www.biogeosciences-discuss.net/8/221/2011/ doi:10.5194/bgd-8-221-2011 © Author(s) 2011. CC Attribution 3.0 License.

Abstract. Natural methane (CH4) emissions from wet ecosystems are an important part of today's global CH4 budget. Climate affects the exchange of CH4 between ecosystems and the atmosphere by influencing CH4 production, oxidation, and transport in the soil. The net CH4 exchange depends on ecosystem hydrology, soil and vegetation characteristics. Here, the LPJ-WHyMe global dynamical vegetation model is used to simulate global net CH4 emissions for different ecosystems: northern peatlands (45°–90° N), naturally inundated wetlands (60° S–45° N), rice agriculture and wet mineral soils. Mineral soils are a potential CH4 sink, but can also be a source with the direction of the net exchange depending on soil moisture content. The geographical and seasonal distributions are evaluated against multi-dimensional atmospheric inversions for 2003–2005, using two independent four-dimensional variational assimilation systems. The atmospheric inversions are constrained by the atmospheric CH4 observations of the SCIAMACHY satellite instrument and global surface networks. Compared to LPJ-WHyMe the inversions result in a~significant reduction in the emissions from northern peatlands and suggest that LPJ-WHyMe maximum annual emissions peak about one month late. The inversions do not put strong constraints on the division of sources between inundated wetlands and wet mineral soils in the tropics. Based on the inversion results we diagnose model parameters in LPJ-WHyMe and simulate the surface exchange of CH4 over the period 1990–2008. Over the whole period we infer an increase of global ecosystem CH4 emissions of +1.11 Tg CH4 yr−1, not considering potential additional changes in wetland extent. The increase in simulated CH4 emissions is attributed to enhanced soil respiration resulting from the observed rise in land temperature and in atmospheric carbon dioxide that were used as input. The long-term decline of the atmospheric CH4 growth rate from 1990 to 2006 cannot be fully explained with the simulated ecosystem emissions. However, these emissions show an increasing trend of +3.62 Tg CH4 yr−1 over 2005–2008 which can partly explain the renewed increase in atmospheric CH4 concentration during recent years.

The goal of this study is to identify trends and variability of global net CH 4 exchange from 10 ecosystems over the last two decades. We also make a first attempt to attribute the variability of CH 4 exchange to different categories of ecosystems on the global scale. In the existing litera-12 ture, biogeochemical processes leading to CH 4 exchanges are commonly treated identically as emissions from "wetlands". Here we try to assess these processes individually by modelling the 14 biogeochemical cycle of the land biosphere, the atmospheric chemistry and transport of emitted CH 4 , and the uncertainties associated with these processes. 16

Uncertainties and variability
The uncertainty in total natural CH 4 emissions is mainly due to lack of knowledge of the ge-18 ographical distribution and interannual variability of CH 4 emissions from wet ecosystems, the largest natural source (Denman et al., 2007). Inverse ("top -down") modelling of atmospheric 20 CH 4 emissions suggests that interannual variability of wet ecosystem emissions is ± 12 Tg yr −1 , which explains about 70% of global emission anomalies over the period 1984-200322 (Bousquet et al., 2006. Interannual variability in CH 4 loss additionally contributes to the uncertainty of the global CH 4 budget and thereby the inferred emissions. In the free troposphere For the "bottom-up" biogeochemical process modelling of CH 4 emissions we apply LPJ-WHyMe v1.3 (hereafter LPJ) for (i) high-northern-latitude peatlands (45 • -90 • N) and for (ii) global min-4 eral soils.
Additionally, LPJ simulates the carbon balance of the peat, thus the soil carbon stock, carbon accumulation and decomposition rates (Wania et al., 2009b). The soil carbon serves as a 12 substrate for methanogenesis parametrised as a fraction of soil heterotrophic respiration (HR). CH 4 is transported to the surface by plant mediated transport, by diffusion through the soil or 14 by ebullition. CH 4 is oxidized under aerobic conditions in the soil layer and during transport (Wania et al., 2010b). 16 (ii) On mineral soils, LPJ simulates natural vegetation dynamics of 10 PFTs, soil hydrology (evapotranspiration, soil moisture), soil temperatures (freezing and thawing) and snow cover 18 (Sitch et al., 2003;Gerten et al., 2004;Wania et al., 2009a) using non-peatland hydrology in the model. The use of peatland and non-peatland hydrology depends on the soil class, but can 20 also be switched off completely. LPJ gives estimates for the land carbon pools and soil HR rates depending on temperature and soil moisture content (Wania et al., 2009b). Versions of 22 LPJ have been applied to study the global carbon cycle in the past (e.g., Joos et al., 2004) and the future (e.g., Sitch et al., 2008). CH 4 soil fluxes from mineral soils are estimated by using 24 relationships between CH 4 emissions and soil HR, soil moisture and temperature simulated by LPJ as described in the next section. 26 The Climate Research Unit (CRU) TS 3.0 climate data set (Mitchell and Jones, 2005) is used to force LPJ. Monthly input data to LPJ are surface air temperature, total precipitation, 28 6 fractional sunshine hours from cloud cover percentage, and number of wet days from the CRU climatology. Additionally, the CRUNCEP data set was used to perform simulations for the 2 period 1990-2008(Viovy N. and Ciais P., 2009. The model spin-up procedure and other input data are described in Wania et al. (2009a) and Wania et al. (2010a). 4 2.2 Inverse atmospheric modelling using four-dimensional data assimilation Inversions of trace gas emissions constrained by surface as well as satellite-borne observations 6 are a powerful tool for the validation of emission distributions and their trends and variability. Inversions based in 4D variational data assimilation (4D-Var, Houweling et al., 1999;Meirink 8 et al., 2008a,b;Pison et al., 2009), which use the adjoint of a chemistry transport model to correct fluxes such that the fit between observed and modelled concentration is improved, are especially powerful.
Here we use two 4D-Var inversion systems, TM5-4Dvar, to test in a diagnostic way the 12 biogeochemically modelled CH 4 emissions by LPJ (Bergamaschi et al., 2007(Bergamaschi et al., , 2009Meirink et al., 2008a,b) and LMDz-SACS (Pison et al., 2009). For the tracer-transport model TM5, 14 the 4D-Var inversion framework optimizes CH 4 flux per model-grid-box per emission category per month. This inversion system has no interactive chemistry, with CH 4 loss parametrised 16 with a prescribed OH field (Bergamaschi et al., 2007). In contrast the LMDz-SACS inversion framework simultaneously inverts three chemical species (methane -CH 4 , carbon monoxid -18 CO, hydrogen -H 2 ) and has interactive chemistry (Pison et al., 2009). While TM5-4Dvar solves for fluxes per category, LMDz-SACS solves for total monthly CH 4 flux per model grid-20 box. The full description of both inversion systems is given in Appendix B.
3 Methane source and sink categories 22 Among the natural sources of CH 4 we focus on biogenic soil emissions and subdivide them by type of ecosystem. Many different global classifications categorise potentially CH 4 -producing 24 ecosystems by hydrology, geomorphology, salinity, soil composition, vegetation, biogeochem-7 istry, or a combination of these (e.g., Matthews and Fung, 1987;Semeniuk and Semeniuk, 1997). Here we use a classification with respect to the processes relevant for the microbial 2 production, oxidation and transport of CH 4 in natural soils. The classification is simplified to three types: peatlands, inundated wetlands (including rice agriculture), and mineral soils. In 4 addition, we simulate CH 4 uptake in mineral soils and prescribe anthropogenic emissions from the EDGAR inventory (EC-JRC/PBL, 2009) as well as some smaller natural CH 4 sources (sec-6 tion 3.5). In the following sections 3.1 -3.4 we explain the calculation and parametrisation of sources and sinks in more detail. We test two emission scenarios, SC1 and SC2 (summarized in 8 Table 1 and discussed in section 4) for their consistency with the observed CH 4 concentrations.

