Evaluation of CH4MOD_wetland and Terrestrial Ecosystem Model (TEM) used to estimate global CH₄ emissions from natural wetlands

Wetlands are the largest and most uncertain natural sources of atmospheric methane (CH₄). Several process-based models have been developed to quantify the magnitude and estimate spatial and temporal variations in CH₄ emissions from global wetlands. Reliable models are required to estimate global wetland CH₄ emissions. This study aimed to test two process-based models, CH4MOD_wetland and Terrestrial Ecosystem Model (TEM), against the CH₄ flux measurements of marsh, swamp, peatland and coastal wetland sites across the world; specifically, model accuracy and generality were evaluated for different wetland types and in different continents, and then the global CH₄ emissions from 2000 to 2010 were estimated. Both models showed similar high correlations with the observed seasonal/annual total CH₄ emissions, and the regression of the observed versus computed total seasonal/annual CH₄ emissions resulted in R² values of 0.81 and 0.68 for CH4MOD_wetland and TEM, respectively. The CH4MOD_wetland produced accurate predictions for marshes, peatlands, swamps and coastal wetlands, with model efficiency (EF) values of 0.22, 0.52, 0.13 and 0.72, respectively. TEM produced good predictions for peatlands and swamps, with EF values of 0.69 and 0.74, respectively, but it could not accurately simulate marshes and coastal wetlands (EF <0). There was a good correlation between the simulated CH₄ fluxes and the observed values on most continents. However, CH4MOD_wetland showed no correlation with the observed values in South America and Africa. TEM showed no correlation with the observations in Europe. The global CH₄ emissions for the period 2000–2010 were estimated to be 105.31 ± 2.72 Tg yr⁻¹ by CH4MOD_wetland and 134.31 ± 0.84 Tg yr⁻¹ by TEM. Both models simulated a similar spatial distribution of CH₄ emissions globally and on different continents. Marshes contribute 36 %–39 % of global CH₄ emissions. Lakes/rivers and swamps are the second and third greatest contributors, respectively. Other wetland types account for only approximately 20 % of global emissions. Based on the model applicability, if we use the more accurate model, i.e., the one that performs best as evidenced by a higher model efficiency and a lower model bias, to estimate each continent and wetland type, we obtain a new assessment of 116.99–124.74 Tg yr⁻¹ for the global CH₄ emissions for the period 2000–2010. Our results imply that performance at a global scale may conceal model uncertainty. Efforts should be made to improve model accuracy for different wetland types and regions, particularly hotspot regions, to reduce the uncertainty in global assessments.

1 Introduction

Atmospheric methane (CH₄) is the second most prevalent human-induced greenhouse gas (GHG) after carbon dioxide (CO₂). Its radiative forcing effect is 28 times greater than that of CO₂ on a 100-year horizon (Myhre et al., 2013). The radiative forcing attributed to CH₄ has been re-evaluated by the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5) and was reported to be almost twice as high as the value reported in the Fourth Assessment Report (AR4), with values of 0.97 W m⁻² versus 0.48 W m⁻², respectively (Myhre et al., 2013). This estimate considers that the emission of CH₄ leads to an increase in ozone production, stratospheric water vapor and CO₂, which can affect its own lifetime (Boucher et al., 2009; Myhre et al., 2013; Shindell et al., 2012).

The growth rate of the atmospheric CH₄ concentration has varied in different historical periods. There was an exponential increase from preindustrial times to the 1980s. The growth rate decreased after the 1980s and was close to zero from 1999 to 2006; then, the growth rate resumed strong growth in the period of 2007–2017 (Dlugokencky et al., 2009, 2016; Nisbet et al., 2019). However, the causes that drive the variations in growth rate remain unclear due to the uncertainties in estimating CH₄ emissions and sinks (Ghosh et al., 2015; Saunois et al., 2016; Nisbet et al., 2019; Dalsøren et al., 2016).

Integrated at the global scale, wetlands are the largest and most uncertain source of CH₄ emitted to the atmosphere (Kirschke et al., 2013; Saunois et al., 2016). These emissions represent approximately 30 % of the total CH₄ input (Saunois et al., 2016). Bottom-up and top-down approaches are popular methods for estimating global CH₄ emissions from natural wetlands. Top-down approaches are based on inverse models (e.g., Bousquet et al., 2006; Fraser et al., 2013; Meirink et al., 2008; Tsuruta et al., 2017; Bruhwiler et al., 2014), which determine “optimal” surface fluxes that best fit atmospheric CH₄ observations given an atmospheric transport model including chemistry, prior estimates of fluxes and their uncertainties (Kirschke et al., 2013). Bottom-up approaches use process-based models that describe the relationship between the environmental factors and the processes of CH₄ production, oxidation and emission using mathematical equations (e.g., Li et al., 2010; Zhu et al., 2013, 2014; Zhang et al., 2002; Walter and Heimann, 2000; Tian et al., 2015; Riley et al., 2011; Meng et al., 2012; Zhuang et al., 2006).

