Large and increasing methane emissions from Eastern Amazonia derived from satellite data, 2010-2018

We use a global inverse model, satellite data and flask measurements to estimate methane (CH4) emissions from South America, Brazil and the basin of the Amazon river for the period 2010 – 2018. We find that emissions from Brazil have risen during this period, most quickly in the Eastern Amazon basin, and that this is concurrent with increasing surface 20 temperatures in this region. Brazilian CH4 emissions rose from 49.8 ± 5.4 Tg yr-1 in 2010 – 2013 to 55.6 ± 5.2 Tg yr-1 in 2014 – 2017, with the wet season of December – March having the largest positive trend in emissions. Amazon basin emissions grew from 41.7 ± 5.3 Tg yr-1 to 49.3 ± 5.1 Tg yr-1 during the same period. We derive no significant trend in regional emissions from fossil fuels during this period. We find that our posterior distribution of emissions within South America is significantly and consistently changed from our prior estimates, with the strongest emission sources being in the far north of the continent 25 and to the south and south-east of the Amazon basin, at the mouth of the Amazon river and nearby marsh, swamp and mangrove regions. We derive particularly large emissions during the wet season of 2013/14, when flooding was prevalent over larger regions than normal within the Amazon basin. We compare our posterior CH4 mole fractions, derived from posterior fluxes, to independent observations of CH4 mole fraction taken at five lowerto mid-tropospheric vertical profiling sites over the Amazon and find that our posterior fluxes outperform prior fluxes at all locations. In particular the large emissions from the 30 eastern Amazon basin are shown to be in good agreement with independent observations made at Santarém, a location which has long displayed higher mole fractions of atmospheric CH4 in contrast with other basin locations. We show that a bottomup wetland flux model can neither match the variation in annual fluxes, nor the positive trend in emissions, produced by the

temperatures in this region. Brazilian CH4 emissions rose from 49.8 ± 5.4 Tg(CH4)/yr in 2010 -2013 to 55.6 ± 5.2 Tg(CH4)/yr in 2014 -2017, with the wet season of December -March having the largest positive trend in emissions. We derive no significant trend in regional emissions from fossil fuels during this period. We find that our posterior distribution of emissions within South America is significantly and consistently changed from our prior estimates, with the strongest emission sources being in the far north of the continent and to the south and south-east of the Amazon Basin, near the mouth 25 of the Amazon and in other wetland regions. We derive particularly large emissions during the wet season of 2013/14, when flooding was prevalent over larger regions than normal within the Amazon Basin. We compare our posterior CH4 mole fractions, derived from posterior fluxes, to independent observations of CH4 mole fraction taken at five lower to mid tropospheric vertical profiling sites over the Amazon and find that our posterior fluxes outperform prior fluxes at all locations. In particular the large emissions from the eastern Basin are shown to be in good agreement with independent 30 observations made at Santarém, a location which has long displayed higher mole fractions of atmospheric CH4 in contrast with other Basin locations. We show that a bottom-up flux model cannot match the variation in annual fluxes, nor the positive trend in emissions, produced by the inversion. Our results show that the Amazon alone was responsible for 24 ± 18% of the total global increase in CH4 flux during the study period, and it may contribute further in future due to its sensitivity to temperature changes.

Introduction
Methane (CH4), a strong greenhouse gas emitted from a variety of anthropogenic and natural sources, is second only to carbon dioxide (CO2) in its importance regarding the anthropogenic radiative forcing contributing to Earth's climate change (Myhre et al., 2013). Much of the CH4 that is emitted into the atmosphere is destroyed through reaction with the hydroxyl (OH) radical and other smaller sinks, but a net positive imbalance means that the atmospheric burden of CH4 has been 40 increasing steadily since preindustrial times (e.g. Rubino et al., 2019). With an atmospheric lifetime of approximately 9 years (Prather et al., 2012), CH4 is a potentially important species for short-term gains in mitigation of anthropogenic climate change (Shindell et al., 2012). However, the magnitude of the global sources of CH4 to the atmosphere, and of its sinks once in the atmosphere, are still not well quantified . The geographical distribution and sectoral attribution of methane emissions, and the inter-annual variation of these sources, are also uncertain (Saunois et al., 2016;Schaefer, 2019). 45 This leads to difficulties in assessing potential emission mitigation strategies, hampering our ability to meet and assess the criteria for limiting the global temperature increase put forward as part of the Paris climate agreement (Nisbet et al., 2019).
The atmospheric methane burden is now approximately 2.5 times higher than it was in 1750 (Rubino et al., 2019). The global mean burden stabilised between 2000 and 2007, after which it began increasing again (Nisbet et al., 2016). 50 Concerningly, the rate of increase of the atmospheric burden has accelerated since 2014 (Nisbet et al., 2019). This suggests that CH4 emissions have been increasing at an accelerated rate during the past decade, but our understanding of how emissions are changing is complicated by the following: (1) attributing a potential emission increase to a particular region and/or sector is complex, leading to conflicting hypotheses regarding the changing fluxes (e.g. Nisbet et al., 2016;Worden et al., 2017;Monks et al., 2018;Schaefer, 2019;Lan et al., 55 2019;Jackson et al., 2020); (2) the uncertainty surrounding the distribution and variation of tropospheric OH means that variations in this major atmospheric sink of methane might also have played some role in the stabilisation and renewed rise Rigby et al., 2017;Turner et al., 2017;McNorton et al., 2018); and, (3) whilst rising atmospheric mole fractions of many greenhouse gases signify increasing anthropogenic influence, the 60 changing isotopic signature of atmospheric CH4 as the burden rises appears to indicate that fossil fuel emissions are not the main contributors to the increase, and that other sectors could be responsible (Schaefer et al., 2016;Nisbet et al., 2019;Fujita et al., 2020), including anthropogenic agricultural emissions. However, it has been argued that increasing fossil fuel emissions could still be reconciled with the observed isotopic signature Howarth, 2019).