Northern peatlands
Peatlands are important ecosystems in northern high latitudes. Peat is an accumulation of partially decomposed organic matter. Peat layer growth and decomposition depend principally on 12 its composition and the water table position, and to a lesser extent also on soil temperature (Rouse et al., 1997). Peatland formation in the arctic, boreal and alpine regions is spatially and 14 temporally influenced by the occurrence of permafrost (Robinson and Moore, 2000). LPJ simulates these physical processes directly and in addition, CH 4 production, oxidation and transport 16 to the atmosphere (Wania et al., 2010b). The production term is calculated proportional to HR in the soil, where HR depends on the mass (M i ) of each individual LPJ carbon pool (CP) in g 18 C and the associated turnover rate k i (yr −1 ): Turnover rates (k 10 i ) at 10 • C for the exudates, aboveground and below-ground litter carbon pools, and fast and slow soil carbon pools are given in (Wania et al., 2010b). The turnover rates 22 increase with temperature (R T ) (Lloyd and Taylor, 1994) and have normalised exponential dependency with soil water content (R moist ) (Fang and Moncrieff, 1999) for mineral soils. For 24 peatlands R moist is set to a constant of 0.35 (Wania et al., 2010a). CH 4 emissions are sensitive to the carbon ratio of CH 4 to CO 2 production (r C[CH 4 ]/C[CO 2 ] ), which is applied to anoxic 8 conditions but weighted by the volumetric fraction of air if a layer is not completely anoxic (Wania et al., 2010b). Due to some uncertainty of this value, we test two different values for r C[CH 4 ]/C[CO 2 ] based on comparisons with site data, namely 25% (SC1) and 20% (SC2) (Wania et al., 2010a). The latter resulting in a lower methane production in peatlands.

4
In our classification, peatlands include bogs, fens and mires, and are found predominantly in the Northern Hemisphere with an estimated maximum area of 2.99 -3.20×10 6 km 2 (Matthews 6 and Fung, 1987;Aselmann and Crutzen, 1989). As detailed in section 6 of Wania et al. (2010a) and section 3 of Wania et al. (2009a), we calculate the fractional peatland cover for the circum-8 polar region (45 • -90 • N) from organic soil carbon content derived from the IGBP-DIS 5'×5' resolution map (Global Soil Data Task Group, 2000). The assumption in this approach is that 10 high organic soil carbon content indicates areas where peat has been or currently is accumulating. After a comparison with the wetland area derived from a multiple satellite approach 12 (Prigent et al., 2007) the peatland fractions derived from the IGBP-DIS data were further downscaled by a factor of 0.38 (Wania et al., 2010a) . The resulting fractional peatland area (f peat ) 14 between 45 • and 90 • N is about 2.05 × 10 6 km 2 . In addition, flux strengths are multiplied by a factor of 0.75 to take account of the micro-topographic heterogeneity found in peatlands (Wania 16 et al., 2010a). Total peatland emissions for the circumpolar region are 51.1 Tg CH 4 (SC1) and 38.5 Tg CH 4 (SC2) for the year 2004 (Fig. 1a). lands (i.e. fractional cover of rice agriculture) are included in the data set (Prigent et al., 2007). To separate naturally inundated wetlands from rice agriculture, we concatenate fractional inun-10 dation data with a map of fractional rice cover (Leff et al., 2004). Rice agricultural areas are concentrated in SE Asia and are an important net source of atmospheric CH 4 of 31 -112 Tg 12 CH 4 (Denman et al., 2007). However, more recent estimates point to somewhat lower emissions of 14.8 to 41.7 Tg CH 4 from rice agriculture (Yan et al., 2009). The fractional rice cover 14 as given by Leff et al. (2004) is considered as an annual maximum extent (f ricemax ). To get the monthly rice extent (f rice ), we truncate the fractional rice cover to the fractional inundation 16 (f inund ) for each 1 • × 1 • grid cell (i) and each month (m). The remaining fractional area is then assumed to represent the fractional cover of naturally inundated wetlands (f natwet ): 20 With this separation it is possible to discriminate between CH 4 emissions from naturally inundated wetlands and irrigated rice agriculture. LPJ does not simulate CH 4 emissions in a process-22 based way for temperate, sub-tropical and tropical ecosystems. Nevertheless, LPJ dynamically simulates natural vegetation distribution (trees and grass), gross and net primary productivity, 24 soil HR and related carbon pools (Sitch et al., 2003). Assuming that these natural soils develop anoxic conditions when being flooded, we expect that a fraction of carbon, respired in the soil, 26 is released as CH 4 instead of as CO 2 for the period of inundation. For CH 4 emissions from 10 naturally inundated wetlands and rice agriculture we directly modify the carbon conversion ratio r C[CH 4 ]/C[CO 2 ] to include CH 4 oxidation, transport and general flux tuning. The modified 2 ratio is thus lower than for peatland emissions, and set to 2.40% (SC1) and 4.15% (SC2) (see Appendix A) in agreement with previous estimates (Christensen et al., 1996). CH 4 emissions 4 (e inund ) per m 2 and month (m) in grid cell i are then derived from soil HR as where E natwet and E rice are total emissions from naturally inundated wetlands and rice agriculture, respectively, per grid cell with the area A i (in m 2 ). The global parametrisation can be checked against regional emission inventories for rice agriculture in SE Asia (Fig. 2). The parametrised fluxes agree well with emission distributions and total estimates from the EDGAR 12 data base (EC-JRC/PBL, 2009) with the exception that emissions in North-Western India along the Himalayan foothills are missing; this is due to a geographical mismatch of fractional rice 14 cover (Leff et al., 2004) and precipitation input data, with low precipitation in the CRU input data preventing vegetation growth and CH 4 production in LPJ in the Indo-Gangetic Plain. 16 In summary, we simplify the classification of global wet ecosystems by latitude to prevent double counting of areas and emissions. Wetlands north of 45 • N are considered to be peatlands 18 (f natwet = 0), whereas wetlands south of 45 • N are classified as inundated wetlands (f peat = 0). The reason for the 45 • N cut-off line is that LPJ-WHyMe was specifically developed for 20 methane emissions from peatlands in cold areas (either high latitude or high altitude), of which the majority is found north of 45 • N. Since LPJ-WHyMe has not been tested yet for wetlands 22 other than this kind of peatland, we chose to use the 45 • N cut-off as a boundary between simulating CH 4 emissions within LPJ-WHyMe for northern peatlands and using a more generic 24 correlation approach (eqns. 4-6) for the rest of the inundated areas. Additionally, seasonal rice agriculture is calculated from global fractional inundation and global rice cover extent. The

