Rain-fed pulses of methane from East Africa during 2018–2019 contributed to atmospheric growth rate

Two CH₄ satellite datasets are used in this study to examine changes over East Africa: TROPOMI and GOSAT. TROPOMI measures, amongst others, short wave infrared (SWIR) radiances around 2.3 µm. To calculate column-averaged CH₄ concentrations (XCH₄) a physics-based retrieval is used, making use of the Oxygen A-band around 760 nm to infer information about the scattering properties of the atmosphere. The satellite is in a Sun-synchronous orbit with an overpass time of 13:30 local solar time. TROPOMI has a swath width of 2600 km and a ground pixel of 7 × 7 km². We use the scientific CH₄ data product (Lorente et al 2020), which implements the RemoTec retrieval algorithm (Butz et al 2011, Hu et al 2016), and use bias-corrected XCH₄ in our analysis, with the bias correction supplied in the data (Lorente et al 2020). Only data with a quality flag of greater than 0.5 are used, which filters data for clouds, high solar and viewing zenith angles and high aerosol loading. In addition we filter the data based on where the SWIR surface albedo is in the range 0.05–0.30, the aerosol optical thickness (AOT) is less than 0.1, and the standard deviation of surface topography in each retrieved pixel is less than 60 m. Initial tests revealed some anomalously low value data values surviving through these filters, so we further filter the data based on where the difference between observed values and the simulated model background levels within the domain are greater than −5 ppb. We use data over the spatial domain of our model simulations which covers −36 to +20^∘ N and −20 to 55^∘ E.

GOSAT is in a Sun-synchronous orbit with a local equator crossing time of 13:00, resulting in global coverage every three days (Kuze et al 2009). We use data from the Thermal And Near infrared Sensor for carbon Observations—Fourier Transform Spectrometer (TANSO-FTS) that measures short wave infrared (SWIR) radiances between 0.76 and 2.0 µm at a resolution of 0.3 cm⁻¹ (Parker et al 2011). We use the University of Leicester's (UoL) v9 GOSAT-OCPR proxy XCH₄ product (Parker et al 2020a). This retrieval algorithm uses a different approach to TROPOMI to calculate XCH₄, relying on the co-retrieved CO₂ column. Taking the ratio between CH₄ and CO₂ can account for factors that impact the retrievals, such as aerosol and cloud scattering. An advantage of this proxy approach is that it is more robust than the full physics approach in the presence of clouds and aerosols. However, the proxy approach does rely on having good knowledge of the true CO₂ distribution in the atmosphere, and errors in this could impact on the derived XCH₄ values. As such, both the full physics and proxy approaches have their strengths and weaknesses. Due to the greater spatial coverage of TROPOMI the number of successful retrievals is far greater than from GOSAT (figure S1). Both datasets have been validated against ground-based data from the Total Carbon Column Observing Network (TCCON) (Wunch et al 2011). For both datasets we average the data into $0.25^{\circ} \times 0.3125^{\circ}$ pixels to create a set of super-observations at a resolution consistent with the GEOS-Chem atmospheric transport model.

Rainfall data in this study are taken from the TAMSAT (v3.1) monthly precipitation dataset (Maidment et al 2014, Tarnavsky et al 2014, Maidment et al 2017). The data are based on high-resolution thermal-infrared observations and available from 1983 to the present. Rainfall anomalies reported in this work are relative to the 1983–2012 monthly mean. To examine changes in terrestrial water storage we use data from the GRACE Follow-On (GRACE-FO) mission (Landerer et al 2020). The GRACE-FO dataset provides monthly liquid water equivalent (LWE) estimates at a resolution of 3^∘ × 3^∘. Anomalies are relative to the consistent longer-term GRACE dataset 2002–2017 monthly means.

We use the GEOS-Chem chemistry transport model to simulate the emissions and transport of atmospheric methane (Bey et al 2001, Turner et al 2015). The model was run in a nested configuration with a spatial resolution of 0.25$^{\circ}\times0.3125^{\circ}$ driven by offline meteorology fields from the GEOS-FP analysis product. The model has 47 vertical levels. The temporal resolution of the meteorology fields was hourly and the transport and chemistry time steps were five and ten minutes respectively. The nested domain covered −36 to 20^∘ N and −20 to 55^∘ E, encompassing sub-Saharan Africa.

