Abstract
Atmospheric methane observations are used to test methane emission inventories as the sum of emissions should correspond to observed methane concentrations. Typically, concentrations are inversely projected to a net flux through an atmospheric chemistrytransport model. Current methods to partition net fluxes to underlying sectorbased emissions often scale fluxes based on the relative weight of sectors in a prior inventory. However, this approach imposes correlation between emission sectors which may not exist. Here we present a Bayesian optimal estimation method that projects inverse methane fluxes directly to emission sectors while accounting uncertainty structure and spatial resolution of prior fluxes and emissions. We apply this method to satellitederived fluxes over the U.S. and at higher resolution over the Permian Basin to demonstrate that we can characterize a sectorbased emission budget. This approach provides more robust comparisons between different topdown estimates, critical for assessing the efficacy of policies intended to reduce emissions.
Introduction
As the second highest greenhouse gas (GHG) contributor to global radiative forcing, understanding the global budget of methane (CH_{4}) is a top climate priority^{1,2}. Methane is emitted from a variety of anthropogenic and natural emission sectors, including oil and gas operations, waste management, coal mining, agriculture, wetlands, and fires among others^{3}. Article 14 of the Paris Agreement^{4} requires participating countries to report progress towards achieving their climate mitigation goals, or nationally determined contributions. Reporting progress, including any changes in the CH_{4} budget, necessitates inventorying all possible emission sources. CH_{4} emission inventories can be constructed from “bottomup” or derived from “topdown” observations. Bottomup accounting relies on a knowledge of activity data and emission factors for anthropogenic sectors and/or detailed processedbased models that predict CH_{4} emissions based on a set of environmental factors for natural emission sectors. By aggregating an ensemble of bottomup inventories and processmodels, Saunois et al.^{5} calculated a global methane budget for 2008–2017 and estimated total emissions of 594–880 TgCH_{4} a^{−1}, with 113–154 TgCH4 a^{−1} from fossil fuels, 191–223 TgCH_{4} a^{−1} from agriculture and waste, 26–40 TgCH_{4} a^{−1} from biomass and biofuel burning, 102–182 TgCH_{4} a^{−1} from wetlands, and 143–306 TgCH_{4} a^{−1} from other natural sources. Uncertainty and bias in bottomup CH_{4} emissions in some geographic regions may be caused by imprecise emission factors and activity data that are not readily available at necessary spatial and temporal scales, or by processbased models that perform poorly due to a host of environmental factors (e.g., wetland models rely on wetland inundation maps, biogeochemical process parameterizations, and knowledge of carbon availability^{6,7}).
Atmospheric observations of CH_{4}, combined with an atmospheric transport model and a regularizing statistical approach, provide a topdown constraint on the global CH_{4} budget. Typically referred to as “inversions” or topdown inventories, these methods estimate CH_{4} fluxes by assimilating tower, aircraft, or satellitebased CH_{4} measurements^{8,9,10,11,12,13,14}. Generally, these topdown methods only estimate total fluxes (i.e., sum of all emission sector contributions) explicitly, and may rely on using prior ratios or relative weights (RWs) between source categories to partition fluxes to specific source sectors. Topdown inverse models may be driven by different regularization or prior conditions, complicating direct comparison between an ensemble of inventories. Therefore, these partitioning approaches are prone to error when comparing with bottomup inventories if the prior distribution of emissions is biased, or if different sectors have different uncertainties. For example, Saunois et al. compared an ensemble of 22 topdown CH_{4} global inversions to an ensemble of bottomup inventories, and found the bottomup estimate to be 30% higher than the topdown ensemble mean. The study attributes much of this total discrepancy to large differences in nonwetland natural sources (e.g., lakes and rivers, oceans seeps, termites, geologic sources, wild animals, etc.). However, when integrating emissions over the whole globe, they find that other source categories from bottomup and topdown approaches are consistent within their reported uncertainties (fossil fuel topdown: 81–131 Tg a^{−1}, bottomup: 113–154 Tg a^{−1}; agriculture+waste topdown: 207–240 Tg a^{−1}, bottomup: 191–223 Tg a^{−1}; and biomass+biofuel burning topdown: 22–36 Tg a^{−1}, bottomup: 26–40 Tg a^{−1}). Reported uncertainties reflect the range of estimates among distinct bottomup and topdown inventories across various emission sectors. Saunois et al. also relied on using RWs to partition topdown fluxes to individual bottomup sectors. An explicit approach for comparison between topdown and bottomup inventories, and between independent topdown based inventories, is needed to reduce this uncertainty in the global CH_{4} budget by sector and by region.
