$ emission estimates, prior knowledge of wetland CH${}_{\mathrm{4}}$ emissions and their associated uncertainty is critical in the formulation of Bayesian atmospheric CH${}_{\mathrm{4}}$ inversions. Atmospheric inversions combine CH${}_{\mathrm{4}}$ measurements from surface, aircraft and satellites (Wecht et al., 2014a; Jacob et al., 2016) and the prior probability of the magnitude and uncertainty characteristics of CH${}_{\mathrm{4}}$ emissions (Bousquet et al., 2011; Pison et al., 2013; Fraser et al., 2013; Turner et al., 2015). Typically, CH${}_{\mathrm{4}}$ inversions do not explicitly formulate wetland CH${}_{\mathrm{4}}$ emission uncertainty correlations; rather, prior wetland CH${}_{\mathrm{4}}$ uncertainty correlations are either absent or implicitly prescribed through space–time correlation lengths on CH${}_{\mathrm{4}}$ emissions. However, inter-model similarities reveal significant levels of emergent correlations in the timing, magnitude and spatial variability of wetland CH${}_{\mathrm{4}}$ emissions. For example, the Wetland CH${}_{\mathrm{4}}$ Inter-comparison of Models Project (WETCHIMP) model ensemble (Melton et al., 2013) reveals varying levels of spatial and temporal agreement between models; these correlations stem from large-scale patterns in biogeochemical process controls (such as temperature, inundation and carbon cycling). Given the relatively large WETCHIMP CH${}_{\mathrm{4}}$ emission uncertainties (the model range is typically 150–300 % of the model mean over major wetland areas and greater elsewhere), this prior “biogeochemical covariance” can potentially amount to a critical constraint on atmospheric CH${}_{\mathrm{4}}$ inversions; such a covariance structure can be incorporated into an atmospheric inversion cost function (Michalak et al., 2005) or as a means of improving the attribution of posterior CH${}_{\mathrm{4}}$ fluxes to wetland CH${}_{\mathrm{4}}$ emissions (Wecht et al., 2014b).

Here we propose a process-informed wetland CH${}_{\mathrm{4}}$ emission and uncertainty dataset for atmospheric chemistry and transport modelling (WetCHARTs) based on multiple terrestrial biosphere models, wetland extent scenarios and CH${}_{\mathrm{4}}$ : C temperature dependencies. In contrast to a conventional process-based model inter-comparison approach, our wetland CH${}_{\mathrm{4}}$ emission ensemble members are derived by exhaustively combining a range of temperature, carbon and wetland extent parameterizations. An advantage of our approach is that it provides a prior probability distribution of biogeochemical process control uncertainty. Top-down CH${}_{\mathrm{4}}$ emission estimates can then be used to quantify (a) the probability of individual ensemble members and (b) the combined probability distribution of carbon models, CH${}_{\mathrm{4}}$ : C temperature dependencies and wetland extent scenarios.

We formulate a full (2009–2010) and extended (2001–2015) estimate of wetland CH${}_{\mathrm{4}}$ emission magnitude and its associated biogeochemical covariance structure, based on knowledge of the global wetland CH${}_{\mathrm{4}}$ source and the primary biogeochemical process controls. We validate and compare the wetland CH${}_{\mathrm{4}}$ emissions ensemble against a suite of regional flux estimates; we use a global atmospheric chemical transport model (GEOS-Chem; Bey et al., 2001) to evaluate the CH${}_{\mathrm{4}}$ emissions ensemble mean relative to existing wetland CH${}_{\mathrm{4}}$ emission models (Sects. 2 and 3). Finally, we summarize the strengths and limitations of our wetland emissions ensemble and outline its potential applications in global atmospheric inversion frameworks (Sect. 4).

Wetland CH${}_{\mathrm{4}}$ model ensemble

The wetland CH${}_{\mathrm{4}}$ emissions ensemble provides CH${}_{\mathrm{4}}$ fluxes and the associated uncertainty estimates based on four wetland extent parameterizations, nine terrestrial biosphere models of heterotrophic respiration and three CH${}_{\mathrm{4}}$ : C temperature parameterizations. Global monthly 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ emissions and their associated uncertainty structure span 2009–2010 (full ensemble, henceforth FE); we also evaluate a subset of the model ensemble spanning 2001–2015 (extended ensemble, henceforth EE). We validate FE and EE emissions against a range of regional CH${}_{\mathrm{4}}$ emission estimates. Finally, we incorporate FE, EE and existing wetland emission inventories into GEOS-Chem and evaluate the atmospheric CH${}_{\mathrm{4}}$ simulations against 104 surface CH${}_{\mathrm{4}}$ measurement sites.

Wetland CH${}_{\mathrm{4}}$ emissions and uncertainty

We derive wetland CH${}_{\mathrm{4}}$ emissions $F$ (mg CH${}_{\mathrm{4}}$ m${}^{-\mathrm{2}}$ day${}^{-\mathrm{1}})$ at time $t$ and location $x$ as $$F(t,x)=sA\left(t,,,x\right)R\left(t,,,x\right)q_{\mathrm{10}}^{\frac{T(t,x)}{\mathrm{10}}},$$ where $A(t,x)$ is the wetland extent fraction, $R(t,x)$ is the C heterotrophic respiration per unit area at time $t$, $q_{\mathrm{10}}^{T(t,x)/\mathrm{10}}$is the temperature dependence of the ratio of C respired as CH${}_{\mathrm{4}}$ (where $q_{\mathrm{10}}$ is the relative CH${}_{\mathrm{4}}$ : C respiration for a 10 ${}^{\circ }$C increase and $T(t,x)$ is the surface skin temperature) and $s$ is a global scaling factor. This empirical parameterization provides first-order constraints on the role of carbon, water and temperature variability in the global spatial and temporal variability of wetland CH${}_{\mathrm{4}}$ emissions. Variants of the Eq. (1) parameterization have been used within a range of wetland CH${}_{\mathrm{4}}$ emission models (e.g. Hodson et al., 2011; Pickett-Heaps et al., 2011; Bloom et al., 2012; Melton et al., 2013).

