The main three bottom-up global inventories covering all anthropogenic emissions are from the United States Environmental Protection Agency, USEPA (2012, 2006), the Greenhouse gas and Air pollutant Interactions and Synergies (GAINS) model developed by the International Institute for Applied Systems Analysis (IIASA) (Höglund-Isaksson, 2012) and the Emissions Database for Global Atmospheric Research (EDGARv4.1, 2010; EDGARv4.2FT2010, 2013). The latter is an inventory compiled by the European Commission Joint Research Centre (EC-JRC) and Netherland's Environmental Assessment Agency (PBL). These inventories report the major sources of anthropogenic methane emissions: fossil fuel production, transmission and distribution; livestock (enteric fermentation and manure management); rice cultivation; solid waste and waste water. However, the level of detail provided by country and by sector varies between inventories, as these inventories do not consider the same number of geographical regions and source sectors (Höglund-Isaksson et al., 2015, see their Table 5.2). In these inventories, methane emissions for a given region/country and a given sector are usually calculated as the product of an activity level, an emission factor for this activity and an abatement coefficient to account for regulations implemented to control emissions if existing (see Eq. 5.1 of Höglund-Isaksson et al., 2015; IPCC, 2006). The integrated emission models USEPA and GAINS provide estimates every 5 or 10 years for both past and future periods, while EDGAR provides annual estimates only for past emissions. These datasets differ in their assumptions and the data used for the calculation; however, they are not completely independent as they follow the IPCC guidelines (IPCC, 2006). While the USEPA inventory adopts the emissions reported by the countries to the UNFCCC, EDGAR and the GAINS model produced their own estimates using a consistent approach for all countries. As a result, the latter two approaches need large country-specific information or, if not available, they adopt IPCC default factors or emission factors reported to UNFCCC (Olivier et al., 2012; Höglund-Isaksson, 2012). Here, we also integrate the Food and Agriculture Organization (FAO) dataset, which provides estimates of methane emissions at country level but only for agriculture (enteric fermentation, manure management, rice cultivation, energy usage, burning of crop residues and of savannahs) and land use (biomass burning) (FAO, 2016). It will hereafter be referred as FAO-CH4. FAO-CH4 uses activity data from the FAOSTAT database as reported by countries to National Agriculture Statistical Offices (Tubiello et al., 2013) and mostly the Tier 1 IPCC methodology for emission factors (IPCC, 2006), which depend on geographic location and development status of the country. For manure, the necessary country-scale temperature was obtained from the FAO global agro-ecological zone database (GAEZv3.0, 2012).

We use the following versions of these inventories: version EDGARv4.2FT2010 that provides yearly gridded emissions by sectors from 2000 to 2010 (Olivier and Janssens-Maenhout, 2012; EDGARv4.2FT2010, 2013), version 5a of the GAINS model (Höglund-Isaksson, 2012) that assumes current legislation for air pollution for the future, the revised estimates of 2012 from the USEPA (2012), and finally the FAO emission database accessed in April 2016. Further details of the inventories used in this study are provided in Table 1. Overall, only EDGARv4.2FT2010 and GAINS provide gridded emission maps by sectors, and only EDGAR provides gridded maps on a yearly basis, which explains why this inventory is the most used in inverse modelling. These inventories are not all regularly updated. For the purpose of this study, the estimates from USEPA and GAINS have been linearly interpolated to provide yearly values, as provided by the EDGAR inventory. We also use the EDGARv4.2FT2012 data, which is an update of the time series of the country total emissions until 2012 (Rogelj et al., 2014; EDGARv4.2FT2012, 2014). This update has been developed based on EDGARv4.2FT2010 and uses IEA energy balance statistics (IEA, 2013) and NIR/CRF of UNFCCC (2013), as described in part III of IEA's CO2 book by Olivier and Janssens-Maenhout (2014).

For this study, engaged before the update of EDGARv4.2 inventory up to 2012, we built our own update from 2008 up to 2012 using FAO emissions to quantify CH4 emissions from enteric fermentation, manure management and rice cultivation (described above) and BP statistical review of fossil fuel production and consumption (http://www.bp.com/) to update CH4 emissions from coal, oil and gas sectors. In this inventory, called EDGARv4.2EXT, methane emissions after 2008 are set up equal to the FAO emissions (or BP statistics) of year t times the ratio between the mean EDGAR CH4 emissions (EEDGARv4.2) over 2006–2008 and the mean value of FAO emissions (VFAO in the following equation) (or BP statistics) over 2006–2008. For each emission sector, the country-specific emissions (EEDGARv4.2ext) in year (t) are estimated following Eq. (1): EEDGARv4.2EXT(t)==VFAOt×13∑i=20062008(EEDGARv4.2(i)/VFAO(i)). Other sources than those aforementioned are kept constant at the 2008 level. This extrapolation approach is necessary and often performed by top-down inversions to define prior emissions, because, up to now, global inventories such as sector-specific emissions in EDGAR database have not been updated on a regular basis. EC-JRC released, however, their update up to 2012 (EDGARv4.2FT2012) containing country total emissions, which allows evaluation of our extrapolation approach. The extrapolated global totals of EDGARv4.2EXT are within 1 % of EDGARv4.2FT2012.

Based on the ensemble of inventories detailed above, anthropogenic emissions are ∼ 352 [340–360] Tg CH4 yr-1 for the decade 2003–2012 (Table 2, including biomass and biofuel burning). For the 2000–2009 period, anthropogenic emissions are estimated at ∼ 338 [329–342] Tg CH4 yr-1. This estimate is consistent, albeit larger and with a smaller uncertainty range than Kirschke et al. (2013) for the 2000–2009 decade (331 Tg CH4 yr-1 [304–368]). Such differences are due to the different sets of inventories gathered. The range of our estimate (∼ 5 %) is smaller then the range reported in the AMAP assessment report (∼ 20 %) both because the latter was reporting more versions of the different inventories and projections, and because it was for the particular year 2005 and not for a decade as here.

Figure 2 presents the global methane emissions of anthropogenic sources (excluding biomass and biofuel burning) estimated and projected by the different inventories between 2000 and 2020. The inventories consistently estimate that about 300 Tg of methane was released into the atmosphere in 2000 by anthropogenic activities. The main discrepancy between the inventories is observed in their trend after 2005 with the lowest emissions projected by USEPA and the largest emissions estimated by EDGARv4.2FT2012. The increase in CH4 emissions is mainly determined from coal mining, whose activity increased considerably in China from 2002 to 2012 (see Sect. 3.1.3).

Despite relatively good agreement between the inventories on total emissions from year 2000 onwards, large differences can be found at the sector and country levels (IPCC, 2014). Some of these discrepancies are detailed in the following sections.

For the fifth IPCC Assessment Report, four representative concentration pathways (RCPs) were defined RCP8.5, RCP6, RCP4.5 and RCP2.6 (the latter is also referred to as RCP3PD, where “PD” stands for peak and decline). The numbers refer to the radiative forcing by the year 2100 in W m-2. These four independent pathways developed by four individual modelling groups start from the identical base year 2000 (Lamarque et al., 2010) and have been harmonized with historical emissions up to 2005. An interesting feature is the fact that global emission inventories track closer to methane emissions in the most carbon-intensive scenario (RCP8.5) and that all other RCP scenarios remain below the inventories. This suggests the tremendous challenge of climate mitigation that lies ahead, particularly if current trajectories need to change to be consistent with pathways leading to lower levels of global warming (Fig. 2).

Most of the methane anthropogenic emissions related to fossil fuels come from the exploitation, transportation, and usage of coal, oil and natural gas. This geological and fossil type of emission (see natural source section) is driven by human activity. Additional emissions reported in this category include small industrial contributions such as production of chemicals and metals, and fossil fuel fires. Spatial distribution of methane emissions from fossil fuel is presented in Fig. 3 based on the mean gridded maps provided by EDGARv4.2FT2010 and GAINS over the 2003–2012 decade.

Global emissions of methane from fossil fuels and other industries are estimated from three global inventories in the range of 114–133 Tg CH4 yr-1 for the 2003–2012 decade with an average of 121 Tg CH4 yr-1 (Table 2), but with a large difference in the rate of change depending on inventories. It represents on average 34 % (range 32–39 %) of the total global anthropogenic emissions.

During mining, methane is emitted from ventilation shafts, where large volumes of air are pumped into the mine to keep methane at a rate below 0.5 % to avoid accidental inflammation. To prevent the diffusion of methane in the mining working atmosphere, boreholes are made in order to evacuate methane. In countries of the Organization for Economic Co-operation and Development (OECD), methane recuperated from ventilation shafts is used as fuel, but in many countries it is still emitted into the atmosphere or flared, despite efforts for coal-mine recovery under the UNFCCC Clean Development Mechanisms (http://cdm.unfccc.int). Methane emissions also occur during post-mining handling, processing, and transportation. Some CH4 is released from coal waste piles and abandoned mines. Emissions from these sources are believed to be low because much of the CH4 would likely be emitted within the mine (IPCC, 2000).

Almost 40 % (IEA, 2012) of the world's electricity is produced from coal. This contribution grew in the 2000s at the rate of several per cent per year, driven by Asian production where large reserves exist, but has stalled from 2011 to 2012. In 2012, the top 10 largest coal producing nations accounted for 88 % of total world emissions for coal mining. Among them, the top three producers (China, USA and India) produced two-thirds of the total (CIA, 2016).

Global estimates of methane emissions from coal mining show a large variation, in part due to the lack of comprehensive data from all major producing countries. The range of coal mining emissions is estimated at 18–46 Tg of methane for the year 2005, the highest value being from EDGARv4.2FT2010 and the lower from USEPA.

As announced in Sect. 3.1.2, coal mining is the main source explaining the differences observed between inventories at global scale (Fig. 2). Indeed, such differences are explained mainly by the different CH4 emission factors used for calculating the fugitive emissions of the coal mining in China. Coal mining emission factors depend strongly on the type of coal extraction (underground mining emitting up to 10 times more than surface mining), the geological underground structure (very region-specific) and history (basin uplift), and the quality of the coal (brown coal emitting more than hard coal). The EDGARv4.2FT2012 seems to have overestimated by a factor of 2 the emission factor for the coal mining in China and allocated this to very few coal mine locations (hotspot emissions). A recent county-based inventory of Chinese methane emissions also confirms the overestimate of about +38 % with total anthropogenic emissions estimated at 43 ± 6 Tg CH4 yr-1 (Peng et al., 2016). Also, assimilating also 13CH4 data, Thompson et al. (2015) showed that their prior (based on EDGARv4.2FT2010) overestimated the Chinese methane emissions by 30 %; however, they found no significant difference in the coal sector estimates between prior and posterior. EDGARv4.2 follows the IPCC guidelines 2006, which recommends region-specific data. However, the EDGARv4.2 inventory compilation used the European averaged emission factor for CH4 from coal mine production in substitution for missing data, which seems to be twice too high in China. This highlights that significant errors on emission estimates may result from inappropriate use of some emission factor and that applying “Tier 1” for coal mine emissions is not accurate enough, as stated by the IPCC guidelines. The upcoming new version of EDGARv4.3.2 will revise this down and distribute the fugitive CH4 from coal mining to more than 80 times more coal mining locations in China.

For the 2003–2012 decade, methane emissions from coal mining are estimated at 34 % of total fossil-fuel-related emissions of methane (41 Tg CH4 yr-1, range of 26–50), consistent with the AMAP report when considering the evolution since 2005. An additional very small source corresponds to fossil fuel fires (mostly underground coal fires, ∼ 0.1 Tg yr-1, EDGARv4.2FT2010).

Natural gas is comprised primarily of methane, so any leaks during drilling of the wells, extraction, transportation, storage, gas distribution, and incomplete combustion of gas flares contribute to methane emissions (Lamb et al., 2015; Shorter et al., 1996). Fugitive permanent emissions (e.g. due to leaky valves and compressors) should be distinguished from intermittent emissions due to maintenance (e.g. purging and draining of pipes). During transportation, leakage can occur in gas transmission pipelines, due to corrosion, manufacturing, welding, etc. According to Lelieveld et al. (2005), the CH4 leakage from gas pipelines should be relatively low; however, distribution networks in older cities have increased leakage, especially those with cast-iron and unprotected steel pipelines. Recent measurement campaigns in different cities in the USA and Europe also revealed that significant leaks occur in specific locations (e.g. storage facilities, city gates, well and pipeline pressurization/depressurization points) along the distribution networks to the end-users (Jackson et al., 2014a; McKain et al., 2015). However, methane emissions can vary a lot from one city to another depending in part on the age of city infrastructure (i.e. older cities on average have higher emissions). Ground movements (landslides, earthquakes, tectonic movements) can also release methane. Finally, additional methane emissions from the oil industry (e.g. refining) and production of charcoal are estimated to be a few Tg CH4 yr-1 only (EDGARv4.2, 2011). In many facilities, such as gas and oil fields, refineries and offshore platforms, venting of natural gas is now replaced by flaring with a partial conversion into CO2; these two processes are usually considered together in inventories of oil and gas industries.

Methane emissions from oil and natural gas systems also vary greatly in different global inventories (46 to 98 Tg yr-1 in 2005; Höglund-Isaksson et al., 2015). The inventories rely on the same sources and magnitudes regarding the activity data. Thus, the derived differences result from different methodologies and parameters used, including both emission and activity factors. Those factors are country- or even site-specific, and the few field measurements available often combine oil and gas activities (Brandt et al., 2014) and remain largely unknown for most major oil- and gas-producing countries. Depending on the country, the emission factors reported may vary by 2 orders of magnitude for oil production and by 1 order of magnitude for gas production (Table 5.5 of Höglund-Isaksson et al., 2015). The GAINS estimate of methane emissions from oil production is 4 times higher than EDGARv4.2FT2010 and USEPA. For natural gas, the uncertainty is also large (factor of 2), albeit smaller than for oil production. The difference in these estimates comes from the methodology used. Indeed, during oil extraction, the gas generated can be either recovered (re-injected or utilized as an energy source) or not recovered (flared or vented to the atmosphere). The recovery rates vary from one country to another (being much higher in the USA, Europe and Canada than elsewhere), and accounting for country-specific rates of generation and recovery of associated gas might lead to an amount of gas released into the atmosphere 4 times higher during oil production than when using default values (Höglund-Isaksson, 2012). This difference in methodology explains, in part, why GAINS estimates are higher than EDGARv4.2FT2010 and USEPA. Another challenge lies in determining the amount of flared or vented unrecovered gas, with venting emitting CH4, whereas flaring converts all or most methane (often > 99 %) to CO2. The balance of flaring and venting also depends on the type of oil: flaring is less common for heavy oil wells than conventional ones (Höglund-Isaksson et al., 2015). Satellite images can detect flaring (Elvidge et al., 2009, 2016) and may be used to verify the country estimates, but such satellites cannot currently be used to estimate the efficiency of CH4 conversion to CO2.

