Atmospheric observations show accurate reporting and little growth in India’s methane emissions

Changes in tropical wetland, ruminant or rice emissions are thought to have played a role in recent variations in atmospheric methane (CH4) concentrations. India has the world’s largest ruminant population and produces ~ 20% of the world’s rice. Therefore, changes in these sources could have significant implications for global warming. Here, we infer India’s CH4 emissions for the period 2010–2015 using a combination of satellite, surface and aircraft data. We apply a high-resolution atmospheric transport model to simulate data from these platforms to infer fluxes at sub-national scales and to quantify changes in rice emissions. We find that average emissions over this period are 22.0 (19.6–24.3) Tg yr−1, which is consistent with the emissions reported by India to the United Framework Convention on Climate Change. Annual emissions have not changed significantly (0.2 ± 0.7 Tg yr−1) between 2010 and 2015, suggesting that major CH4 sources did not change appreciably. These findings are in contrast to another major economy, China, which has shown significant growth in recent years due to increasing fossil fuel emissions. However, the trend in a global emission inventory has been overestimated for China due to incorrect rate of fossil fuel growth. Here, we find growth has been overestimated in India but likely due to ruminant and waste sectors.

Supplementary Figure 8: Percentage difference between three chemical transport models/inversion methods for XCO 2 , which was used to derive proxy XCH 4 . Metric is defined as the maximum difference between ensemble members and the median model divided by the median model. Due to increasing mole fractions over time and slightly different sampling each year, only 2012 is shown for clarity, rather than an average over 2010-2015. GOSAT: Dry air column-averaged CH 4 mole fractions, XCH 4 , have been measured since 2009 in the shortwave infrared (SWIR) using the TANSO-FTS instrument on the GOSAT (Greenhouse gases Observing Satellite) satellite 2 . GOSAT measurements are valuable for flux quantification because the satellite has a repeat cycle of 3 days and has sensitivity to the near-surface. Soundings are typically separated by approximately ~100 km with diameter ~10 km. Therefore, the satellite can provide spatial and temporal coverage for regions of the world that are not well-covered by surface monitoring.
In this study, we used XCH 4 data from the University of Leicester version 6 proxy retrieval (http://www.leos.le.ac.uk/GHG/data/) 3,4 from 2010-2015. We used only those data that have been filtered for clouds and passed the quality flag and these data were averaged into 0.5° bins. Retrievals occurring over significant topography (defined as over 1000 metres above sea level) were removed from the analysis to minimize any artefacts that may arise in the retrieval algorithm due to the compressed column as well as due to errors in the transport model over these regions. Ocean glint-mode observations also were not used to avoid any systematic errors that may exist between the two modes of operation. Supplementary Figure 7 shows While two methods are available (full-physics and proxy), the proxy retrieval was chosen for South Asia because we expect there to be impact from atmospheric scattering due to clouds and aerosols. XCH 4 was derived by multiplying the measured XCH 4 /XCO 2 ratio from a non-scattering atmosphere by a model XCO 2 field (equation 1). Because XCO 2 varies much less in the atmosphere than does XCH 4 , it is assumed that the errors in XCO 2 are of much smaller magnitude than the errors due to incompletely quantified atmospheric scattering in the full-physics retrieval. This is particularly important for regions of high aerosol loading and/or cloud cover. (1) The main limitation of the proxy method is that it relies on accuracy of the model XCO 2 field, which has assimilated surface CO 2 observations using a chemical transport model. Proxy XCH 4 has been used in a variety of global and regional studies 6-12 and has been validated against surface, aircraft and TCCON measurements 3,4,13,14 .
Firstly, we investigated the range of XCO 2 between the three models, which are found in version 6 of the GOSAT product. Supplementary Figure 8 shows the percentage range of these models relative to the median model (which varies for each retrieval). The smallest and largest differences, as a percentage, occur in winter and summer, respectively. The maximum difference, which affects a small subset of retrievals, is on the order of 1%. To assess the sensitivity of national emissions to the range of XCO 2 fields, we have repeated our inversions by scaling each XCH 4 retrieval up and down by the percentage difference in Supplementary Figure 8 and these results are discussed further in Supplementary Note 3 We have also compared MACC-II model output 15 against CO 2 mole fractions measured at Cape Rama, India (CRI) and on the CARIBIC aircraft (Supplementary Figure 9). It is important to note that because CO 2 measurements are sparse in India and this analysis is limited to two datasets, this comparison is presented to show the potential impact of this error, rather than to account for it. Because the inversion results will only be sensitive to spatial gradients in the CO 2 model error (uniform differences would be absorbed into boundary conditions and offset terms), there is insufficient information to know what effect this error would have on the inversion results. The main difference between CRI data and model CO 2 occurs during the post-monsoon and winter. While the model simulates the minimum of the seasonal cycle in November, this is not evident in the CRI data. At these times, the model is as much as 8% lower than the observed value. All other times of year appear consistent between model and observations. This discrepancy could be due to the resolution of the model or due to the prior flux field used. The CARIBIC data shows that the model could be biased slightly low (<0.1%) but this is less pronounced than the CRI comparison.
Because the aircraft samples in the upper troposphere, spatial gradients are much smaller than at the surface and therefore, it is expected that the differences between the model and observations will not be as large. This data comparison suggests that results from October-January should be interpreted with greater caution. However, it should be noted that the effect on the column will be much less extreme than what is shown at the surface. Our results during October-January do not show any extreme changes relative to other times of the year.
Previous studies have also found systematic errors between GOSAT retrievals and ground-based FTIR measurements from TCCON (Total Carbon Column Observing Network) but TCCON data is not available in South Asia 3,16 . Systematic errors between GOSAT and TCCON have typically been less than 5 ppb. In addition, studies have found biases between GOSAT and model derived column mole fractions 7 , which could additionally be due to systematic errors in the chemical transport model. These errors have exhibited a latitudinal dependence on the order of 20 ppb, peaking in the tropics. Many inverse studies account for a systematic error in the inversion between satellite and calibrated data, which we also incorporate into our inverse method.