65
In general, anthropogenic emissions of CH4 from fossil fuels, agriculture and waste are better constrained than natural emissions, particularly in bottom-up inventories . The majority of natural emissions come from wetlands, with smaller contributions from inland freshwaters, oceans, termites, wild animals and geological seeps. There are also small but significant emissions from biomass burning, which are sometimes counted separately from other anthropogenic emissions despite often being due to agricultural land clearing (van der Werf et al., 2017). 70 Wetlands are the largest single-sector contributors to the global methane flux  and the basin of the Amazon river in South America, covering an area of approximately 6,000,000 km 2 , is a significant contributor to the global wetland CH4 emission budget Bloom et al., 2017). The majority of the Basin is within the borders of Brazil. There are also a number of other significant wetland sources within South America, and often 75 significant contributions from fires during warmer, drier years (van der Werf et al., 2017). Recent studies have suggested that there is also a direct contribution from trees in the Amazon, although there is likely some overlap with wetland fluxes in some inventories (Pangala et al., 2017). In fact, the contribution of each of these sources of CH4, along with their regional distribution and variance over time, is still relatively uncertain. Earlier estimates of CH4 emissions from the Amazon Basin ranged from 4 to 92 Tg(CH4)/yr (Melack et al., 2004;do Carmo et al., 2006;Kirschke et al., 2013), but recently estimates 80 have converged somewhat, e.g. 31.6 -41.1 Tg/yr , 42.7 ± 5.6 Tg /yr (including tree flux) (Pangala et al., 2017) and 44.4 ± 4.8 Tg yr (Ringeval et al., 2014). The global wetland total was recently estimated to be 148 ± 25 Tg(CH4)/yr from bottom-up estimates and 159 -200 Tg(CH4)/yr from top-down models , which implies that if the majority of the emissions from the Amazon are from wetlands, then the region contributes up to ~30% of the global CH4 wetland flux. 85 Many studies have attempted to estimate national CH4 emissions rather than from ecosystems such as the Amazon, partly as it will likely be easier for countries to put in place emission reduction protocols on a national basis. Some recent studies have therefore reported emission totals for the country of Brazil. The synthesis of  used a suite of top-down models to find a wide range of 47.3 -78.2 Tg(CH4)/yr total emissions from all sources within Brazil during the period 2008 90 -2017. Natural sources made up 26.9 -53.8 Tg(CH4)/yr of this total. Janardanan et al. (2020) used a global inversion to constrain total Brazilian emissions to 56.2 ± 10 Tg(CH4)/yr in the period 2011-2017. However, Tunnicliffe et al. (2020) used a high-resolution regional inversion to find much smaller emissions from the country, calculating total Brazilian emissions of 33.6 ± 3.6 Tg(CH4)/yr, with wetlands making up 13.0 ± 1.9 Tg(CH4)/yr of this total. The relatively large range of estimates produced by these studies, some of which make use of the same observational datasets, is indicative of the difficulties 95 inherent in using top-down methods to assess surface emissions of CH4 from within the poorly monitored continent of South America. However, in order to best understand the global methane budget and its sources, it is still vital that the significant contribution of South American emissions is evaluated and attributed.
In order to best unite these estimates, regular observation of atmospheric methane over South America is necessary. The 100 Thermal And Near infrared Sensor for carbon Observations -Fourier Transform Spectrometer (TANSO-FTS) instrument on https://doi.org/10.5194/acp-2020-1136 Preprint. Discussion started: 10 November 2020 c Author(s) 2020. CC BY 4.0 License. the GOSAT satellite (Kuze et al., 2009) is particularly advantageous, as it is sensitive far down into the troposphere and has been providing regular global coverage of atmospheric CH4 continuously since April 2009 (Parker et al., 2020a). This decade of uninterrupted global coverage allows for understanding of methane variations over a much longer time period than many of the other available datasets, particularly in the tropics. 105 In this paper we use CH4 observations from GOSAT along with flask measurements both from within and outside the Amazon Basin to provide an almost complete 10-year record of methane emissions from South America, beginning in 2009.
We use the TOMCAT chemical transport model and its inverse model, INVICAT, to quantify emissions and their uncertainties during this decade. Ours aims are to 1) assess the geographical distribution of South American CH4 emissions, with focus on the country of Brazil and the Amazon Basin ecosystem; 2) examine how these emissions have changed during 110 the previous decade; and 3) investigate why any changes to natural emissions might have occurred. We describe the observations used and the modelling methodology in Section 2. We show our results and discuss our findings in Section 3 and Section 4, respectively.

Observations
We assimilate both in-situ flask observations and GOSAT satellite retrievals of CH4 into the inverse model. We also hold back a set of observations made as part of regular flask-based aircraft monitoring campaign within the Amazon Basin since 2010, for validation of our results.

Surface flask observations 120
We assimilate global long-term surface data of CH4 provided by the National Oceanic and Atmospheric Administration's Global Monitoring Laboratory (NOAA GML) (Table A4). We use data from 56 background monitoring sites, the locations of which are shown in Figure 1. Whole air samples in flasks are collected weekly to biweekly at each site, and CH4 is measured using gas chromatography with a flame ionization detection method (Dlugokencky et al., 2018). Data from these sites is assimilated in order to constrain the background variations in CH4 mole fractions at the Earth's surface. The 125 observations made at these locations have high accuracy but are generally located in regions that are not near significant sources of CH4. There is also a relative lack of regular observations in tropical regions, where CH4 emissions are significant and uncertain. This means that these observations can provide accurate values for background CH4 values but are not usually able to provide accurate regional CH4 distributions in those areas that require the most constraint.