Bayesian estimation of emissions from fluxes
We propose a comprehensive Bayesian framework that derives topdown gridded CH_{4} emissions (\(\hat{{{{{{\bf{z}}}}}}}\)) and their error covariance (\(\hat{{{{{{\bf{Z}}}}}}}\)) from inverse fluxes (\(\hat{{{{{{\bf{x}}}}}}}\)) and their error covariance (\(\hat{{{{{{\bf{S}}}}}}}\)) without reliance on RWs from an inventory. This framework takes the following form:
Where the vector \(\hat{{{{{{\bf{x}}}}}}}\) represents inverse CH_{4} fluxes on a spatial grid with full error characterization given by the covariance matrix \(\hat{{{{{{\bf{S}}}}}}}\), x_{A} is the vector of prior fluxes, I is the identity matrix, S_{A} is the prior error covariance matrix, and M is an aggregation matrix that sums emissions to fluxes (Methods section: Eqs. 8–9). The posterior emission error covariance matrix \(\hat{{{{{{\bf{Z}}}}}}}\) is calculated explicitly given M, S_{A}, \(\hat{{{{{{\bf{S}}}}}}}\), and prior emissions error covariance matrix Z_{A}:
The full derivation is documented in the Methods section and Supporting Information (SI) and has been mechanically verified and tested using simulated emissions, concentrations, and fluxes. This approach has been previously described for atmospheric trace gas retrievals^{15} but is here modified and applied for flux comparisons.
Applying Eqs. 1 and 2 allows for the ability to “swap priors.” This means that no matter what prior was used in an initial flux inversion, we can swap it with a different prior emissions vector \({{{{{{\bf{z}}}}}}}_{{{{{{\rm{A}}}}}}}\), which can include sectorbased information. This a critical component for comparison between two different topdown inventories, as it removes error that may arise from choice of prior. Another advantage of this Bayesian approach is that fluxes are partitioned according to not just the prior emission state z_{A}, but also according prior uncertainties on those emissions (i.e., \({{{{{{\bf{Z}}}}}}}_{{{{{{\rm{A}}}}}}}\)). For example, if we have evidence that a particular emission sector is wellcharacterized in bottomup inventories (i.e., a tight prior uncertainty), Eqs. 1–2 take this knowledge into account when optimizing emissions. In this sense, our approach is similar to updated methods that use prior ratios with prior error variance when computing RWs used for partitioning^{16}. However, any RWbased approach still assumes that correlation exists between emission sectors at the gridlevel, creating a relationship that may bias results depending on prior construction.
The main caveat of this Bayesian approach is that an explicit representation of \(\hat{{{{{{\bf{S}}}}}}}\) is needed. For analytical flux inversions, this matrix has a closedform representation that is computed as part of the inversion. For adjointbased inversions^{10,11}, a closedform representation of \(\hat{{{{{{\bf{S}}}}}}}\) is not directly computed. Though the error covariance can still be estimated^{17,18}, this is computationally expensive and is generally not done. Analytical frameworks are best suited for direct comparison between inverse products due to explicit error characterization.
Results
In what follows, we show examples of this approach using a previously performed 2010–2015 GOSAT flux inversion of the Continental United States^{9} (CONUS). We also apply Eqs. 1 and 2 to a previously performed 2018–2019 TROPOMI flux inversion over the Permian Basin in western Texas, southern New Mexico^{19}. We compare with the partitioned results from the 2010–2015 GOSAT inversion to show how projection to a common prior can be used to assess regional emission trends since uncertainties from different priors and different spatial resolutions of the flux inversions are removed.
Partitioning CH_{4} fluxes over CONUS
We apply the partitioning algorithm described by Eqs. 1 and 2 to a 2010–2015 0.50 × 0.625° resolution North American CH_{4} flux inversion^{6} performed using Greenhouse Gas Observing Satellite^{20} (GOSAT) dry air column mixing ratios of CH_{4}. We partition emission to seven distinct sectors at 0.1 × 0.1° resolution: oil, gas, coal, livestock, waste management, wetlands, and other emission sources (soil, fire, etc.). For oil, gas, and coal prior emissions and uncertainties, we use a global inventory for 2016 based on national reports to the United Nations Framework Convention on Climate Change (Scarpelli et al.^{21}). For wetland prior emissions and uncertainties, we take the ensemble average and standard deviation of wetland models that were found to be the highest performing when compared with a global CH_{4} flux inversion^{22}. For all other emission sectors, we use the 2012 EPA gridded CH_{4} inventory as the prior estimate^{23}, and assume a one standard deviation uncertainty equal to 50% of the mean value. For illustration and computational tractability, the prior error covariances are represented as diagonal matrices.