In our approach, wetland CH${}_{\mathrm{4}}$ emission statistics within and across 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ grid cells are derived based on an ensemble of wetland CH${}_{\mathrm{4}}$ emission simulations. The 324-member FE is based on 3 CH${}_{\mathrm{4}}$ : C temperature dependencies, 9 heterotrophic respiration configurations, 4 wetland extent scenarios and 3 global-scale factor configurations (3 $\times$ 9 $\times$ 4 $\times$ 3 $=$ 324); the 18-member EE ensemble is a subset of FE based on data availability during 2001–2015 (see Table 1 for details).

Wetland CH${}_{\mathrm{4}}$ model ensemble configurations.

Parameter	Description		Ensemble configurations
$s$	Global scaling factor		3 configurations: emissions are scaled such that 2009–2010 emissions amount to 124.5, 166 or 207.5 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}}$
$A$	Wetland extent		2 spatial extent parameterization (scaled using GLOBCOVER and GLWD)
			2 temporal variability parameterizations (SWAMPS inundation extent${}^{*}$ and ERA-Interim precipitation)
$R$	Heterotrophic respiration		8 MsTMIP terrestrial C models${}^{*}$
			CARDAMOM terrestrial C cycle analysis
$q_{\mathrm{10}\left(c\right)}$	Temperature-dependent CH${}_{\mathrm{4}}$ respiration fraction		3 CH${}_{\mathrm{4}}$ : C temperature parameterizations; $q_{\mathrm{10}\left(c\right)}=[\mathrm{1},\mathrm{2},\mathrm{3}]$

${}^{*}$ These datasets are only used in the 2009–2010 “full ensemble” (FE).

The heterotrophic respiration configurations are derived from eight terrestrial biosphere models used in the Multi-scale Synthesis and Terrestrial Model Intercomparison Project (MsTMIP BG1 simulations; see Huntzinger et al. (2013) and Wei et al. (2014) for the model and experiment details) and the global CARbon DAta MOdel fraMework (CARDAMOM) terrestrial carbon analysis (Bloom et al., 2016). V1.0 outputs from the MsTMIP are available for the period 1900–2010 (Huntzinger et al., 2016), and the CARDAMOM analysis was extended to span 2001–2015 based on the Bloom et al. (2016) methodology (see Appendix A for details). Since MsTMIP and CARDAMOM respiration estimates vary intrinsically as a function of temperature, $q_{\mathrm{10}}$ only accounts for the temperature dependence of the fraction of C respired as CH${}_{\mathrm{4}}$. We prescribe three CH${}_{\mathrm{4}}$ : C temperature dependencies (Table 1) which are broadly equivalent to a $\pm$50 % range on the CH${}_{\mathrm{4}}$ : CO${}_{\mathrm{2}}$ temperature dependence reported by Yvon-Durocher et al. (2014).

Here we use two spatial ($i=\mathrm{1},\mathrm{2}$) and two temporal ($j=\mathrm{1},\mathrm{2}$) wetland extent parameterization approaches to represent the uncertainty associated with the role of hydrology in wetland CH${}_{\mathrm{4}}$ emissions. Each temporal and spatial wetland extent parameterization, $A_{i,j}(t,x)$, is derived as $$A_{i,j}\left(t,,,x\right)=w_i\left(x\right)h_{i,j}\left(t,,,x\right),$$ where $w_i\left(x\right)$ represents the wetland extent fraction, and $h_{i,j}(t$,$x)$ represents the temporal variability relative to $w_i\left(x\right)$. $w_{\mathrm{1}}\left(x\right)$ is the sum of all GLOBCOVER wetland and freshwater land cover types (all flooded, waterlogged and inland water body land cover types; Bontemps et al., 2011); $w_{\mathrm{2}}\left(x\right)$ is the Global Lakes and Wetlands Database (GLWD) maximum recorded wetland and freshwater body extent map by Lehner and Döll (2004).

For $h_{\ast ,j}(t$,$x)$, we use (a) the Surface WAter Microwave Product Series (SWAMPS) multi-satellite surface water product (Schroeder et al., 2015; $j=\mathrm{1}$) and the (b) monthly ERA-Interim precipitation ($j=\mathrm{2}$). For $i=\mathrm{1}$ ($i=\mathrm{2}$), $h_{i,j}(t,x)$ is normalized such that the mean (maximum) $h_{i,j}(t,x)$ is equal to 1. In order avoid physically unrealistic outcomes, we derive $A_{\mathrm{1},j}(t,x)$ as min$\mathit{\left\{}{w}_{\mathrm{1}}\right(x\left)h_{\mathrm{1},j}\right(t,x),\mathrm{1}\mathit{\}}$, where the “min$\mathit{\left\{}\mathit{\right\}}$” function represents the minimum between the two bracketed values.

We note that the two hydrological proxies provide contrasting advantages and disadvantages. Satellite-retrieved surface water extent provides an observation-based constraint on the spatial and temporal extent of wetlands and freshwater bodies. While our temporal scaling of static wetland and freshwater extent mitigates the role of spatial biases in satellite-retrieved inundation, vegetation cover remains a major confounding variable in satellite-constrained wetland extent (Schroeder et al., 2015). Moreover, satellites cannot directly observe subsurface soil saturation, even though these soils amount to significant CH${}_{\mathrm{4}}$ fluxes to the atmosphere (Turetsky et al., 2014). On the other hand, precipitation does not provide a direct constraint on the wetland and freshwater extent; however, it provides an aggregate constraint on ecosystem hydrological variability and wall-to-wall coverage across the globe. We henceforth refer to $F$ as “wetland CH${}_{\mathrm{4}}$ emissions”; however, we recognize that lakes, rivers and reservoirs account for $\sim$ 20 % of the total wetland and freshwater body extent (Lehner and Döll, 2004). We discuss the implications of including non-wetland freshwater bodies in Sect. 4.

For each of the 324 FE configurations ($c=\mathrm{1}$–324) and 18 EE configurations ($c=\mathrm{1}$–18), we derive $s_c$ such that $$s_c={\displaystyle \frac{G}{{\mathrm{\Sigma }}_t{\mathrm{\Sigma }}_x{F}_c(t,x)a\left(x\right)\frac{\mathrm{\Delta }t}{n}}},$$ where $F_c(t,x)$ are the $c$th ensemble member emissions at grid cell $x$ and time $t$, $a\left(x\right)$ is the area of grid cell $x$, $\mathrm{\Delta }t$ is the time step (1 month), $n$ is the number of years and $G$ is the global total CH${}_{\mathrm{4}}$ emitted from wetlands. We derive $s_c$ such that the FE and EE ensemble members amount to a mean global annual flux of $G=\mathrm{124}$.5, 166 or 207.5 g CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}}$ during 2009–2010. The prescribed range of total wetland CH${}_{\mathrm{4}}$ emissions spans the Saunois et al. (2016) mean 2000–2009 top-down wetland CH${}_{\mathrm{4}}$ emission estimates (166 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}}$; 125–204 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}})$.