GOSAT observations
We also assimilate column-averaged dry-air mole fractions of CH4 (XCH4) from the University of Leicester Proxy retrieval scheme v7.2 for GOSAT (Parker et al., 2011(Parker et al., , 2020a. This dataset has been used in the past in forward modelling studies to assess wetland CH4 emissions using the TOMCAT model (Parker et al., , 2020b. The GOSAT Proxy scheme uses the ratio of the retrieved XCO2 and XCH4, together with model-based estimates of XCO2, in order to reduce the effects of 135 atmospheric scattering and improve coverage of XCH4 retrievals. This is particularly true in tropical land regions where the prevalence of cloudy pixels often restricts the successful direct retrieval of XCH4. GOSAT XCH4 retrievals have been used previously in a number of forward and inverse modelling studies (Fraser et al., 2013;McNorton et al., 2016;Feng et al., 2017;Miller et al., 2019). The observations are regularly validated against independent data, including CH4 observations made as part of the Total Carbon Column Observing Network (TCCON, Wunch et al., (2011)), although unfortunately none 140 of the measurement sites included as part of this network are located within the Amazon region. Webb et al. (2016) compared GOSAT XCH4 to vertical profile observations of CH4 taken over the Amazon Basin and found that the two agreed within their respective errors.
Before assimilation, GOSAT observations were averaged onto the model grid. Both sun-glint observations over the oceans 145 and nadir observations over land were included in the inversion. All XCH4 values measured by the satellite during one model timestep in the same grid cell were averaged using a weighted mean according to their uncertainties. The largest number of observations combined into a single value was 32, and the mean number was 4.7 over land and 6.0 over oceans. Within the Amazon Basin, the mean number of observations combined was 3.8. Figure 1 shows an example monthly distribution of observations used in the inversion. For accurate comparison between the retrieved XCH4 and those simulated by the model, 150 the GOSAT averaging kernels were averaged similarly to the XCH4 and applied to the model vertical profiles. This meant that the adjoint code for this process was also produced for this study. Retrievals for which the model and satellite surface pressure differed by more than 50 hPa were rejected.
Due to a range of potential error sources in both the atmospheric transport model and the GOSAT retrievals, there is a 155 persistent bias between them, which varies with latitude. We quantified this bias by comparing the results of a previous inversion, in which only the surface flask observations had been assimilated for the full 2009-2018 period, to the GOSAT XCH4. We applied the averaging kernels to the three-dimensional (3-D) CH4 output from the flask data inversion and calculated the model -observation zonal mean bias ( ), in parts per billion (ppb), as a function of latitude ( ), over this period: 160 where is equal to the latitude of the observation in degrees north. Positive values of ( ) indicate positive observation bias relative to the model. Across the tropics (30°S -30°N), the derived bias varies between 2.8 and 8.8 ppb. Further south, the bias reaches values up to 13.4 ppb. In the analysis below we add the estimated bias value to the simulated XCH4 values in the 165 inversion after the averaging kernels are applied.

Amazonian aircraft profiles
Since 2010, aircraft-borne flask air observations of a number of species, including CH4, CO2 and carbon monoxide (CO) have been made at five locations within the Amazon Basin (shown in Figure 1  Measurements were only ever made concurrently at four locations, as the measurements at Tefé were started in 2013, to replace those made at Tabatinga up to 2012. We therefore combine observations made at these locations and refer to them as 175 TAB/TEF throughout this manuscript. Both sites are located in the north-west of the Amazon Basin and sample similar air masses. Flights are undertaken at approximately biweekly intervals above each site up to an altitude of ~4.4 km, and 0.7 L flasks were filled every 300-500m to produce vertical profiles. All measurements were taken between 12:00 and 13:00 local time, when the boundary layer is fully developed. The flasks were analysed for CH4 mole fractions at the high-precision gas analytics laboratory at IPEN and INPE, following the NOAA GML approach, including rigorous calibration to the World 180 Meteorological Organization (WMO) CH4 mole fraction scale. The measurement locations were chosen in order to sample the dominant tropospheric airstream across the Basin. For more information about these measurements, see Gatti et al. (2014) and Basso et al. (2016).