For the 2003–2012 decade, methane emissions from upstream and downstream natural oil and gas sectors are estimated to represent about 65 % of total fossil CH4 emissions (79 Tg CH4 yr-1, range of 69–88, Table 2), with a lower uncertainty range than for coal emissions for most countries.

Production of natural gas from the exploitation of hitherto unproductive rock formations, especially shale, began in the 1980s in the US on an experimental or small-scale basis. Then, from early 2000s, exploitations started at large commercial scale. Two techniques developed and often applied together are horizontal drilling and hydraulic fracturing. The shale gas contribution to total natural gas production in the United States reached 40 % in 2012, growing rapidly from only small volumes produced before 2005 (EIA, 2015). Indeed, the practice of high-volume hydraulic fracturing (fracking) for oil and gas extraction is a growing sector of methane and other hydrocarbon production, especially in the US. Most recent studies (Miller et al., 2013; Moore et al., 2014; Olivier and Janssens-Maenhout, 2014; Jackson et al., 2014b; Howarth et al., 2011; Pétron et al., 2014; Karion et al., 2013) albeit not all (Allen et al., 2013; Cathles et al., 2012; Peischl et al., 2015) suggest that methane emissions are underestimated by inventories and agencies, including the USEPA. For instance, emissions in the Barnett Shale region of Texas from both bottom-up and top-down measurements showed that methane emissions from upstream oil and gas infrastructure were 90 % larger than estimates based on the USEPA's inventory and corresponded to 1.5 % of natural gas production (Zavala-Araiza et al., 2015). This study also showed that a few high emitters, neglected in the inventories, dominated emissions. Moreover these high emitting points, located on the conventional part of the facility, could be avoided through better operating conditions and repair of malfunctions. It also suggests that emission factor of conventional and non-conventional gas facilities might not be as different as originally thought (Howarth et al., 2011). Field measurements suggest that emission factors for unconventional gas are higher than for conventional gas, though the uncertainty, largely site-dependent, is large, ranging from small leakage rate of 1–2 % (Peischl et al., 2015) to widely spread rates of 3–17 % (Caulton et al., 2014; Schneising et al., 2014). For current technology, the GAINS model has adopted an emission factor of 4.3 % for shale-gas mining, still awaiting a clear consensus across studies.

This category includes methane emissions related to livestock (enteric fermentation and manure), rice cultivation, landfills, and waste-water handling. Of all types of emission, livestock is by far the largest emitter of CH4, followed by waste handling and rice cultivation. Field burning of agricultural residues was a minor source of CH4 reported in emission inventories. The spatial distribution of methane emissions from agriculture and waste handling is presented in Fig. 3 based on the mean gridded maps provided by EDGARv4.2FT2010 and GAINS over the 2003–2012 decade.

Global emissions for agriculture and waste are estimated at 195 Tg CH4 yr-1 (range 178–206, Table 2), representing 57 % of total anthropogenic emissions.

Domestic livestock such as cattle, buffalo, sheep, goats, and camels produce a large amount of methane by anaerobic microbial activity in their digestive systems (Johnson et al., 2002). A very stable temperature (39 ∘C), a stable pH (6.5–6.8) in their rumen, and constant flow of plants (cattle graze many hours per day) induce a production of metabolic hydrogen, used by methanogenic Archaea together with CO2 to produce methane. The methane and carbon dioxide are released from the rumen mainly through the mouth of multi-stomached ruminants (eructation, ∼ 87 % of emissions) or absorbed in the blood system. The methane produced in the intestines and partially transmitted through the rectum is only ∼ 13 %. There are about 1.4 billion cattle globally, 1 billion sheep, and nearly as many goats. The total number of animals is growing steadily (http://faostat3.fao.org), although the number is not linearly related to the CH4 emissions they produce; emissions are strongly influenced by the total weight of the animals and their diet. Cattle, due to their large population, large size, and particular digestive characteristics, account for the majority of enteric fermentation CH4 emissions from livestock, particularly, in the United States (USEPA, 2016). Methane emissions from enteric fermentation are also variable from one country to another as cattle experience water-limited conditions that highly vary spatially and temporally (especially in the tropics).

In addition, when livestock or poultry manure are stored or treated in systems that promote anaerobic conditions (e.g. as a liquid/slurry in lagoons, ponds, tanks, or pits), the decomposition of the volatile solids component in the manure tends to produce CH4. When manure is handled as a solid (e.g. in stacks or drylots) or deposited on pasture, range, or paddock lands, it tends to decompose aerobically and produce little or no CH4. Ambient temperature, moisture, and manure storage or residency time affect the amount of CH4 produced because they influence the growth of the bacteria responsible for CH4 formation. For non-liquid-based manure systems, moist conditions (which are a function of rainfall and humidity) can promote CH4 production. Manure composition, which varies with animal diet, growth rate, and type, including the animal's digestive system, also affects the amount of CH4 produced. In general, the greater the energy contents of the feed, the greater the potential for CH4 emissions. However, some higher-energy feeds also are more digestible than lower quality forages, which can result in less overall waste excreted from the animal (USEPA, 2006).

In 2005, global methane emissions from enteric fermentation and manure are estimated in the range of 96–114 Tg CH4 yr-1 in the GAINS model and USEPA inventory, respectively, and in the range of 98–105 Tg CH4 yr-1 suggested by Kirschke et al. (2013). They are consistent with the FAO-CH4 estimate of 102 Tg CH4 yr-1 for 2005 (Tubiello et al., 2013).

Here, for the 2003–2012 decade, based on all the databases aforementioned, we infer a range of 97–111 Tg CH4 yr-1 for the combination of enteric fermentation and manure with a mean value of 106 Tg CH4 yr-1 (Table 2), about one-third of total global anthropogenic emissions.

This sector includes emissions from managed and non-managed landfills (solid waste disposal on land), and waste-water handling, where all kinds of waste are deposited, which can emit significant amounts of methane by anaerobic decomposition of organic material by microorganisms. Methane production from waste depends on pH, moisture and temperature. The optimum pH for methane emission is between 6.8 and 7.4 (Thorneloe et al., 2000). The development of carboxylic acids leads to low pH, which limits methane emissions. Food or organic waste, leaves and grass clippings ferment quite easily, while wood and wood products generally ferment slowly, and cellulose and lignin even more slowly (USEPA, 2010b).

Waste management is responsible for about 11 % of total global anthropogenic methane emissions in 2000 at global scale (Kirschke et al., 2013). A recent assessment of methane emissions in the US accounts landfills for almost 26 % of total US anthropogenic methane emissions in 2014, the largest contribution of any CH4 source in the United States (USEPA, 2016). In Europe, gas control is mandatory on all landfills from 2009 onwards, following the ambitious objective raised in the EU Landfill Directive (1999) to reduce the landfilling of biodegradable waste by 65 % below the 1990 level by 2016. This is attempted through source separation and treatment of separated biodegradable waste in composts, bio-digesters and paper recycling. This approach is assumed more efficient in terms of reducing methane emissions than the more usual gas collection and capture. Collected biogas is either burned by flaring or used as fuel if it is pure enough (i.e. the content of methane is > 30 %). Many managed landfills have the practice to apply cover material (e.g. soil, clay, sand) over the waste being disposed of in the landfill to prevent odour, reduce risk to public health, as well as to promote microbial communities of methanotrophic organisms (Bogner et al., 2007). In developing countries, very large open landfills still exist, with important health and environmental issues in addition to methane emissions (André et al., 2014).

Waste water from domestic and industrial sources is treated in municipal sewage treatment facilities and private effluent treatment plants. The principal factor in determining the CH4 generation potential of waste water is the amount of degradable organic material in the waste water. Waste water with high organic content is treated anaerobically and that leads to increased emissions (André et al., 2014). The large and fast urban development worldwide, and especially in Asia, could enhance methane emissions from waste if adequate policies are not designed and implemented rapidly.

The inventories give robust emission estimates from solid waste in the range of 28–44 Tg CH4 yr-1 in the year 2005, and waste water in the range 9–30 Tg CH4 yr-1 given by GAINS model and EDGAR inventory.

In this study, global emissions of methane from landfills and waste are estimated in the range of 52–63 Tg CH4 yr-1 for the 2003–2012 period with a mean value of 59 Tg CH4 yr1, about 18 % of total global anthropogenic emissions.

Most of the world's rice is grown on flooded fields (Baicich, 2013). Under these shallow-flooded conditions, aerobic decomposition of organic matter gradually depletes most of the oxygen in the soil, resulting in anaerobic conditions under which methanogenic Archaea decompose organic matter and produce methane. Most of this methane is oxidized in the underlying soil, while some is dissolved in the floodwater and leached away. The remaining methane is released to the atmosphere, primarily by diffusive transport through the rice plants, but also methane escapes from the soil via diffusion and bubbling through floodwaters (USEPA, 2016; Bridgham et al., 2013).

The water management systems used to cultivate rice are one of the most important factors influencing CH4 emissions and is one of the most promising approach to mitigate the CH4 emissions from rice cultivation (e.g. periodical drainage and aeration not only causes existing soil CH4 to oxidize but also inhibits further CH4 production in soils (Simpson et al., 1995; USEPA, 2016; Zhang et al., 2016). Upland rice fields are not flooded and, therefore, are not believed to produce much CH4. Other factors that influence CH4 emissions from flooded rice fields include fertilization practices (i.e. the use of urea and organic fertilizers), soil temperature, soil type (texture and aggregated size), rice variety and cultivation practices (e.g. tillage, seeding, and weeding practices) (USEPA, 2011, 2016; Kai et al., 2011; Yan et al., 2009; Conrad et al., 2000). For instance, methane emissions from rice paddies increase with organic amendments (Cai et al., 1997) but can be mitigated by applying other types of fertilizers (mineral, composts, biogas residues, wet seeding) (Wassmann et al., 2000). Some studies have suggested that decreases in microbial emissions, particularly due to changes in the practice of rice cultivation, could be responsible for a ∼ 15 Tg CH4 yr-1 decrease over the period from 1980s to 2000s (Kai et al., 2011).

The geographical distribution of the emissions is assessed by global (USEPA, 2006, 2012; EDGARv4.2FT2010, 2013) and regional (Peng et al., 2016; Chen et al., 2013; Chen and Prinn, 2006; Yan et al., 2009; Castelán-Ortega et al., 2014; Zhang et al., 2014) inventories or by land surface models (Spahni et al., 2011; Zhang and Chen, 2014; Ren et al., 2011; Tian et al., 2010, 2011; Li et al., 2005; Pathak et al., 2005). The emissions show a seasonal cycle, peaking in the summer months in the extratropics associated with the monsoon and land management. Similar to emissions from livestock, emissions from rice paddies are influenced not only by extent of rice field area (equivalent to the number of livestock) but also by changes in the productivity of plants as these alter the CH4 emission factor used in inventories.

The largest emissions are found in Asia (Hayashida et al., 2013), with China (5–11 Tg CH4 yr-1; Chen et al., 2013; Zhang et al., 2016) and India (∼ 3–5 Tg CH4 yr-1; Bhatia et al., 2013) accounting for 30 to 50 % of global emissions (Fig. 3). The decrease of CH4 emissions from rice cultivation over the past decades is confirmed in most inventories, because of the decrease in rice cultivation area, the change in agricultural practices, and a northward shift of rice cultivation since 1970s (e.g. Chen et al., 2013). Furthermore, recent studies revealed that, together, high carbon dioxide concentrations and warmer temperatures predicted for the end of the twenty-first century will about double the amount of methane emitted per kilogramme of rice produced (van Groenigen et al., 2013).

Based on global inventories only, global methane emissions from rice paddies are estimated in the range 24–36 Tg CH4 yr-1 for the 2003–2012 decade, with a mean value of 30 Tg CH4 yr-1 (Table 2), about 9 % of total global anthropogenic emissions. The lower estimate (24 Tg CH4 yr-1) is provided by FAO-CH4 inventory (Tubiello et al., 2013), which is based on a mix of FAO statistics for crop production and IPCC guidelines.

This category includes all the combustion processes: biomass (forests, savannahs, grasslands, peats, agricultural residues) and biofuels in the residential sector (stoves, boilers, fireplaces). Biomass and biofuel burning emits methane under incomplete combustion conditions, when oxygen availability is insufficient such as charcoal manufacture and smouldering fires. The amount of methane that is emitted during the burning of biomass depends primarily on the amount of biomass, the burning conditions, and the material being burned. At the global scale, biomass and biofuel burning lead to methane emissions of 27–35 Tg CH4 yr-1 with an average of 30 Tg CH4 yr-1 (2003–2012 decade, Table 2), of which 30–50 % is biofuel burning (Kirschke et al., 2013).

In this study, we use the large-scale biomass burning (forest, savannah, grassland and peat fires) from specific biomass burning inventories and the biofuel burning contribution for the inventories (USEPA, GAINS and EDGAR).

The spatial distribution of methane emissions from biomass burning over the 2003–2012 decade is presented in Fig. 3 and is based on the mean gridded maps provided by EDGARv4.2FT2010 and GAINS for the biofuel burning, and based on the mean gridded maps provided by the biomass burning inventories presented thereafter.