Recent studies related to the bottom-up approach have used an ensemble of process-based models driven by the same climate forcing to estimate the global CH₄ emissions from natural wetlands. For example, the Wetland and Wetland CH₄ Intercomparison of Models Project (WETCHIMP) used 10 land surface models and estimated global CH₄ emissions of 190 ± 76 Tg CH₄ yr⁻¹ for the 1993–2004 period (Melton et al., 2013). In the following year, Kirschke et al. (2013) assessed a large emission range of 142–287 Tg CH₄ yr⁻¹ from 1980 to 2010. Saunois et al. (2016) and Poulter et al. (2017) estimated global emissions of 153–227 Tg CH₄ yr⁻¹ for the decade 2003–2012 and 184 ± 22 Tg CH₄ yr⁻¹ for the decade 2000–2012 using ensemble process-based models (Poulter et al., 2017). Saunois et al. (2016) suggested that approximately 70 % of the uncertainty was due to model structures and parameters.

Natural wetland ecosystems are greatly heterogeneous on a global scale. Wetlands vary widely by continent with respect to area and type (Kingsford et al., 2016; Keddy, 2010). Some wetland types have higher emissions, while some emit less CH₄; this difference is because the processes of controls on CH₄ cycling differ among wetland types (Bridgham et al., 2013). For example, sedge-dominated marshes or fens often emit higher CH₄ fluxes, because sedges can increase methanogenic substrates as part of their plant productivity and promote CH₄ transportation through their soft aerenchyma and lacunae tissues (McEwing et al., 2015; Jitka et al., 2017; Bhullar et al., 2013; Joabsson and Christensen, 2001; Kwon et al., 2017; King et al., 2002; Chanton, 2005). Bog soils with anaerobic incubations emit little CH₄ due to the particularly high CO₂:CH₄ ratios of the end products of anaerobic carbon (Bridgham et al., 1998; Galand et al., 2010; Keller and Bridgham, 2007). Coastal wetlands with high salinity usually emit less CH₄ than other wetlands, because the sulfate in seawater inhibits CH₄ production (Bartlett et al., 1985; Delaune et al., 1983; Li et al., 2016; Poffenbarger et al., 2011).

Model evaluation is a core part of model development and testing (Bennett et al., 2013). Based on the model evaluation, the modeler must be confident that the model will fulfill its purpose (Bennett et al., 2013; Rykiel, 1996). If applying process-based models for global-scale CH₄ estimations, it is necessary to evaluate its performance in different wetland types and regions. This process is also helpful for confirming the source of uncertainties and improving the model. However, previous studies have always focused on global assessments and have overlooked model performance in different wetland types or regions, which may have induced high uncertainties (Poulter et al., 2017; Saunois et al., 2016; Kirschke et al., 2013; Melton et al., 2013). CH4MOD_wetland (Li et al., 2010) and the Terrestrial Ecosystem Model (TEM) (Zhuang et al., 2004, 2007, 2013; Melillo et al., 1993) are two established process-based models that can be used to simulate regional and global wetland CH₄ emissions. Both models have been validated at specific sites (Zhu et al., 2013; Li et al., 2010, 2017). However, we do not have information on the accuracy and applicability of the models for different wetland types and on different continents. The objectives of this study were to comprehensively evaluate the model performances of CH4MOD_wetland and TEM for different wetland types and on different continents and then to use the models to estimate global CH₄ emissions from natural wetlands.

2 Methods and materials

The performance evaluation should clearly depend on the model objectives (Bennett et al., 2013). The models considered in this study aim to estimate the annual emissions from global wetlands. Therefore, the accuracy and applicability of the model in simulating annual CH₄ emissions for different wetland types and continents are very important in a performance evaluation. Several process-based models have been developed in recent decades (Xu et al., 2016). Some models are simple semiempirical models that focus on the biochemical processes of CH₄ production, oxidation and emission, e.g., Walter's model (Walter et al., 1996; Walter and Heimann, 2000), CASA (Potter, 1997) and CH4MOD_wetland (Li et al., 2010). This kind of model requires simple inputs and parameters and is easily extrapolated to a regional scale. Other models are based on more complex land ecosystem models coupled to the CH₄ processes module, such as Community Land Model 4 Methane model (CLM4Me), Organising Carbon and Hydrology In Dynamic Ecosystems model (ORCHIDEE), Sheffield Dynamic Global Vegetation Model (SDGVM) and Terrestrial Ecosystem Model (TEM). These models describe complex ecosystem processes and require more inputs and parameters. In this study, we chose CH4MOD_wetland and TEM to compare the model performance of a simple easy-to-run model and a sophisticated land ecosystem model. Moreover, both models have been validated at the site scale, but no comprehensive accuracy analysis in different continents or for various wetland types has been done before. We collected CH₄ flux measurements from 43 wetlands spanning the main wetland types in the world from peer-reviewed literature (Table 1). A set of statistical methods was used to comprehensively evaluate the performance of CH4MOD_wetland and TEM in different wetland types and on different continents. Finally, we extrapolated both models to estimate the global CH₄ emissions from 2000 to 2010.

Table 1 Description of observation wetland sites. Observation period (YYYY.MM).

^a Tundra. ^b Tidal marsh. ^c Mangrove. ^d We used the reported average yearly CH₄ flux of the experimental year or period from the literature. ^e Wetland sites used for calibration. ^f These swamp sites do not have plant species
information in the literature.