Wet mineral soils
Mineral soils that are not inundated can still be a net CH 4 source. Soils with a low soil moisture 2 content imply oxic conditions which allow bacteria to consume CH 4 (Curry, 2007, and refs. therein). Relatively high soil moisture content however limits the availability of oxygen, a situ-4 ation which enables methanogenic archaea to produce CH 4 and suppresses CH 4 consumption. Above a certain soil moisture threshold, a part of CH 4 generated within the soil can diffuse 6 through the soil layer into the atmosphere without being oxidised. This has been observed during field experiments in different soil-vegetation systems (see references in Table A.1). e.g. in 8 tropical forests (Yan et al., 2008), in savanna grass lands (Sanhueza and Donoso, 2006) and in rice fields between growing seasons (Xuh et al., 2003). Therefore, we propose an additional global source of 10 CH 4 from wet mineral soils. We test two soil moisture thresholds above which CH 4 emissions can occur: 85% (SC1) and 95% (SC2) of water holding capacity (whc). These thresholds cor-12 respond to a fraction of 0.28 to 0.49 (SC1) and 0.31 to 0.55 (SC2) of water filled pore space (WFP th ), depending on soil type, field capacity, permanent wilting point and porosity. In sa-14 vannas, a switch from methane sink to methane source was found at a WFP th of about 0.2 (Otter and Scholes, 2000). The WFP th is fulfilled in LPJ for large areas in the boreal and trop- 16 ical region in SC1 and predominantly in the tropics in SC2 (Fig. 1d). The additional source could thus contribute to the high CH 4 emissions in the tropics as inferred from satellite data 18 (Frankenberg et al., 2008). The fractional area for wet mineral soils (f wetsoil ) is given by: 20 For each grid cell i and month m with soil moisture above the threshold the fraction of wet mineral soil is determined by subtracting the fraction of peatland, inundated wetland and rice 22 agriculture to prevent double counting of emission areas. f wetsoil is set to zero when soil moisture is below the threshold. Net exchange in wet mineral soils is calculated similarly as for 24 inundated wetlands. But since the oxidation in the partially oxic soils is higher than in inundated soils, we set the carbon conversion ratio r C[CH 4 ]/C[CO 2 ] to a value of 0.52% (see Appendix

26
A). The CH 4 emission rates per unit area (e wetsoil ) and total emissions per grid cell (E wetsoil ) 12 from wet mineral soils are calculated by: where emissions scale with the difference of the actual fraction of water filled pore space 4 (WFP) to the threshold fraction (WFP th ). Maximum annual CH 4 emission rates range from ∼ 1 g CH 4 m −2 yr −1 in Europe, North America, Africa and SE Asia in agreement with field 6 studies (e.g., Yan et al., 2008) to ∼ 5g CH 4 m −2 yr −1 in northern South America and Indonesia. Thus, CH 4 emission rates per m 2 and year are at least an order of magnitude smaller than 8 in naturally inundated wetlands or peatlands. Nevertheless, the large areas of wet mineral soils sum up to a globally significant source with total annual emissions of 93.0 Tg CH 4 yr −1 (SC1) and 57.8 Tg CH 4 yr −1 (SC2) in 2004 (Fig. 1d).

Soil uptake 12
Atmospheric CH 4 is biologically consumed in near-surface soils. Global soils account for 28 Tg CH 4 yr −1 , or ∼ 5% of the total CH 4 sink, with an uncertainty range of 9 -47 Tg CH 4 14 yr −1 (Curry, 2007). The soil consumption of CH 4 occurs via oxidation by aerobic bacteria, or methanotrophs, within 3 -15cm soil depth. The CH 4 consumption is determined by the micro- 16 bial oxidation rate within the soil and the transport of atmospheric CH 4 into the soil. The two processes themselves depend most importantly on soil moisture, but also on soil temperature 18 and soil texture. Here we use the uptake scheme of Curry (2007) applied to LPJ output that allows for a monthly estimate of the global CH 4 soil sink. In this scheme the CH 4 uptake is 20 parametrised for grid cell i and month m as where g 0 = 586.7 mg CH 4 ppmv −1 s d −1 m −2 cm −1 is a constant factor that converts the surface concentrations C 0 expressed in parts per million by volume (ppmv) into an uptake flux J (mg CH 4 m −2 d −1 ) (Curry, 2007). We do not account for inhibition of CH 4 uptake in cultivated lands and thus leave this factor constant at r cult,i,m = 1. However, we do scale CH 4 2 uptake by the fractional area not covered by peatlands, naturally inundated wetlands and rice agriculture with This means that CH 4 uptake and CH 4 emissions from wet mineral soils share the same 6 fractional grid cell area. But resulting fluxes are mutually exclusive through the level of soil moisture, which is a key variable in the parametrisation of the CH 4 effective soil diffusion 8 (D soil in cm 2 s −1 ) and the CH 4 oxidation rate (k in s −1 ) as calibrated by Curry (2007). As an input for the parametrisation we use soil moisture and soil temperature directly calculated in LPJ at 10 cm soil depth. Annual fluxes multiplied with grid cell area are shown in Fig. 1e.
The total global soil uptake for 2004 is 38.1 Tg CH 4 , which is within the range of previous 12 estimates (Curry, 2007;Ridgwell et al., 1999).