Inverse model set-up 185
The TOMCAT model is a global 3-D Eulerian offline chemical transport model (CTM) (Chipperfield, 2006;Monks et al., 2017). It has been used in a number of previous studies of atmospheric composition and transport (e.g. Wilson et al., 2016;McNorton et al., 2016;Parker et al., 2018). We use the INVICAT inverse model , which is based on the TOMCAT model. INVICAT uses a variational scheme based on 4D-Var methods used in Numerical Weather Prediction (NWP) and has been used in the past to constrain emissions of other species Monks et al., 2018;190 Thompson et al., 2019;Tian et al., 2020). The inverse method employed by INVICAT is described in depth in these previous references.
In this study, the forward and adjoint model simulations were carried out at 5.6° horizontal resolution, with 60 vertical levels up to 0.1 hPa. The model time step was 30 minutes. The meteorology was taken from the European Centre for  Range Weather Forecasts (ECMWF) ERA-Interim reanalyses (ERA-I, Dee et al. (2011)). The inversions were carried out for each year separately and each completed 40 minimisation iterations. The inversion for each year was actually run for 14 months up to the end of February for the following year, with the final two months being discarded from the results. This was in order to better constrain fluxes during the final months of each year. Each inversion therefore overlapped with the following one for two months but was initialized using 3-D fields provided from the correct date in the previous year, so that 200 total CH4 burden was conserved across years.
For the assimilated surface observations, the model output was linearly interpolated to the correct longitude, latitude and altitude, at the nearest model timestep. For the averaged GOSAT observations, the model mole fractions were interpolated to the correct longitude and latitude at the nearest time step, before the GOSAT averaging kernels were applied to the model 205 output to give an XCH4 value comparable with GOSAT. GOSAT observations were given an uncorrelated uncertainty equal to 2.5 times the supplied retrieval error, which ranged from 3.5 to 25.8 ppb, in order to account for representation error and observation correlations removed by the averaging of the retrievals, as in Chevallier (2007). This inflation value was based on the mean number of observations combined in each grid cell. In short sensitivity tests, the magnitude of posterior emissions was not sensitive to this inflation factor once it was larger than 2, although the posterior error estimate was 210 affected. This choice gave a mean GOSAT XCH4 uncertainty value of 24.4 ppb. NOAA observations were given uncorrelated errors of 3 ppb plus representation error. For these observations, representation error was estimated as the mean difference across the 8 grid cells surrounding the cell containing the observation location.
Prior emissions were given grid cell uncertainties of 250%, but also included spatial and temporal correlations. Although 215 inversions such as this do not directly allow for sectorial analysis of emissions, we used the off-diagonal values of the prior covariance matrix to provide some information of this nature. Similar to Meirink et al., (2008), we split out prior and posterior solutions into the anthropogenic fossil fuel emissions assumed to be strongly correlated in time (FF), and emissions with strong seasonal cycles from natural, agricultural and biomass burning sources (NAT + AGR + BB) by imposing prior temporal correlations on the FF contributions. FF emissions in each grid cell in each month were correlated with emissions 220 from the same grid cell in other mo nths w ith an exponential correlation time scale of 9.5 months (equivale nt to a consecutive-month correlation of 0.9). Both NAT + AGR + BB and FF sectors had spatial correlations imposed with normal distributions and correlation length scales of 500km. This gives global uncertainty of approximately 70 Tg(CH4)/yr. The sectors which make up the NAT + AGR + BB and FF emissions are explained in Section 2.2.2.

225
We produced estimates for each year's posterior emission covariance error matrix using the L-BFGS method (Nocedal, 1980) and updates suggested by Bousserez et al. (2015). This uses multiple iterations in order to estimate the inverse of the hessian (the second derivative) of the cost function, and does not include the off-diagonal elements of the posterior covariance matrix, so the posterior errors described in this manuscript are likely to be upper limits (Bousserez et al., 2015).

Prior emissions and chemical sinks of CH4
Prior emissions were taken from a range of widely available bottom-up models and inventories. Anthropogenic emissions were originally taken from the EDGAR v4.2 FT 2010 inventory (Olivier et al., 2012) and scaled as in McNorton et al. (2018). Biomass burning emissions were taken from GFEDv4.2 (van der Werf et al., 2017). The JULES model (Clark et al., 2011) was used to provide wetland fluxes, in a configuration described in McNorton et al. (2016). Rice emissions were taken 235 from Yan et al., (2009) and are scaled as in Patra et al. (2011). Remaining natural sources were included as in Wilson et al. (2016). The surface soil sink due to methanotrophs was from the Soil Methanotrophy Model (MeMo, Murguia-Flores et al., imposed in the prior uncertainty matrix and made up the FF category, whilst the remaining emissions (NAT + AGR + BB) had no prior temporal correlations imposed. Prior totals for each source type within South American regions are shown in 240 Table 1. Atmospheric OH fields, based on those provided within the TransCom CH4 study (Patra et al., 2011) were taken from Spivakovsky et al. (2000) and scaled downwards by 8% in accordance with Huijnen et al. (2010). These vary from month to month but do not vary between years. Montzka et al. (2011) suggested that variability in annual OH mole fractions is small, but some recent research has suggested the possibility of a declining trend in OH since 2004 (Rigby et al., 2017;Turner et al., 2017), although this trend had a high level of uncertainty. A trend, or any significant year-to-year variability, in 245 OH which was not included in our analysis, would affect our conclusions, but for now we do not have enough evidence to include any potential variations. Stratospheric loss fields due to reactions with atomic chlorine (Cl) and excited oxygen atoms (O 1 D) varied on a monthly and annual basis and were taken from a previous full chemistry simulation from TOMCAT (Monks et al., 2017). Loss in the troposphere through reaction with chlorine was not included in these simulations. 250