2.1 Model overview

2.1.1 CH4MOD_wetland

The CH4MOD_wetland model is a process-based biogeophysical model used to simulate the processes of CH₄ production, oxidation and emission from natural wetlands (Li et al., 2010). The model was established based on CH4MOD, which is used to predict CH₄ emissions from rice paddies (Huang et al., 1997). In CH4MOD_wetland, we focused on the differences in the supply of methanogenic substrates between natural wetlands and rice paddies. Methanogenic substrates are derived from root exudates, the decomposition of plant litter and soil organic matter. The methane production rates were determined based on the methanogenic substrates and the influence of environmental factors, including soil temperature, soil texture and soil redox potential. Additionally, we incorporated the influence of salinity on CH₄ production to improve the model performance for coastal wetlands (Li et al., 2016). Inputs to the CH4MOD_wetland model include the daily air and soil temperature, water table depth, annual aboveground net primary productivity (ANPP), soil sand fraction, soil organic matter, bulk density, and soil salinity. The outputs are the daily and annual CH₄ production and emissions. We used the TOPMODEL hydrological model to simulate the water table depth as the inputs of CH4MOD_wetland (Bohn et al., 2007; Li et al., 2015, 2019a; Zhu et al., 2013; Beven and Kirkby, 1979).

The main parameters that must be calibrated in CH4MOD_wetland include the vegetation index (VI), which was used to quantify the different capacities for producing root exudates of the various plant species; the fraction of plant-mediated transport available (T_veg); the fraction of CH₄ oxidized during plant-mediated transport (P_ox); the proportion of belowground net primary productivity (BNPP) to the total net primary productivity (NPP) (f_r); the fraction of nonstructural component in plant litter (F_N) (Table S1 in the Supplement); and the empirical constant of the influence of salinity. The model parametrization and main parameters are described in Sect. S1 in the Supplement.

2.1.2 TEM

TEM is another process-based biogeochemical model that couples carbon, nitrogen, water and heat processes in terrestrial ecosystems to simulate ecosystem carbon and nitrogen dynamics (Melillo et al., 1993; Zhuang et al., 2007, 2013). The methane dynamics module was first coupled within TEM by Zhuang et al. (2004) to explicitly simulate the process of methane production (methanogenesis), oxidation (methanotrophy) and transport between the soil and the atmosphere. Methane production is assumed to occur only in saturated zones and is regulated by organic substrate, soil thermal conditions, soil pH, and soil redox potentials; methane oxidation, which occurs in the unsaturated zone, depends on the soil methane and oxygen concentrations, temperature, moisture and redox potential. Methane transport is described by three pathways in TEM: (1) diffusion through the soil profile, (2) plant-aided transport and (3) ebullition. TEM has also been coupled with TOPMODEL (Zhu et al., 2013). The model calibration of the TEM is well documented in Sect. S2 and Table S2.

2.2 Site information and data sources

2.2.1 Site information

We collected data from 43 wetland sites across the world (Table 1). The wetland sites included 6 marsh sites, 25 peatland sites, 8 swamp sites and 4 coastal wetland sites. Among the wetland sites, 7 sites are distributed in Europe (EU), 11 sites are distributed in Asia (AS), 2 sites are distributed in Africa (AF), 4 sites are distributed in South America (SA) and 19 sites are distributed in North America (NA). The observations were from the late 1980s to the 2010s. The observation periods covered either a growing season or a whole year (Table 1). We calculated the total amount of CH₄ emissions during the growing season or the whole year as the observed seasonal/annual CH₄ emissions. For most of the wetland sites, the total amount of seasonal/annual CH₄ emissions during the observation period was calculated by summing the daily observations. Gaps in the CH₄ emission measurements were filled by linear interpolation between two adjacent days of observations. For a few wetland sites, the observed seasonal/annual CH₄ emissions were directly obtained from the literature. More details about the location, vegetation and observation periods are described in Table 1.

2.2.2 Wetland map

The global wetland distributions of different wetland types were based on the Global Lakes and Wetlands Database (GLWD-3; https://www.worldwildlife.org/publications/global-lakes-and-wetlands-database-lakes-and-wetlands-grid-level-3, last access: 10 October 2019) (Lehner and Döll, 2004) (Fig. 1). According to GLWD-3, the wetland types include (1) lakes, (2) reservoirs, (3) rivers (we combined lakes, reservoirs and rivers as a single wetland type, hereafter referred to as lakes/rivers), (4) freshwater marsh and floodplain (hereafter referred to as marsh), (5) swamp forest and flooded forest (hereafter referred to as swamp), (6) coastal wetland, (7) saline wetland (we combined coastal wetland and saline wetland as a single wetland type, hereafter referred to as coastal wetland), (8) bog, fen and mire (hereafter referred to as peatland), (9) intermittent wetland and (10) no-specific wetland. All of the observed sites (Table 1) are distributed on the wetland map (Fig. 1).

Figure 1 Wetland site distribution (Table 1) and global wetland maps of GLWD-3 (Lehner and Döll, 2004).