14
In order to close the global CH 4 budget, we prescribe additional CH 4 sources. Included are emissions from coal mining, oil and gas production and transport, ruminants, biomass burning 16 (includes natural source), and waste deposits, as given in the EDGAR emission data base (EC-JRC/PBL, 2009). Additionally, we prescribe small natural sources of oceanic (Lambert and 18 Schmidt, 1993) and geologic (Etiope et al., 2008;Neef et al., 2010) origin, as well as emissions from termites (Sanderson, 1996). We used the same ocean/geological emission distribution as in 20 Bergamaschi et al. (2007). The tropospheric and stratospheric CH 4 sinks are either prescribed or directly calculated within the atmospheric chemistry transport models. The global total of 22 other sources is 319 Tg yr −1 for the year 2004 for both scenarios (Table 2). We did not include aerobic CH 4 emissions from plants (Keppler et al., 2006) in our study 24 as they are likely to contribute only 0.2-1.0 Tg CH 4 yr −1 to the global CH 4 budget (Bloom et al. (2010) and reply to comment by F. Keppler in Spahni et al. (2011)). After the comple-26 tion of our simulations, another new source of CH 4 emissions was found, namely emissions 14 from tank bromeliads in the canopy of tropical montane forests (Martinson et al., 2010), who estimated that about 1.2 Tg CH 4 yr −1 are emitted from this source. In our opinion, this small 2 source would have had no significant impact on the outcome of our study. We would like to note that what is claimed to be a new source of CH 4 emissions by Gauci et al. (2010) and Rice 4 et al. (2010), is in fact implicitly included in our modelling approach of 'naturally inundated wetlands' and 'wet mineral soils'. Our modelling approach, as others before (e.g. Christensen 6 et al., 1996) relate CH 4 fluxes to heterotrophic respiration and soil moisture without assuming any specific transport pathway; CH 4 may escape via diffusion, ebullition or plant-mediated 8 transport. Therefore, we did not omit this potentially large source of CH 4 emissions.

10
Results of this study are twofold. First, we highlight the results of the biogeochemical modelling of natural net CH 4 exchange using the LPJ dynamical global vegetation model. Second,  and wet mineral soils, not only scales the CH 4 fluxes per m 2 , but also impacts on the latitudinal distribution and global total of natural CH 4 emissions. Since the attribution to different source nitudeaverages of local flux rates and the global CH 4 budget (Appendix A). TThese two criteria put a constraint on the parameters r C[CH 4 ]/C[CO 2 ] for SC1 and SC2, which are given in Table  1 and resultinglead to global fluxes as given in Table 2. SC1 is a scenario with large emissions from boreal peatlands and wet mineral soils. Emissions from naturally inundated wetlands 4 and rice agriculture are comparatively small. Northern peatland CH 4 emissions are calculated directly from an initial version of Wania et al. (2010a). Soil uptake is calculated directly from 6 LPJ model output and the uptake scheme by Curry (2007). In SC2 emissions are dominated by naturally inundated wetlands (+70% compared to SC1) and rice agriculture (+70%) in the 8 tropics and sub-tropics (Fig. 3). Peatland emissions (-25%) and wet mineral soil emissions (-38%) are substantially reduced in SC2. As a consequence the two scenarios mainly differ for thein latitudinal gradient of emissions and hence gradient of atmospheric concentration, but not as much in the temporal patternseasonal cycle.