Figure 1a shows the inverse fluxes \(\hat{{{{{{\bf{x}}}}}}}\) over CONUS, optimized emissions \(\hat{{{{{{\bf{z}}}}}}}\) for the gas and livestock sectors when applying Eqs. 1–2, and the change in emissions compared to the emission prior \((\hat{{{{{{\bf{z}}}}}}}{{{{{{\bf{z}}}}}}}_{{{{{{\rm{A}}}}}}})\). Other optimized emission sectors are shown in Fig. S1. Optimized oil and gas emissions show large changes from the prior at the basin scale in several major producing basins and shale plays: Permian (New Mexico/Texas; ΔCH_{4}: gas/oil = 0.40/0.76 TgCH4 a^{−1}), Eagle Ford (southern Texas; ΔCH_{4} = 0.14/0.11 TgCH_{4} a^{−1}), Haynesville (Texas/Louisiana; ΔCH_{4} = 0.18/0.0 TgCH_{4} a^{−1}), Barnett Shale (Texas; ΔCH_{4} = 0.21/0.0 TgCH_{4} a^{−1}), Anadarko (Oklahoma; ΔCH_{4} = 0.52/0.05 TgCH_{4} a^{−1}), and the Appalachia Basin (Ohio/Pennsylvania/West Virginia; ΔCH_{4} = 0.20/0.0 TgCH_{4} a^{−1}). The Niobrara (Wyoming/Colorado; ΔCH_{4} = 0.05/0.0 TgCH4 a^{−1}) region shows much less or no change from the prior. For the Bakken (Montana/North Dakota; ΔCH_{4} = 0.00/−0.06 TgCH_{4} a^{−1}), the small CH_{4} flux enhancement observed in Fig. 1a is partitioned entirely to the oil sector, but produces emissions lower than the prior, which contrasts with the increasing production reported in the basin^{24}. This emission reduction may be due to an overestimate in the prior inventory or a difficulty in GOSAT sampling over that region, factors which could be verified with additional study. Posterior livestock emissions show increases from the prior that are distributed across the central and eastern United States, and a 0.07 TgCH4 a^{−1} decrease in emissions over central California.
A major advantage of using Eqs. 1–2 to estimate emissions from independent inverse fluxes is that any discrepancies in flux priors can be accounted for when partitioning to a common emission prior. We show this through a sensitivity study whose results are summarized in Fig. 2. Here, we use two inverse flux products: (1) the 2010–2015 GOSAT CONUS flux shown in Fig. 1a, and (2) inverse fluxes from 2010–2015 GOSAT recomputed by using the EDGAR v5.0 emission inventory for oil, gas, waste, and livestock sectors for 2015^{25}. Differences in these prior inventories are summarized in Table S1 and Figs. S2–S3. Global inventories like EDGAR use consistent bottomup aggregation methods across countries, which sometimes leads to discrepancies when compared with national inventories for particular sectors^{21}. We employ two methods for optimizing/partitioning inverse fluxes: (1) using optimal estimation (OE) from this study’s Eqs. 1–2 to optimize the same emissions prior from Fig. 1, and (2) by computing RWs of each emission sector in the flux prior and partitioning the inverse fluxes to emissions using these weights. In Fig. 2 we show that by employing OE on each separate inverse flux, we get the exact same answer for optimized emissions. This is a result of the \(({{{{{\bf{I}}}}}}\hat{{{{{{\bf{S}}}}}}}{{{{{{\bf{S}}}}}}}_{A}^{1})\) term from Eq. 1. Sometimes called the averaging kernel matrix^{26}, this term represents the spatial resolution of the flux estimate, or alternatively, the degree of smoothing of the flux prior to the estimate. Since the averaging kernel is applied to the prior swap term \(({{{{{{\bf{x}}}}}}}_{{{{{{\rm{A}}}}}}}{{{{{\bf{M}}}}}}{{{{{{\bf{z}}}}}}}_{{{{{{\rm{A}}}}}}})\), we account for discrepancies between flux and emission priors as a condition of emission optimization, so the choice of flux prior is immaterial. However, Fig. 2 shows the result when a relative weighting scheme is used for partitioning. The livestock to gas ratio in EDGAR is higher than in the EPA and Scarpelli et al. inventories. The result is that RWpartitioned livestock emissions are higher and gas emissions are lower when using EDGAR RWs (13.2 TgCH_{4} a^{−1} and 5.8 TgCH_{4} a^{−1}, respectively) than when using EPAScarpelli RWs (11.3 TgCH_{4} a^{−1} and 8.1 TgCH_{4} a^{−1}, respectively). Similarly, wetland and waste emissions differ depending on which prior are used for RWs. Using prior ratios assumes total correlation between emission sectors in each grid cell as each sector’s RW depends on emissions from other sectors. Therefore, we see that different assumptions on the flux prior can complicate comparison even between inverse fluxes that use the same atmospheric observations and transport model.