The global wetland area (excluding rivers) was estimated by the “Global Review of Wetland Resources and Priorities for Wetland Inventory (GRoWI)” as 530–570 Mha (Spiers, 1999). We used an average value, as the wetland area excluded rivers in this study. The global wetland area of rivers was based on GLWD-3. Therefore, we assumed that the global wetland area was 584 Mha, which represented the wetland area for the period from 2000 to 2010. The cartography-based GLWD-3 data provide a global distribution of natural wetlands at a 30 s resolution. Then, we aggregated the merged map up to $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ (latitude × longitude) grids. The wetland area (excluding rivers) in each pixel was adjusted by the ratio of the global wetland area estimated by GRoWI and by GLWD-3.

2.2.3 Driver data

The input climate data for the models include the daily air temperature, precipitation, cloudiness and vapor pressure. The historical daily climate data were developed from the latest monthly datasets of the Climatic Research Unit (CRU TS 3.10) of the University of East Anglia in the United Kingdom (Harris et al., 2014).

The soil properties needed by the CH4MOD_wetland model include soil texture (percentage of sand in the soil), bulk density, soil organic carbon content, soil temperature and soil moisture. The additional information needed by TEM includes the percentage of silt and clay in the soil, soil pH, and site elevation. The soil texture data were derived from the soil map of the Food and Agriculture Organization (FAO) (FAO, 2012). The soil organic carbon content and the reference bulk density of wetland soils were retrieved from the Harmonized World Soil Database (HWSD) (FAO, 2008) by masking the HWSD with the Global Lakes and Wetlands Database (GLWD) (Lehner and Döll, 2004). The daily soil temperature data were estimated by TEM from spatially interpolated climate data. The daily soil moisture driving CH4MOD_wetland coupled with TOPMODEL was developed from the monthly dataset (http://www.cpc.ncep.noaa.gov/soilmst/leaky_glb.htm. last access: 10 October 2019) by temporal linear interpolation (Fan and van den Dool, 2004). The soil pH was also derived from the global soil property dataset of the International Geosphere-Biosphere Programme (IGBP) (Carter and Scholes, 2000).

The vegetation map of the IGBP was referenced to specify the vegetation parameters for CH4MOD_wetland (Table S1) and TEM. The map was derived from the IGBP Data and Information System (DIS) DISCover Database (Belward et al., 1999; Loveland et al., 2000). The 1 km × 1 km DISCover dataset was reclassified into the TEM vegetation classification scheme and then aggregated into $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ grids. The annual ANPP used to drive CH4MOD_wetland was from the output of TEM.

For CH4MOD_wetland, a high-resolution topographic wetness index dataset (Marthews et al., 2015) was used to calculate the changes in the water table. Global salinity data were obtained from the World Ocean Atlas 2009 (Antonov et al., 2010). We also used 1 km × 1 km global elevation data derived from the Shuttle Radar Topography Mission (SRTM) (Farr et al., 2007). The above data were resampled to $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ grids to match the resolution of the other input data.

2.3 Model evaluation

We compared the observed seasonal/annual CH₄ emissions from the wetland sites (Table 1) and the simulated CH₄ emissions at the $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ grid scale for the same period (described in Sect. 2.4). The statistics include the determination coefficient (R²), the root-mean-square error (RMSE), the mean deviation (RMD), the model efficiency (EF) and the coefficient of determination (CD) were used to evaluate model performance on a global scale, a continental scale and for each wetland type. Because of the limited number of sites in Africa and South America, we combined the two continents together.

Two simulations with the same RMSE values may not be considered equivalent, because the distribution of the error among the sources may not be the same (Allen and Raktoe, 1981). We further analyzed the source of the model errors by decomposing it into three components: the mean bias from the modeling procedure (U_M), the errors due to regression (U_R) and the errors due to random disturbances (U_E) (Allen and Raktoe, 1981). The detailed description and the equations used to calculate these statistics are described in Sect. S3.

2.4 Model extrapolation

CH4MOD_wetland and TEM were used to simulate the CH₄ emissions from global wetlands at a spatial resolution of $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$. We established spatially explicit data for climate, soils, vegetation, land use and other environmental inputs at a $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ spatial resolution to facilitate the models at the global scale. Both process-based models were conducted for the period of 1980–2010 in each pixel to simulate the temporal spatial variations in CH₄ fluxes. In this study, we focused only on the total CH₄ emissions for the period 2000–2010, because we assumed that the wetland map represented the distribution of natural wetlands during this period. The total CH₄ emissions from the natural wetlands, excluding the lakes/rivers in each pixel, were calculated as the product of the CH₄ fluxes and the gridded wetland area. To make an overall global or continental CH₄ emissions assessment, we evaluated the CH₄ emissions from lakes/rivers using the IPCC Tier 1 method based on the CH₄ emissions factor (IPCC, 1996) and the area of lakes/rivers in each pixel.