We attribute the uncertainty in the timing and magnitude of $F(t,x)$ (namely, the maximum CH${}_{\mathrm{4}}$ emission month, mean CH${}_{\mathrm{4}}$ emissions and CH${}_{\mathrm{4}}$ emission variability) to carbon decomposition, wetland extent and CH${}_{\mathrm{4}}$ : C temperature dependence uncertainty. The derivation of the “dominant uncertainty” within each zonal band (i.e. the dominance of carbon, water or temperature as the dominant source of uncertainty) is fully described in Appendix C.

GEOS-Chem atmospheric CH${}_{\mathrm{4}}$ simulations

We evaluate the FE and EE wetland CH${}_{\mathrm{4}}$ emission means against the World Data Centre for Greenhouse Gases (WDCGG) CH${}_{\mathrm{4}}$ measurement sites by incorporating these into the 4${}^{\circ }$ $\times$ 5${}^{\circ }$ resolution GEOS-Chem atmospheric chemical and transport model (version 10.01; acmg.seas.harvard.edu/geos). We benchmark the FE and EE runs against GEOS-Chem simulations with the GEOS-Chem wetland CH${}_{\mathrm{4}}$ emissions inventory (Pickett-Heaps et al., 2011; 2009–2010 derivation described in Turner et al., 2015; henceforth GC) and the Bloom et al. (2012) satellite-constrained wetland emissions (henceforth BL), as these emission estimates have been used in a range of atmospheric chemical transport model simulations (Fraser et al., 2013; Turner et al., 2015; and Wilson et al., 2016, amongst others). We perform each GEOS-Chem forward run for the period 2009–2010 with a 4-year (2005–2009) spin-up period. The non-wetland CH${}_{\mathrm{4}}$ sources in GEOS-Chem consist of biofuel, fossil fuel, livestock, waste, rice (EDGAR v4.2; European Commission, 2011), fires (Global Fire Emissions Database version 4; van der Werf et al., 2010), soil C sinks and termites (Fung et al., 1991). The non-wetland CH${}_{\mathrm{4}}$ fluxes are the same in each run, with the exception of the rice source in run BL (as global wetland and rice emissions are treated as one source by Bloom et al., 2012). While model CH${}_{\mathrm{4}}$ surface concentrations are strongly influenced by wetland CH${}_{\mathrm{4}}$ magnitude, timing and distribution (Bloom et al., 2012; Meng et al., 2015), comparisons between GEOS-Chem outputs and surface CH${}_{\mathrm{4}}$ measurements may also be affected by errors in non-wetland CH${}_{\mathrm{4}}$ emissions and in transport. However, Wecht et al. (2012) and Turner et al. (2015) show that the GEOS-Chem emissions and transport provide an unbiased representation of the observed latitudinal background. The global inversion of Turner et al. (2015) using GEOS-Chem emissions as prior further shows no large errors in non-wetland emissions that would confound the analysis presented here.

For each of the four runs (FE, EE, GC and BL), we use the Wecht et al. (2014b) 1 January 2005 initial conditions for atmospheric CH${}_{\mathrm{4}}$ concentrations in GEOS-Chem. For each simulation, we performed a 4-year spin-up period (2005–2009) using 2009 emissions to reduce the potential inconsistency between the initial conditions and the global distribution of wetland CH${}_{\mathrm{4}}$ emissions; this spin-up ensures that the relative variations in 1 January 2009 CH${}_{\mathrm{4}}$ concentrations for each run are broadly consistent with each emission scenario. We save GEOS-Chem atmospheric CH${}_{\mathrm{4}}$ concentrations every 3 h. We compare the mean monthly GEOS-Chem output against all WDCGG sites (104 sites with monthly 2009–2010 data in total). For each site, the nearest 4${}^{\circ }$ $\times$ 5${}^{\circ }$ GEOS-Chem grid cell is used for comparison. We note that the GEOS-Chem analysis outlined here is not a direct validation of FE and EE; rather, it provides supporting evidence for the plausibility of FE and EE emissions relative to existing wetland CH${}_{\mathrm{4}}$ emission datasets.

Results, comparison and validation

Top row (a, b): 2009–2010 full model ensemble (FE; left) and extended model ensemble (EE; right) mean wetland CH${}_{\mathrm{4}}$ emissions. Middle row (c, d): 2009–2010 emissions from the GEOS-Chem wetland CH${}_{\mathrm{4}}$ emissions inventory (GC; left) model and the satellite-constrained estimates by Bloom et al. (2012) (BL; right). Bottom row (e, f): mean 2009–2010 FE 5th–95th percentile range (left) and uncertainty factor (5th–95th percentile range normalized by mean 2009–2010 emissions; right).

[Figure omitted. See PDF]

Mean wetland CH${}_{\mathrm{4}}$ emission zonal profiles: full ensemble mean (FE: red line) and corresponding range (pink area); extended ensemble mean (EE; black line) and corresponding range (grey area); GEOS-Chem emissions inventory (GC; dashed blue line); Bloom et al. (2012) emissions (BL; dashed orange line).

[Figure omitted. See PDF]

Seasonally averaged 2009–2010 wetland CH${}_{\mathrm{4}}$ emissions for this study (full ensemble: FE; extended ensemble: EE), Bloom et al. (2012) wetland emissions (BL) and GEOS-Chem wetland emissions inventory (GC) across North and South America (left column; 180–35${}^{\circ }$ W), Europe and Africa (centre column; 35${}^{\circ }$ W–55${}^{\circ }$ E) and Asia and Oceania (right column; 55–180${}^{\circ }$ E). The emissions for each region are reported as total monthly fluxes across 5${}^{\circ }$ latitude bins. The black dotted line denotes the maximum emission month within each 5${}^{\circ }$ latitude bin.