Fire is the most important disturbance event in terrestrial ecosystems at the global scale (van der Werf et al., 2010) and can be of either natural (typically ∼ 10 %, ignited by lightning strikes or started accidentally) or anthropogenic origin (∼ 90 %, deliberately initiated fires) (USEPA, 2010a, chap. 9.1). Anthropogenic fires are concentrated in the tropics and subtropics, where forests, savannahs and C4 grasslands are burned to clear the land for agricultural purposes or to maintained pasturelands. In addition there are small fires associated with agricultural activity, such as field burning and agricultural waste burning, which are often undetected by commonly used remote-sensing products. Among the species emitting during biomass burning, carbon monoxide is a pertinent tracer for biomass burning emissions (Pechony et al., 2013; Yin et al., 2015).

Usually the biomass burning emissions are estimated using following Eq. (2) (or similar): E(xt)=A(x,t)×B(x)×FB×EF, where A(x,t) is the area burned, B(x) the biomass loading (depending on the biomes) at the location, FB the fraction of the area burned (or the efficiency of the fire depending of the vegetation type and the fire type) and EF the emission factor (mass of the considered species/mass of biomass burned). Depending on the approach, these parameters are derived using satellite data and/or biogeochemical model, or more simple equations.

The Global Fire Emission Database (GFED) is the most widely used global biomass burning emission dataset and provides estimates from 1997. In this review, we use both GFED3 (van der Werf et al., 2010) and GFED4s (Giglio et al., 2013; Randerson et al., 2012). GFED is based on the Carnegie–Ames–Stanford approach (CASA) biogeochemical model and satellite-derived estimates of burned area, fire activity and plant productivity. From November 2000 onwards, these three parameters are inferred from the MODerate resolution Imaging Spectroradiometer (MODIS) sensor. For the period prior to MODIS, burned area maps were derived from the Tropical Rainfall Measuring Mission (TRMM) Visible and Infrared Scanner (VIRS) and Along-Track Scanning Radiometer (ATSR) active fire data and estimates of plant productivity derived from Advanced Very High Resolution Radiometer (AVHRR) observations during the same period. GFED3 has provided biomass burning emission estimates from 1997 to 2011 at a 0.5∘ resolution on a monthly basis. The last versions of GFED (GFED4, without small fires, and GFED4s, with small fires) are available at a higher resolution (0.25∘) and on a daily basis from 2003 to 2014. Compared to GFED3, the main difference comes from the use of additional maps of the burned area product (MCD64A1) leading to a full coverage of land surface in GFED4 (Giglio et al., 2013). The particularity of GFED4s burned area is that small fires are accounted for (Randerson et al., 2012). Indeed small fires occur in several biomes (croplands, wooded savannahs, tropical forests) but are below the detection limit of the global burned area products. Yet the thermal anomalies they generate can be detected by MODIS for instance. Randerson et al. (2012) have shown that small fires increase burned area by approximately 35 % on the global scale leading to a 35 % increase of biomass burning carbon emissions when small fires were included in GFED3. Also it is worth noting that, between GFED3 and GFED4, the fuel consumption was lowered to better match observations (van Leeuwen et al., 2014) and that emission factor changes are substantial for some species and some biomes. Indeed global methane emissions are 25 % lower in GFED4 than in GFED3 mainly because of the new emission factors updated with Akagi et al. (2011).

The Fire INventory from NCAR (FINN, Wiedinmyer et al., 2011) provides daily, 1km resolution estimates of gas and particle emissions from open burning of biomass (including wildfire, agricultural fires and prescribed burning) over the globe for the period 2003–2014. FINNv1 uses MODIS satellite observations for active fires, land cover and vegetation density. The emission factors are from Akagi et al. (2011), the estimated fuel loading are assigned using model results from Hoelzemann et al. (2004), and the fraction of biomass burned is assigned as a function of tree cover (Wiedinmyer et al., 2006).

The Global Fire Assimilation System (GFAS, Kaiser et al., 2012) calculates biomass burning emissions by assimilating Fire Radiative Power (FRP) observations from MODIS at a daily frequency and 0.5∘ resolution and is available for the time period 2000–2013. After correcting the FRP observations for diurnal cycle, gaps etc., it is linked to dry matter combustion rate using Wooster et al. (2005) and CH4 emission factors from Andreae and Merlet (2001).

For FAO-CH4, yearly biomass burning emissions are based on burned area data from the Global Fire Emission Database v.4 (GFED4; Giglio et al., 2013). For forest, the GFED4 burned forest area is an aggregate of burned area in the following MODIS land cover classes (MCD12Q1, Hansen et al., 2000): evergreen needle-leaf, evergreen broadleaf, deciduous needle-leaf, deciduous broadleaf, and mixed forest. For “humid tropical forest”, burned area is obtained by overlapping GFED4 burned forest area data with the relevant FAO-FRA Global Ecological Zones (GAEZv3.0, 2012). For “other forest”, it is obtained by difference between other categories. FAO-CH4 biomass burning emissions are available from 1990 to 2014 (Table 1).

The differences in the biomass burning emission estimates arise from various difficulties among them the ability to represent and know the geographical and meteorological conditions and the fuel composition that highly impact the combustion completeness and the emission factors. Also methane emission factors vary greatly according to fire type, ranging from 2.2 g CH4 kg-1 dry matter burned for savannah and grassland fires up to 21 g CH4 kg-1 dry matter burned for peat fires (van der Werf et al., 2010).

Tian et al. (2016) estimated CH4 emissions from biomass burning during the 2000s (top-down, 17 ± 8 Tg C yr-1; bottom-up, 15 ± 5 Tg C yr-1). In this study, biomass burning emissions are estimated at 18 Tg CH4 yr-1 [15–21] for the decade 2003–2012, about 5 % of total global anthropogenic emissions.

Biomass that is used to produce energy for domestic, industrial, commercial, or transportation purposes is hereafter called biofuel burning. A largely dominant fraction of methane emissions from biofuels comes from domestic cooking or heating in stoves, boilers and fireplaces, mostly in open cooking fires where wood, charcoal, agricultural residues or animal dung are burnt. Although more than 2 billion people, mostly in developing and emerging countries, use solid biofuels to cook and heat their homes on a daily basis (André et al., 2014), methane emissions from biofuel combustion have not yet received the attention it should have to estimate its magnitude. Other much smaller contributors include agricultural burning (∼ 1–2 Tg yr-1) and road transportation (< 1 Tg yr-1). Biofuel burning estimates are gathered from USEPA, GAINS and EDGAR inventories.

In this study, biofuel burning is estimated to contribute 12 Tg CH4 yr-1 [10–14] to the global methane budget, about 3 % of total global anthropogenic emissions.

Wetlands are generally defined as ecosystems in which water saturation or inundation (permanent or not) dominates the soil development and determines the ecosystem composition (USEPA, 2010a). Such a broad definition needs to be refined when it comes to methane emissions. In this work, we define wetlands as ecosystems with inundated or saturated soils where anaerobic conditions lead to methane production (USEPA, 2010a; Matthews and Fung, 1987). This includes peatlands (bogs and fens), mineral wetlands (swamps and marshes), and seasonal or permanent floodplains. It excludes exposed water surfaces without emergent macrophytes, such as lakes, rivers, estuaries, ponds, and dams (addressed in the next section), as well as rice agriculture (see Sect. 3.1.4., rice cultivation paragraph). Even with this definition, one can consider that part of the wetlands could be considered as anthropogenic systems, being affected by human-driven land-use changes (Woodward et al., 2012). In the following we keep the generic denomination wetlands for natural and human-influenced wetlands.

A key feature of wetland systems producing methane is anaerobic soils, where high water table or flooded conditions limit oxygen availability and create conditions for methanogenesis. In anoxic conditions, organic matter can be degraded by methanogens that produce CH4. The three most important factors influencing methane production in wetlands are the level of anoxia (linked to water table), temperature and substrate availability (Wania et al., 2010; Valentine et al., 1994; Whalen, 2005). Once produced, methane can reach the atmosphere through a combination of three processes: molecular diffusion, plant-mediated transport and ebullition. On its way to the atmosphere, methane can be partly or completely oxidized by a group of bacteria, called methanotrophs, which use methane as their only source of energy and carbon (USEPA, 2010a). Concurrently, methane from the atmosphere can diffuse into the soil column and be oxidized (see Sect. 3.3.4).

Land surface models estimate CH4 emissions through a series of processes, including CH4 production, CH4 oxidation and transport and are further regulated by the changing environmental factors (Tian et al., 2010; Xu et al., 2010; Melton et al., 2013). In these models, methane emissions from wetlands to the atmosphere are computed as the product of an emission density (which can be negative; mass per unit area and unit time) multiplied by a wetland extent (see the model intercomparison studies by Melton et al., 2013, and Bohn et al., 2015). The CH4 emission density is represented in land surface models with varying levels of complexity. Many models link CH4 emission to net primary production (NPP) though production of exudates or litter and soil carbon to yield heterotrophic respiration estimates. A proportion of the heterotrophic respiration estimate is then taken to be CH4 production (Melton et al., 2013). The oxidation of produced (and becoming atmospheric) methane in the soil column is then either represented explicitly (e.g. Riley et al., 2011; Grant and Roulet, 2002), or just fixed proportionally to the production (Wania et al., 2013).

In land surface models, wetland extent is either prescribed (from inventories or remote-sensing data) or computed using hydrological models accounting for the fraction of grid cell with flat topography prone to high water table (e.g. Stocker et al., 2014; Kleinen et al., 2012), or from data assimilation against remote-sensed observations (Riley et al., 2011). Mixed approaches can also be implemented with tropical extent prescribed from remote sensing and northern peatland extent explicitly computed (Melton et al., 2013). Wetland extent appears to be a large contributor to uncertainties in methane emissions from wetlands (Bohn et al., 2015). For instance, the maximum wetland extent on a yearly basis appeared to be very different among land surface models in Melton et al. (2013), ranging from 7 to 27 Mkm2. Passive and active remote-sensing data in the microwave domain have been used to retrieve inundated areas, as with the Global Inundation Extent from Multi-Satellites product (GIEMS, Prigent et al., 2007; Papa et al., 2010). These remote-sensed data do not exactly correspond to wetlands, as all flooded areas are not wetlands (in methane emission sense) and some wetlands (e.g. northern bogs) are not always flooded. Inundated areas also include inland water bodies (lakes, ponds, estuaries) and rice paddies, which have to be filtered out to compute wetland emissions. Overall, current remote sensing of wetlands tends to underestimate wetland extent partly because of signal deterioration over dense vegetation and partly because microwave signals only detect water above or at the soil surface and therefore do not detect emitting peatlands that are not inundated (Prigent et al., 2007). For example, the Global Lakes and Wetlands Dataset (GLWD) (Lehner and Döll, 2004) estimates between 8.2 and 10.1 Mkm2 of wetlands globally, while remote-sensing inundation area is smaller, i.e. ∼ 6 Mkm2 (Prigent et al., 2007). Some ancillary data used in the GIEMS processing are not available after 2007 and have prevented so far the extension of the dataset after 2007.

Integrated at the global scale, wetlands are the largest and most uncertain source of methane to the atmosphere (Kirschke et al., 2013). An ensemble of land surface models estimated the range of methane emissions of natural wetlands at 141–264 Tg CH4 yr-1 for the 1993–2004 period, with a mean and 1σ value of 190 ± 39 Tg CH4 yr-1 (Melton et al., 2013). Kirschke et al. (2013) assessed a consistently large emission range of 142–287 Tg CH4 yr-1, using the Melton et al. (2013) land surface models and atmospheric inversions. These emissions represent about 30 % of the total methane source. The large range in the estimates of wetland emissions results from difficulties in defining wetland CH4-producing areas as well as in parameterizing terrestrial anaerobic sources and oxidative sinks (Melton et al., 2013; Wania et al., 2013).

In this work, following Melton et al. (2013), 11 land surface models (Table 1) computing net CH4 emissions have been run under a common protocol with a 30-year spin-up (1901–1930) followed by a simulation until the end of 2012 forced by CRU-NCEP v4.0 reconstructed climate fields. Atmospheric CO2 influencing NPP was also prescribed in the models, allowing the models to separately estimate carbon availability for methanogenesis. In all models, the same wetland extent (SWAMPS-GLWD) has been prescribed. The SWAMPS-GLWD is a monthly global wetland area dataset, which has been developed to overcome the aforementioned issues and combines remote-sensing data from Schroeder et al. (2015) and GLWD inventory in order to develop a monthly global wetland area dataset (Poulter et al., 2016). Briefly, GLWD was used to set the annual mean wetland area, to which a seasonal cycle of fractional surface water was added using data from the Surface WAter Microwave Product Series Version 2.0 (SWAMPS) (Schroeder et al., 2015). The combined GLWD-SWAMPS product leads to a maximum annual wetland area of 10.5 Mkm2 (8.7 Mkm2 on average, about 5.5 % of than global land surface). The largest wetland areas in the SWAMPS-GLWD are in Amazonia, the Congo Basin, and the western Siberian lowlands, which in previous studies have appeared to be strongly underestimated by several inventories (Bohn et al., 2015). However, wetlands above 70∘ N appear under-represented in GLWD as compared to Sheng et al. (2004) and Peregon et al. (2008). Indeed, approximately half of the global natural wetland area lies in the boreal zone between 50 and 70∘ N, while 35 % can be found in the tropics, between 20∘ N and 30∘ S (Matthews and Fung, 1987; Aselmann and Crutzen, 1989). Despite the lower area extent, the higher per-unit area methane emissions of tropical wetlands results in a larger wetland source from the tropics than from the boreal zone (Melton et al., 2013).

The average emission map from wetlands for 2003–2012 built from the 11 models is plotted in Fig. 3. The zones with the largest emissions reflect the GLWD database: the Amazon basin, equatorial Africa and Asia, Canada, western Siberia, eastern India, and Bangladesh. Regions where methane emissions are robustly inferred (i.e. regions where mean flux is larger than the standard deviation of the models) represent 80 % of the total methane flux due to natural wetlands. Main primary emission zones are consistent between models, which is clearly favoured by the common wetland extend prescribed. But still, the different sensitivity of the models to temperature can generate substantial different patterns, such as in India. Some secondary (in magnitude) emission zones are also consistently inferred between models: Scandinavia, continental Europe, eastern Siberia, central USA, and tropical Africa. Using improved regional methane emission datasets (such as studies over North America, Africa, China, and Amazon) can enhance the accuracy of the global budget assessment (Tian et al., 2011; Xu and Tian, 2012; Ringeval et al., 2014; Valentini et al., 2014).