12
Highest net CH 4 exchange per grid cell and m2 is simulated in the tropics (SE Asia, Central Africa, South America) and the boreal regions (Siberia, Scandinavia, Eastern Canada, Alaska) 14 as shown in Fig. 3. with maximum values of more than 60 Gg CH4 grid-cell-1 month-1, corresponding to 5 -15 g CH4 m-2 month-1 at these high-emission locations. Beside these high-emission regions, 16 the model suggests large areas where CH 4 is emitted at a much smaller rate (see also depends on the distribution of source areas that are used in the calculation of total fluxes. As mentioned in section 3, these areas are based on the organic soil carbon map for peatlands, the 28 satellite data of inundated areas for naturally inundated wetlands and simulated soil moisture for 16 wet mineral soil emissions and soil uptake. The impact of these distributions of both scenarios are shown in the bottom plot of Fig. 4b. Highest fluxes with ∼ 7 Tg CH 4 yr −1 per degree lat-2 itude are found around the equator. High fluxes over a broader latitudinal range are also found over the sub-tropics and the northern high latitudes. Both emission distributions modelled by 4 LPJ are very similar to the latitudinal distribution jointly estimated from SCIAMACHY CH 4 concentrations and gravity space borne data by Bloom et al. (2010). The latitudinal distribution 6 of simulated CH 4 fluxes (Fig. 4b) shows only small year-to-year variability, which is understandable since in this study emission from peatland and inundated wetland areas only vary 8 seasonally and not interannually. It has been shown that both emissions per m 2 and emission areas may vary independently, and total CH 4 emission variability in space and time is regulated 10 by both (Ringeval et al., 2010).
A remarkable deviation tofrom the latitudinal distrubtion estimated by Bloom et al. (2010) 12 is found at high latitudes, where LPJ emissions peak between 50 -60 • N and 60 -70 • N; the latter peak cannot be constrained by the SCIAMACHY observations due to the instrument's  ual source types, we use this system first to evaluate the two emission scenarios SC1 and SC2 (simulated by LPJ for the year 2004) against the observations. The prior differences between 26 the two scenarios, in terms of individual sources and sinks, are outlined in Table 1. The prior and posterior (optimised) relative contributions per source type are shown in Fig. 5 for both scenarios. During the optimisation global gross emissions increase from 533 to 576 Tg CH 4 yr −1 , and the relative source contributions change as well.
In Fig. 5 it can be seen that the relative strength of anthropogenic (∼ 55%; coal mining, oil & gas, ruminants, biomass burning, waste and rice) to natural emissions (∼ 45%; inun-4 dated wetlands, wet mineral soils, northern peatlands, termites, and ocean and geologic) is not strongly affected by the optimisation. However, the source distribution within these large 6 categories changes under optimisation. For anthropogenic sources, e.g., CH 4 emissions are shifted from oil & gas (-4%) to increased emissions from ruminants (+6%) and waste (+1% to 8 +2%). Table 3 shows that the changes for the individual categories are substantial in absolute values. The optimisation also considerably reduces the associated error by 51% and 29% 10 for oil & gas and ruminants, respectively.The transfer of emissions from oil & gas to domestic ruminants seems to be largely due to the difference in where these sources are on the planet: 12 oil & gas emissions are heavily located in the Northern Hemisphere (e.g. Russia / Siberia), while domestic ruminant emissions are strong in South America. Focusing only on anthropogenic cat-14 egories, the shift from oil & gas to ruminants largely reflects the increase in tropical / southern hemisphere emissions required by the inversion. 16 In regard to the natural sources the optimisation affects mainly northern peatlands. Peatland CH 4 emissions are halvedreduced from 10% to 5% in SC1 and from 7% to a similar value5% in 18 SC2 , where prior emissions were already smaller. The opposite is true for CH 4 emissions from wet mineral soils and inundated wetlands (including rice agriculture). Although the relative 20 contribution of emissions from these source types differ substantially between SC1 and SC2, the difference is preserved in the optimisation. This means that the inversion can not clearly 22 differentiate between emissions from wet mineral soils or inundated wetlands. Emissions from these two categories have a large spatial overlap, such that after transport in the atmosphere, 24 surface concentrations and satellite measurements can not help to distinguish between them.
However, the inversion produces additional information on the spatial pattern for the LPJ 26 simulated net exchange. Fig. 6 shows the change in CH 4 emissions and uptake, following optimisation, at grid cell resolution. Fluxes from northern peatlands are reduced in Scandinavia, mineral soils) are greatly reduced in western South America, Bangladesh-India and Indonesia ( Fig. 6b -d). On the other hand, the optimisation suggests a net increase in eastern South 2 American and central African emissions. The latter is partly achieved by a reduction in soil uptake (from 39 to 26 Tg CH 4 yr −1 globally, Fig. 6e). Emission changes for the optimisation 4 of SC2 are given in Table 3 for each category and for each geographical region, as defined in the TRANSCOM3 model intercomparison experiment (Gurney et al., 2002). It can be seen from 6 Table 3 and Fig. 6f that the net change in CH 4 emissions is relatively spread out over various regions, but with strong reductions/increases in particular regions. Table 3 also shows the globally 8 integrated uncertainty reduction of the inversion per emission category. Regions with a large relative increase are the North American temperate and the North African region, while emissions from 10 the Eurasian boreal region are strongly decreased. Note that the regional adjustments in Table 3 are for the total CH 4 emissions per region and thus also include substantial changes in 12 anthropogenic sources (Fig. 5). As mentioned above most important are increased emissions from ruminants (+40%), mostly confined to North-and South America and decreased emissions 14 from oil & gas industry (-23%) mostly confined to Eurasia. Regional changes in anthropogenic emissions as derived from TM5-4Dvar inversions using surface and SCIAMACHY CH 4 ob- 16 servations have recently been presented by Bergamaschi et al. (2009). Table 3 also shows the globally integrated uncertainty reduction of the inversion per emission category. Although, the 18 overall reduction in uncertainty is considerable (66%), the atmospheric inversions can only give us estimates of how the LPJ fluxes should be corrected. The observations on its own are not 20 a strong enough constraint to distinguish between all sources, because they have significant spatial overlap. As a result from the inversion both scenario seem to be consistent with the 22 observational constraint. In summary, the strongest constraint imposed by TM5-4Dvarthe observations on the LPJ-24 derived source and sink fluxes is the consistent reduction on northern peatland emissions, in both scenarios SC1 and SC2 to about 5% of total emissions (Fig. 5). Therefore, we evaluate (resulting from the two inversions) in Fig. 7. The two inversions show quite large differences in the seasonality of posterior CH 4 fluxes, both being consistent with the atmospheric concentra-2 tions. While the LMDz-SACS-optimised posterior emissions have a similar seasonal duration to the prior fluxes, the TM5-4Dvar-optimised posterior emission season is considerably shorter.

4
The difference between the inversion systems is inherent to their set-up. Temporal correlations in TM5-4Dvar are category-dependent and are not being applied to seasonally varying emis-6 sions such as from the peatlands at high northern latitudes. In LMDz-SACS the total of all sources is constrained over eight-day periods, with no time correlation between these periods. 8 This clearly shows how uncertain seasonal fluxes by inversions are. On the other hand, both inversion results have a maximum correction in July and reductions in August-October (Fig. 7b).

10
This meanssuggests that northern peatland emissions apparently reach their maximum about one month earlier than simulated by LPJ.

12
The reasons for this disagreement between LPJ and the observed concentrationsthe emissions inferred from observations by inversion lies in the separation of the three different pathways 14 for CH 4 to escape to the atmosphere. In LPJ, CH 4 fluxes from plant mediated transport and diffusion are highest in July, while ebullition fluxes from peatlands peak late in the season, 16 causing the overall CH 4 emissions to wrongly peak in September/October. In a revised version of LPJ we change the parametrisation of the ebullition, such that all excess CH 4 is emitted  nual variabilityanomaly (12 month running mean) of CH 4 exchange of LPJ is smaller (± 7.1 Tg CH 4 yr −1 ) than the inversion estimate (± 10.6 Tg CH 4 yr −1 with constant and ± 11.5 Tg CH 4 24 yr −1 with variable OH). The LPJ emission variability mainly reflects the variability of local fluxes due to climate changevariability and does not incorporate the variability of source area.

26
Including this variability for naturally inundated wetlands for the years 1993-2000(Prigent et al., 2007, the largest natural source of the LPJ categories, leads to a significantly different evolution of interannual emissions (Fig. 8a). Its variability is larger and the trend is negligible. This result confirms the finding of Ringeval et al. (2010) who show that the source area vari-2 ability and flux variability for inundated wetlands are not necessary linearly related, and both are important for CH 4 emission variability.