[Figure omitted. See PDF]

A comparison between the mean annual regional wetland CH${}_{\mathrm{4}}$ emission estimates of (1) Glagolev et al. (2011), (2) Pickett-Heaps et al. (2011), (3) Chang et al. (2014) and (4) Melack et al. (2004) and the global wetland emission datasets by Bloom et al. (2012) (BL), the GEOS-Chem wetland CH${}_{\mathrm{4}}$ emissions inventory (GC), this study (full ensemble: FE; extended ensemble: EE) and the range of WETCHIMP models (Melton et al., 2013). Wetland emissions (horizontal axis) correspond to the mean annual totals within the regions shown in the inset map.

[Figure omitted. See PDF]

Mean full ensemble (FE) global wetland emissions are largely accounted for by three high-latitude regions, three tropical regions and subtropical southeast Asia (Fig. 1). North American, Scandinavian and Siberian median (5th–95th percentiles) CH${}_{\mathrm{4}}$ fluxes amount to 10 % (3–30 %), 2 % (1–6 %) and 2 % (1–6 %) of global emissions, respectively. Amazon wetland emissions (29 %; 20–37 %) account for the largest tropical emission source, followed by the Indonesian archipelago (13 %; 7–23 %) and central Africa (12 %; 7–23 %). Subtropical southeast Asian emissions account for 5 % (1–10 %). High-latitude ($>$ 50${}^{\circ }$ N) and tropical emissions amount to 12 % (5–31 %) and 66 % (43–83 %) of global wetland CH${}_{\mathrm{4}}$ emissions, respectively. Gridded FE uncertainties (shown as the 5th–95th percentile ranges; Fig. 1e) are largely comparable in magnitude to FE emissions (Fig. 1a). Relative FE uncertainties (shown as the ratio of the 90 % confidence range to mean emissions in Fig. 1f) are the lowest in high-emission areas, notably the wetland regions in the Amazon and Congo basins, North America and western Eurasia.

Mean FE and extended ensemble (EE) CH${}_{\mathrm{4}}$ emission patterns exhibit close agreement across all tropical continents and the high northern-latitude wetland regions (Fig. 1a and b). The comparison between zonal mean emissions (Fig. 2) reveals differences of less than 1 Tg/yr/${}^{\circ }$ lat between FE and EE. On a continental scale, FE and EE emission patterns are in broad agreement with the Pickett-Heaps et al. (2011) wetland CH${}_{\mathrm{4}}$ emissions (GC; Fig. 1c) and the Bloom et al. (2012) emissions (BL; Fig. 1d). High-latitude FE and EE emissions peak roughly between 45 and 60${}^{\circ }$ N (Fig. 2), which is in agreement with GC and BL emission peaks ($\sim$ 60 and $\sim$ 50${}^{\circ }$ N, respectively) and tropical emissions for all four emission dataset peaks within 0–5${}^{\circ }$ S. The FE zonal mean is comparable to the BL in the near-equatorial tropics and significantly lower (with respect to the FE model ensemble 90 % confidence range) everywhere else; the FE zonal mean is comparable to GC in high-latitude and temperate regions, but significantly lower than GC in the tropics and the Southern Hemisphere.

The global wetland CH${}_{\mathrm{4}}$ emission inter-annual variability range of the FE (2009–2010) and EE (2001–2015) emission models, normalized relative to 2009 emissions; the WETCHIMP (Melton et al., 2013) model ensemble inter-annual variability is normalized relative to 2001–2004 mean emissions.

[Figure omitted. See PDF]

The dominant uncertainty attribution of the maximum CH${}_{\mathrm{4}}$ emission month (left), magnitude (centre) and seasonal variability (right) to carbon decomposition, temperature CH${}_{\mathrm{4}}$ : C dependence ($q_{\mathrm{10}})$ and wetland extent parameterization within 5${}^{\circ }$ latitude bins. The derivation of dominant uncertainties is described in Appendix C.

[Figure omitted. See PDF]

All CH${}_{\mathrm{4}}$ emission models show similar patterns in the temporal distribution of CH${}_{\mathrm{4}}$ emissions in high-latitude and temperate regions (with CH${}_{\mathrm{4}}$ emissions peaking between July and September; Fig. 3). We note that the larger CH${}_{\mathrm{4}}$ fluxes in the BL emissions over Asia and Oceania are due to rice paddy CH${}_{\mathrm{4}}$ emissions. All emission models exhibit high-latitude ($>$ 50${}^{\circ }$ N) maximum CH${}_{\mathrm{4}}$ emissions between June and August. In tropical South America (0–20${}^{\circ }$ S), FE and EE emissions peak between February and April, which is comparable to BL (February–March) and overall earlier than GC (5–20${}^{\circ }$ S emission peak in September). There is considerable disagreement between northern tropical African emission variability amongst all models, with 0–15${}^{\circ }$ N emissions peaking in February (GC), April–October (FE, EE) and September–November (BL). Subtropical Asian FE and EE emissions (20–30${}^{\circ }$ N) peak in June–July, earlier than BL emissions (August–September) and comparable to GC emissions (June).

We compare the mean FE and EE (2009–2010) wetland emissions against a range of independent wetland CH${}_{\mathrm{4}}$ regional emission estimates (Fig. 4). Emissions from the Siberian wetlands (Glagolev et al., 2011), the Hudson Bay lowlands (Pickett-Heaps et al., 2011) and the Amazon River basin (Melack et al., 2004) are within the 25th–75th percentile estimates of FE and EE wetland CH${}_{\mathrm{4}}$ emissions. Alaskan wetland emissions (Chang et al., 2014; May–September) are higher (2.1 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}})$ but within the 5th–95th percentile range of FE and EE wetland CH${}_{\mathrm{4}}$ emission estimates. With the exception of the Amazon River basin estimates, the FE and EE emission uncertainty estimates are larger than the Melton et al. (2013) wetland CH${}_{\mathrm{4}}$ emission model (WETCHIMP 1993–2004) range. BL (2009–2010) and GC (2009–2010) estimates are also within all regional 5th–95th percentile ranges. We note the temporal mismatch between the modelled and regional wetland CH${}_{\mathrm{4}}$ emission estimates in Fig. 4; however, based on a range of process model approaches (e.g. Bloom et al., 2010; Melton et al., 2013), we expect the inter-annual variation in wetland CH${}_{\mathrm{4}}$ emissions to be substantially smaller than the FE and EE estimate uncertainty. For example, the maximum-to-minimum ratios of WETCHIMP 1993–2004 annual emissions are $\le$ 5.1 across the three extratropical regions and $\le$ 1.4 in the Amazon River basin; in contrast, FE and EE uncertainty intervals span factors of 5.8–156.3 in the extratropics and 2.3–3.9 in the Amazon River basin.