The resulting global flux range for natural wetland emissions is 153–227 Tg CH4 yr-1 for the 2003–2012 decade, with an average of 185 Tg CH4 yr-1 with a 1σ standard deviation of 21 Tg CH4 yr-1 (Table 2).

This category includes methane emissions from freshwater systems (lakes, ponds, rivers) and from brackish waters of estuaries. Methane emissions from fresh waters and estuaries occur through a number of pathways including (1) continuous or episodic diffusive flux across water surfaces, (2) ebullition flux from sediments, (3) flux mediated through the aerenchyma of emergent aquatic macrophytes (plant transport) in littoral environments, and also for reservoirs, (4) degassing of CH4 in the turbines, and (5) elevated diffusive emissions in rivers downstream of the turbines especially if water through the turbines is supplied from anoxic CH4-rich water layers in the reservoir (Bastviken et al., 2004; Guérin et al., 2006, 2016). It is very rare that complete emission budgets include all these types of fluxes. For methodological reasons many past and present flux measurements only account for the diffusive flux based on short-term flux chamber measurements where non-linear fluxes were often discarded. At the same time, diffusive flux is now recognized as a relatively small flux component in many lakes, compared to ebullition and plant fluxes (in lakes with substantial emergent macrophyte communities). The two latter fluxes are very challenging to measure, both typically being associated with shallow near-shore waters and having high spatiotemporal variability. Ebullition can also occur more frequently in areas with high sediment organic matter load and is by nature episodic with very high fluxes occurring over time frames of seconds followed by long periods without ebullition.

Freshwater contributions from lakes were first estimated to emit 1–20 Tg CH4 yr-1 based on measurements in two systems (Great Fresh Creek, Maryland, and Lake Erie; Ehhalt, 1974). A subsequent global emission estimate was 11–55 Tg CH4 yr-1 based on measurements from three arctic lakes and a few temperate and tropical systems (Smith and Lewis, 1992), and 8–48 Tg CH4 yr-1 using extended data from all of the lake rich biomes (73 lakes; Bastviken et al., 2004). Combining results from Bastviken et al. (2004) and Bastviken et al. (2011), Kirschke et al. (2013) reported a range of 8–73 Tg CH4 yr-1. Gradually, methane emissions from reservoirs and rivers have also been included in the most recent global estimate from fresh waters of 103 Tg CH4 yr-1, including emissions from non-saline lakes, reservoirs, ponds and rivers (data from 473 systems; Bastviken et al., 2011). Improved stream and river emission estimates of 27 Tg CH4 yr-1 were recently suggested (Stanley et al., 2016). Importantly, the previous estimates of inland water fluxes are not independent. Instead they represent updates from increasing data quantity and quality. It should also be noted that issues regarding spatiotemporal variability are not considered in consistent ways at present (Wik et al., 2016a; Natchimuthu et al., 2015).

Present data do not allow for separating inland water fluxes over the different time periods investigated in this paper. The global estimates provided are therefore assumed to be constant for this study. Here we combine the latest estimates of global freshwater CH4 emissions (Bastviken et al., 2011) with a more recent regional estimate for latitudes above 50∘ N at present (Wik et al., 2016b) and new extrapolations for tropical river emissions (Borges et al., 2015; Sawakuchi et al., 2014) and streams (Stanley et al., 2016). High-latitude lakes include both post-glacial lakes and thermokarst lakes (water bodies formed by thermokarst), the latter having larger emissions per square metre but smaller regional emissions than the former because of smaller areal extent (Wik et al., 2016b). Water body depth, sediment type, and ecoclimatic region are the key factors explaining variation in methane fluxes from lakes (Wik et al., 2016b).

Altogether, these studies consider data from more than 900 systems, of which ∼ 750 are located north of 50∘ N. In this context we only consider fluxes from open waters assuming that plant-mediated fluxes are included in the wetland emission term. The average total estimated open water emission including the recent estimates from smaller streams is 122 Tg CH4 yr-1. The uncertainty is high with a coefficient of variation ranging from 50 to > 100 % for various flux components and biomes (Bastviken et al., 2011) resulting in a minimum uncertainty range of 60–180 Tg CH4 yr-1. The present data indicate that lakes or natural ponds, reservoirs, and streams/rivers account for 62, 16 and 22 % of the average fluxes, respectively (given the large uncertainty the percentages should be seen as approximate relative magnitudes only).

Potentially, the emissions from reservoirs should be allocated to anthropogenic emissions (not done here). Regarding lakes and reservoirs, tropical (< 30∘ latitude) and temperate (30–50∘ latitude) emissions represent 49 and 33 % of the flux, respectively, with 18 % left for regions above 50∘ latitude. For comparison, approximately 40 % of the inland water surface area is found above 50∘ latitude in the Northern Hemisphere and 34 % of the area is situated between 20∘ S and 20∘ N (Verpoorter et al., 2014). Ebullition typically accounted for 50 to more than 90 % of the flux from the water bodies, while contributions from ebullition appear lower from rivers, although this is currently debated (e.g. Crawford et al., 2014). Several aspects will need consideration to reduce the remaining uncertainty in the freshwater fluxes, including the generation of flux measurement that is more representative in time and space and an update of global lake area databases (e.g. GLOWAB, Verpoorter et al., 2014).

Significant amounts of methane, produced within the Earth's crust, naturally migrate to the atmosphere through tectonic faults and fractured rocks. Major emissions are related to hydrocarbon production in sedimentary basins (microbial and thermogenic methane), through continuous exhalation and eruptions from onshore and shallow marine gas/oil seeps and through diffuse soil microseepage (after Etiope, 2015). Specifically, six source categories have been considered. Five are onshore sources: mud volcanoes (sedimentary volcanism), gas and oil seeps (independent of mud volcanism), microseepage (diffuse exhalation from soil in petroleum basins), geothermal (non-volcanic) manifestations and volcanoes. One source is offshore: submarine seepage (several types of gas manifestation at the seabed). Figure 4a shows the areas and locations potentially emitting geological methane, showing diffuse potential microseepage regions, macroseepage locations (oil–gas seeps, mud volcanoes) and geothermal/volcanic areas (built from Etiope, 2015), which represent more than 1000 emitting spots.

Studies since 2000 have shown that the natural release to the Earth's surface of methane of geological origin is an important global greenhouse gas source (Etiope and Klusman, 2002; Kvenvolden and Rogers, 2005; Etiope et al., 2008; USEPA, 2010a; Etiope, 2012, 2015). Indeed, the geological source is in the top-three natural methane sources after wetlands (and with freshwater systems) and about 10 % of total methane emissions, of the same magnitude or exceeding other sources or sinks, such as biomass burning, termites and soil uptake, considered in recent IPCC assessment reports (Ciais et al., 2013).

In this study, the following provided estimates were derived by bottom-up approaches based on (a) the acquisition of thousands of land-based flux measurements for various seepage types in many countries, and (b) the application of the same procedures typically used for natural and anthropogenic gas sources, following upscaling methods based on the concepts of “point sources”, “area sources”, “activity” and “emission factors”, as recommended by the air pollutant emission guidebook of the European Environment Agency (EMEP/EEA, 2009). Our estimate is consistent with a top-down global verification, based on observations of radiocarbon-free (fossil) methane in the atmosphere (Etiope et al., 2008; Lassey et al., 2007b), with a range of 33–75 Tg CH4 yr-1.

As a result, in this study, the global geological methane emission is estimated in the range of 35–76 Tg CH4 yr-1 (mean of 52 Tg CH4 yr-1), with 40 Tg CH4 yr-1 [30–56] for onshore emissions (10–20 Tg CH4 yr-1 for mud volcanoes, 3–4 Tg yr-1 for gas–oil seeps, 10–25 Tg yr-1 for microseepage, 2–7 Tg CH4 yr-1 for geothermal/volcanic manifestations) and 12 Tg CH4 yr-1 [5–20] for offshore emissions through marine seepage (Rhee et al., 2009; Berchet et al., 2016; Etiope, 2012; see Sect. 3.2.6 for offshore contribution explanations).

Termites are important decomposer organisms, which play a very relevant role in the cycling of nutrients in tropical and subtropical ecosystems (Sanderson, 1996). The degradation of organic matter in their gut, by symbiotic anaerobic microorganisms, leads to the production of CH4 and CO2 (Sanderson, 1996). The upscaling approaches which have been used to quantify the contribution of termites to global CH4 emissions (Sanderson, 1996; Sugimoto et al., 1998; Bignell et al., 1997) are affected by large uncertainties, mainly related to the effect of soil and mound environments on net CH4 emissions; the quantification of termite biomass for each ecosystem type; and the impact of land-use change on termite biomass. For all these factors, uncertainty mainly comes from the relatively small number of studies compared to other CH4 sources. In Kirschke et al. (2013) (see their Supplement), a reanalysis of CH4 emissions from termites at the global scale was proposed and CH4 emissions per unit of surface were estimated as the product of termite biomass, termite CH4 emissions per unit of termite mass and a scalar factor expressing the effect of land-use/land-cover change. The latter two terms were estimated from published literature reanalysis (Kirschke et al., 2013, Supplement). A climate zoning (following the Köppen–Geiger classification) was applied to updated climate datasets by Santini and Di Paola (2015) and was adopted to take into account different combinations of termite biomass per unit area and CH4 emission factor per unit of termite biomass. In the case of tropical climate, first termites' biomass was estimated by a simple regression model representing its dependence on gross primary productivity (Kirschke et al., 2013, Supplement), whereas termites' biomass for forest and grassland ecosystems of the warm temperate climate and for shrublands of the Mediterranean subclimate were estimated from data reported by Sanderson (1996). CH4 emission factor per unit of termite biomass was derived from published literature and was estimated equal to 2.8 mg CH4 g-1 termite h-1 for tropical ecosystems and Mediterranean shrublands (Kirschke et al., 2013) and 1.7 mg CH4 g-1 termite h-1 for temperate forests and grasslands (Fraser et al., 1986). Emissions were scaled up in GIS environment and annual CH4 fluxes computed for the three periods 1982–1989, 1990–1999 and 2000–2007 representative of the 1980s, 1990s and 2000s, respectively. CH4 emissions showed only little interannual and interdecadal variability (0.1 Tg CH4 yr-1) and strong regional variability with tropical South America and Africa being the main sources (36 and 30 % of the global total emissions, respectively) due to the extent of their natural forest and savannah ecosystems (Fig. 4b). For the 2000s, a global total of 8.7 ± 3.1 Tg CH4 yr-1 (range 3–15 Tg CH4 yr-1) was obtained. This value is close to the average estimate derived from previous upscaling studies which report values spanning from 2 to 22 Tg CH4 yr-1 (Ciais et al., 2013).

In this study, we adopt a value of 9 Tg CH4 yr-1 (range 3–15 Tg CH4 yr-1, Table 2).

As for domestic ruminants, wild ruminants eruct or exhale methane through the microbial fermentation process occurring in their rumen (USEPA, 2010a). Global emissions of CH4 from wild animals range from 2–6 Tg CH4 yr-1 (Leng, 1993) to 15 Tg CH4 yr-1 (Houweling et al., 2000). The global distribution of CH4 emissions from wild ruminants is generally estimated as a function of the percentage and type of vegetation consumed by the animals (Bouwman et al., 1997). However, as suspected, numerous and various wild animals live partly hidden in the forests, savannahs, etc., challenging the assessment of these emissions.

The range adopted in this study is 2–15 Tg CH4 yr-1 with a mean value of 10 Tg CH4 yr-1 (Table 2).

Possible sources of oceanic CH4 include the following: (1) leaks from geological marine seepage (see also Sect. 3.2.3); (2) production from sediments or thawing subsea permafrost; (3) emission from the destabilization of marine hydrates and (4) in situ production in the water column, especially in the coastal ocean because of submarine groundwater discharge (USEPA, 2010a). Once at seabed, methane can be transported through the water column by diffusion in a dissolved form (especially in the upwelling zones) or by ebullition (gas bubbles, e.g. from geological marine seeps), for instance, in shallow waters of continental shelves. Among these different origins of oceanic methane, hydrates have attracted a lot of attention. Methane hydrates (or sometimes called clathrates) are ice-like crystals formed under specific temperature and temperature conditions (Milkov, 2005). The stability zone for methane hydrates (high pressure, ambient temperatures) can be found in the shallow lithosphere (i.e. < 2000 m depth), either in the continental sedimentary rocks of polar regions or in the oceanic sediments at water depths greater than 300 m (continental shelves, sediment–water interface) (Kvenvolden and Rogers, 2005; Milkov, 2005). Methane hydrates can be either of biogenic origin (formed in situ at depth in the sediment by microbial activity) or of thermogenic origin (non-biogenic gas migrated from deeper sediments and trapped due to pressure/temperature conditions or due to some capping geological structure such as marine permafrost). The total stock of marine methane hydrates is large but uncertain, with global estimates ranging from hundreds to thousands of Pg CH4 (Klauda and Sandler, 2005; Wallmann et al., 2012).