4
In our approach the LPJ emission variability is composed of the different source and sink categories (Fig. 8b). Emissions from northern peatlands and naturally inundated wetlands 6 have a similar, but anti-correlated contribution to total source variability. Emissions from wet mineral soils show a larger variability and explain most of the total interannual variability. 8 This can be partly explained by the soil moisture content. Its variability does not only affect the CH 4 fluxes, but also the source area of wet mineral soils. However, areas of northern peatlands and inundated wetlands are not affected bydirectly related to soil moisture availability in our study. Soil uptake has a comparably small variability, but is steadily increasing and thus 12 slightly compensating the trend in emissions. The main reason for the increase in total CH 4 emissions is an increase in LPJ heterotrophic respiration. HR and CH 4 emissions are highly 14 correlated (R 2 = 0.91). CH 4 emissions go along with increasing temperature (R 2 = 0.55) and precipitation (R 2 = 0.83) over land as shown in Fig. 8c. Soil moisture content alone is not 16 well correlated to CH 4 emissions (R 2 = 0.22). A good predictor for the simulated global CH 4 emissions is a normalised index that combines global mean surface temperature (30%) and 18 global mean precipitation (70%) over land, favouring high CH 4 emissions under warm and wet conditions (R 2 = 0.94). cal processes including the land carbon cycle. The inversion systems then constrain emissions by calculation of atmospheric transport and CH 4 loss from the observed atmospheric CH 4 con-2 centration distribution. Despite these various observational and modelling constraints, there remains more than one solution for the global source and sink distribution. Finding "the opti-4 mal solution" certainly takes a lot of effort at all scales. In this study we present two emission scenarios, SC1 and SC2 (Table 1, Table 2, Fig. 5) that are evaluated and used to revise LPJ 6 emissions accordingly. Though northern peatland emissions for both scenarios are within the range of earlier estimates of 31 to 106 Tg CH 4 yr −1 (Zhuang et al., 2004). Our atmospheric 8 inversion suggest northern peatland emissions being even lower than SC2 (39 Tg CH 4 yr −1 ) in 2004, at about 28 Tg CH 4 yr −1 . This is in line with an earlier inversion study (Chen and Prinn, 2006), in which a prior estimate of 43 ±65 Tg yr −1 for northern hemisphere emissions was reduced to 33 ± 18 Tg CH 4 yr −1 . This study also estimated 223 Tg yr −1 for remaining 12 wet ecosystem methane sources (including rice agriculture), which is more or less in line with our estimate of 204 Tg yr −1 .
14 Tropical sources have already been increased from SC1 to SC2 by a 70% larger carbon conversion rate and tuning parameter r C[CH 4 ]/C[CO 2 ] for natural inundated wetlands and rice 16 agriculture. Since the two categories use the same parametrisation, an additional increase in r C[CH 4 ]/C[CO 2 ] , as suggested by the TM5-4Dvar inversion, leads to slightly overestimated CH 4 18 emission from rice agriculture (Fig. 2). This limits CH 4 emissions from inundated wetlands from 45 • N to 60 • S to ∼ 80 Tg yr −1 . Additionally, wet mineral soils contribute ∼ 50 Tg CH 4 20 yr −1 summing to a total of ∼ 130 Tg CH 4 yr −1 in this latitudinal band. A separation of these two source categories by the inversion has proven to be difficult. Overall, the reduced northern 22 peatland source and the good agreement with emissions from rice agriculture suggest SC2 is a more plausible scenario than SC1.

LPJ trends 4
All natural LPJ flux categories show an increase over the years 1990 -2008 (Fig. 8b) that becomes +1.03 Tg CH 4 yr −1 without rice emissions and +1.11 Tg CH 4 yr −1 when rice emissions 6 are included. The increase in simulated CH 4 emissions is attributed to enhanced soil respiration resulting from the observed rise in land temperature and in atmospheric carbon dioxide 8 that were used as input. With a conversion factor of 2.78 Tg CH 4 ppbv −1 this implies an atmospheric CH 4 increase of 0.37 and 0.4 ppbv yr −1 , respectively, which is very small and about  Fig. 8a,b). This in-22 crease contributes to the observed maximum in atmospheric growth rate in 2007 (Dlugokencky et al., 2009). The biggest contribution of the 2008 -2004 difference in simulated CH 4 emissions 24 (17.33 Tg CH 4 ) comes from wet mineral soils (56.6%). Northern peatland emissions (24.3%) and emissions from inundated wetlands including rice agriculture (20.5%) contributed similarlyequally, but less, to the rise in CH 4 . The LPJ run also suggests that only post 2005 are the inter-annual emission anomalies of peatlands and inundated wetlands not anti-correlated ( Fig.   28 24 8b). The compensation of emissions through an increased soil sink is small (-1.4%).
This source attribution agrees with the finding of Dlugokencky et al. (2009), namely that despite the emission increase at Arctic latitudes, the largest increase in atmospheric CH 4 concentrations in 2007 happened in the tropics. This is represented in the simulation by the dominant 4 contribution in 2006/2007 from the low-latitude sources wet mineral soils, inundated wetlands and rice agriculture (Fig. 8b). Emission fluxes could be modulated by variations in global wet-6 land area, which is partly considered by the change in emission area of wet mineral soils that seem to play an important role for the inter-annual CH 4 variability. The inclusion of a hydro-8 logical module in a DGVM, that calculates wetland area in addition to wetland fluxes, as well as more observational constraints are needed to properly address questions on long term trends 10 in global CH 4 emissions.

Summary and conclusions
12 In a multiple model approach we derive estimates for global CH 4 emissions and uptake in organic and mineral soils that are in agreement with atmospheric observations. We show that the 14 global CH 4 source category usually summarised in the literature as "wetlands" can be usefully broken down into process-defined subcategories: northern peatlands, naturally inundated wet- sulting global CH 4 emission distribution by latitude for the LPJ categories generally compares well to reconstructions based on concentration and gravity field satellite observations (Bloom lite observations, the LPJ peatland emissions are relatively high compared to the estimate by Bloom et al. (2010).
From the results of two different sets of parametrisations of our biogeochemical model (SC1, SC2) supplemented with other natural and anthropogenic sources, we derive two global CH 4 4 budgets that are tested against atmospheric observations, both ground based and space borne, using the TM5-4Dvar and LMDz-SACS inversion systems. The inversions show that CH 4 6 emissions predicted by LPJ for northern peatlands are overestimated in total, and reach their annual maximum about one month late. Although the two derived scenarios SC1 and SC2 8 differ substantially in the emissions attributed to wet mineral soils and inundated wetlands, the inversion results can not readily distinguish between these two emission categories. This study thus clearly points out the need for further observational constraints and an improved inversion system.