We aggregated the gridded values and obtained the annual mean CH₄ emissions from each wetland type and each continent by CH4MOD_wetland combined with the IPCC Tier1 method (hereafter referred to as Method A) and TEM combined with the IPCC Tier1 method (hereafter referred to as Method B). In addition to the two global assessments Method A and Method B, we made two other assessments of global CH₄ emissions by choosing the more accurate model (Method C and Method D). Based on the model performance evaluation (Sect. 2.3), we found a more accurate model for each wetland type and each continent. In the Method C approach, we chose the CH₄ emissions from each continent simulated by the more accurate model. In Oceania, we used the average simulated result by CH4MOD_wetland and the TEM, because there was no wetland site on this continent (Table 1). We summed the CH₄ emissions from all continents and made an assessment of the global CH₄ emissions. In the Method D approach, we chose the CH₄ emissions from marsh, peatland, swamp and coastal wetlands simulated by the more accurate model. The CH₄ emissions are from intermittent wetlands and nonspecific wetlands (no-specific wetlands), which were used as the average result by CH4MOD_wetland and TEM. The CH₄ emissions from lakes/rivers were based on the IPCC Tier 1 method. We summed the CH₄ emissions from all wetland types and assessed the global CH₄ emissions.

3 Results

3.1 Model evaluation

3.1.1 Model evaluation for global wetland sites

Figure 2 shows the correlation of the modeled versus observed total amount of seasonal/annual CH₄ emissions by CH4MOD_wetland (Fig. 2a) and the TEM (Fig. 2b). The regression of the observed versus computed total seasonal/annual CH₄ emissions by CH4MOD_wetland (Fig. 2a) resulted in an R² of 0.81, with a slope of 1.17 and an intercept of −1.93 g m⁻² (n=58, p<0.001). The regression of the observed versus computed total seasonal/annual CH₄ emissions by TEM (Fig. 2b) resulted in an R² of 0.68, with a slope of 0.74 and an intercept of 4.77 g m⁻² (n=58, p<0.001). These results indicated that the variations in the CH₄ emissions between sites and in different years could be delineated by both process-based models.

Figure 2 Regression of simulated against observed total amount of seasonal/annual CH₄ emissions from global wetland sites by CH4MOD_wetland (a) and TEM (b). The horizontal bars are the standard errors from the sampling replicates at the wetland site. The red line is the regression line of simulated vs. observed between modeled and observed values. The blue line is the prediction correspondence. The dashed line is the 1 : 1 line.

Table 2 Model performance for CH4MOD_wetland and TEM for different continents and wetland types.

NS represents no significant correlation.

The statistics of the model performance of seasonal/annual CH₄ emissions (Table 2) indicated that both process-based models had the capability to simulate seasonal/annual CH₄ emissions from natural wetlands on a global scale (EF =0.65 for CH4MOD_wetland and EF =0.68 for TEM). However, a discrepancy still existed between the simulated and observed seasonal/annual CH₄ emissions (RMSE =67.00 % for CH4MOD_wetland and RMSE =63.58 % for TEM). For CH4MOD_wetland, the source of the errors was mainly from the regression error and random error, while for TEM the errors were mainly due to random disturbances (Table 2). Both models slightly overestimated the seasonal/annual CH₄ emissions on a global scale, with RMD values of ∼4 % (Table 2).

3.1.2 Model evaluation for different continents

We further analyzed the model predictions by CH4MOD_wetland and TEM among different continents (Fig. 3, Table 2). There was a good correlation between the simulated seasonal/annual CH₄ emissions and the observed values on most of the continents by the two models. The R² varied between 0.35 (Fig. 3e) and 0.94 (Fig. 3c) for CH4MOD_wetland and between 0.26 (Fig. 3d) and 0.80 (Fig. 3h) for TEM. The CH4MOD_wetland model yielded more accurate predictions in Asia and North America, with EFs of 0.93 and 0.57, respectively (Fig. 3b and a, Table 2), than in South America and Africa (EF <0 in Table 2) (Fig. 3g). TEM yielded more accurate predictions in North America and South America/Africa than CH4MOD_wetland, with EF values of 0.76 and 0.53, but performed poorly in Europe (EF <0 in Table 2). CH4MOD_wetland underestimated the observed emissions (RMD $=-\mathrm{12.64}$ %) in Asia and Europe (RMD $=-\mathrm{29.91}$ %) (Table 2). TEM overestimated the CH₄ emissions in South America/Africa (RMD =15.31 %) and slightly underestimated the CH₄ emissions in North America (RMD $=-\mathrm{2.86}$ %) (Table 1). Random error was the main contributor to the model errors in Asia and Europe in CH4MOD_wetland and in Asia, North America and South America/Africa in TEM (Table 2). However, the regression error contributed most to the model errors in North America in CH4MOD_wetland (Table 2).

Figure 3 Regression of simulated against observed total amount of seasonal/annual CH₄ emissions from North American wetland sites by CH4MOD_wetland (a) and TEM (b), from Asian wetland sites by CH4MOD_wetland (c) and TEM (d), from European wetland sites by CH4MOD_wetland (e) and TEM (f), and from South American and African wetland sites by CH4MOD_wetland (g) and TEM (h). The horizontal bars are the standard errors from the sampling replicates at the wetland site. The blue line is the prediction correspondence. The dashed line is the 1 : 1 line.