The full ensemble (FE) spatial and temporal error covariance, summarized as a monthly error correlation across boreal and Arctic ($>$ 55${}^{\circ }$ N) temperate (23–55${}^{\circ }$ N), tropical (23${}^{\circ }$ S–23${}^{\circ }$ N) and Southern Hemisphere ($<$ 23${}^{\circ }$ S) latitudes. A correlation between two location-and-time indices indicates the degree to which models consistently over- or under-predict wetland CH${}_{\mathrm{4}}$ emissions relative to the ensemble mean. The non-zero off-diagonal correlation patterns emerge as a function of varying biogeochemical commonalities across ensemble members, such as wetland CH${}_{\mathrm{4}}$ dependencies on temperature, carbon availability and wetland extent. Negative correlations between tropical and Northern Hemisphere extratropical (i.e. temperate, boreal and Arctic) wetlands emerge as a function of a global constraint on wetland CH${}_{\mathrm{4}}$ emissions (166 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}}$ $\pm$ 25 %).

[Figure omitted. See PDF]

FE and EE ensemble models exhibit a median of 6.7 and a 7.2 % increase in global emissions between 2009 and 2010 (Fig. 5). BL and GC changes from 2009 to 2010 ($+$1.8 and $+$3.2 %) are within the FE uncertainty range ($-$2.6 to $+$13.4 %). Uncertainties in the WETCHIMP inter-annual variations (IAV; relative to the 2001–2004 model means) are larger than EE IAV uncertainty throughout 2001–2015 (relative to 2009) and smaller than the FE change uncertainty for 2009 to 2010. For the 2003–2013 period, BL IAV is generally lower or within the range of EE IAV. In comparison to regional top-down constraints, we find that the regional EE IAV is comparable to the Miller et al. (2016) 2012–2014 annual Alaskan wetland emission variability (coefficient of variation observed $=$ 4.9 %, EE $=$ 4.2–6.9 %), and within the Wilson et al. (2016) constraints on the change in annual Amazon wetland emissions for 2010 to 2011 (coefficient of variation: observed $<$ 20 %; EE $=$ 0.5–2.9 %).

On a zonal basis, the “dominant uncertainty”, i.e. the dominant source of uncertainty within each band, in mean CH${}_{\mathrm{4}}$ emissions and the timing of maximum CH${}_{\mathrm{4}}$ emissions is almost completely dominated by carbon decomposition and wetland extent uncertainties (Fig. 6). Seasonal variability in CH${}_{\mathrm{4}}$ emissions is also largely dominated by carbon and extent uncertainties, although the temperature CH${}_{\mathrm{4}}$ : C dependence is the dominant source of uncertainty in temperate latitudes. At latitudes $>$ 20${}^{\circ }$ N, wetland extent is the dominant source of uncertainty in mean CH${}_{\mathrm{4}}$ emissions, while the temperature CH${}_{\mathrm{4}}$ : C dependence accounts for $<$ 5 % of the dominant uncertainty attribution. Across tropical latitudes (23${}^{\circ }$ S–23${}^{\circ }$ N) and northern high latitudes ($>$ 45${}^{\circ }$ N), carbon decomposition is the dominant source of uncertainty in the timing of wetland CH${}_{\mathrm{4}}$ emissions.

We summarize the FE global error covariance structure as an error correlation matrix between mean monthly 2009–2010 emissions across boreal and Arctic ($>$ 55${}^{\circ }$ N) temperate (23–55${}^{\circ }$ N), tropical (23${}^{\circ }$ S–23${}^{\circ }$ N) and Southern Hemisphere ($<$ 23${}^{\circ }$ S) latitudes (Fig. 7); the error correlation matrix quantitatively summarizes similarities in the spatial and temporal patterns between ensemble members relative to the ensemble mean (see Appendix B for a description and interpretation). The FE error correlation matrix highlights the positively correlated ensemble member CH${}_{\mathrm{4}}$ emissions within each region, with larger correlations (generally Pearson's $r>\mathrm{0}$.8) between emissions separated by 1–2 months. Tropical emissions exhibit the largest overall temporal correlations ($r>\mathrm{0}$.5). Tropical emissions exhibit negative correlations against temperate emissions ($r<-\mathrm{0}$.3) and boreal and Arctic CH${}_{\mathrm{4}}$ emissions ($r<-\mathrm{0}$.1).

Mean 2009–2010 observed and GEOS-Chem forward model run CH${}_{\mathrm{4}}$ concentrations (with FE, EE, BL and GC wetland emissions) are broadly consistent on a latitudinal basis. The observed and modelled zonal atmospheric CH${}_{\mathrm{4}}$ concentration anomaly (relative to mean global 2009–2010 CH${}_{\mathrm{4}}$ concentrations) is shown in Fig. 8 (the zonal profile root mean square errors (RMSEs) are 6.5, 6.6, 8.4 and 9.2 ppb for FE, EE, BL and GC relative to the observed CH${}_{\mathrm{4}}$ anomaly zonal profile). Within the primary wetland CH${}_{\mathrm{4}}$ emission latitudes (10${}^{\circ }$ S–80${}^{\circ }$ N; Fig. 2), all mean CH${}_{\mathrm{4}}$ model estimates are within the mean standard deviation of observed CH${}_{\mathrm{4}}$, except for GC at $>$ 60${}^{\circ }$ N and all models at 80${}^{\circ }$ N.