A priori CH₄ emissions within the nested domain mostly followed those used in previous published work (Maasakkers et al 2019), with anthropogenic emissions for 2012 from EDGAR v4.3.2, and wetland emissions for 2015 from the WetCHARTs dataset (Bloom et al 2017). Wetland emissions varied monthly and the anthropogenic emissions included a monthly varying seasonal cycle for manure emissions and rice. Daily biomass burning emissions for 2018–2019 were taken from the GFAS database (Kaiser et al 2012). Offline loss fields were included accounting for oxidation by the hydroxyl and chlorine radicals, as well as soil absorption (Wecht et al 2014).

Boundary conditions to the nested domain were created from a global run of the GEOS-Chem model run at 2^∘ × 2.5^∘, with 3 h output fields. These fields were sampled at the time and location of GOSAT data, and column average XCH₄ values compared to the data after applying the GOSAT averaging kernels. To create more realistic boundary condition fields that were consistent with the observed zonal distribution of CH₄, these model fields were then scaled to match the GOSAT zonal monthly mean concentrations within a domain of −50^∘ E to 100^∘ E, encompassing data over the Atlantic and Indian oceans. GOSAT data from within the nested African domain were not included in this zonal mean comparison to avoid using these data twice.

The zonal mean of each 2^∘ latitudinal grid cell in the model is thus consistent with the equivalent zonal mean of GOSAT data. However, the vertical and longitudinal distributions of the field are reliant on the a priori emissions fields and GEOS-Chem model transport. The boundary conditions thus represent a reasonable approximation to the true atmospheric field at the boundaries of the nested domain and are consistent with the satellite data which is important to negate any global or latitudinal biases in the data or model. The same boundary conditions were used for both TROPOMI and GOSAT inversions, although due to an offset between the two satellites the boundary conditions from TROPOMI were adjusted according to the global mean ratio between the two satellite datasets each month. The impact of using alternative boundary condition fields on our results is explored in the supplementary material (available online at stacks.iop.org/ERL/16/024021/mmedia).

To estimate CH₄ emissions we use an Ensemble Kalman Filter (EnKF) system. Here, we use a variant called the Local Ensemble Transform Kalman Filter (LETKF) (Hunt et al 2007) which has been applied in various previous atmospheric inversions to infer emissions (Miyazaki et al 2012, Liu et al 2016). We briefly describe the LETKF here, and refer the reader to Hunt et al (2007) for an in-depth description.

The LETKF transforms a background ensemble of emissions $\mathbf{x}^{b}$ with k ensemble members into an analysis ensemble $\mathbf{x}^{a}$. The emissions ensemble is defined by its mean ${\bar{\mathbf{x}}^{\mathbf{b}}}$ and ensemble perturbations $\mathbf{X^{b}}$ given by:

$$\begin{equation} \mathbf{X^{b}} = \mathbf{x^{b}} - {\bar{\mathbf{x}}^{\mathbf{b}}}. \end{equation} \tag{ 1 }$$

The GEOS-Chem model is used to create the operator, H that propagates the emissions ensemble into the observation space following:

$$\begin{equation} \mathbf{y^{b}} = H \cdot \mathbf{x^{b}}. \end{equation} \tag{ 2 }$$

Ensemble perturbations in the measurement space, $\mathbf{Y^{b}}$, are similarly defined to the emissions space as:

$$\begin{equation} \mathbf{Y^{b}} = \mathbf{y^{b}} - {\bar{\mathbf{y}}^{\mathbf{b}}}. \end{equation} \tag{ 3 }$$