3.1.3 Model evaluation for different wetland types

Figure 4 shows the regressions of the simulated values against the observed total amount of seasonal/annual CH₄ emissions from the different wetland types. Regression analysis indicated that both models showed good performance in modeling seasonal/annual CH₄ emissions from the peatland sites (Fig. 4c and d). TEM showed a better model efficiency and a lower RMSE and RMD than the CH4MOD_wetland (Table 2) for peatland. For the other wetland types, CH4MOD_wetland showed good performance in simulating the seasonal/annual CH₄ emissions from coastal wetlands (EF =0.72), followed by marshes (EF =0.22) and swamps (EF =0.13) (Table 2). TEM showed poor performance for the marsh sites (EF $=-\mathrm{0.42}$) and coastal wetlands (EF $=-\mathrm{2.26}$) (Table 2); however, it showed good performance for the swamp sites (EF =0.74). There was no significant correlation (p>0.05) between the modeled and observed seasonal/annual CH₄ emissions from the marsh sites (Fig. 4b) and coastal wetland sites (Fig. 4h).

Figure 4 Regressions of simulated against observed total amount of seasonal/annual CH₄ emissions from marsh sites by CH4MOD_wetland (a) and TEM (b), from peatland sites by CH4MOD_wetland (c) and the TEM (d), from swamp sites by CH4MOD_wetland (e) and TEM (f), and from coastal wetland sites by CH4MOD_wetland (g) and TEM (h). The horizontal bars are the standard errors from the sampling replicates at the wetland site. The blue line is the prediction correspondence. The dashed line is the 1 : 1 line.

The errors by CH4MOD_wetland were mainly due to the regression error for marsh and peatland (Table 2). For coastal wetlands, the model bias contributed 24 %, the regression error contributed 30 %, and the random error contributed 47 % to the model errors (Table 2). The errors by the TEM were mainly due to the random error in peatland and swamps (Table 2).

3.2 Global CH₄ emissions from natural wetlands

3.2.1 Spatial pattern of global CH₄ emissions

The distribution of the simulated annual mean CH₄ fluxes and total CH₄ emissions for the period 2000–2010 showed similar patterns in CH4MOD_wetland and TEM (Fig. 5). The simulated latitudinal contributions of CH₄ fluxes were consistent between the two models (Sect. S4, Fig. 5a and b). Large emissions were found in South America, southern Africa, and near the border of Canada and the United States (Fig. 5c and d). The latitudinal sums of CH₄ emissions indicated that the strongest contribution came from the tropical zone (Fig. 5c and 5d). The latitudinal band of 10–0^∘ S contributed 22.77 and 23.23 Tg yr⁻¹ CH₄ in CH4MOD_wetland and TEM, which accounted for 22 % and 18 % of the global emissions, respectively. A secondary large peak was simulated in the 40–50^∘ N latitudinal band, with values of 14.64 and 16.66 Tg yr⁻¹ CH₄ according to CH4MOD_wetland and TEM, respectively. Generally, both models simulated a common decline in CH₄ emissions from lower latitudes to higher latitudes (Fig. 5c and d). The largest peak in CH₄ emissions was modeled in the 60–50^∘ W meridional band, with values of 11.63 Tg yr⁻¹ in CH4MOD_wetland (Fig. 5c) and 13.83 Tg yr⁻¹ in TEM (Fig. 5d). This peak corresponded to the longitudes of the Amazon in South America. Both models simulated secondary peaks in the 30–40^∘ E meridional band (Fig. 5c and d), which corresponded to the longitudes of the Congo in Africa.

Figure 5 Spatial pattern of annual mean CH₄ fluxes for 2000–2010, with latitudinal and longitudinal distributions of annual mean CH₄ fluxes by CH4MOD_wetland (a) and TEM (b). Spatial pattern of annual mean CH₄ emissions for 2000–2010, with latitudinal and longitudinal distributions of annual mean CH₄ emissions by CH4MOD_wetland (c) and TEM (d). The CH₄ fluxes and emissions are aggregated in steps of 10^∘.

3.2.2 CH₄ emissions from different continents and wetland types

Table 3 provides an overview of the CH₄ emissions from different continents and wetland types simulated by CH4MOD_wetland and TEM. A comparison of simulated CH₄ fluxes from different continents by CH4MOD_wetland and TEM showed that the three highest fluxes were modeled in South America, Africa and Asia (Table 3). TEM simulated higher CH₄ fluxes in Europe than in North America, but the CH4MOD_wetland simulations showed the opposite. For Oceania, the two models simulated similar fluxes.

Table 3 CH₄ simulations by CH4MOD_wetland and TEM for different continents and wetland types. All units are Tg CH₄ yr⁻¹ ±1σ, where the standard deviation represents the interannual variation in the model estimates.

^∗ The IPCC Tier 1 method was used to estimate the CH₄ emissions from lakes and rivers. The CH₄ emission factor was from IPCC (1996).

Both models simulated the same sequence of CH₄ fluxes: swamp, marsh, intermittent wetland, no-specific wetland, coastal wetland and peatland (Table 3). The simulated annual mean CH₄ fluxes from intermittent wetlands were almost equivalent in both models. For other wetland types, TEM simulated higher CH₄ fluxes than the CH4MOD_wetland model (Table 3). Both models simulated peak emissions in summer and lower emissions in winter for all wetland types except swamps (Fig. S1). Since large areas of swamps are distributed in the Southern Hemisphere (Fig. 1), higher and lower CH₄ emissions were simulated during March to May and June to August, respectively (Fig. S1).