CARIBIC:
We used data from the CARIBIC passenger aircraft program, including 24 winter (Jan-Mar, Nov-Dec) flights over South Asia (http://www.caribic-atmospheric.com). Summer flights over South Asia were discontinued after 2008 but have been used to estimate monsoon-related biogenic production of CH 4 from South Asia 21 . These flights took place either from Frankfurt, Germany to Chennai, India or Frankfurt to Bangkok, Thailand. Methane mole fractions were measured in flight using a Los Gatos Research Fast Greenhouse Gas Analyzer at 1 Hz frequency. The precision with 10s averaging is approximately 1 ppb 22 and measurements were calibrated using whole air samples on the NOAA-2004 calibration scale 23 . To minimize the effect of contaminated samples from airports, air was only sampled at altitudes with pressure below 650 hPa. For this study, measurements were averaged into 1 minute periods.

Supplementary Note 2:
Model uncertainties derived through the hierarchical inversion method were derived for each 5-day period per surface site or per 5-degree latitude bin for satellite/aircraft observations. Average model uncertainties on GOSAT retrievals were found to be 14.0 ppb, with the largest uncertainties in July -September (mean model error of 20.4 ppb) and the lowest in January-April (mean of 9.0 ppb). These uncertainties are reflected in the national emissions presented in (manuscript Fig. 1A) which shows the smallest and largest uncertainties at these times. There are two possible reasons for the larger errors during the summer: firstly, emissions and therefore concentrations above background are higher so absolute uncertainties are also larger and secondly, the summer period coincides with the Indian summer monsoon, a period of intense and deep convection, which is likely to not be as well-represented in the model. Offsets between satellite and calibrated data were also derived along with an analysis of the systematic errors associated with incorrect specification of the CO 2 model used to derive the proxy XCH 4 data used in this study.
Offsets between GOSAT and surface/aircraft observations were also estimated in the inversion, assuming that the surface/aircraft data were `unbiased' through anchoring to a calibration scale. We estimated one offset value for the domain per month but this could be extended to represent a physical process such as bias in the XCO 2 field. On average, summertime offsets (defined as y gosat, obs -y gosat, mod ) were +38.6 ppb (largely based on GOSAT and DJI data). Wintertime offsets, in contrast (largely based on GOSAT and CARIBIC), were -8.1 ppb. The offset that is estimated is a combination of offset due to the observations themselves as well as any offset due to the model and both of these are expected to vary seasonally. The model component also includes any error in the assumption of the prior model mole fractions being correct above maxlev. We have found that changes in the representation of the mole fractions above the maximum level in NAME (i.e. assuming either prior model mole fractions or assuming a constant mole fraction between maxlev-20) had little effect on the emissions that were derived and were largely absorbed into the offset term. The changes to the offset term in these two scenarios suggests that a significant component of the offset term could be due to the representation of the upper atmosphere in the model. Supplementary Note 3 discusses the effect of using GOSAT data alone versus the full set of observations and the effect of changes to the model stratosphere.