Mean 2009–2010 CH${}_{\mathrm{4}}$ measurements and model CH${}_{\mathrm{4}}$ zonal anomalies ($\mathrm{\Delta }$CH${}_{\mathrm{4}})$ relative to the mean 2009–2010 global CH${}_{\mathrm{4}}$ concentration. The black dots denote the mean WDCGG network observed CH${}_{\mathrm{4}}$ concentrations within 5${}^{\circ }$ latitude bins; the grey envelope denotes the mean 2009–2010 standard deviation across all sites within 5${}^{\circ }$ latitude bins. The coloured symbols and error bars denote the GEOS-Chem equivalent model concentration statistics based on the FE and EE ensembles (this study), Bloom et al. (2012) (BL) and the GEOS-Chem emissions inventory (GC) wetland CH${}_{\mathrm{4}}$ emission datasets.

[Figure omitted. See PDF]

The symbol colours denote the monthly de-trended CH${}_{\mathrm{4}}$ Pearson's $r$ correlation (left column) for the model observations and RMSE (right column) for the FE (top row) and EE (bottom row) wetland CH${}_{\mathrm{4}}$ emissions (monthly CH${}_{\mathrm{4}}$ observations are from the WDCGG measurement site network). The $y$ axis denotes the difference between FE and EE and the model runs with Bloom et al. (2012) wetland CH${}_{\mathrm{4}}$ emissions (BL) and the GEOS-Chem wetland CH${}_{\mathrm{4}}$ emissions inventory (GC).

[Figure omitted. See PDF]

The median site-level correlation (Pearson's $r$) between the observed and modelled de-trended CH${}_{\mathrm{4}}$ concentrations (Fig. 9) is the highest for BL (0.75), followed by EE (0.74), FE (0.73) and GC (0.72). The median RMSEs between the observed and modelled de-trended CH${}_{\mathrm{4}}$ concentrations for FE (11.78 ppb) and EE (11.89 ppb) are lower than BL (12.42 ppb) and GC (13.27 ppb). FE and EE improvements (relative to GC and BL Pearson's $r$ and RMSE) are primarily in Northern Hemisphere high latitudes ($>$ 50${}^{\circ }$ N; Fig. 9). In the Southern Hemisphere extratropical latitudes ($<$ 23${}^{\circ }$ S), FE and EE exhibit a comparable performance relative to GC, while BL outperforms both FE and EE.

Discussion

Model limitations

Densely vegetated wetland areas are likely to amount to a large component of the global wetland CH${}_{\mathrm{4}}$ sources; a high carbon density (and high temperatures in the case of tropical wetlands) results in high CH${}_{\mathrm{4}}$ emissions under inundated conditions. However, satellite-derived observations of surface water area (Schroeder et al., 2015) are ill-equipped to observe densely vegetated wetland areas, as the passive microwave sensors become increasingly sensitive to vegetation moisture within high-biomass ecosystems (Sippel et al., 1994). For example, FE estimates of Amazon River basin wetland CH${}_{\mathrm{4}}$ emissions amount to 16–29 % (5th–95th percentiles) of the global wetland emissions source; the high biomass density in this region (Saatchi et al., 2011) may be a significant source of inundation area bias. Therefore, while we incorporate prior information on the mean and maximum wetland extent to scale the satellite-derived inundation fraction, we anticipate that errors in seasonal and inter-annual inundation variability are likely to be larger within densely vegetated wetland areas. We are optimistic that current and upcoming missions such as SMAP and BIOMASS (Entekhabi et al., 2010; Le Toan et al., 2011) combined with data integration approaches (Schroeder et al., 2015; Fluet-Chouinard et al., 2015) can potentially provide the additional constraints required to extend current inundation datasets and to improve current surface inundation detection capabilities.

The MsTMIP model ensemble provides a first-order estimate of the magnitude and variability of C decomposition within each 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ grid cell. Here we highlight four potentially major sources of error: (a) differences in aerobic : anaerobic turnover rates of major (labile and recalcitrant) C pools, (b) systematic differences in wetland and non-inundated area carbon uptake within each 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ grid cell, (c) systematic differences in dead organic matter C stocks and accumulation between wetland and non-inundated areas and (d) lateral flows of C into (or out of) wetland areas. Top-down estimates of seasonal and inter-annual terrestrial CO${}_{\mathrm{2}}$ fluxes (e.g. Liu et al., 2014) could be used to independently assess the validity of heterotrophic respiration from the MsTMIP models and CARDAMOM. In turn, top-down CH${}_{\mathrm{4}}$ and CO${}_{\mathrm{2}}$ flux retrievals and a range of in situ and regional-scale CH${}_{\mathrm{4}}$ flux estimates (Schriel-Uijl et al., 2011; Chang et al., 2014; and Budishchev et al., 2014, amongst others) can be combined to assess whether our empirical parameterization is able to capture regional, seasonal and inter-annual wetland CH${}_{\mathrm{4}}$ emission variability and their link to the broader terrestrial carbon cycle. Finally, in succession to the eddy covariance tower-site analyses of CO${}_{\mathrm{2}}$ respiration dependence on temperature (Mahecha et al., 2010), we anticipate that CH${}_{\mathrm{4}}$ eddy covariance measurements will provide critical site-level constraints on the temperature dependence of wetland CH${}_{\mathrm{4}}$ emissions.

Rice paddies likely amount to $<$ 20 % of wetland CH${}_{\mathrm{4}}$ emissions, and the majority of rice paddy areas are implicitly excluded from our analysis. GLOBCOVER distinguishes between natural and irrigated water bodies, and GLWD explicitly excludes rice paddy extents in China (which alone account for a large portion of global rice paddy CH${}_{\mathrm{4}}$ emissions). However, satellite-based inundation fraction retrievals are unable to distinguish the temporal variability in co-located agriculture and natural wetland inundation extent. Moreover, a 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ carbon cycle model resolution may be insufficient to resolve spatial differences in wetland and agricultural C cycling. The inadvertent inclusion of co-located rice CH${}_{\mathrm{4}}$ emissions is therefore a potential source of bias in our approach. We note that the distinction between wetland and rice CH${}_{\mathrm{4}}$ emissions has yet to be consistently addressed in global wetland CH${}_{\mathrm{4}}$ emission quantification efforts (see Bloom et al., 2010; Hodson et al., 2011; Melton et al., 2013, and the references therein).