The global CH₄ emissions simulated by TEM were 29 Tg yr⁻¹ higher than those simulated by CH4MOD_wetland (Table 3). This difference depended on the differences in the CH₄ fluxes and on the wetland area. The simulated results showed that half of this difference was attributed to marshes. South America contributed 30 % to this difference, because the simulated CH₄ fluxes differed greatly between TEM and CH4MOD_wetland (Table 3).

The two models simulated similar spatial distributions of the CH₄ emissions among different wetland types and continents (Table 3). Marshes emit higher CH₄ fluxes and have the largest area. Thus, marshes were the greatest contributor to global CH₄ emissions and contributed 36 %–39 % to global CH₄ emissions (Table 3). Lakes/rivers and swamps were the second and third contributors, respectively (Table 3). The CH₄ emissions from peatlands, coastal wetlands, intermittent wetlands and no-specific wetlands accounted for only approximately 20 % of the global emissions (Table 3).

Although North America accounted for 36 % of the global wetland area, it contributed only 22 %–23 % to global emissions (Table 3). In contrast, the wetland area in South America accounted for 15 % of the global area and contributed 25 %–26 % to global CH₄ emissions. Asia and Africa also accounted for approximately 20 % of global emissions. The lowest area and emissions were found in Oceania (Table 3).

3.2.3 Global CH₄ estimations

The global CH₄ emissions for the period 2000–2010 were estimated to be 105.31 ± 2.72 Tg yr⁻¹ by Method A and 134.31 ± 0.84 Tg yr⁻¹ by Method B. Based on the evaluation of model performance (Table 2), CH4MOD_wetland yielded the most accurate predictions for Asia and Europe, and TEM yielded the most accurate predictions for North America and South America/Africa. Using this combination, the global CH₄ emissions were estimated to be 124.74 ± 1.22 Tg by Method C. Similarly, in Method D, CH4MOD_wetland was used for simulations in marshes and coastal wetlands, and TEM was used for simulations in peatlands and swamps; as a result, the global wetland CH₄ emissions were estimated to be 116.99 ± 2.23 Tg.

4 Discussion

4.1 Generality of CH4MOD_wetland and TEM

A lack of correspondence between the model output and observations could be partly due to the observed flux data, e.g., the inevitable gap-filling of missing data points to determine the seasonal/annual total emissions (Kramer et al., 2002). The results showed differences between the observed and simulated CH₄ emissions by both CH4MOD_wetland and TEM on a global scale (Fig. 2) and a continental scale (Fig. 3) and for different wetland types (Fig. 4). The reliability of the observed flux data is not under discussion in this study. We evaluated only the model accuracy and applicability across different wetland types and continents.

On a global scale, both models fulfilled the criteria of sufficient accuracy for the ability to predict CH₄ fluxes (Table 2). However, this fuzzy analysis may miss some real model performance. For the model applicability on different continents, CH4MOD_wetland performed best in Asia, followed by North America and Europe. It performed poorly in South America/Africa, where swamps are more common (Table 2). TEM performed best in North America, followed by South America/Africa and Asia. It performed poorly in Europe (Table 2). Each continent has different main wetland types; thus, the model applicability for different continents depended on its applicability for different types. CH4MOD_wetland is suitable for marshes, peatlands and coastal wetlands, but it cannot be applied in swamps (Table 2). This limitation may be because in CH4MOD_wetland only a semiempirical logistic model is used to simulate plant growth (Li et al., 2010). This characteristic may induce large uncertainties in simulating the growth of forests in swamps (Table 1). However, TEM uses the carbon and nitrogen dynamics module (CNDM) to describe the effects of photosynthesis, respiration, decomposition and nutrient cycling on NPP (Melillo et al., 1993). Compared with CH4MOD_wetland, TEM performed well in simulating NPP in various vegetation types (Melillo et al., 1993). According to the model evaluation, TEM was suitable for swamps and peatlands but had large uncertainties in marshes and coastal wetlands (Table 2). This pattern may be because TEM focuses on two major wetland types: boreal tundra and forest wetland (Zhuang et al., 2004). The biochemical processes in TEM may be suitable for peatlands (tundra) and swamps (forest wetland) but not suitable for marshes. For coastal wetlands, TEM did not consider the inhibition of salinity on CH₄ production (Poffenbarger et al., 2011; Bartlett et al., 1987) and greatly overestimated the CH₄ fluxes (Table 2). CH4MOD_wetland introduced the influence of salinity on CH₄ production and had good performance for coastal wetlands (Table 2).