Bottom-up model
We also use a simple bottom-up (B-U) model to estimate CH4 emissions from climatological driving data, so that we can investigate the causes of variations in CH4 emissions derived in the inversion. The B-U model calculates wetland CH4 emissions based on the method used in Bloom et al. (2017), in which the CH4 emissions in a grid cell, , at time, , are dependent on climatological factors as follows: 255 where ( , ) is the flux of CH4 in molecules cm -2 s -1 , ( , ) is the wetland fraction, ( , ) is the heterotrophic respiration of carbon per unit area, ( , ) is the surface temperature in°C, and is the relative CH4:C ratio of respiration for a 10°C 260 change in temperature. Finally, is a scaling factor. We use monthly mean values for each element of Eq. (2) and interpolate all parameters to the TOMCAT model grid for comparison with the inversion results.
We take from the CASA-GFED v4.1 data product (Randerson et al., 2015), which runs up to 2016, and gridded 2m temperature from the NOAA/NCEP Global Historical Climatology Network version 2 and the Climate Anomaly Monitoring System GHCN Gridded V2 data provided by the NOAA Physical Sciences Laboratory (https://psl.noaa.gov/, Fan and Dool, 265 (2008)). We estimate using a combination of two products. We take a climatology of wetland fraction ( ) from the JULES land surface model version that was used to produce the prior emissions used in the inversion . We then use measurements of gravity anomalies made on the twin GRACE satellite mission, ( , ) as a proxy for variations in the soil moisture, as in (e.g.) Bloom et al. (2010) and Gloor et al. (2018). We then apply scaling factors and to give wetland fraction as follows: 270 This makes the assumption that anomalies in the gravity anomaly ( , ) are linearly related to wetland fraction anomalies, which may not be the case. The distributions and variations of the GRACE gravity anomalies and surface temperature are 275 discussed in Section 4. We create an ensemble of B-U estimates for , letting the scaling factors and and the temperature response function vary within reasonable limits, and varying appropriately so that each member gives the same mean total emissions over 2010 -2017, equal to the mean posterior value produced by the inversion. We are interested only in the variations in time and space produced by the B-U model, rather than the absolute value. We let vary between 1 and 3, based on experimental bounds and previous bottom-up studies of methane emissions (Yvon-Durocher et al., 2014;280 Bloom et al., 2017), we let vary between 0.8 and 1.2, and we let vary in such a way that the overall wetland fraction does not vary by more than 20%, depending on the value of . Since there is no data for 2017 given for the heterotrophic respiration, we use a climatology made up from the preceding seven years applied to that year. We also create an 'optimised' B-U model, in which we use a curve-fitting procedure to choose values of , , and which best fit, in least-squares terms, the results from the inversion for the monthly and spatial mean values over the whole Amazon, for all months within 285 the wet season over 2010 -2017. For this B-U model, we consider only the wet season NAT + AGR + BB emissions within the Amazon Basin, which we assume to be almost entirely from wetlands.
The equation that our B-U model is based on is commonly used in other studies which estimate wetland fluxes of CH4 (e.g. Clark et al., 2011;Melton et al., 2013;Bloom et al., 2017), but our application of the driving climate variables is fairly 290 simple relative to these previous works. This method is sufficient for this work as the purpose of the B-U model is to investigate the possibility of reproducing the inversion results, and if they can be reproduced, to learn how and why the CH4 wetland emissions change according to the input variables.

Average distribution of emissions 295
Average GOSAT XCH4 over South America since 2009 show that XCH4 column mole fractions were largest over the west of the continent, particularly in the northwest (Figure 2). Using the a priori emission distribution in TOMCAT leads the model to underestimate XCH4 in the northeast and far north of the continent and in the outflow into the Atlantic Ocean.  38.9 ± 11.7 10.6 ± 9.2 38.2 ± 11.4 10.6 ± 9.2 39.9 ± 5.3 9.9 ± 0.9 45.7 ± 5.1 9.9 ± 0.9

Temporal variations of CH4 emissions 340
The annual total prior emissions in Brazil are consistent over time (Figure 4), with a mean value of 48.6 ± 14.9 Tg(CH4)/yr. annual uncertainty as we assume that posterior uncertainty for each year is strongly correlated with that in other years. This There is a significant positive trend over the whole time period (2010 -2018) of 1.37 ± 0.69 Tg(CH4)/yr 2 (p < 0.05), driven by the NAT + AGR + BB emissions category, although the distribution is more of a step-change from 2014 onwards.

350
Posterior emissions in Brazil peak in February and September (Figure 4b) and represent the wet-season and dry-season peaks, most likely due to contributions from the local seasonal cycles of wetland emissions and biomass burning emissions, depending on the location. The peak monthly emission rate of 66.2 ± 8.2 Tg(CH4)/yr is in February, before lower rates of emission during the shoulder season of April to July. This February peak corresponds to a peak in precipitation across the

-c) Total annual (red lines) prior and posterior emissions of CH4 (Tg(CH4)/yr) in three Brazilian subregions; the western Brazilian Amazon (WBrAm), the eastern Brazilain Amazon (EBrAm) and non-Amazon Brazil (NonAmBr). Prior and posterior emissions during the wet season (December -March, brown lines) and the dry season (August -October, maroon lines) are also shown. Shading represents the posterior uncretainties for each region derived in the inversion. (d-f) Monthly mean prior and posterior emissions for the period 2009 -2018 (Tg(CH4)/yr) for the three sub-regions. Shading shows the standard deviation of the monthly means.
https://doi.org/10.5194/acp-2020-1136 Preprint. Discussion started: 10 November 2020 c Author(s) 2020. CC BY 4.0 License. the non-Amazon region of Brazil (NonAmBr), emissions decrease slightly between over these years (from 17.5 ± 3.0 Tg(CH4)/yr to 16.4 ± 2.9 Tg(CH4)/yr). The Amazon regions of Brazil display the two-peak seasonal cycle of CH4 emissions, 365 although this is much more pronounced in the east. This is most likely due to the significant effect of biomass burning within the arc of deforestation in the south-east of the Basin that usually occurs during these months. Emissions are largest in NonAmBr during the dry season, possibly due to fires in savanna regions.
We also display total emissions for each subregion during the wet season (December -March) and the dry season (August -370 October). These periods were defined using the GPCP precipitation data, as periods when the average monthly precipitation during 2009 -2018 within the Basin was more than 7 mm day -1 and less than 3 mm day -1 , respectively. In both WBrAm and EBrAm, the trends for the 2009 -2018 period are largest in the wet season. This suggests that trends in wetland emissions might be responsible for the rising CH4 emissions.

Tefé (TAB/TEF). Model output has been interpolated to observations locations and altitudes, before both were averaged into monthly means and into altitude bins of 3km and above (a-d) and 1.5km and below (e-h). Dotted vertical lines show the zero line, whilst dashed vertical lines show prior and posterior mean model -observation bias.
https://doi.org/10.5194/acp-2020-1136 Preprint. Discussion started: 10 November 2020 c Author(s) 2020. CC BY 4.0 License.