Comparing inverse fluxes in the Permian Basin
As seen from Figs. 1 and S1, the Permian Basin shows large increases in CH4 emissions for both oil and gas sectors compared to the prior inventory. Production data from the Energy Information Administration^{24} (EIA) indicates that between 2010 and 2019, the Permian increased oil production by 190% and gas production by 150%. We expect that fugitive emissions from these sectors would increase proportionately to the production increases, but the actual posterior estimates do not differentiate between intentional (planned) and unintentional (unplanned) emissions. The sensitivity study in Fig. 2 shows that the application of Eqs. 1–2 can be used to compare distinct inverse fluxes when a common emissions prior is used. We compare the 2010–2015 GOSAT flux product with a Permian 0.25° × 0.325° flux product^{19} based on May 2018–March 2019 TROPOspheric Monitoring Instrument^{27} (TROPOMI). We partition fluxes to the same sectors as Fig. 1, but at a finer 0.1 × 0.1° grid resolution.
Figure 3a, b shows the 2010–2015 GOSAT flux inversion and the 2018–2019 TROPOMI flux inversion over the Permian Basin, respectively. Within the Permian domain (black line in Fig. 3), the GOSAT inverse flux estimate is 2.01 ± 0.01 TgCH_{4} a^{−1} and the TROPOMI inverse flux estimate is 2.68 ± 0.5 TgCH_{4} a^{−1}, though these inversions were performed over different time periods and used different flux priors and prior error covariance matrices to constrain their solutions. The GOSAT and TROPOMI inverse products show distinct spatial patterns. The TROPOMI inversion shows two main regions of elevated CH4 flux, which correspond to the Delaware Basin on the western side of the Permian (west Texas, southeast New Mexico) and the Midland Basin on the eastern side of the Permian. The GOSAT inversion shows a more distributed region of flux enhancement across the Permian. The difference in spatial distribution of these CH_{4} flux maps could be due to the more limited GOSAT spatial observing coverage over the Permian and the coarser spatial resolution over which the 2010–2015 GOSAT flux inversion was performed. However, given that the two flux inversions are asynchronous and that oil and gas production increased dramatically between 2015 and 2018, spatial differences in CH_{4} fluxes could also be the result of changing infrastructure throughout the basin.
Figure 3d, e show the optimized emissions for the oil, gas, and livestock sectors in the Permian, respectively, and compared to the prior inventory (Fig. 3c). Optimized emissions are derived from GOSAT and TROPOMI fluxes (Fig. 3a, b) that were swapped with a consistent prior (Fig. 3c) using Eqs. 1 and 2. The partitioned GOSAT emissions continue to show more distributed oil and gas CH_{4} emissions across the basin when compared to the partitioned TROPOMI emissions, which are mostly concentrated to the Delaware and Midland Basins. Figure 4 shows the aggregated topdown oil and gas emissions in the Permian compared with reported trends from EIA production reports. The mean 2010–2015 oil and gas production in the Permian was 1.3 million bbl per day and 5.0 million Mcf per day, respectively. This increased to 3.8 million bbl per day and 12.7 million Mcf per day between May 2018 and March 2019 mean production, respectively^{24}. Though not linear with increased production, comparing partitioned 2010–2015 GOSAT oil and gas emissions and 2018–2019 TROPOMI emissions, we see a 0.52 ± 0.29 TgCH_{4} a^{−1} change in CH_{4} emissions, representing a 29% increase. Optimized emissions summed across all sectors increased by 0.42 ± 0.33 TgCH_{4} a^{−1}, with the increase driven mostly by gas, some contribution by oil, and offset by a small decrease in livestock emissions. Originally reported posterior flux estimates showed a 0.67 ± 0.5 TgCH_{4} a^{−1} difference between TROPOMI and GOSAT inversions, larger than the 0.42 ± 0.33 TgCH_{4} a^{−1} difference we quantify here after reprojection to a common prior. This difference discrepancy between topdown estimates corresponds explicitly to differences in flux priors used in the original inversions that is now accounted for with this approach. Therefore, we can quantify how much the choice of flux prior directly impacts any quantified changes between topdown inventories.