If the production of methane at seabed can be of importance, for instance, marine seepages emit up to 65 Tg CH4 yr-1 globally at seabed level (USEPA, 2010a); more uncertain is the flux of oceanic methane reaching the atmosphere. For example, bubble plumes of CH4 from the seabed have been observed in the water column but not detected in the Arctic atmosphere (Westbrook et al., 2009; Fisher et al., 2011). A large part of the seabed CH4 production and emission is oxidized in the water column and does not reach the atmosphere (James et al., 2016). There are several barriers preventing methane from being expelled to the atmosphere. From the bottom to the top, gas hydrates and permafrost serve as a barrier to fluid and gas migration towards the seafloor (James et al., 2016). First, on centennial to millennium timescales, trapped gases may be released when permafrost is perturbed and cracks or through Pingo-like features. At present, microbial processes are the most important control on methane emissions from marine environments. Aerobic oxidation in the water column is a very efficient sink, which allows very little methane even from established and vigorous gas seep areas or even gas well blowouts such as the Deepwater Horizon from reaching the atmosphere. Anaerobic methane oxidation, first described by Reeburgh and Heggie (1977), coupled to sulfate reduction controls methane losses from sediments to the overlying water (Reeburgh, 2007). Methane only escapes marine sediments in significant amounts from rapidly accumulating sedimentary environments or via advective processes such as ebullition or groundwater flow in shallow shelf regions. Anaerobic methane oxidation was recently demonstrated to be able to keep up with the thaw front of thawing permafrost in a region that had been inundated within the past 1000 years (Overduin et al., 2015). Second, the oceanic pycnocline is a physical barrier limiting the transport of methane (and other species) towards the surface. Third, another important mechanism stopping methane from reaching the ocean surface is the dissolution of bubbles into the ocean water. Although bubbling is the most efficient way to transfer methane from the seabed to the atmosphere, the fraction of bubbles actually reaching the atmosphere is very uncertain and critically depends on emission depths (< 100–200 m, McGinnis et al., 2015) and on the size of the bubbles (> 5–8 mm; James et al., 2016). Finally, surface oceans are aerobic and contribute to the oxidation of dissolved methane (USEPA, 2010a). However, surface waters can be more supersaturated than the underlying deeper waters, leading to a methane paradox (Sasakawa et al., 2008). Possible explanations involve upwelling in areas with surface mixed layers covered by sea ice (Damm et al., 2015) or methane produced within the anoxic centre of sinking particles (Sasakawa et al., 2008), but more work is needed to correct such an apparent paradox.

All published estimates agree that contemporary global methane emissions from oceanic sources are only a small contributor to the global methane budget, but the range of estimates is relatively large from 1 to 35 Tg CH4 yr-1 when summing geological and other emissions (e.g. Rhee et al., 2009; Etiope, 2015; USEPA, 2010a). For geological emissions, the most used value is 20 Tg yr-1, relying on expert knowledge and literature synthesis proposed in a workshop reported in Kvenvolden et al. (2001); the authors of this study recognized that this first estimation needs to be revised. Since then, oceanographic campaigns have been organized, especially to sample bubbling areas. For instance, Shakhova et al. (2010, 2014) infer 8–17 Tg CH4 yr-1 emissions just for the East Siberian Arctic Shelf (ESAS), based on the extrapolation of numerous but local measurements, and possibly related to melting seabed permafrost (Shakhova et al., 2015). Because of the highly heterogeneous distribution of dissolved CH4 in coastal regions, where bubbles can reach the atmosphere, extrapolation of in situ local measurements to the global scale can be hazardous and lead to biased global estimates. Indeed, using very precise and accurate continuous atmospheric methane observations in the Arctic region, Berchet et al. (2016) showed that Shakhova's estimates are 4–8 times too large to be compatible with atmospheric signals. This recent result suggests that the current estimate of 20 Tg yr-1 for the global emissions due to geological seeps emissions to the atmosphere in coastal oceans is too large and needs revision. Applying crudely the Berchet et al. (2016) abatement factor leads to emissions as low as less than 5 Tg CH4 yr-1.

More studies are needed to sort out this discrepancy and we choose to report here the full range of 5–20 Tg CH4 yr-1 for marine geological emissions, with a mean value of 12 Tg CH4 yr-1.

Concerning non-geological ocean emissions (biogenic, hydrates), the most common value found in the literature is 10 Tg CH4 yr-1 (Rhee et al., 2009). It appears that most studies rely on the work of Ehhalt (1974), where the value was estimated on the basis of the measurements done by Swinnerton and co-workers (Lamontagne et al., 1973; Swinnerton and Linnenbom, 1967) for the open ocean, combined with purely speculated emissions from the continental shelf. Based on basin-wide observations using updated methodologies, three studies found estimates ranging from 0.2 to 3 Tg CH4 yr-1 (Conrad and Seiler, 1988; Bates et al., 1996; Rhee et al., 2009), associated with supersaturations of surface waters that are an order of magnitude smaller than previously estimated, both for the open ocean (saturation anomaly ∼ 0.04, see Rhee et al., 2009, Eq. 4) and for the continental shelf (saturation anomaly ∼ 0.2). In their synthesis indirectly referring to the original observations from Lambert and Schmidt (1993), Wuebbles and Hayhoe (2002) use a value of 5 Tg CH4 yr-1. Proposed explanations for discrepancies regarding sea-to-air methane emissions in the open ocean rely on experimental biases in the former study of Swinnerton and Linnenbom (1967) (Rhee et al., 2009). This may explain why the Bange et al. (1994) compilation cites a global source of 11–18 Tg CH4 yr-1 with a dominant contribution of coastal regions. Here, we report a range of 0–5 Tg CH4 yr-1, with a mean value of 2 Tg CH4 yr-1.

Concerning more specifically atmospheric emissions from marine hydrates, Etiope (2015) points that current estimates of methane air–sea flux from hydrates (2–10 Tg CH4 yr-1 in e.g. Ciais et al., 2013, or Kirschke et al., 2013) originate from the hypothetical values of Cicerone and Oremland (1988). No experimental data or estimation procedures have been explicitly described along the chain of references since then (Lelieveld et al., 1998; Denman et al., 2007; Kirschke et al., 2013; IPCC, 2001). It was recently estimated that ∼ 473 Tg CH4 was released in the water column over 100 years (Kretschmer et al., 2015). Those few Tg per year become negligible once consumption in the water column has been accounted for. While events such as submarine slumps may trigger local releases of considerable amounts of methane from hydrates that may reach the atmosphere (Etiope, 2015; Paull et al., 2002), on a global scale, present-day atmospheric methane emissions from hydrates do not appear to be a significant source to the atmosphere.

Overall, these elements suggest the necessity to revise to a lower value the current total oceanic methane source to the atmosphere. Summing biogenic, geological and hydrate emissions from oceans leads to a total oceanic methane emission of 14 Tg CH4 yr-1 (range 5–25). Refining this estimate requires performing more in situ measurements of atmospheric and surface water methane concentrations and of bubbling areas and would require the development of process-based models for oceanic methane linking sediment production and oxidation, transport and transformation in the water column and atmospheric exchange (James et al., 2016).

Permafrost is defined as frozen soil, sediment, or rock having temperatures at or below 0 ∘C for at least two consecutive years (ACIA, 2005; Arctic Research Commission, 2003). The total extent of permafrost zones of the Northern Hemisphere is about 15 % of the land surface, with values around 15 million square kilometres (Slater and Lawrence, 2013; Levavasseur et al., 2011; Zhang et al., 1999). Where soil temperatures have passed the 0 ∘C mark, thawing of the permafrost at its margins occurs, accompanied by a deepening of the active layer (Anisimov and Reneva, 2006) and possible formation of thermokarst lakes (Christensen et al., 2015). A total of 1035 ± 150 Pg of carbon can be found in the upper 3 m or permafrost regions, or ∼ 1300 Pg of carbon (1100 to 1500) Pg C for all permafrost (Hugelius et al., 2014; Tarnocai et al., 2009).

The thawing permafrost can generate direct and indirect methane emissions. Direct methane emissions rely on the release of the methane contained in the thawing permafrost. This flux to the atmosphere is small and estimated to be at maximum 1 Tg CH4 yr-1 at present (USEPA, 2010a). Indirect methane emissions are probably more important. They rely on the following: (1) methanogenesis induced when the organic matter contained in thawing permafrost is released; (2) the associated changes in land surface hydrology possibly enhancing methane production (McCalley et al., 2014); and (3) the formation of more thermokarst lakes from erosion and soil collapsing. Such methane production is probably already significant today and could be more important in the future associated with a strong positive feedback to climate change. However, indirect methane emissions from permafrost thawing are difficult to estimate at present, with no data yet to refer to, and in any case they largely overlap with wetland and freshwater emissions occurring above or around thawing areas.

Here, we choose to report here only the direct emission range of 0–1 Tg CH4 yr-1, keeping in mind that current wetland, thermokarst lakes and other freshwater methane emissions already likely include a significant indirect contribution originating from thawing permafrost. For the next century, it has been recently estimated that 5–15 % of the terrestrial permafrost carbon pool is vulnerable to release in the form of greenhouse gases, corresponding to 130–160 Pg C. The likely progressive release in the atmosphere of such an amount of carbon as carbon dioxide and methane will have a significant impact on climate change trajectory (Schuur et al., 2015). The underlying methane hydrates represent a substantial reservoir of methane, estimated up to 530 000 Tg of CH4 (Ciais et al., 2013). Present and future emissions related to this reservoir are very difficult to assess at the moment and require more studies.

A series of recent studies define three distinct pathways for the production and emission of methane by living vegetation. First, plants produce methane through an abiotic photochemical process induced by stress (Keppler et al., 2006). This pathway was criticized (e.g. Dueck et al., 2007; Nisbet et al., 2009), and although numerous studies have since confirmed aerobic emissions from plants and better resolved its physical drivers (Fraser et al., 2015), global estimates still vary by 2 orders of magnitude (Liu et al., 2015) meaning any potential implication for the global methane budget remains highly uncertain. Second, plants act as “straws”, drawing methane produced by microbes in anoxic soils (Rice et al., 2010; Cicerone and Shetter, 1981). Third, the stems of living trees commonly provide an environment suitable for microbial methanogenesis (Covey et al., 2012). Static chambers demonstrate locally significant through-bark flux from both soil-based (Pangala et al., 2013, 2015), and tree-stem-based methanogens (Wang et al., 2016). These studies indicate trees are a significant factor regulating ecosystem flux; however, estimates of biogenic plant-mediated methane emissions at broad scales are complicated by overlap with methane consumption in upland soil and production in wetlands. Integrating plant-mediated emissions in the global methane budget will require untangling these processes to better define the mechanisms, spatio-temporal patterns, and magnitude of these pathways.

OH radicals are produced following the photolysis of ozone (O3) in the presence of water vapour. OH is destroyed by reactions with CO, CH4, and non-methane volatile organic compounds, but since OH exists in photochemical equilibrium with HO2, the net effect of CH4 oxidation on the HOx budget also depends on the level of NOx (Lelieveld et al., 2002) and other competitive oxidants. Considering its very short lifetime (a few seconds, Lelieveld et al., 2004), it is not possible to estimate global OH concentrations directly from observations. Observations are generally carried out within the boundary layer, while the global OH distribution and variability are more influenced by the free troposphere (Lelieveld et al., 2016). A series of experiments were conducted by several chemistry-climate models and chemistry transport models participating in the Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP) to study the long-term changes in atmospheric composition between 1850 and 2100 (Lamarque et al., 2013). For the year 2000, the multimodel mean (14 models) global mass-weighted OH tropospheric concentration is 11.7 ± 1.0 × 105 molec cm-3 (range 10.3–13.4×105 molec cm-3, Voulgarakis et al., 2013), consistent with the estimates of Prather et al. (2012) at 11.2 ± 1.3 × 105 molec cm-3. However, it is worth noting that, in the ACCMIP estimations, the differences in global OH are larger between models than between pre-industrial, present and future emission scenario simulations. Indeed Lelieveld et al. (2016) suggest that tropospheric OH is buffered against potential perturbations from emissions, mostly due to chemistry and transport connections in the free troposphere, through transport of oxidants such as ozone. Besides the uncertainty on global OH concentrations, the OH distribution is highly discussed. Models are often high biased in the Northern Hemisphere leading to a NH / SH OH ratio greater than 1 (Naik et al., 2013). A methane inversion using a NH / SH OH ratio higher than 1 infers higher methane emissions in the Northern Hemisphere and lower in the tropics and in the Southern Hemisphere (Patra et al., 2014). However, there is recent evidence for parity in interhemispheric OH concentrations (Patra et al., 2014), which needs to be confirmed by other observational and model-derived estimates.

OH concentrations and their changes can be sensitive to climate variability (e.g. Pinatubo eruption, Dlugokencky et al., 1996), to biomass burning (Voulgarakis et al., 2015) and to anthropogenic activities. For instance, the recent increase of the oxidizing capacity of the troposphere in South and East Asia, associated with increasing NOx emissions and decreasing CO emissions (Mijling et al., 2013; Yin et al., 2015), possibly enhances CH4 consumption and therefore limits the atmospheric impact of increasing emissions (Dalsøren et al., 2009). Despite such large regional changes, the global mean OH concentration was suggested to have changed only slightly over the past 150 years (Naik et al., 2013). This is due to the concurrent increases of positive influences on OH (water vapour, tropospheric ozone, nitrogen oxides (NOx) emissions, and UV radiation due to decreasing stratospheric ozone) and of OH sinks (methane burden, carbon monoxide and non-methane volatile organic compound emissions and burden). However the sign and integrated magnitude (from 1850 to 2000) of OH changes is uncertain, varying from -13 to +15 % among the ACCMIP models (mean of -1 %, Naik et al., 2013). Dentener et al. (2003) found a positive trend in global OH concentrations of 0.24 ± 0.06 % yr-1 between 1979 and 1993, mostly explained by changes in the tropical tropospheric water vapour content. Accurate methyl chloroform atmospheric observations together with estimates of its emissions (Montzka and Fraser, 2003) allow an estimate of OH concentrations and changes in the troposphere from the 1980s. Montzka et al. (2011) inferred small interannual OH variability and trends (typical OH changes from year to year of less than 3 %) and attributed previously estimated large year-to-year OH variations before 1998 (e.g. Bousquet et al., 2005; Prinn et al., 2001) to overly large sensitivity of OH concentrations inferred from methyl chloroform measurements to uncertainties in the latter's emissions. However, Prinn et al. (2005) also showed lower post-1998 OH variability that they attributed to the lack of strong post-1998 El Niño events. For the ACCMIP models providing continuous simulations over the past decades, OH interannual variability ranged from 0.4 to 0.9 %, consistent but lower than the value deduced from methyl chloroform measurements. However these runs take into account meteorology variability but not emission interannual variability (e.g. from biomass burning) and thus are expected to simulate lower OH interannual variability than in reality. As methyl chloroform has reached very low concentrations in the atmosphere, in compliance with the application of the Montreal Protocol and its amendments, a replacement compound is needed to estimate global OH concentrations. Several hydrochlorofluorocarbons and hydrofluorocarbons have been tested (Miller et al., 1998; Montzka et al., 2011; Huang and Prinn, 2002) to infer OH but do not yet provide equivalent results to methyl chloroform.