4.2 Reducing uncertainties in global estimations

The estimates of global wetland CH₄ emissions had large ranges in previous studies (Zhu et al., 2015). The estimates by process-based models ranged from 92 Tg yr⁻¹ (Cao et al., 1996) to 297 Tg yr⁻¹ (Gedney et al., 2004) during the period of 1980–2012. Recently, an ensemble of process-based models driven by the same climatic data has commonly been used to estimate global wetland CH₄ emissions (Melton et al., 2013; Kirschke et al., 2013; Poulter et al., 2017; Saunois et al., 2016). However, the uncertainties in the model mean estimation range from 12 % (Poulter et al., 2017) to 40 % (Melton et al., 2013). The uncertainty mainly comes from the wetland distribution and model structure and parameters (Saunois et al., 2016). Estimating accurate wetland extent and its seasonal and annual variations is a major challenge in present studies. The global estimations of wetland area ranged from 4.3 to 12.9 Mha during the period of 1990 to 2005 (Melton et al., 2013). The wetland extent of 9.2 Mha from the GLWD excluded water bodies, and this value was ∼40 % higher than the wetland area used in this study. That is, this difference was the main reason for the lower global estimations determined in this study than those reported in previous works (Zhu et al., 2015; Melton et al., 2013; Poulter et al., 2017; Saunois et al., 2016). Improving the accuracy of wetland extent and temporal variations is important in reducing uncertainties in global wetland CH₄ estimations.

In addition to wetland area, the model structure and parameters accounted for ∼70 % of the total uncertainties (Saunois et al., 2016). The results of the accuracy analysis showed that for CH4MOD_wetland regression bias accounted for 61 % of the model errors in peatland and mean bias accounted for 22 % of the RMSE in swamp; for TEM, mean bias and regression bias accounted for 29 % and 42 %, respectively, of the model errors in coastal wetland (Table 2). This result indicated that there were still uncertainties in the modeling procedure, e.g., in the model mechanism or in parameterization (Zhang et al., 2017; Allen and Raktoe, 1981). In the existing process-based models, which are not limited to CH4MOD_wetland and TEM, some important procedures should be focused on to reduce the bias due to the model mechanism. For example, the mechanism of the freeze–thaw cycle is important in process-based models (Wei and Wang, 2017) because of the large contribution of CH₄ released during the nongrowing season in some frozen regions (Friborg et al., 1997; Huttunen et al., 2003; Mastepanov et al., 2008; Zona et al., 2016). In addition, quantifying CH₄ ebullition is important but difficult due to the uncertainty in estimates of CH₄ emissions from peatlands (Stanley et al., 2019). Moreover, although the importance of plants in CH₄ biogeochemical processes has been reported in many studies, better modeling and characterization of plant community structure is needed (Bridgham et al., 2013). Finally, most of the present process-based models do not have the ability to simulate CH₄ exchange from water bodies, such as lakes, rivers and reservoirs, although such water bodies contribute significantly to the global budget (Deemer et al., 2016). The use of the IPCC Tier method inevitably induces large uncertainties in the global estimates. The above mechanisms should be incorporated into existing process-based models to reduce the uncertainties in the current assessment.

The observational data that are related to processes of and controls on CH₄ production, consumption and transport also limit the model calibration and validation. The flux data of 43 wetland sites used for model performance in this study are quite limited and do not represent all climatic, soil, hydrologic and vegetation conditions across global natural wetlands (Table 1). The observations in this study used both the chamber method and the eddy covariance method (Aubinet et al., 2012), which are widely used for CH₄ observations (Table 1). There are differences in measuring CH₄ fluxes between the two methods (Chaichana et al., 2018). The eddy covariance method may underestimate the fluxes (Twine et al., 2000; Sachs et al., 2010), while the chamber method may overestimate the fluxes (Werle and Kormann, 2001). These differences may introduce uncertainties to model calibration and validation. Furthermore, both process-based models were evaluated on an annual basis rather than on a daily scale. The validation of seasonal variation was not performed in this study, partly because we cannot obtain the daily step data for some of the sites. Fine temporal validation against more flux datasets, especially fluxes by eddy covariance experiments, and intermediate variables that control the CH₄ process are necessary in future studies (Wei and Wang, 2017).

5 Conclusion

Two process-based models, CH4MOD_wetland and TEM, were used to simulate annual CH₄ emissions from different wetland types and continents, and their performances were evaluated. Model validation showed that both models could simulate variations between different wetland sites and years. The statistical analysis of model performance showed that CH4MOD_wetland was capable of simulating CH₄ emissions from marshes, peatlands, swamps and coastal wetlands, while TEM was capable of simulating CH₄ emissions from peatlands and swamps (model efficiency >0). CH4MOD_wetland performed well in Asia, Europe and North America, while TEM performed well in North America, Asia, South America and Africa. The models were then used to estimate global wetland CH₄ emissions. The CH₄ simulations of both models had good agreement in terms of the latitudinal and meridional bands. The global CH₄ emissions for the period 2000–2010 were estimated to be 105.31 ± 2.72 Tg yr⁻¹ by CH4MOD_wetland and 134.31 ± 0.84 Tg yr⁻¹ by the TEM. If we used a more accurate model to estimate each continent and wetland type based on the models' generality, the estimated global CH₄ emissions would be 116.99–124.74 Tg yr⁻¹ for the period 2000–2010. The uncertainty in global wetland CH₄ assessments by the process-based model approach comes from the inaccuracy of the wetland mapping area, the modeling procedure and the observational limitations. Future research on accurately mapping wetlands, improving model mechanisms and parametrization, and using more observations to evaluate model performance would improve global estimations.