Though consistent with the increase in gas production, the quantified 0.42 TgCH_{4} a^{−1} increase in oil and gas emissions from 2010–2015 to 2018–2019 could also be due to observing constraints and/or biases in satellite retrievals. For example, Qu et al.^{28} performed global 2 × 2.5° CH_{4} inversions for 2019 using GOSAT and TROPOMI using the same prior and atmospheric chemistrytransport model, and derived Permian net fluxes of 2.36 TgCH_{4} a^{−1} and 2.43 TgCH_{4} a^{−1}, respectively, showing consistency between signals observed by these independent remote sensing platforms for this region. Therefore, we conclude that the trends observed in Fig. 3 are likely due to changes in gas operations, and not bias in observing systems. However, in other global regions where the surface is less bright and homogeneous than the Permian, flux results derived from TROPOMI and GOSAT inversion may not agree^{28}.
Discussion
Having robust intercomparison methods in place are needed for interpreting the ever increasing number of atmospheric observations of CH_{4}, particularly with regard to the expected launches of several CH_{4}observing satellites in the 2020s^{29}. As described in this study, inverse fluxes derived from these satellite observations will depend on the observations themselves, the chemical transport model, and the prior constraint. Having the ability to directly quantify how these terms influence emissions is needed for diagnosing uncertainty in the estimated methane budget. Ultimately, this information can be combined to provide better global understanding of methane emissions and finer and more policyrelevant spatial and temporal scales. Likewise, the Bayesian framework we describe in this study can be applied to other atmospheric gas fluxes like carbon dioxide (CO_{2}). Partitioning between biospheric and anthropogenic CO_{2} emissions remains highly uncertain^{30}, so incorporating this framework that directly optimizes emission sectors could be useful for reconciling the budget. This approach does require a move from traditional ensemble or adjointbased inversions, created to reduce cost of this computationally expensive problem, to an analytic or optimal estimation inversion as an explicit representation of the posterior covariance is required and this covariance is not easily calculated from ensemble or 4Dvar methods.
While our estimates account for the spatial resolution and error associated with the inversion of observations to fluxes, we do not explicitly account for error in model transport and chemistry. These errors could be important when comparing emissions between seasons or in regions where transport is poorly modeled such as in the tropics. For example, a study of global carbon monoxide emissions, a trace gas that (like CH_{4}) is affected by reaction with the hydroxl radical (OH) and transport, found that convective mass flux in the tropics is likely responsible for errors in emissions^{31}. While our approach can account for this error if a corresponding posterior covariance is provided, we emphasize that studies that both characterize and mitigate this part of the flux error budget are needed to better use satellite observations of methane and of other trace gases. Ultimately, improved emission and error characterization from topdown information will allow for better updates and comparisons with bottomup inventories, which can guide progress towards CH_{4} mitigation.
Methods
In this section we derive a method to estimate CH_{4} emissions from atmospheric observations. The SI contains and alternate derivation (Section S1) and conceptual examples to further clarify the mechanics of priorswapping. Emissions can be represented as a vector i.e., \(({{{{{\bf{z}}}}}}\in {{\mathbb{R}}}^{m})\) that contains both sectoral and spatial information. Atmospheric observations i.e., \(({{{{{\bf{y}}}}}}\in {{\mathbb{R}}}^{p})\) often represent concentrations of CH_{4} observed by surface, satellite, airborne, or some other observing system. Generally, atmospheric inversions do not directly optimize CH_{4} sectoral emissions from observations, and instead optimize CH_{4} fluxes i.e., \(({{{{{\bf{x}}}}}}\in {{\mathbb{R}}}^{n})\), which represent the summation of CH_{4} emissions within a grid cell. In the flux inversion setup, atmospheric CH_{4} observations y are used to estimate CH_{4} fluxes x. We estimate this optimal state by finding the mode of the posterior flux distribution p (xy), or \(\hat{{{{{{\bf{x}}}}}}}\). A transformation or Jacobian matrix K can be derived from atmospheric transport simulations (e.g., GEOSChem), such that we can represent the relationship between fluxes and observations:
Where n represents noise. We apply Bayes Theorem to estimate the posterior distribution p (xy):
Where p (yx) is the maximum likelihood given by Eq. 3, and \(p({{{{{\bf{x}}}}}})\) is the prior distribution. If we assume that p (yx) and p(x) are Gaussian distributions, and y and x_{A} represent the modes of those respective distributions, then the mode of the posterior distribution \(\hat{{{{{{\bf{x}}}}}}}\) has a closed form solution, known as the Maximum A Posteriori (MAP; Rodgers, 2000) solution:
For policyrelevance and CH_{4} budget quantification, we really wish to optimize emissions using atmospheric observations, i.e., we want to compute the explicit posterior representation \(p({{{{{\bf{z}}}}}}\,\,{{{{{\boldsymbol{y}}}}}})\) without resimulation of an atmospheric transport model. The relationship between z and x is simple aggregation, and can represented by matrix M:
If z and x share the same grid resolution (i.e., if \({{{{{\bf{z}}}}}}\in {{\mathbb{R}}}^{m}\) and \({{{{{\bf{x}}}}}}\in {{\mathbb{R}}}^{n}\) and s is the number of emission sectors, then m = ns), the matrix \({{{{{\bf{M}}}}}}\in {{\mathbb{R}}}^{n\times m}\) is represented with the following terms:
If z and x pertain to different grids, the relationship M is defined by the geographic area (Ω) and intersections (\(\cap\)) of grid cells:
Since M is simply a summation matrix, we assume there is no noise associated with its application. Using M, we can update Eqs. 5 and 6 to find the optimal emission state vector \(\hat{{{{{{\bf{z}}}}}}}\) and its posterior error covariance \(\hat{{{{{{\bf{Z}}}}}}}\):
Equations 10 and 11 provide an explicit closed form solution for \(\hat{{{{{{\bf{z}}}}}}}\), which is sufficient for emission optimization without reliance on RWs. Therefore, application of Eqs. 10 and 11 into existing inverse frameworks would provide posterior emission estimates constrained by atmospheric observations. However, these equations require the computation of the matrix (KM), which can be large in cases where many atmospheric observations exist. An exactly equivalent solution is possible with just the products of the flux inversion, specifically \(\hat{{{{{{\bf{S}}}}}}}\), \(\hat{{{{{{\bf{x}}}}}}}\), \({{{{{{\bf{S}}}}}}}_{{{{{{\rm{A}}}}}}}\), and \({{{{{{\bf{x}}}}}}}_{{{{{{\rm{A}}}}}}}\). We show one derivation below and provide an alternative derivation in the SI:
Equation 5 can be shown to have an equivalent form that is often used in atmospheric trace gas retrievals^{26}:
Where A is the averaging kernel matrix \({{{{{\bf{A}}}}}}=\frac{\partial \hat{{{{{{\bf{x}}}}}}}}{\partial {{{{{\bf{x}}}}}}}\):
And G is the Gain Matrix \({{{{{\bf{G}}}}}}=\frac{\partial \hat{{{{{{\bf{x}}}}}}}}{\partial {{{{{\bf{y}}}}}}}\):
And x are the true atmospheric fluxes. Therefore, Eq. 12 shows that the optimal solution \(\hat{{{{{{\bf{x}}}}}}}\) is a combination of the truth, smoothed by some prior and includes noise. From Eq. 12, we can create a flux to posterior flux operation H, given that our relationship M is known:
The operation H allows for emissions to be smoothed with an averaging kernel, allowing for direct comparison with \(\hat{{{{{{\bf{x}}}}}}}\). With this operator relationship, we treat \(\hat{{{{{{\bf{x}}}}}}}\) as an observable. The error covariance \(\hat{{{{{{\bf{S}}}}}}}\) includes both smoothing (\({{{{{{\bf{S}}}}}}}_{{{{{{\rm{s}}}}}}}\)) and measurement error (S_{m}):
For flux partitioning, we want to isolate the S_{m} error component \(\hat{{{{{{\bf{x}}}}}}}\), as the H operator already accounts for smoothing via the averaging kernel. The matrix S_{m} can be represented^{32} using G and S_{y}:
While S_{s} has the following representation:
We can combine Eqs. 17 and 19 to get an alternate form of S_{m} that does not require S_{y} and G explicitly:
Using S_{m} for observational error covariance, we can apply the MAP solution to derive \(\hat{{{{{{\bf{z}}}}}}}\) and \(\hat{{{{{{\bf{Z}}}}}}}\):