Supplementary Note 3:
Five sensitivity studies were performed to assess the effect on the top-down results of (1) the prior emissions; (2) XCO 2 model error; (3) GOSAT tracks; (4) differences between using the full set of surface, aircraft, and satellite data and satellite data alone; (5) errors in the model stratosphere.
(1) A sensitivity study was performed to assess the influence that the prior seasonal cycle had on the estimated emissions. For the prior of this sensitivity study, we used EDGAR v4.2FT2010 (including rice), GFED v3.1 and natural sources and did not use the seasonally varying rice prior used in the main results. Though the prior did not contain a summertime monsoon signature, the same summertime maximum emerged, showing that the results are largely data-driven and not heavily influenced by the prior (Supplementary Figure 31). Emissions derived with this prior largely fall within the uncertainties of the top-down emissions presented in the main results.
(2) We assessed the effect of errors in the model XCO 2 used to derive the proxy XCH 4 on derived emissions. Supplementary Figure 32 shows national emissions derived using perturbed values of XCH 4 based on differences between three CO 2 models and inversion frameworks. Each retrieval was perturbed high or low based on the model CO 2 range for that retrieval. In this way, differences in the spatial distribution of the model errors (Supplementary Figure 8) were accounted for. While the median emissions of the base case occasionally lie outside of the uncertainties in the perturbed cases, the uncertainties overlap. These findings suggest that the variability between the three CO 2 models does play a role in the emissions derived for the region, however, it should be noted, that all three models are based on a similar set of underlying CO 2 data, which are from global networks and not from South Asia. Furthermore, it is not expected that all retrievals would necessarily be biased consistently in the same direction, so these results reflect the extreme. Larger biases could result from inaccuracies in model CO 2 compared to observed CO 2 , as discussed above. These findings suggest that the uncertainties quantified using this methodology should be extended to account for systematic errors to more completely capture all sources of error.
(3) We assessed the effect of randomly resampling 200 measurements from the GOSAT observations each month. This was done to determine whether the spatial tracks of the GOSAT soundings could impart any systematic bias into the results. Supplementary  Figure 33 shows that the emissions derived with the random resampling are consistent with the results provided in the main text. This provides evidence that the spatial tracks of GOSAT are not introducing artefacts into the results.
(4) We assessed the effect of using the full set of satellite, aircraft and surface data (with offsets estimated and accounted for) versus using GOSAT data alone. This was performed for the year 2012 when all data was available. This was done is determine whether results are influenced by the varying amounts of data available for different time periods. Supplementary Figure 34 shows little difference in national emissions derived using satellite data alone and emissions derived using all available data over 2012 when all sites were operational. The difference between the two is insignificant, suggesting that the effect of offsets have been accounted for in the derived boundary conditions because they are generally spatially uniform across India.
(5) We assessed the effect of errors in the model's simulation of the stratosphere based on findings from Saad et al., 2016 26 . This study found that errors in the model simulation of stratospheric CH 4 could lead to biases in XCH 4 , which could result in errors in derived surface fluxes. To assess the impact of this error, we perturbed the prior boundary conditions by 5% (approximately 100 ppb at the upper levels, as shown by Figure 5 in Saad et al., 2016 26 ). Supplementary Figure 35 shows that the differences in derived fluxes lie within the uncertainties presented in our base case. We propose that these small differences are because this is a regional inversion and any errors in stratosphere-troposphere exchange will be corrected each month from observations in that month. Any errors in the model boundary conditions, which were used as a prior in the inversion, will be minimized as offsets and gradients to the prior field are estimated in the inversion.