24
The geographical distribution of corrections to LPJ fluxes by source and sink categories can greatly improve the understanding and representation of CH 4 relevant processes in biogeochem-26 ical models. Our study also shows that in addition to accurate global wetland extent (Ringeval et al., 2010), a good characterisation of the different wetland types is needed in "bottom-up" 28 biogeochemical process models. It will continue to be very difficult to identify different source 26 and sink changes from atmospheric inversion calculations alone. An iteration of both modelling approaches together with observational constraints, as presented in this study, offers the potential to further constraining global CH 4 emissions from individual sources and their variability on various temporal and spatial scales.  to values reported in Table 1 of Christensen et al., 1996).
In this study we have chosen to vary this parameter as it is considered to be the most uncer-22 tain in our approach of methane emission modelling. Many of the other model parameters are physically or biogeochemically better constrained.

24
For the two scenarios SC1 and SC2 the carbon conversion ratio for wet mineral soils was fixed at 0.52% to match roughly the annual emission range (Table A.1). Together with the 26 27 soil moisture thresholds (85% whc for SC1 and 95% whc for SC2) one can calculate global emissions from wet mineral soils. Knowing the global budget Bousquet et al. (2006) and other 2 LPJ derived sources and sinks one can estimate the global emissions from inundated wetlands. From the global total one can finally deduce the carbon conversion ratio of inundated wetlands 4 to 2.4% (SC1) and 4.15% (SC2). Nevertheless, the carbon conversion ratios remain global tuning numbers that can be changed, as long as criteria (i) and (ii) are fulfilled. For the revised LPJ 6 fluxes as described in section 4.3 the carbon conversion ratios have been retuned to 5.37% for inundated wetlands and 0.67% for wet mineral soils.

A2 Comparison with field measurements
Comparing simulated global CH 4 fluxes with field measurements from individual sites needs in one or the other way an upscaling. This bears the risk for introducing large uncertainties. This is especially true for a trace gas like atmospheric CH 4 that has large emission variability in 12 space and time. We thus compiled a list of studies (Table A.1) that gives a rough estimate how large the potential CH 4 source from wet mineral soils can be. Sites are well distributed over the 14 world and cover different ecosystems. Fluxes were measured directly over non-saturated soils or over the canopy of trees. As indicated in Table A.1 measured fluxes are given for different 16 periods of the year and thus do not always represent the annual mean. LPJ fluxes of wet mineral soils are averaged (i) temporally and (ii) spatially.

18
(i) Annual average LPJ fluxes are comparable to observed annual fluxes. LPJ emission rates in months with maximum emissions are considerably larger and give an upper range for simu-20 lated emissions.
(ii) By comparing site flux data we implicitely assume that measurements are representative 22 for a larger area, in our case for a 1 • × 1 • latitude/longitude grid cell size. For some sites LPJ does simulate zero emissions from wet mineral soils because of very low soil moisture content 24 (see section 3.3). In this case we also calculate the spatial average or the spatial maximum flux for a larger region of the size of ∼ 10 • × 10 • latitude/longitude, indicated in Table A.1 with ra 26 and rm, respectively. Monthly emission rates of up to 140 mg CH 4 m −2 d −1 are measured over 2 months in a trop-ical forest (Teh et al., 2005), albeit a very large proportion is oxidized in the soil. If there would be zero emissions for the other 10 months the annual average flux would still be equivalent 2 to 23.3 mg CH 4 m −2 d −1 . This is more than the range of annual average emissions simulated from LPJ wet mineral soils ( The two inversion systems used in this study were developed independently at two different research centres and thus represent two distinct implementations of a similar technique. The two systems differ in three main ways: the transport model used (and its adjoint), the estimation of background error covariances, and the composition of the control (or optimisation) vector.

12
The TM5-4DVar system is described more closely in Meirink et al. (2008a,b) and the sources cited therein, and the LMDz-SACS system is described in (Pison et al., 2009).

B1.1 Transport Models
The TM5-4DVar system connects emissions to concentrations using the methane-tracer version 16 of the chemical transport model TM5 (Krol et al., 2005;Bergamaschi et al., 2007). In the simulations presented here, TM5 is run on a 6 • × 4 • (longitude-latitude) grid and 25 sigma-18 pressure vertical levels. Transport in TM5 is driven by meteorological fields from 6-hourly forecasts of the ECMWF operational model. The reaction between CH 4 and OH is modelled 20 by a described, constant k[OH] field (Bergamaschi et al., 2007).
In the LMDz-SACS system, the model used to link emissions to concentrations is the off-  (Hauglustaine et al., 2004;Folberth et al., 2005), limited to the main components of the methane oxidation chain (CH 4 , formaldehyde (HCHO), CO and H 2 ). OH radicals are 2 optimised within the same inversion framework using methyl-chloroform (MCF) surface observations. LMDzs grid is 3.75 • × 2.5 • (longitude-latitude) on 19 sigma-pressure levels. The 4 air mass fluxes are pre-computed by the on-line LMDz version nudged to winds from weather analysis.

B1.2 Estimation of background error
In the TM5-4DVar system, errors are estimated by assuming 100%-uncertainty in the flux per 8 emission category per gridcell. In addition, the correlation between gridcells is assumed as a Gaussian function with a decorrelation length of 500 km. Practically, this means (1) that gridcells with greater prior emissions are considered to be more uncertain, and vice versa, and (2) that the inversion's adjustment in each gridcell is dependent on the adjustment in nearby grid-12 cells. For source types that are seasonally varying (wet ecosystems, and biomass burning), no correlation is assumed between monthly fluxes. For the other sources, the temporal correlation 14 is modelled as an exponential decay with a decorrelation time of 9.5 months.
In the LMDz-SACS inversion system, errors for CH 4 fluxes are estimated by assuming an 16 uncertainty in the (total) emission flux of 100% of the maximum flux over the cell and its eight neighbours during each month. This allow local variations in the spatial pattern of the inventory 18 to occur. The correlations between gridcells are modelled as described in Chevallier et al. (2005) with correlation lengths of 500 km on land and no time correlations.