CH${}_{\mathrm{4}}$ production in non-wetland freshwater bodies, such as very small ponds (Holgerson and Raymond, 2016), lakes (Wik et al., 2016) and rivers (Bastviken et al., 2011), is potentially a significant, albeit highly uncertain, term in the global CH${}_{\mathrm{4}}$ budget (Kirschke et al., 2013; Bridgham et al., 2013). Our approach implicitly accounts for non-wetland freshwater body emissions, since their extent is incorporated in grid-cell scaling factors (see Eq. 2). We recognize the challenge in explicitly distinguishing between wetlands and non-wetland freshwater body CH${}_{\mathrm{4}}$ emissions, as well as the associated physical and biogeochemical process controls. The quantitative distinction of CH${}_{\mathrm{4}}$ emissions from wetland and non-wetland freshwater extent remains challenging with the current spatial resolution ($\sim$ 25 km) of surface inundation retrievals (Prigent et al., 2007; Schroeder et al., 2015). Equally, the current global carbon cycle model resolutions ($\ge$ 0.5${}^{\circ }$) are insufficient to resolve spatial variations in heterotrophic processes across $\le$ 1 km of wetland and freshwater land cover definitions (Lehner and Döll, 2004). Contingent on future resolution enhancements in surface inundation and carbon cycle models, we recommend further investigation of the adequate distinction and estimation of non-wetland freshwater CH${}_{\mathrm{4}}$ emissions for atmospheric CH${}_{\mathrm{4}}$ chemical transport modelling applications.

By constraining global emission estimates to the Saunois et al. (2016) model range, our approach does not challenge the global annual CH${}_{\mathrm{4}}$ source and uncertainty; rather, it places constraints on spatial and temporal wetland CH${}_{\mathrm{4}}$ source variability. Since the global uncertainty (166 Tg CH${}_{\mathrm{4}}$ yr${}^{-\mathrm{1}}$; range $=$ $\pm$25 %) is substantially smaller than regional uncertainties (spanning a factor of 2–156; see Fig. 4), new or improved constraints on the global wetland CH${}_{\mathrm{4}}$ source are unlikely to significantly influence our regional CH${}_{\mathrm{4}}$ flux confidence range estimates. We therefore anticipate that wetland CH${}_{\mathrm{4}}$ in situ measurements and associated up-scaling efforts (e.g. Olefeldt et al., 2013; Turetsky et al., 2014; and Sjörgesten et al., 2014, amongst others) will undoubtedly become critical for reducing emission and process uncertainty in future wetland emission model ensembles.

Applications

Based on comparisons against measured CH${}_{\mathrm{4}}$ concentrations and a range of regional and global CH${}_{\mathrm{4}}$ emission estimates (Figs. 2–4, 7–8), we have shown that the FE and EE wetland CH${}_{\mathrm{4}}$ emission ensembles robustly represent the global magnitude and uncertainty of wetland CH${}_{\mathrm{4}}$ emissions. The combined ensemble configurations of inundation extent, carbon decomposition and temperature dependence have provided a characterization of the dominant source of uncertainty in global wetland CH${}_{\mathrm{4}}$ estimates (Fig. 6). The approach outlined here provides a framework for producing prior emission estimates and the associated uncertainty. The error covariance structure, along with the CH${}_{\mathrm{4}}$ observation system capabilities (Wecht et al., 2014b), can be used to devise an optimal strategy for spatially and/or temporally aggregating CH${}_{\mathrm{4}}$ fluxes in an atmospheric inversion framework. The retrieved CH${}_{\mathrm{4}}$ flux from assimilating atmospheric CH${}_{\mathrm{4}}$ observations in an inverse modelling framework (e.g. Fraser et al., 2013) could in turn provide a quantitative constraint on the wetland ensemble; the FE and EE model members can be treated as an ensemble of probable biogeochemical process hypotheses that can be weighted against atmospheric constraints. In contrast to conventional wetland CH${}_{\mathrm{4}}$ emission estimates (Riley et al., 2011; Pickett-Heaps et al., 2011) and model inter-comparisons (Melton et al., 2013), top-down CH${}_{\mathrm{4}}$ flux estimates can constrain the joint probability distribution of FE carbon models, wetland extent parameterizations and temperature dependencies. We note that due to the smaller ensemble size and the use of only one carbon model (see Table 1), the 2001–2015 EE emission variability should be interpreted with caution and, where possible, evaluated against the FE ensemble during the 2009–2010 period.

We anticipate extensions of the FE beyond the 2009–2010 time period, contingent on the extensions of the MsTMIP and SWAMPS datasets beyond 2010 and 2012, respectively. In light of continued satellite CH${}_{\mathrm{4}}$ retrievals from GOSAT (Parker et al., 2011; Butz et al., 2011) and upcoming satellite CH${}_{\mathrm{4}}$ measurements from the TROPOMI on-board ESA Sentinel-5 precursor (Veefkind et al., 2012), we anticipate that the FE and EE datasets will provide key process-based prior knowledge in future atmospheric CH${}_{\mathrm{4}}$ inversions.

The full ensemble (FE) and extended ensemble (EE) datasets (Bloom et al., 2017) are available from the Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC; http://dx.doi.org/10.3334/ORNLDAAC/1502). MsTMIP monthly 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ datasets (Huntzinger et al., 2016) were obtained from http://nacp.ornl.gov/MsTMIP.shtml. ERA-Interim datasets were obtained from http://apps.ecmwf.int/datasets/data/interim-full-mnth. CARDAMOM 2001–2010 heterotrophic respiration outputs are available at http://datashare.is.ed.ac.uk/handle/10283/875; the complete 2001–2015 heterotrophic extension outputs are included in the Supplement. Inundation datasets were obtained from http://wetlands.jpl.nasa.gov. The GLWD dataset was obtained from http://gcmd.gsfc.nasa.gov. The GLOBCOVER dataset was obtained from http://due.esrin.esa.int. The WDCGG data were obtained from http://ds.data.jma.go.jp/gmd/wdcgg. The Surface WAter Microwave Product Series inundation dataset (described by Schroeder et al., 2015) was obtained from http://wetlands.jpl.nasa.gov (accessed on 5 June 2014); European Centre for Medium-Range Weather Forecasts reanalysis (ECMWF ERA-Interim) synoptic monthly means were downloaded from http://apps.ecmwf.int. The code used to generate the FE and EE datasets is included in the Supplement.