We report here a climatological range of 454–617 Tg CH4 yr-1 as in Kirschke et al. (2013) for the total tropospheric loss of methane by OH oxidation in the 2000s.

Approximately 60 Tg CH4 yr-1 enters the stratosphere by cross-tropopause mixing and the Hadley circulation (Reeburgh, 2007). Stratospheric CH4 distribution is highly correlated to the changes in the Brewer–Dobson circulation (Holton, 1986) and may impact Arctic air through subsidence of isotopically heavy air depending on the polar vortex location (Röckmann et al., 2011). In the stratosphere, currently approximately 51 [16–84] Tg CH4 yr-1 (i.e. about 10 [3–16] % of the total chemical loss in the atmosphere) is lost through reactions with excited atomic oxygen O(1D), atomic chlorine (Cl), atomic fluorine (F) and OH (Voulgarakis et al., 2013; Williams et al., 2012). The fraction of the stratospheric loss due to the different oxidants is uncertain, possibly within 20–35 % due to halons, about 25 % due to O(1D), the rest being due to stratospheric OH (Neef et al., 2010). The oxidation of methane in the stratosphere produces significant amounts of water vapour, which has a positive radiative forcing, and stimulates the production of OH through its reaction with atomic oxygen (Forster et al., 2007). Stratospheric methane thus contributes significantly to the observed variability and trend in stratospheric water vapour (Hegglin et al., 2014). Uncertainties in the chemical loss of stratospheric methane are large, due to uncertain interannual variability in stratospheric transport as well as through its chemical interactions with stratospheric ozone (Portmann et al., 2012).

We report here a climatological range of 16–84 Tg CH4 yr-1 as in Kirschke et al. (2013).

Halogen atoms can also contribute to the oxidation of methane in the troposphere. Allan et al. (2005) measured mixing ratios of methane and δ13C–CH4 at two stations in the Southern Hemisphere from 1991 to 2003 and found that the apparent kinetic isotope effect of the atmospheric methane sink was significantly larger than that explained by OH alone. A seasonally varying sink due to atomic chlorine (Cl) in the marine boundary layer of between 13 and 37 Tg CH4 yr-1 was proposed as the explaining mechanism (Allan et al., 2007). This sink was estimated to occur mainly over coastal and marine regions, where NaCl from evaporated droplets of seawater react with NO2 to eventually form Cl2, which then UV dissociates to Cl. However significant production of nitryl chloride (ClNO2) at continental sites has been recently reported (Riedel et al., 2014) and suggests the broader presence of Cl, which in turn would expand the significance of the Cl sink in the troposphere. More work is needed on this potential re-evaluation of the Cl impact on the methane budget.

We report here a climatological range of 13–37 Tg CH4 yr-1 as in Kirschke et al. (2013).

Unsaturated oxic soils are sinks of atmospheric methane due to the presence of methanotrophic bacteria, which consume methane as a source of energy. Wetlands with temporally variable saturation can also act as methane sinks. Dutaur and Verchot (2007) conducted a comprehensive meta-analysis of field measurements of CH4 uptake spanning a variety of ecosystems. They reported a range of 36 ± 23 Tg CH4 yr-1 but also showed that stratifying the results by climatic zone, ecosystem and soil type led to a narrower range (and lower mean estimate) of 22 ± 12 Tg CH4 yr-1. A modelling study by Ridgwell et al. (1999) simulated the sink to be 20–51 Tg CH4 yr-1. Curry (2007) used a process-based methane consumption scheme coupled to a land surface model (and calibrated to field measurements) to obtain a global estimate of 28 Tg CH4 yr-1, with a range of 9–47 Tg CH4 yr-1, which is the result reported in Kirschke et al. (2013). Tian et al. (2016) further updated the CH4 uptake from soil, with the estimate of 30 ± 19 Tg CH4 yr-1. In that model, CH4 uptake was determined by the diffusion rate of methane and oxygen through the uppermost soil layer, which was in turn dependent upon the soil characteristics (e.g. texture, bulk density) and water content (Curry, 2007). Riley et al. (2011) used another process-based model and estimated a global atmospheric CH4 sink of 31 Tg CH4 yr-1. The methane consumption rate was also dependent on the available soil water, soil temperature and nutrient availability. Although not addressed in that model, it should be noted that if the soil water content increases enough to inhibit the diffusion of oxygen, the soil could become a methane source (Lohila et al., 2016). This transition can be rapid, thus creating areas that can be either a source or a sink of methane depending on the season.

Following Curry (2007), and consistent with Tian et al. (2015), we report here a climatological range of 9–47 Tg CH4 yr-1 as in Kirschke et al. (2013).

The global atmospheric lifetime is defined for a gas in steady state as the global atmospheric burden (Tg) of this gas divided by its global total sink (Tg yr-1) (IPCC, 2001). In a case of a gas whose local lifetime is constant in space and time, the atmospheric lifetime equals the decay time (e-folding time) of a perturbation. As methane is not in a steady state, we need to fit with a function that approaches steady state when calculating methane lifetime using atmospheric measurements (Sect. 4.1.1). Global models provide an estimate of the loss of the gas due to individual sinks, which can then be used to derive lifetime due to a specific sink. For example, methane's tropospheric lifetime is determined as global atmospheric methane burden divided by the loss from OH oxidation in the troposphere, sometimes called “chemical lifetime”, while its total lifetime corresponds to the global burden divided by the total loss including tropospheric loss from OH oxidation, stratospheric chemistry and soil uptake. Recent multimodel estimate of the tropospheric methane lifetime is of 9.3 years (range 7.1–10.6; Voulgarakis et al., 2013; Kirschke et al., 2013) and that of the total methane lifetime is 8.2 ± 0.8 years (for year 2000, range 6.4–9.2, Voulgarakis et al., 2013). The model results for total methane lifetime are consistent with, though smaller than, the value reported in Table 6.8 of the IPCC AR5 of 9.1 ± 0.9 years (which was the observationally constrained estimate of Prather et al., 2012) most commonly used in the literature (Ciais et al., 2013) and the steady-state calculation from atmospheric observations (9.3 years, Sect. 4.1.1).

Four observational networks provide globally averaged CH4 mole fractions at the Earth's surface: the Earth System Research Laboratory from US National Oceanic and Atmospheric Administration (NOAA/ESRL, Dlugokencky et al., 1994), the Advanced Global Atmospheric Gases Experiment (AGAGE, Prinn et al., 2000; Cunnold et al., 2002; Rigby et al., 2008), the Commonwealth Scientific and Industrial Research Organisation (CSIRO, Francey et al., 1999) and the University of California Irvine (UCI, Simpson et al., 2012). The data are archived at the World Data Centre for Greenhouse Gases (WDCGG) of the WMO Global Atmospheric Watch (WMO-GAW) programme, including measurements from other sites that are not operated as part of the four networks.

The networks differ in their sampling strategies, including the frequency of observations, spatial distribution, and methods of calculating globally averaged CH4 mole fractions. Details are given in the Supplement of Kirschke et al. (2013). For the global average values of CH4 concentrations presented here, all measurements are made using gas chromatography with flame ionization detection (GC/FID), although chromatographic schemes vary among the labs. Because GC/FID is a relative measurement method, the instrument response must be calibrated against standards. NOAA maintains the WMO CH4 mole fraction scale X2004A; NOAA and CSIRO global means are on this scale. AGAGE uses an independent standard scale (Aoki et al., 1992), but direct comparisons of standards and indirect comparisons of atmospheric measurements show that differences are below 5 ppb (WMO RoundRobin programme). UCI uses another independent scale that was established in 1978 and is traceable to NIST (Simpson et al., 2012) but has not been included in standard exchanges with other networks so differences with the other networks cannot be quantitatively defined. Additional experimental details are presented in the Supplement from Kirschke et al. (2013) and references therein.

In Fig. 1, (a) globally averaged CH4 and (b) its growth rate (derivative of the deseasonalized trend curve) through 2012 are plotted for a combination of the four measurement programmes using a procedure of signal decomposition described in Thoning et al. (1989). We define the annual increase GATM as the increase in the growth rate from 1 January in one year to 1 January in the next year. Agreement among the four networks is good for the global growth rate, especially since ∼ 1990. The long-term behaviour of globally averaged atmospheric CH4 shows a decreasing but positive growth rate (defined as the derivative of the deseasonalized mixing ratio) from the early 1980s through 1998, a near-stabilization of CH4 concentrations from 1999 to 2006, and a renewed period with positive but stable growth rates since 2007. When a constant atmospheric lifetime is assumed, the decreasing growth rate from 1983 through 2006 implies that atmospheric CH4 was approaching steady state, with no trend in emissions. The NOAA global mean CH4 concentration was fitted with a function that describes the approach to a first-order steady state (SS index): [CH4](t)=[CH4]SS-([CH4]SS-[CH4]0)e-t/τ; solving for the lifetime, τ, gives 9.3 years, which is very close to current literature values (e.g. Prather et al., 2012).

On decadal timescales, the annual increase is on average 2.1 ± 0.3 ppb yr-1 for 2000–2009, 3.5 ± 0.2 ppb yr-1 for 2003–2012 and 5.0 ± 1.0 ppb yr-1 for the year 2012. The two decadal values hide a jump in the growth rate after 2006. Indeed, from 1999 to 2006, the annual increase of atmospheric CH4 was remarkably small at 0.6 ± 0.1 ppb yr-1. In the last 8 years, the atmospheric growth rate has recovered to a level similar to that of the mid-1990s (∼ 5 ppb yr-1), before the stabilization period of 1999–2006, as stated in Kirschke et al. (2013).

In the 2000s, two space-borne instruments sensitive to atmospheric methane were put in orbit and have provided atmospheric methane column-averaged dry air mole fraction (XCH4), using either shortwave infrared spectrometry (SWIR) or thermal infrared spectrometry (TIR).

Between 2003 and 2012, the Scanning Imaging Absorption spectrometer for Atmospheric CartograpHY (SCIAMACHY) was operated on board the ESA ENVIronmental SATellite (ENVISAT), providing nearly 10 years of XCH4 sensitive to the atmospheric boundary layer (Burrows et al., 1995; Buchwitz et al., 2006; Dils et al., 2006; Frankenberg et al., 2011). These satellite retrievals were the first to be used for global and regional inverse modelling of methane fluxes (Meirink et al., 2008a; Bergamaschi et al., 2007, 2009). The relatively long time record allowed the analysis of the interannual methane variability (Bergamaschi et al., 2013). However, the use of SCIAMACHY necessitates important bias correction, especially after 2005 (up to 40 ppb from south to north) (Bergamaschi et al., 2009; Houweling et al., 2014; Alexe et al., 2015).

In January 2009, the JAXA satellite Greenhouse Gases Observing SATellite (GOSAT) was launched containing the TANSO-FTS instrument, which observes in the shortwave infrared (SWIR). Different retrievals of methane based on TANSO-FTS/GOSAT products are made available to the community (Yoshida et al., 2013; Schepers et al., 2012; Parker et al., 2011) based on two retrieval approaches: proxy and full physics. The proxy method retrieves the ratio of methane column (XCH4) and carbon dioxide column (XCO2), from which XCH4 is derived after multiplication with transport model-derived XCO2 (Chevallier et al., 2010; Peters et al., 2007; Frankenberg et al., 2006). It intends mostly to remove biases due to light scattering on clouds and aerosols and is highly efficient owing to the small spectral distance between CO2 and CH4 sunlight absorption bands (1.65 µm for CH4 and 1.60 µm for CO2). Because of this, scattering-induced errors are similar for XCO2 and XCH4 and cancel out in the ratio. The second approach is the full-physics algorithm, which retrieves the aerosol properties (amount, size and height) along with CO2 and CH4 columns (e.g. Butz et al., 2011). Although GOSAT retrievals still show significant unexplained biases (possibly also linked to atmospheric transport modelling; Locatelli et al., 2015) and limited sampling in cloud-covered regions and in the high-latitude winter, it represents an important improvement compared to SCIAMACHY both for random and systematic observation errors (see Table S2 of Buchwitz et al., 2016).

Atmospheric inversions based on SCIAMACHY or GOSAT CH4 retrievals have been carried out by different research groups (Monteil et al., 2013; Cressot et al., 2014; Alexe et al., 2015; Bergamaschi et al., 2013; Locatelli et al., 2015). For GOSAT, differences between the use of proxy and full-physics retrievals have been investigated. In addition, joint CO2–CH4 inversions have been conducted to investigate the use of GOSAT retrieved ratios avoiding a model-derived hard constraint on XCO2 (Pandey et al., 2015, 2016; Fraser et al., 2013). Results from some of these studies are reported in Sect. 5 of this paper.

The processes emitting methane discriminate differently its isotopologues (isotopes). The two main stable isotopes of CH4 are 13CH4 and CH3D, and there is also the radioactive carbon isotope 14C–CH4. Isotopic signatures are conventionally given by the deviation of the sample mole ratio (for example, R=13CH4/12CH4 or CH3D / CH4) relative to a given standard (Rstd) relative to a reference ratio, given in per mil as in Eq. (3). δ13CH4 or δDCH4=RRstd-1×1000 For the 13CH4 isotope, the conventional reference standard is known as Vienna Pee Dee Belemnite (VPDB), with Rpdb=0.0112372. The same definition applies to CH3D, with the Vienna Standard Mean Ocean Water (VSMOW) RSMOW=0.00015575. The isotopic composition of atmospheric methane is measured at a subset of surface stations (Quay et al., 1991, 1999; Lowe et al., 1994; Miller et al., 2002; Morimoto et al., 2006; Tyler et al., 2007). The mean atmospheric values are about -47 ‰ for δ13CH4 and -86/-96 ‰ for δD(CH4). Isotopic measurements are made mainly on flask air samples analysed with gas-chromatograph isotope ratio spectrometry for which an accuracy of 0.05 ‰ for δ13CH4 and 1.5 ‰ for δD(CH4) can be achieved (Rice et al., 2001; Miller et al., 2002). These isotopic measurements based on air flask sampling have relatively low spatial and temporal resolutions. Laser-based absorption spectrometers and isotope ratio mass spectrometry techniques have recently been developed to increase sampling frequency and allow in situ operation (McManus et al., 2010; Santoni et al., 2012).