Now we have a direct solution for \(\hat{{{{{{\bf{z}}}}}}}\) and \(\hat{{{{{{\bf{Z}}}}}}}\) derived only from the products of a flux inversion (specifically, \({{{{{{\bf{x}}}}}}}_{{{{{{\rm{A}}}}}}},\,{{{{{{\bf{S}}}}}}}_{{{{{{\rm{A}}}}}}}\), \(\hat{{{{{{\bf{x}}}}}}}\), \(\hat{{{{{{\bf{S}}}}}}}\)), an emissions prior (\({{{{{{\bf{Z}}}}}}}_{{{{{{\rm{A}}}}}}}\)), and an aggregation matrix M. Equations 22 and 23 can be shown to be of the same form as Eqs. 1 and 2 by showing that \({{{{{{\bf{A}}}}}}}^{T}{{{{{{\bf{S}}}}}}}_{{{{{{\rm{m}}}}}}}^{1}={\hat{{{{{{\bf{S}}}}}}}}^{1}\). Decomposing \({{{{{{\bf{A}}}}}}}^{T}\) and recognizing that \(\hat{{{{{{\bf{S}}}}}}}\) and \({{{{{{\bf{S}}}}}}}_{{{{{{\rm{A}}}}}}}\) are symmetric matrices, we have
We first expand Eq. 19 for an alternative expression of S_{m}:
Taking the inverse of S_{m} using the Eq. 27, we have:
The algebra for 28 and 29 are only possible if \(({{{{{\bf{A}}}}}}^{T})^{1}\) exists. For overdetermined systems (i.e., dimension of y » dimension of x), this is generally valid. Multiplying both sides of Eq. 29 by \({{{{{{\bf{A}}}}}}}^{T}\) finishes the proof:
In a similar fashion, we can show that Eqs. 2 and 23 are equivalent if \({{{{{{\bf{A}}}}}}}^{T}{{{{{{\bf{S}}}}}}}_{{{{{{\rm{m}}}}}}}^{1}{{{{{\bf{A}}}}}}={\hat{{{{{{\bf{S}}}}}}}}^{1}{{{{{{\bf{S}}}}}}}_{A}^{1}\). To do this, we multiply Eq. 30 by A:
Data availability
Fossil fuel prior emission inventories are available for download at https://doi.org/10.7910/DVN/HH4EUM. Wetland emission prior inventories are available at (https://doi.org/10.3334/ORNLDAAC/1502). The EPA gridded methane inventory is available for download at https://www.epa.gov/ghgemissions/gridded2012methaneemissions.
Code availability
Code used for this analysis can be found at https://github.com/dcusworth/partition_fluxes_to_emissions.
References
Myhre, G. et al. Anthropogenic and Natural Radiative Forcing Supplementary Material. In: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change Available from www.climatechange2013.org and www.ipcc.ch. (2013).
United Nations Environment Programme and Climate and Clean Air Coalition (2021). Global Methane Assessment: Benefits and Costs of Mitigating Methane Emissions. Nairobi: United Nations Environment Programme (2021).
Kirschke, S. et al. Three decades of global methane sources and sinks. Nature Geoscience 6, 813–823 (2013).
UNFCCC. Adoption of the Paris Agreement. Report No. FCCC/CP/2015/L.9/Rev.1, http://unfccc.int/resource/docs/2015/cop21/eng/l09r01.pdf, 2015
Saunois, M. et al. The global methane budget 2000–2017. Earth Syst. Sci. Data 12, 1561–1623 (2020).
Melton, J. R. et al. Present state of global wetland extent and wetland methane modelling: Conclusions from a model intercomparison project (WETCHIMP). Biogeosciences 10, 753–788 (2013).
Poulter, B. et al. Global wetland contribution to 20002012 atmospheric methane growth rate dynamics. Environ. Res. Lett. https://iopscience.iop.org/article/10.1088/17489326/aa8391 (2013).
Zhang, Y. et al. Attribution of the accelerating increase in atmospheric methane during 2010–2018 by inverse analysis of GOSAT observations. Atmos. Chem. Phys. 21, 3643–3666 (2021).
Maasakkers, J. D. et al. 2010–2015 North American methane emissions, sectoral contributions, and trends: a highresolution inversion of GOSAT satellite observations of atmospheric methane. Atmos. Chem. Phys. 21, 4339–4356 (2021).
Bergamaschi, P. et al. Inverse modelling of European CH4 emissions during 2006–2012 using different inverse models and reassessed atmospheric observations. Atmos. Chem. Phys. 18, 901–920 (2018).
Alexe, M. et al. Inverse modelling of CH4 emissions for 2010–2011 using different satellite retrieval products from GOSAT and SCIAMACHY. Atmos. Chem. Phys. 15, 113–133 (2015).
Yadav, V. et al. Spatio‐temporally resolved methane fluxes from the Los Angeles Megacity. J. Geophys. Res. Atmos. 124, 5131–5148 (2019).
Miller, S. M. et al. Anthropogenic emissions of methane in the United States. Proc. Natl Acad. Sci. USA 110, 20018–20022 (2013).
Ganesan, A. L. et al. Quantifying methane and nitrous oxide emissions from the UK and Ireland using a nationalscale monitoring network. Atmos. Chem. Phys. 15, 6393–6406 (2015).
Rodgers, C. D. & Connor, B. J. Intercomparison of remote sounding instruments. J. Geophys. Res. Atmos. 108, 4116 https://doi.org/10.1029/2002JD002299, D3 (2003).