CARDAMOM extension

CARDAMOM heterotrophic respiration was derived from the Bloom et al. (2016) global terrestrial C cycle 1${}^{\circ }$ $\times$ 1${}^{\circ }$ analysis. CARDAMOM-retrieved C-state and process variables for the period 2001–2010 were used to run the ecosystem carbon balance model DALEC2 (Bloom and Williams, 2015) to span 2001–2015. The 2011–2015 ERA-Interim meteorological drivers and MODIS-burned area were obtained as described by Bloom et al. (2016). The CARDAMOM output consists of 4000 heterotrophic respiration realizations at each monthly time step; for each time step, we use the median CARDAMOM heterotrophic respiration output. We downscale the data to a 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ resolution using a nearest neighbour interpolation.

Error correlation structure

We derive the model ensembles' space–time $n$ $\times$ $n$ error correlation matrix $\mathbf{M}$ as follows: $${\displaystyle }{M}_{ij}=\text{cor}({\mathit{A}}_{i,\ast },{\mathit{A}}_{j,\ast }),$$ where $n$ corresponds to the number of space and time wetland CH${}_{\mathrm{4}}$ emission aggregations, and $i$ and $j$ span 1 to $n$. $A_{i,m}$ and $A_{j,m}$ correspond to the total CH${}_{\mathrm{4}}$ flux for model $m$ within the $i$th and $j$th space–time aggregations (i.e. total wetland CH${}_{\mathrm{4}}$ emissions within a given time and area); ${\mathit{A}}_{i,\ast }$ and ${\mathit{A}}_{j,\ast }$ are 1 $\times$ $N$ vectors, where $N$ is the number of models within the ensemble. The “cor()” operator denotes the Pearson's correlation coefficient between the two bracketed vectors. For Fig. 7, we aggregated model wetland CH${}_{\mathrm{4}}$ emissions for each month across four zonal bands: boreal and Arctic ($>$ 55${}^{\circ }$ N) temperate (23–55${}^{\circ }$ N), tropical (23–23${}^{\circ }$ N) and Southern Hemisphere ($<$ 23${}^{\circ }$ S). A perfect correlation between the $i$th and $j$th indices ($M_{ij}=\mathrm{1}$) indicates that the models are consistently over- or under-predicting CH${}_{\mathrm{4}}$ emissions at times-and-locations $i$ and $j$ relative to the ensemble mean; a perfect anti-correlation ($M_{ij}=-\mathrm{1}$) indicates that the models are consistently over-predicting CH${}_{\mathrm{4}}$ emissions at time-and-location $i$ and consistently under-predicting CH${}_{\mathrm{4}}$ emissions at time-and-location $j$ (relative to the ensemble mean) and vice versa.

Dominant process uncertainty

We quantify the dominant process uncertainty of wetland CH${}_{\mathrm{4}}$ emission state variables ($s=\mathrm{1}$–3; (1) maximum emission month, (2) mean CH${}_{\mathrm{4}}$ emissions and (3) seasonal variability in terms of standard deviation) to wetland emission controls ($e=\mathrm{1}$–3; (1) model carbon decomposition, (2) CH${}_{\mathrm{4}}$ : C temperature dependence and (3) wetland extent parameterization) at location $x$ as follows: $${\displaystyle }{R}_{x,s,e}=\sum _{c=\mathrm{1}}^N{\displaystyle \frac{max\left({\mathit{M}}_{x,s,{\mathit{m}}_c}\right)-min\left({\mathit{M}}_{x,s,{\mathit{m}}_c}\right)}{N}},$$ where $R_{x,s,e}$ is the mean range of state variable $s$ across the ensemble given a fixed emission control $e$. ${\mathit{M}}_{x,s,\ast }$ is a vector of all ensemble member state variables $s$ at location $x$, ${\mathit{m}}_c$ denotes the indices of the ensemble subset driven by $c$th emission control $e$, and $N$ values are the number of configurations for each $e$ (the ensemble configuration details are show in Table 1). The “max()” and “min()” functions denote the maximum and minimum elements of the bracketed vectors. For example, $R_{\mathrm{100},\mathrm{3},\mathrm{1}}$ is the mean range of seasonal CH${}_{\mathrm{4}}$ variability ($s=\mathrm{3}$) for a fixed carbon model configuration ($e=\mathrm{1}$) at the 100th grid cell ($x=\mathrm{100}$). We attribute the zonal dominant uncertainty of state variable $s$ to emission control $e$ as $${\displaystyle }{P}_{z,s,e}={\displaystyle \frac{\sum _{x_z}{r}_{x_z,s,e}{F}_{x_z}}{\sum _{x_z}{F}_{x_z}}}\times \mathrm{100}\mathit{\%},$$ where $x_z$ values are the pixels $x$ within a 5${}^{\circ }$ zonal band $z,F_{x_z}$ is the mean 2009–2010 area-integrated CH${}_{\mathrm{4}}$ flux (Eq. 1 in the main text) and $r_{x_z,s,e}=\mathrm{1}$ if $R_{x_z,s,e}=min\left({\mathit{R}}_{x_z,s,\ast }\right)$; otherwise, $r_{x_z,s,e}=\mathrm{0}$. Therefore, $e$ is the largest source of uncertainty when the mean range in state variable $s$ is the smallest for a fixed $e$. $P_{z,s,e}$ denotes the percentage of zonal band $z$ where emission control $e$ is the greatest source of uncertainty for each $s$.

The Supplement related to this article is available online at https://doi.org/10.5194/gmd-10-2141-2017-supplement.

The authors declare that they have no conflict of interest.

Acknowledgements

Part of this research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. Funding for this study was provided through a NASA Carbon Monitoring System grant (#NNH14ZDA001N-CMS). Funding for the Multi-scale synthesis and Terrestrial Model Intercomparison Project (MsTMIP; http://nacp.ornl.gov/MsTMIP.shtml) activity was provided through a NASA ROSES grant (#NNX10AG01A). Data management support for preparing, documenting and distributing the model driver and output data was performed by the Modeling and Synthesis Thematic Data Center at Oak Ridge National Laboratory (ORNL; http://nacp.ornl.gov) with funding through a NASA ROSES grant (#NNH10AN681). Finalized MsTMIP data products are archived at the ORNL DAAC (http://daac.ornl.gov).Edited by: J. Williams Reviewed by: two anonymous referees