Measurements of δ13CH4 can help to partition the different methanogenic processes of methane: biogenic (-70 to -55 ‰), thermogenic (-55 to -25 ‰) or pyrogenic (-25 to -15 ‰) sources (Quay et al., 1991; Miller et al., 2002; Fisher et al., 2011) or even the methanogenic pathway (McCalley et al., 2014). δD(CH4) provides valuable information on the oxidation by the OH radicals (Röckmann et al., 2011) due to a fractionation of about 300 ‰. Emissions also show substantial differences in δD(CH4) isotopic signatures: -200 ‰ for biomass burning sources vs. -360 to -250 ‰ for biogenic sources (Melton et al., 2012; Quay et al., 1999). 14C–CH4 measurements (Quay et al., 1991, 1999; Lowe et al., 1988) may also help to partition for fossil fuel contribution (radiocarbon-free source). For example, Lassey et al. (2007a) used more than 200 measurements of radioactive 14C–CH4 (with a balanced weight between Northern and Southern hemispheres) to further constrain the fossil fuel contribution to the global methane source emission to 30 ± 2 % for the period 1986–2000.

Integrating isotopic information is important to improve our understanding of the methane budget. Some studies have simulated such isotopic observations (Neef et al., 2010; Monteil et al., 2011) or used them as additional constraints to inverse systems (Mikaloff Fletcher et al., 2004; Hein et al., 1997; Bousquet et al., 2006; Neef et al., 2010; Thompson et al., 2015). Using pseudo-observations, Rigby et al. (2012) found that quantum-cascade-laser-based isotopic observations would reduce the uncertainty in four major source categories by about 10 % at the global scale (microbial, biomass burning, landfill and fossil fuel) and by up to 50 % at the local scale. Although all source types cannot be separated using 13C, D and 14C isotopes, such data bring valuable information to constrain groups of sources in atmospheric inversions, if the isotopic signatures of the various sources can be precisely assessed (Bousquet et al., 2006, Supplement).

Other types of methane measurements are available, which are not commonly used to infer fluxes from inverse modelling (yet) but are used to verify its performance (see e.g. Bergamaschi et al., 2013). Aircraft or balloon-borne in situ measurements can deliver vertical profiles with high vertical resolution. Such observations can also be used to test remote-sensing measurement from space or from the surface and bring them on the same scale as the in situ surface measurements. Aircraft measurements have been undertaken in various regions either during campaigns (Wofsy, 2011; Beck et al., 2012; Chang et al., 2014; Paris et al., 2010) or in a recurrent mode using small aircrafts in the planetary boundary layer (Sweeney et al., 2015; Umezawa et al., 2014; Gatti et al., 2014) and commercial aircrafts (Schuck et al., 2012; Brenninkmeijer et al., 2007; Umezawa et al., 2012, 2014; Machida et al., 2008). Balloons can carry in situ instruments (e.g. Joly et al., 2008; using tunable laser diode spectrometry) or air samplers (e.g. air cores, Karion et al., 2010) up to 30 km height. New technologies have also developed systems based on cavity ring-down spectroscopy (CRDS), opening a large ensemble of new activities to estimate methane emissions such as drone measurements (light version of CRDS), as land-based vehicles for real-time, mobile monitoring over oil and gas facilities, as well as ponds, landfills, livestock, etc.

In October 2006, the Infrared Atmospheric Sounding Interferometer (IASI) on board the European MetOp-A satellite began to operate. Measuring the thermal radiation from Earth and the atmosphere in the TIR, it provides mid-to-upper troposphere columns of methane (representative of the 5–15 km layer) over the tropics using an infrared sounding interferometer (Crevoisier et al., 2009). Despite its sensitivity being limited to the mid-to-upper troposphere, its use in flux inversions has shown consistent results in the tropics with surface and other satellite-based inversions (Cressot et al., 2014).

The Total Carbon Column Observing Network (TCCON) uses ground-based Fourier transform spectrometers to measure atmospheric column abundances of CO2, CO, CH4, N2O and other molecules that absorb sunlight in the near-infrared spectral region (Wunch et al., 2011). As TCCON measurements make use of sunlight, they can be performed throughout the day during clear-sky conditions, with the sun typically 10∘ above the horizon. The TCCON network has been established as a reference for the validation of column retrievals, like those from SCIAMACHY and GOSAT. TCCON data can be obtained from the TCCON Data Archive, hosted by the Carbon Dioxide Information Analysis Center (CDIAC, http://cdiac.ornl.gov/).

An atmospheric inversion for methane fluxes (sources and sinks) optimally combines atmospheric observations of methane and associated uncertainties, a prior knowledge of the fluxes including their uncertainties, and a chemistry transport model to relate fluxes to concentrations (Rodgers, 2000). In this sense, top-down inversions integrate all the components of the methane cycle described previously in this paper. The observations can be surface or upper-air in situ observations, as well as satellite and surface retrievals. Prior emissions generally come from bottom-up approaches such as process-based models or data-driven extrapolations (natural sources) and inventories (anthropogenic sources). The chemistry transport model can be Eulerian or Lagrangian, and global or regional, depending on the scale of the flux to be optimized. Atmospheric inversions generally rely on the Bayes' theorem, which leads to the minimization of a cost function as Eq. (4): J(x)=12y-H(x)TR-1y-H(x)+12(x-xb)TB-1(x-xb), where y is a vector containing the atmospheric observations, x is a state vector containing the methane emissions and other appropriate variables (like OH concentrations or CH4 concentrations at the start of the assimilation window) to be estimated, xb is the prior state of x, and H is the observation operator, here the combination of an atmospheric transport and chemistry model and an interpolation procedure sampling the model at the measurement coordinates. R is the error covariance matrix of the observations and Pb is the error covariance matrix associated with xb. The errors on the modelling of atmospheric transport and chemistry are included in the R matrix (Tarantola, 1987). The minimization of a linearized version of J leads to the optimized state vector xa (Eq. 5): xa=xb+HT×R-1×H+Pb-1-1HT×R-1y-H(x), where Pa is given by Eq. (6) and represents the error covariance matrix associated with xa, and H contains the sensitivities of any observation to any component of state vector x (linearized version of the observation operator H(x)). Pa=HT×R-1×H+Pb-1-1 Unfortunately, the size of the inverse problem usually does not allow computing Pa, which is therefore approximated using the leading eigenvectors of the Hessian of J (Chevallier et al., 2005) or from stochastic ensembles (Chevallier et al., 2007). Therefore, the optimized fluxes xa are obtained using classical minimization algorithms (Chevallier et al., 2005; Meirink et al., 2008b). Alternatively, Chen and Prinn (2006) computed monthly emissions by applying a recursive Kalman filter in which Pa is computed explicitly for each month. Emissions are generally derived at weekly to monthly timescales, and for spatial resolutions ranging from model grid resolution to large aggregated regions. Spatio-temporal aggregation of state vector elements reduces the size of the inverse problem and allows the computation of Pa. However, such aggregation can also generate aggregation errors inducing possible biases in the inferred emissions and sinks (Kaminski et al., 2001). The estimated xa can represent either the net methane flux in a given region or contributions from specific source categories. Atmospheric inversions use bottom-up models and inventories as prior estimates of the emissions and sinks in their setup, which make bottom-up and top-down approaches generally not independent.

A group of eight atmospheric inversion systems using global Eulerian transport models were used in this synthesis. Each inversion system provides from 1 to 10 inversions, including sensitivity tests varying the assimilated observations (surface or satellite) or the inversion setup. This represents a total of 30 inversion runs with different time coverage: generally 2000–2012 for surface-based observations, 2003–2012 for SCIAMACHY-based inversions and 2009–2012 for GOSAT-based inversions (Table 3). When multiple sensitivity tests were performed we use the mean of this ensemble not to overweight one particular inverse model. Bias correction procedures have been developed to assimilate SCIAMACHY (Bergamaschi et al., 2009, 2013; Houweling et al., 2014) and GOSAT data (Cressot et al., 2014; Houweling et al., 2014; Locatelli et al., 2015; Alexe et al., 2015). These procedures can lead to corrections from several parts per billion and up to several tens of parts per billion (Bergamaschi et al., 2009; Locatelli et al., 2015). Although partly due to transport model errors, the large corrections applied to satellite total column CH4 data question the comparably low systematic errors reported in satellite validation studies using TCCON (Dils et al., 2014; CCI-Report, 2016). It should also be noticed that some satellite-based inversions are in fact combined satellite and surface inversions as they use either instantaneous in situ data simultaneously (Bergamaschi et al., 2013; Alexe et al., 2015) or annual mean surface observations to correct satellite bias (Locatelli et al., 2015). Nevertheless, these inversions are still referred to as satellite-based inversions.

General characteristics of the inversion systems are provided in Table 3. Further detail can be found in the referenced papers. Each group was asked to provide gridded flux estimates for the period 2000–2012, using either surface or satellite data, but no additional constraints were imposed so that each group could use their preferred inversion setup. This approach is appropriate for our purpose of flux assessment but not necessarily for model intercomparison. We did not require posterior uncertainty from the different participating groups, which may be done for the next release of the budget. Indeed chemistry transport models have some limitations that impact on the inferred methane budget, such as discrepancies in interhemispheric transport, stratospheric methane profiles and OH distribution. We consider here an ensemble of inversions gathering a large range of chemistry transport models, through their differences in vertical and horizontal resolutions, meteorological forcings, advection and convection schemes and boundary layer mixing; we assume that this model range is sufficient to cover the range of transport model errors in the estimate of methane fluxes. Each group provided gridded monthly maps of emissions for both their prior and posterior total and for sources per category (see the categories Sect. 2.3). Results are reported in Sect. 5. Atmospheric sinks were not analysed for this budget, which still relies on Kirschke et al. (2013) for bottom-up budget and on a global mass balance for top-down budget (difference between the global source and the observed atmospheric increase).

The last year of reported inversion results is 2012, which represents a 4-year lag with the present. Satellite observations are linked to operational data chains and are generally available within days to weeks after the recording of the spectra. Surface observations can lag from months to years because of the time for flask analyses and data checks in (mostly) non-operational chains. With operational networks such as ICOS in Europe, these lags will be reduced in the future. In addition, the final 6 months of inversions are generally ignored (spun down) because the estimated fluxes are not constrained by as many observations as the previous months. Finally, the long inversion runs and analyses can take up to months to be performed. For the next global methane budget the objective is to represent more recent years by reducing the analysis time and shortening the in situ atmospheric observation release.

At the global scale, the total emissions inferred by the ensemble of 30 inversions are 558 Tg CH4 yr-1 [540–570] for the 2003–2012 decade (Table 4), with a higher value of 568 Tg CH4 yr-1 [542–582] for 2012. Global emissions for 2000–2009 (552 Tg CH4 yr-1) are consistent with Kirschke et al. (2013), and the range of uncertainties for global emissions (535–566) is in line as well with that of Kirschke et al. (2013) (526–569), although 8 out of the 30 inversions presented here (∼ 25 %) are different. The latitudinal breakdown of emissions inferred from atmospheric inversions reveals a dominance of tropical emissions at 359 Tg CH4 yr-1 [339–386], representing 64 % of the global total. Thirty-two per cent of the emissions are from the midlatitudes and 4 % from high latitudes (above 60∘ N).

The picture given by the bottom-up approaches is quite different with global emissions of 736 Tg CH4 yr-1 [596–884] for 2003–2012 (Table 2). This estimate is much larger than top-down estimates. The bottom-up estimate is given by the sum of individual anthropogenic and natural processes, with no constraint on the total. As noticed in Kirschke et al. (2013), such a large global emissions rate is not consistent with atmospheric constraints brought by OH optimization and is very likely overestimated. This overestimation likely results from errors in the estimation of natural sources and sinks: extrapolation or double counting of some natural sources (e.g. wetlands, inland waters), or estimation of atmospheric sink terms. The anthropogenic sources are much more consistent between bottom-up and top-down approaches (Sect. 5.1.2).

The global methane budget for five source categories (see Sect. 2.3) for 2003–2012 is presented in Fig. 5 and Table 2. Top-down estimates attribute about 60 % of the total emissions to anthropogenic activities (range of 50–70 %) and 40 % to natural emissions. As natural emissions from bottom-up models are much larger, the anthropogenic vs. natural emission ratio is more balanced for bottom-up (∼ 50 % each). A predominant role of anthropogenic sources of methane emissions is strongly supported by the ice core and atmospheric methane records. The data indicate that atmospheric methane varied around 700 ppb during the last millennium before increasing by a factor of 2.6 to ∼ 1800 ppb. Accounting for the decrease in mean lifetime over the industrial period, Prather et al. (2012) estimate from these data a total source of 554 ± 56 Tg CH4 in 2010 of which about 64 % (352 ± 45 Tg CH4) are of anthropogenic origin, very consistent estimates with our synthesis.

For 2003–2012, the top-down and bottom-up derived estimates of respectively 167 Tg CH4 yr-1 (range 127–202) and 185 Tg CH4 yr-1 (range 153–227) are statistically consistent. Mean wetland emissions for the 2000–2009 period appear similar, albeit slightly smaller than found in Kirschke et al. (2013): 166 Tg CH4 yr-1 in this study vs. 175 Tg CH4 yr-1 in Kirschke et al. (2013) for top-down (-4 %) and 183 Tg CH4 yr-1 in this study vs. 217 Tg yr-1 in Kirschke et al. (2013) for bottom-up (-15 %). Note that more inversions (top-down) and more wetland models (bottom-up) were used in this study. Inversions have difficulty in separating wetlands from other sources so that uncertainties on top-down wetland emissions remain large. In this study, all bottom-up models were forced with the same wetland extent and climate forcings (Poulter et al., 2016), with the result that the amplitude of the range of emissions of 151–222 for 2000–2009 has narrowed by a third compared to the previous estimates from Melton et al. (2013) (141–264) and from Kirschke et al. (2013) (177–284). This suggests that differences in wetland extent explain about a third (30–40 %) of the former range of the emission estimates of global natural wetlands. The remaining range is due to differences in model structures and parameters. It is also worth noting that bottom-up and top-down estimates differ less in this study (∼ 17 Tg yr-1 for the mean) than in Kirschke et al. (2013) (∼ 30 Tg yr-1), although results from many more models are reported here. For top-down inversions, natural wetlands represent 30 % on average of the total methane emissions but only 25 % for bottom-up models (because of higher total emissions inferred by bottom-up models).