Comparison to independent observations 375
Observations of CH4 made during flights within the Basin between 2010 and 2018 were used to independently check the performance of the prior and posterior emission distributions in the model ( Figure 6, Table 2). For the observations made at altitudes higher than 3km, which represents the free troposphere above the Amazon, the performance of the posterior emissions is significantly improved compared to the prior at all locations. The absolute value of the model -observation bias is reduced to below 6 ppb at all sites, and the correlation between the model and the observations increases at all locations. 380 However, the posterior performance against observations made in the boundary layer, at altitudes below 1.5 km, is generally worse than the prior performance. At the western sites, RBA and TAB/TEF, the mean bias in the model increases by approximately 15 ppb, although the correlation improves, particularly at TAB/TEF. At ALF, the correlation decreases slightly, and the mean bias increases by a large amount (31 ppb). Finally, at SAN, the performance improves significantly by both measures, with the mean bias being reduced from -47.8 ppb to -15.2 ppb. There are no significant trends in the model -385 aircraft residual biases in 2010 -2017, except at TAB/TEF below 1.5km. This site has a posterior residual bias trend of +2.1 ppb/year, but this may have been caused by the change in the flight location halfway through the study period.
The improved performance at SAN is significant, as the high mole fractions of CH4 sampled at this location relative to expectations given its location situated close to the eastern coast have been previously noted (Miller et al., 2007;Basso et al., 2016;Wilson et al., 2016). The prior model therefore leads to a large negative bias at SAN, particularly near the surface. The 390 posterior distribution of emissions, with a region of significant emissions to the south and east of the Basin, significantly reducing the model -observation difference at SAN. The model still underestimates methane mole fractions at this site even after the improvement, however, which might still be due to remaining bias or model representation uncertainty. The fact that ALF is also located near these significant emissions leads to degradation in the model performance within the boundary layer, which was previously better at ALF than at SAN. The capability of assimilation of GOSAT XCH4 to improve 395 performance at both of these locations might have been reduced due to the relatively coarse model grid. Webb et al. (2016) found that comparisons between the flight-based observations and a previous version of the GOSAT XCH4 used in this study showed that the GOSAT values were larger than equivalents estimated using the flight data at ALF, but that the discrepancy was much smaller at SAN. This being the case, it is not surprising that the model in which the GOSAT data has been assimilated has difficulties in matching the flight observations at both locations at once. Since we assimilated XCH4 from 400 GOSAT, which is mostly representative of the troposphere, it is expected that the model performance is improved at all locations when compared to observations made at the higher altitudes. This also indicates good model representation of inflow of CH4 to the Basin from elsewhere. However, the fact that the posterior comparisons are generally degraded close to the surface, apart from at SAN, mean that the local sources close to these sites might be overestimated at this model resolution, that there are errors in the model's representation of vertical mixing, or that there remains a positive bias in the 405 assimilated retrievals from GOSAT in this region. Generally, however, the temporal variation and mean bias in the model is much improved after the assimilation of GOSAT XCH4.