B1.3 Control Vector and Cost function Minimisation
In both systems, the adjoint of the respective transport models is used to minimise a cost 22 function, that is a function of the misfit between predicted and observed CH 4 concentrations (Chevallier et al., 2005). The set of model parameters with respect to which the cost function 24 is minimised is called the control vector. In the TM5-4DVar system, the control vector is comprised of monthly methane fluxes per gridcell per source, plus the initial 3-D CH 4 concentration field. The minimisation of the cost function is performed using the Lanczos algorithm.
In the LMDz-SACS system, the control vector contains 8-day average net surface fluxes 2 per grid cell of CO, CH 4 , MCF, and H 2 , the 8-day average production of HCHO per model column, 8-day average OH column-average concentration over four latitude bands (Bousquet 4 et al., 2005), and the initial-time concentration fields of CO, CH 4 , MCF, and H 2 . In this system the cost function and the norm of its gradient (computed by the adjoint) are minimised with the 6 algorithm M1QN3 (Gilbert and Lemaréchal, 1989).
Since the LMDz-SACS system does not provide an optimisation by category per se, we select 8 grid cells with more than 90% peatland emissions in the prior. Then we assume that optimised peatland emissions fully account for changes in the posterior. This works quite well because 10 peatlands are localised in the boreal region with little interference from other source types. Emissions from selected grid cells are then rescaled to reflect total peatland emissions. The 12 downside is that the information of total peatland emissions is lost, but we gain information on the temporal optimisation for a single category. Resulting emission anomalies are shown in Fig (Dlugokencky et al., 2005) with the factors provided by GLOBALVIEW-CH 4 (2009).

24
In the TM5-4DVar system, station observations are averaged into 3-hourly bins and compared to mean concentrations per 3-hour time step, resulting in a total of 7912 station observations 26 of CH 4 concentration assimilated from 2003 to 2005. The uncertainty assumed per station observation is estimated by a 3 ppb measurement error, plus a representativeness error estimate based on the differences in modelled concentrations in neighbouring gridcells (Bergamaschi et al., 2005). This yields observation error varying from 3 to several tens of ppb.

2
In the LMDz-SACS system, measurements are thinned so that only one measurement per grid cell per dynamical time step (30 minutes) is kept. Moreover, after comparison with the 4 prior forward simulation, measurements for which the matching simulated concentration is out of a range of 3 σ (see below for the specification of the error) were not used as constraints for the 6 inversion. The observations of MCF concentrations are measurements made by NOAA/GMD and AGAGE surface networks (Montzka et al., 2000;Conway et al., 1994;Prinn et al., 2000). Assuming that the synoptic variability is an approximation of the transport errors, we used values of variances from Prinn et al. (2005) or from the NOAA (depending on the station); when 12 no such data was available for a station, the uncertainties associated with the measurements were used, with a minimum threshold of ± 1.2 ppt for MCF and ± 3 ppb for CH 4 (Pison et al.,  Frankenberg et al., 2005Frankenberg et al., , 2006Frankenberg et al., , 2008. In both systems, column-average concentrations are assimilated. The measurement error is set at 2% of the value of the observation (Frankenberg 22 et al., 2006). The selection criteria provided with the data include solar zenithal angles less than 70 and a simple cloud filter plus other selection criteria to eliminate back-scan pixels, poor fits 24 of RMS of CH 4 or CO microwindows and ensure minimal signal and number of pixels for the fit. In the TM5-4Dvar system, observations are averaged over 1 • × 1 • grid boxes and three-hour 26 assimilation windows before being compared to the modelled concentrations. An error of 1.5% is assumed (see also Bergamaschi et al., 2009;Meirink et al., 2008a), and between 2003 and 2005, a total of 749314 observations are assimilated. In the LMDz-SACS system the measurements selected according to the above criteria are averaged over each model grid cell during 2 each time step and the observation error is taken to be the quadratic sum of the measurement error (2%) and the chemistry-transport model error (arbitrary set to 10%). A further sampling 4 of the measurements for which the difference with the first-guess is less than one sigma finally had 923415 constraints for the inversion. Acknowledgements. We would like to thank Catherine Prigent for providing the global inundation data set. This work was conducted within and supported by the EU-Project HYMN (Hydrogen, Methane, 8 and Nitrous Oxide: Trend variability, budgets, and interactions with the biosphere; GOCE-037048). Additional financial support was provided by the Oeschger Centre for Climate Change Research and the Swiss National Science Foundation (Switzerland), the NERC funded programme QUEST -Quantifying and Understanding the Earth System (UK).

24
And-Diagnostics-Laboratory Global-Air-Sampling-Network, J.  3 and 4). The area assigned to the category wet mineral soils and their CH 4 emissions were considerably reduced in SC2 by raising the threshold of soil moisture (SM), expressed as fraction of the water holding capacity (whc), above which CH 4 emissions can occur. * Here, the carbon conversion factor also includes oxidation of CH 4 to CO 2 during transport to the surface and general flux tuning.      (Bloom et al., 2010). The colour shaded areas represent the 2 σ band of the interannual variability over the last two decades for the corresponding scenario.   Table 3. Regional total (anthropogenic+natural) emission changes following the optimisation of the LPJ (SC2) scenario using TM5-4Dvar for the year 2004 at a 1 • × 1 • resolution. Given are prior and posterior total fluxes by region (top) and category (bottom) in Tg CH 4 yr −1 , their associated error and the estimated uncertainty reduction. The 13 land regions are given as defined in the TRANSCOM3 model intercomparison experiment (Gurney et al., 2002), while the 14 th region, referred to as "Oceans", combines the 10 oceanic regions of TRANSCOM3.   Table A.1 Comparison of site data and LPJ (SC2) CH 4 fluxes for individual years from studies that classify as a 'wet mineral soil' source. Fluxes are given in mg CH 4 m −2 d −1 . Note that fluxes have been measured for individual ecosystems (sec. forest = secondary forest) and represent the average or the maximum (m) flux observed for a certain period (annual/seasonal). LPJ fluxes are given for the annual average and for the month with maximum emissions in the respective year, as well as for different spatial extents: either for the flux of the grid cell at the site (s) , for the average flux of grid cells representing a larger region (ra) or for the maximum flux of grid cells representing a larger region (rm). Uncertainty in measured mean flux or flux variability is very large i.e. ± 20% or more.    Figure 1, but with CH 4 fluxes per area instead of per grid cell for the individual categories (a-e) and the category weighted net flux (f ). Note that here emissions from individual categories (a-e) are non-area-weighted and none of the subplots (a-f ) take account of the decreasing grid cell area with higher latitude.