Shen, L. et al. Unravelling a large methane emission discrepancy in Mexico using satellite observations. Remote Sens. Environ. 260, 112461 (2021).
Meirink, J. F., Bergamaschi, P. & Krol, M. C. Fourdimensional variational data assimilation for inverse modelling of atmospheric methane emissions: method and comparison with synthesis inversion. Atmos. Chem. Phys. 8, 6341–6353 (2008).
Bousserez, N. & Henze, D. K. Optimal and scalable methods to approximate the solutions of largescale Bayesian problems: theory and application to atmospheric inversion and data assimilation. Q. J. R. Meteorol. Soc. 144, 365–390 (2018).
Zhang, Y. et al. Quantifying methane emissions from the largest oilproducing basin in the United States from space. Sci. Adv. 6, eaaz5120 (2020).
Kuze, A. et al. Update on GOSAT TANSOFTS performance, operations, and data products after more than 6 years in space. Atmos. Meas. Tech. 9, 2445–2461 (2016).
Scarpelli, T. R. et al. A global gridded (0.1× 0.1) inventory of methane emissions from oil, gas, and coal exploitation based on national reports to the United Nations Framework Convention on Climate Change. Earth System Science. Data 12, 563–575 (2020).
Ma, S. et al. Satellite constraints on the latitudinal distribution and temperature sensitivity of wetland methane emissions. AGU Adv. 2, e2021AV000408 (2021).
Maasakkers, J. D. et al. Gridded national inventory of US methane emissions. Environ. Sci. Technol. 50, 13123–13133 (2016).
EIA. Drilling Productivity Report, URL https://www.eia.gov/petroleum/drilling/, Last Accessed 2 Feb 2021 (2020).
Crippa, M. et al. Fossil CO2 and GHG emissions of all world countries  2019 Report, EUR 29849 EN, Publications Office of the European Union, Luxembourg, ISBN 9789276111009, 10.2760/687800, JRC117610. (2019).
Rodgers, C. D. Inverse methods for atmospheric sounding: theory and practice (Vol. 2). World Scientific (World Scientific Publishing Co. Pte. Ltd, 2000).
Veefkind, J. P. et al. TROPOMI on the ESA Sentinel5 Precursor: A GMES mission for global observations of the atmospheric composition for climate, air quality and ozone layer applications. Remote Sens. Environ. 120, 70–83 (2012).
Qu, Z. et al. Global distribution of methane emissions: a comparative inverse analysis of observations from the TROPOMI and GOSAT satellite instruments. Atmos. Chem. Phys. Discussions 21, 14159–14175 (2021).
Crisp, D. et al. A constellation architecture for monitoring carbon dioxide and methane from space. Prepared by the CEOS Atmospheric Constellation Greenhouse Gas Team, Version, 1(8), https://ceos.org/document_management/Virtual_Constellations/ACC/Documents/CEOS_ACVC_GHG_White_Paper_Version_1_20181009.pdf (2018).
Friedlingstein, P. et al. Global carbon budget 2020. Earth Syst. Sci. Data 12, 3269–3340 (2020).
Jiang, Z. et al. Impact of model errors in convective transport on CO source estimates inferred from MOPITT CO retrievals. J. Geophys. Res. Atmos. 118, 2073–2083 (2013).
Bowman, K. W. et al. Tropospheric emission spectrometer: retrieval method and error analysis. IEEE Trans. Geosci. Remote Sens. 44, 1297–1307 (2006).
Acknowledgements
Portions of this work research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration (80NM0018D0004). Some of the work was supported by NASA’s Carbon Monitoring System program. Yuzhong Zhang acknowledges funding by NSFC (42007198). Government sponsorship acknowledged.
Author information
Authors and Affiliations
Contributions
D.H.C. and J.R.W. designed the study. D.H.C. performed the analysis and wrote the manuscript. D.H.C., Y.Y., J.R.W., A.A.B., and K.B. developed/derived the algebraic equations. S.M., J.D.M., C.E.M., T.R.S., Z.Q., D.J.J., and Y.Z. provided and guided interpretation and integration of prior and posterior inventories. All authors provided feedback and edits on the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Peer review information Communications Earth & Environment thanks Luke Western, Hossein Maazallahi and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Joshua Dean and Clare Davis. Peer reviewer reports are available.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Cusworth, D.H., Bloom, A.A., Ma, S. et al. A Bayesian framework for deriving sectorbased methane emissions from topdown fluxes. Commun Earth Environ 2, 242 (2021). https://doi.org/10.1038/s43247021003126
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s43247021003126
This article is cited by

Predictability of fossil fuel CO2 from air quality emissions
Nature Communications (2023)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.