Bottom-up model results
The inversion suggests that CH4 emissions have been increasing from Amazon regions throughout the 2010s, but it is not easy to determine the source sectors responsible for this rise. The largest increases over time occur during the wet season ( Figure 5), when wetland emissions dominate the atmospheric signal, so it seems most likely that changes to these emissions 415 are driving the increase. Wetland emissions are sensitive to temperature, precipitation (which drives wetland area) and carbon availability in the soil , so we examined these driving factors to see how they varied during the previous decade. other event that stands out is the prolonged flooded period running from though the wet season of 2013/14, during which rainfall in the south-west of the Basin was up to twice as much as usual (Espinoza et al., 2014). This flooded period did not coincide with a significant El Niño -Southern Oscillation (ENSO) period but was likely caused by warm conditions in the Subtropical South Atlantic.  that year. We felt that this was justified since the temperature and water table depths also had only very small anomalies during that season. As might be expected, the temperature and gravity anomalies in the wet season were strongly negatively correlated (r=-0.66), since hot and dry conditions are often linked.
The temperature trend in the Amazon was positive throughout almost the entire Basin (Figure 8a), being strongest to the far 445 west and in the south east. The trend in the wetland fraction (Figure 8b) was more heterogeneous, with positive trends in the west contrasting with strong negative trends across the east of the Basin. For both of these variables, the trends are strongly impacted by the hot, dry conditions in the wet season of 2015/16.
The geographical distribution of the NAT + AGR + BB wet season CH4 emission trend produced by the inversion (Figure  450 8c) is positive across the north west and south east of the Basin, with a fairly similar distribution to the locations with https://doi.org/10.5194/acp-2020-1136 Preprint. Discussion started: 10 November 2020 c Author(s) 2020. CC BY 4.0 License. positive temperature trends. The positive emission trends in the north west are also collocated with an area with wettening trends. However, the regions to the east and south with strong positive emission trends are in the region in which wetland fraction had been decreasing as temperatures increased. This suggests that the emissions were more sensitive to the increasing temperature than to the decrease in wetland fraction or in heterotrophic respiration (not shown). 455 We ran the B-U model multiple times, varying the temperature response and the GRACE anomaly scaling variables within their bounds in order to produce a range of likely values for CH4 flux from the Basin. We also used a curve fitting program in order to best reproduce the INVICAT results using the B-U model (Figures 8d and 8e). The B-U model combines the three driving variables, but the strong anti-correlation between the temperature and wetland fractions mean that this model 460 does not produce strong variations in emissions, since the two tend to cancel out. Using the optimised B-U model produces inversion results (2.9 Tg(CH4)/yr) is relatively large compared to the standard deviation, however, meaning that the B-U model results almost always remain within the posterior inversion uncertainty. The exception to this is the wet season of 2014, when the inversion results produce larger emissions than in any other year (20.1 ± 2.7 Tg) and this feature is unable to be reproduced in the B-U model. As discussed, the wet season of 2014 was subject to extreme precipitation and widespread flooding in the Basin (Espinoza et al., 2014), and the GRACE gravity anomalies are large throughout this period (Figure 7), 470 whilst heterotrophic respiration was high and temperatures were relatively cool (although warmer than in 2011 and 2012).
Despite these conditions which seem favourable to CH4 emission, the B-U model does not produce emissions significantly larger than any other year. The discrepancy between the inversion and B-U model results is discussed further in Section 4.
We also show here the wet season emissions within the Basin from the WetCHARTS emission dataset , which use a similar method to estimate wetland emissions that used in our B-U model. The values in Figure   Tg(CH4)/yr. This increase was found to be entirely due to the NAT + AGR + BB emissions within the Amazon region.
This increase between the two periods is very similar to that found by Tunnicliffe et al., (2020), although the total emissions found in our study are larger than their finding of 33.3 ± 3.7 Tg(CH4)/yr. They removed a model -satellite bias of 22 ± 8 485 ppb from the GOSAT observations used in their study, which is much larger than our bias of 3 -9 ppb removed from XCH4 over the Amazon. This larger bias removal, coupled with the different model transport of their regional inversion, could explain the smaller emissions that they derive. The positive biases in our posterior CH4 relative to aircraft observations within the boundary layer also suggests that our emissions may be overestimated. However, we note the absence of TOMCAT forward model to represent the atmospheric transport, so is not fully independent from our results. Our findings here are within the range of these models, albeit towards the lower end. The majority of these top-down studies used either the same GOSAT and surface observation data used in our study, or some variation of it. The fact that the derived emissions using similar observation data can vary so much highlights the inherent uncertainties still remaining in top-down studies of CH4 emissions, with differences in model transport, chemistry representation, inversion methodology, bias correction and 500 error assumptions all contributing to differences in results.
The increase in emissions from 2014 onwards that we derive coincides with a faster rate of increase in the observed surface mole fraction of CH4 (Nisbet et al., 2019). Unfortunately, the extent that the increase in observed mole fractions in the atmosphere is driven by increasing Amazon emissions is difficult to constrain without more extensive knowledge of the atmospheric chemical loss of CH4. Our global inversion, using repeating OH values each year, indicates that the increase of 505 5.8 ± 5.2 Tg(CH4)/yr from Amazon emissions is responsible for 24 ± 18% of the global total increase in emissions between 2010-13 and 2014-17, which was 24.1 ± 15.0 Tg(CH4)/yr.
The Amazon emissions derived in this study for 2010 and 2011 (41.6 ± 5.3 Tg(CH4)/yr) are a little above the higher limit of those found in our previous study using the flight observations only (31.6 -41.1 Tg(CH4)/yr, Wilson et al., (2016)). This 510 indicates that using the vertical profile data only to calculate Basin-wide emission totals may lead to a small underestimation of the total compared to using satellite data. This discrepancy is supported by the positive bias seen in this study within the boundary layer at most of the sites when comparing the posterior model output to the in situ flight observations. However, the emission totals are fairly similar across the different methodologies, with the caveat that the same transport model was used for both findings. 515 Our comparisons to the independent observations taken during flights within the Amazon highlight both some success and some remaining issues with our results. Assimilating the GOSAT data leads to an improvement compared to the prior in the mean bias and correlation at all four locations when observations made above the boundary layer are considered. However, the posterior comparison to observations made close to the surface are inferior to the prior comparison at three of the 520 locations. It seems that improving the performance compared to the GOSAT data throughout the troposphere is at the expense of reducing performance at the surface. There could, therefore, be transport errors in the inverse model, possibly in the boundary layer transport. It is possible also that the relatively coarse resolution of the inversion leads to poorer comparisons to the boundary layer observation. Finally, as stated by Webb et al. (2016), comparisons between the flight observations and GOSAT at the Alta Floresta (ALF) site, which displays the worst posterior performance in the model, are 525 also not as good as at other locations. Despite the increased posterior bias in the boundary layer at three of the sites, the improved performance at Santarém suggests that the significant emissions close to the mouth of the Amazon derived by the inversion are potentially a realistic feature, consistent with the previous in situ data-based flux estimates of Miller et al. (2007) and Basso et al. (2016). However, the degradation in performance at Alta Floresta, also in the east of the Basin, suggests that the strong emissions do not extend as far south as in our model posterior. We will in the future produce 530 inversions at higher resolution to investigate this feature further.
Due to computational constraints, we could not carry out inversions for the entire GOSAT period using a higher horizontal resolution than the one chosen for our inverse model, but to examine the sensitivity of our results to the model resolution, we ran an inversion for 2010 at 2.8° horizontal grid resolution ( Figure S3), averaging the GOSAT XCH4 onto this model grid. 535 We did not split the results into different source sectors, instead deriving total CH4 surface flux. Otherwise the model set-up was identical to the 5.6° inversions of the main study. Many of the features of the posterior solution are identical to those of the coarser grid, with higher emissions from the region to the south and east of the Amazon river, and a decrease in emissions from the south of Brazil, near the densely populated cities. However, there is no decrease in emissions to the west of the Amazon Basin, as consistently seen when using the coarser model grid. Total derived emissions for Brazil and for the 540 Amazon Basin are similar when using the 2.8° and the 5.6° grids, however. We derive total posterior emissions for Brazil in 2010 of 49.9 Tg(CH4)/yr using the coarser grid, and 51.4 Tg(CH4)/yr using the finer grid.
Our derived positive trends are largest during the wet season within the east Amazon, indicating that increasing flux from wetland sources are most likely responsible for the increase in total emissions. However, attempting to reproduce these 545 trends, and the interannual variations, using a B-U model was largely unsuccessful. Although the B-U model mainly stayed within the uncertainty derived in the inversion, it was unable to capture a large increase in emissions in the wet season of 2014. This indicates either that the variation produced in INVICAT was exaggerated, that uncertainty in the B-U model setup and input data led to this inability to match the inversion results, or some combination of these factors. It is also possible that biomass burning CH4 flux has increased in the region outside of the dry season (e.g. Silva Junior et al. (2019) Tunnicliffe et al., (2020) used a regional model in which the chemical sink of CH4 was not a factor, and found similar levels of interannual variably to those produced here.
Meanwhile, our B-U model was much simpler than full land surface models and used only one input source for each set driving of data. The fact that wetland fraction and temperature were strongly anti-correlated meant that the model was not 560 able to produce significant emission variations from year to year when the two were included in the model. In the future we plan to use a more complex land surface model for comparisons such as this, but our use of the JULES model to produce our prior emissions inventory meant that it would have been inappropriate for post hoc comparisons here. The independent WetCHARTS results, however, also produced very different results to those of our inversion.