The discrepancy between top-down and bottom-up budgets is the largest for the natural emission total, which is 384 Tg CH4 yr-1 [257–524] for bottom-up and only 231 Tg CH4 yr-1 [194–296] for top-down over the 2003–2012 decade. Processes other than natural wetlands (Fig. 5), namely freshwater systems, geological sources, termites, oceans, wild animals, wildfires, and permafrost, explain this large discrepancy. For the 2003–2012 decade, top-down inversions infer non-wetland natural emissions of 64 Tg CH4 yr-1 [21–132], whereas the sum of the individual bottom-up emissions is 199 Tg CH4 yr-1 [104–297]. The two main contributors to this large bottom-up total are freshwater (∼ 60 %) and geological emissions (∼ 20 %), both of which have large uncertainties without spatially explicit representation. Because of the discrepancy, this category represents 10 % of total emissions for top-down inversions but 27 % for bottom-up approaches.

Improved area estimates of freshwater emissions would be beneficial. For example, stream fluxes are difficult to assess because of the high-expected spatial variability and very uncertain areas of headwater streams where methane-rich groundwater may be rapidly degassed. There are also uncertainties in the geographical distinction between wetlands, small lakes (e.g. thermokarst lakes), and floodplains that will need more attention to avoid double counting. In addition, major uncertainty is still associated with representation of ebullition. The intrinsic nature of this large but very locally distributed flux highlights the need for cost-efficient high-resolution techniques for resolving the spatio-temporal variations of these fluxes. In this context of observational gaps in space and time, freshwater fluxes are considered underestimated until measurement techniques designed to properly account for ebullition become more common (Wik et al., 2016a). On the contrary, global estimates for freshwater emissions rely on upscaling of uncertain emission factors and emitting areas, with probable overlapping of wetland emissions (Kirschke et al., 2013), which may also lead to an overestimate. More work is needed, based on both observations and process modelling, to overcome these uncertainties.

For geological emissions, relatively large uncertainties come from the extrapolation of only a subset of direct measurements to estimate the global fluxes. Moreover, marine seepage emissions are still widely debated (Berchet et al., 2016), and particularly diffuse emissions from microseepage are highly uncertain. However, summing up all fossil-CH4-related sources (including the anthropogenic emissions) leads to a total of 173 Tg CH4 yr-1 [149–209], which is about 31 % [25–35 %] of global methane emissions. This result is consistent with 14C atmospheric isotopic analyses inferring a 30 % contribution of fossil-CH4 to global emissions (Lassey et al., 2007b; Etiope et al., 2008). All non-geological and non-wetland land source categories (wild animals, wildfires, termites, permafrost) have been evaluated at a lower level than in Kirschke et al. (2013) and contribute only 23 Tg CH4 yr-1 [9–36] to global emissions. From a top-down point of view, the sum of all natural sources is more robust than the partitioning between wetlands and other natural sources. To reconcile top-down inversions and bottom-up estimates, the estimation and proper partition of methane emissions from wetlands and freshwater systems should receive high priority.

Total anthropogenic emissions are found statistically consistent between top-down (328 Tg CH4 yr-1, range 259–370) and bottom-up approaches (352 Tg CH4 yr-1, range 340–360), although top-down average is about 7 % smaller than bottom-up average over 2003–2012. The partition of anthropogenic emissions between agriculture and waste, fossil fuel extraction and use, and biomass and biofuel burning also shows good consistency between top-down and bottom-up approaches (Table 2 and Fig. 7). For 2003–2012, agriculture and waste contributed 188 Tg CH4 yr-1 [115–243] for top-down and 195 Tg CH4 yr-1 [178–206] for bottom-up. Fossil fuel emissions contributed 105 Tg CH4 yr-1 [77–133] for top-down and 121 Tg CH4 yr-1 [114–133] for bottom-up. Biomass and biofuel burning contributed 34 Tg CH4 yr-1 [15–53] for top-down and 30 Tg CH4 yr-1 [27–35] for bottom-up. Biofuel methane emissions rely on very few estimates at the moment (Wuebbles and Hayhoe, 2002; GAINS model). Although biofuel is a small source globally (∼ 12 Tg CH4 yr-1), more estimates are needed to allow a proper uncertainty assessment. Overall for top-down inversions the global fraction of total emissions for the different source categories are 33 % for agriculture and waste, 20 % for fossil fuels, and 6 % for biomass and biofuel burnings. With the exception of biofuel emissions, the global uncertainty of anthropogenic emissions appears to be smaller than that of natural sources but with asymmetric uncertainty distribution (mean significantly different than median). In poorly observed regions, top-down inversions rely on the prior estimates and bring little or no additional information to constrain the (often) spatially overlapping emissions (e.g. in India, China). Therefore, the relative agreement between top-down and bottom-up may indicate the limited capability of the inversion to separate the emissions and should therefore be treated with caution. Although the uncertainty range of some emissions has been decreased in this study compared to Kirschke et al. (2013) (e.g. oceans, termites, geological), there is no uncertainty reduction in the regional budgets because of the larger range reported for emissions from freshwater systems.

At regional scale, for the 2003–2012 decade (Table 4 and Fig. 6), total methane emissions are dominated by Africa with 86 Tg CH4 yr-1 [73–108], tropical South America with a total of 84 Tg CH4 yr-1 [65–101], and South East Asia with 73 Tg CH4 yr-1 [55–84]. These three (mainly) tropical regions represent almost 50 % of methane emissions worldwide. The other high-emitting source regions are China (58 Tg CH4 yr-1 [51-72]), central Eurasia and Japan (46 Tg CH4 yr-1 [38–54]), contiguous USA (41 Tg CH4 yr-1 [34–49]), Russia (38 Tg CH4 yr-1 [31–44]), India (39 Tg CH4 yr-1 [37–46]) and Europe (28 Tg CH4 yr-1 [21–34]). The other regions (boreal and central North America, temperate South America, Oceania, oceans) contribute between 7 and 20 Tg CH4 yr-1. This budget is consistent with Kirschke et al. (2013) within the large ranges around the mean emissions, although larger emissions are found here for South America, South East Asia, and Europe and lower emissions are found for Africa, North America and China. The regions with the largest changes are usually the least constrained by the surface networks.

The different inversions assimilated either satellite- or ground-based observations. It is of interest to determine whether these two types of data provide consistent surface emissions. To do so, we computed global, hemispheric and regional methane emissions using satellite-based inversions and ground-based inversions separately for the 2010–2012 time period, which is the longest time period for which results from both GOSAT satellite-based and surface-based inversions were available. At the global scale, satellite-based inversions infer significantly higher emissions (+12 Tg CH4 yr-1, p=0.04) than ground-based inversions. At the regional scale, emissions varied between the satellite-based and surface-based inversions, although the difference is not statistically significant due to too few inversions and some outliers making the ensemble not robust enough. Yet the largest differences (satellite-based minus surface based inversions) are observed over the tropical region: tropical South America +11 Tg CH4 yr-1; southern Africa +6 Tg CH4 yr-1; India -6 Tg CH4 yr-1; and over China -7 Tg CH4 yr-1. Satellite data provide more constraints on fluxes in tropical regions than surface-based inversions, due to a much larger spatial coverage. It is therefore not surprising that most differences between these two types of observations are found in the tropical band. However, such differences could also be due to the larger systematic errors of satellite data as compared to surface networks (Dils et al., 2014). In this context, the way the stratosphere is treated in the atmospheric models used to produce atmospheric methane columns from remote-sensing measurements (e.g. GOSAT or TCCON) seems important to further investigate (Locatelli et al., 2015; Monteil et al., 2011; Bergamaschi et al., 2009). Recent papers have developed methodologies to extract tropospheric partial column abundances from the TCCON data (Saad et al., 2014; Wang et al., 2014). Such partitioning could help explain the discrepancies between atmospheric models and satellite data.

The analysis of the regional methane budget per source category (Fig. 7) can be performed both for bottom-up and top-down approaches but with limitations. A complementary view of the methane budget is also available as an interactive graphic produced using data visualization techniques (http://lsce-datavisgroup.github.io/MethaneBudget/). Moving the mouse over regions, processes or fluxes reveals their relative weights in the global methane budget and provides the mean values and the minimum–maximum ranges of their contributions (mean [min, max]). The total source estimates from the bottom-up approaches are further classed into finer subcategories. This graphic shows that there is good consistency between top-down and bottom-up approaches in the partition of anthropogenic emissions between agriculture and waste, fossil fuel extraction and use, and biomass and biofuel burning, and it also highlights the disequilibrium between top-down (left) and bottom-up (right) budgets, mainly due to natural sources. On the bottom-up side, some natural emissions are not (yet) available at regional scale (oceans, geological, inland waters). Therefore, the category “others” is not shown for bottom-up results in Fig. 7 and is not regionally attributed in the interactive graphic. On the top-down side, as already noted, the partition of emissions per source category has to be considered with caution. Indeed, using only atmospheric methane observations to constrain methane emissions makes this partition largely dependent on prior emissions. However, differences in spatial patterns and seasonality of emissions can still be constrained by atmospheric methane observations for those inversions solving for different sources categories (see Sect. 2.3).

Wetland emissions largely dominate methane emissions in tropical South America, boreal North America, southern Africa, temperate South America and South East Asia, although agriculture and waste emissions are almost as important for the last two regions. Agriculture and waste emissions dominate in India, China, contiguous USA, central North America, Europe and northern Africa. Fossil fuel emissions dominate in Russia and are close to agriculture and waste emissions in the region called central Eurasia and Japan. In China, fossil fuel emissions are on average close, albeit smaller, than agriculture and waste emissions. Comparison between bottom-up and top-down approaches shows good consistency, but one has to consider the generally large error bars, especially for top-down inversions. The largest discrepancy occurs for wetland emissions in boreal North America where bottom-up models infer larger emissions (32 Tg CH4 yr-1) than top-down inversions (13 Tg CH4 yr-1). Indeed, one particular bottom-up model infers a 61 Tg CH4 yr-1 emission for this region, largely above estimates from other models, which lie between 15 and 45 Tg CH4 yr-1. Top-down models results are consistent with the climatology proposed by Kaplan (2002), whereas bottom-up models are more in line, albeit larger, than the climatology of Matthews and Fung (1987), who infer about 30 Tg CH4 yr-1 for boreal North America. Interestingly, the situation is different for Russia where top-down and bottom-up approaches show similar mean emissions from natural wetlands (mostly boreal, ∼ 13–14 Tg CH4 yr-1), consistently with Kaplan (2002) but not with Matthews and Fung (1987), who infer almost 50 Tg CH4 yr-1 for Russia. Wetland emissions from Russia appear very uncertain, as also found by Bohn et al. (2015) for western Siberia. Wetland emissions from tropical South America are found more consistent in this work than in Kirschke et al. (2013), where top-down inversions showed 2 times less emission than bottom-up models. The larger number of bottom-up models (11 against 3) and top-down inversions (30 against 8) are plausible causes explaining the improved agreement in this tropical region, poorly constrained by the surface networks (Pison et al., 2013).

Anthropogenic emissions remain close between top-down and bottom-up approaches for most regions, again with the possibility that part of this agreement is due to the lack of information brought by atmospheric observations to top-down inversions for some regions. One noticeable exception is the lower emissions for China as compared to the prior, visible also in Fig. 6. A priori anthropogenic emissions for China are mostly provided by the EDGARv4.2 inventory. Starting from prior emissions of 67 Tg CH4 yr-1 [58–77], the mean of the atmospheric derived estimates for China is 58 Tg CH4 yr-1 [51–72], corresponding to a -14 % difference of the Chinese emissions. A t test performed for the available estimates suggests that the mean posterior total emission for China is different from the prior emission at the 95 % confidence level. Several atmospheric studies have already suggested a possible overestimation of methane emissions from coal in China in the EDGARv4.2 inventory (Bergamaschi et al., 2013; Kirschke et al., 2013; Tohjima et al., 2014; Umezawa et al., 2014). Indeed, comparing the results of top-down inversions to EDGARv4.2 inventory (maximum of bottom-up estimates for China in Fig. 7), fossil fuel emissions are reduced by 33 % from 30 to 20 Tg CH4 yr-1 (range 9–30) and agriculture and waste emissions are reduced by 27 % from 37 to 27 Tg CH4 yr-1 (range 16–37). This result is consistent with a new inventory for methane emissions from China based on county-scale data (43 ± 6 Tg yr-1), indicating that coal-related methane emissions are 37 % (-7 Tg yr-1) lower than reported in the EDGARv4.2 inventory (Peng et al., 2016) (see also Sect. 3.1.2). Thompson et al. (2015) showed that their prior (based on EDGARv4.2) overestimated the Chinese methane emissions by 30 %; however, they found no significant difference in the coal sector estimates between prior and posterior and attribute the difference to rice emissions. It demonstrates that inversions are capable of verifying regional emissions when biases in the inventories are substantial, as in the case of China.

In contrast to the Chinese estimates, emissions inferred for Africa and especially southern Africa are significantly larger than in the prior estimates (Fig. 6). For example, for southern Africa, the mean of the inversion ensemble is 44 Tg CH4 yr-1 [37–53], starting at a mean prior of 36 Tg CH4 yr-1 [27–35]. This is a 25 % increase compared to mean prior estimates for southern Africa. A t test performed for the available estimates suggests that the mean posterior for southern Africa is different from the prior at the 98 % confidence level. An increase of northern African emissions is also inferred from the ensemble of inversions but is less significant.

For all other regions, emission changes compared to prior values remain within the first and third quartiles of the distributions. In particular, contiguous USA (without Alaska) is found to emit 41 Tg CH4 yr-1 [34–49], which is close to the prior estimates. Top-down and bottom-up estimates are consistent for anthropogenic sources in this region. Only natural wetlands are lower as estimated by top-down models (9 Tg CH4 yr-1 [6–13]) than by bottom-up models (13 Tg CH4 yr-1 [6–23]).