565
The performance of the B-U model compared to the inverse model suggests conflicting hypotheses. The positive trend in emissions produced in INIVCAT was concurrent with increasing temperatures across much of the Amazon. This indicates that the temperature response of wetland emissions in the region might be high. However, the fact that the B-U model was unable to produce significantly larger emissions during the 2014 wet season, as were produced by the inversion, despite large wetland fraction and heterotrophic respiration at the time, indicates that the wetland fraction response might also be high, 570 and potentially non-linear. Comparing the results from the B-U model for 2012 and 2014 is instructive, as 2014 had higher heterotrophic respiration and temperature, and a similar (but slightly higher) mean wetland fraction. However, the B-U emission totals for these two years were very similar. Although the observed mean gravity anomalies were similar, they were characterised differently, with prolonged positive anomalies throughout 2013/14, but a short and intense positive anomaly during the end of the 2012 wet season. This suggests that emissions could be a function of the period of time for which the 575 soil is saturated. It should be noted that Tunnicliffe et al. (2020) also derived large CH4 fluxes during this wet season, but they were allocated to anthropogenic sources rather than wetlands using their methodology, likely due to differences in the transport model and sector allocation method. Increased complexity in the B-U model and examination of correlations between inversely-derived fluxes and potential wetland flux drivers are both necessary for future comparisons, and for now it is not possible to determine definitively the cause of the trend in CH4 emissions in the Amazon Basin. 580

Conclusions
Our global inversion of CH4 emissions using satellite data and surface observations allowed us to quantify changes in South American emissions over the period 2009 -2018. We found that emissions increased during this period, particularly during the wet season of December -March. Total Brazilian emissions rose from 49.8 ± 5.4 Tg(CH4)/yr in 2010 -2013 to 55.6 ± 5.2 Tg(CH4)/yr in 2014 -2017, whilst natural emissions from the Amazon Basin (from all countries) ), an area of 6.9 million 585 km 2 on this model grid, rose from 38.2 ± 5.3 Tg(CH4)/yr to 45.6 ± 5.2 Tg(CH4)/yr. We show that there was significant emission from the south and east of the Basin throughout this period, and that the positive trends were largest in the east Brazilian Amazon. We derive particularly large emissions during the 2013/14 wet season, a period during which there were widespread flooding. It is significant that our inversions show improved performance at Santarém due to the large emissions in the east of the Basin, similar to previous aircraft-based studies (Miller et al., 2007;Basso et al., 2016). Indeed, based on 590 the remaining negative model-observation bias at that location, it is possible that CH4 emissions affecting that location could be even larger. However, it appears that the Alta Floresta site is overly affected by these large emissions in our analysis, indicating that the southerly extent of the large emissions might be too great.
However, attempting to reproduce these trends in a simple bottom-up model were unsuccessful, mainly due to strong anti-595 correlations between the wetland fraction and the temperature within the Basin leading to little variation in annual wet season emissions. This suggests that the complexity of the model must be increased in order to fully represent the relationship between carbon availability, wetland fraction and soil temperature. Our B-U model, and other models , suggest a negative trend in emissions from driving conditions, but this is at odds with our findings and those of others. This suggests that temperature has a strong role to play in wetland emissions of CH4 in the Amazon region, since this 600 has also had an increasing trend over the past decade. It is also important to consider the role of wetland variability, however. For the inverse model the contribution of how sinks of CH4 in the atmosphere might have varied should also be considered.
The results of our inversion are in agreement with previous studies (e.g. Janardanan et al., (2020)), and within the range provided by . However, our posterior emissions from Brazil are significantly larger than those produced 605 by Tunnicliffe et al. (2020) using a similar observational data set, showing the importance of model transport in inversion results.
Our results show that the Amazon Basin was responsible for 24 ± 18% of the total global increase in CH4 emissions during the last decade, and it could contribute further in future due to its sensitivity to increasing temperature. Our study shows the 610 benefit of using satellite data to inform on emissions of CH4, particularly in poorly sampled tropical regions, along with the benefits of long-term satellite missions to produce large-scale, consistent datasets. As the satellites and models improve, we can further refine our estimates of emissions from the important and changing role of South American ecosystems on global methane variability.