The background ensemble, $\mathbf{x}^{b}$, is updated to an analysis ensemble $\mathbf{x}^{a}$, based on the data, $\mathbf{y}^{o}$, and calculated by:

$$\begin{equation} \bar{\mathbf{x}^{\mathbf{a}}} = {\bar{\mathbf{x}}^{\mathbf{b}}} + \mathbf{X^{b}} \cdot \tilde{\mathbf{P}}^{\mathbf{a}} \cdot (\mathbf{Y^{b}})^\textrm{T} \mathbf{R^{-1}} \cdot (\mathbf{y}^{o} - {\bar{\mathbf{y}}^{\mathbf{b}}}), \end{equation} \tag{ 4 }$$

where $\mathbf{R}$ is the observation error covariance matrix, and $\mathbf{R^{-1}}$ its matrix inverse. $\tilde{\mathbf{P}}^{\mathbf{a}}$ is the local analysis error covariance matrix estimated in the ensemble space, and given by:

$$\begin{equation} {\tilde{\mathbf{P}}^{\mathbf{a}}} = \left[ (\mathbf{Y^{b}})^{T} \mathbf{R^{-1}} \mathbf{Y^{b}} + (k-1)\mathbf{I} \right]^{-1}. \end{equation} \tag{ 5 }$$

The updated ensemble perturbations can be calculated according to:

$$\begin{equation} \mathbf{X^{a}} = \mathbf{X^{b}} \left[ (k-1) {\tilde{\mathbf{P}}^{\mathbf{a}}} \right]^{1/2}. \end{equation} \tag{ 6 }$$

Finally, the a posteriori analysis error covariance matrix is equal to:

$$\begin{equation} \mathbf{P^{a}} = \mathbf{X^{a}} (\mathbf{X^{a}})^{T} \frac{1}{k-1}. \end{equation} \tag{ 7 }$$

The diagonals of $\mathbf{P^{a}}$ represent the uncertainties of the a posteriori emissions. In our implementation of the LETKF, the state vector, $\mathbf{x}$, describes a vector of scale factors applied to the a priori emissions field, with each term in $\mathbf{x}$ representing a box of horizontal dimension $0.5^{\circ}\times0.625^{\circ}$. Our a posteriori emissions are calculated by multiplying the a posteriori scale factors by the a priori emission magnitudes. Unless otherwise stated, all results represent total CH₄ emissions from the sum of all sources.

We used an ensemble of 140 members, a temporal assimilation window of 15 days, and lag window of 1 month. A priori emission uncertainties of each grid cell were set to be 200% of the value of each grid cell. In the LETKF, we applied a spatial localization of 800 km that limits the set of observations that are used in the analysis of each state vector element to be within this distance. The diagonal terms of the measurement covariance matrix, $\mathbf{R}$, were formed from a combination of the measurement retrieval a posteriori uncertainties provided in the TROPOMI and GOSAT data files and a modelling uncertainty of 4 ppb, based on the global mean bias of the satellite columns versus TCCON data (Lorente et al 2020, Parker et al 2020a). We defined off-diagonal terms of $\mathbf{R}$, to follow an exponential spatial covariance that decayed with a length scale of 50 km. We tested the impact of the inversion parameter definitions on our results through a series of sensitivity tests (see supplementary material).

Between December 2017 and December 2019 there were two notable positive rainfall anomalies: the long rains season of MAM 2018, and the short rains season of OND 2019 (figure 1(a)). The long rains anomaly in 2018, peaked in April with 86 mm month⁻¹, relative to the 1983–2012 mean. The short rains anomaly in 2019, peaked in October with 81 mm month⁻¹. The standard deviation of monthly anomalies over the 38-year self-consistent data record was 17 mm month⁻¹ so that both rainfall anomalies exceed the 3σ level, approximately equivalent to a once in a 30-year event. The 30-year mean rainfall peak in April was 88 mm month⁻¹, and 67 mm month⁻¹ in October, showing that these anomalies represent at least double the normal amount of rainfall across the region.

Zoom In Zoom Out Reset image size

These two large seasonal anomalies had contrasting spatial distributions. The MAM 2018 anomalies were principally over Kenya, Tanzania and Ethiopia (figure 1(b)), while the anomalies during OND 2019 were also over Somalia, Uganda and South Sudan (figure 1(c)). In contrast the wet seasons in OND 2018 and MAM 2019 were close to climatological mean values, with mean anomalies of −3 mm month⁻¹ in both OND 2018 and MAM 2019.

A similar, albeit smoother, pattern of anomalies is seen in liquid water equivalent (LWE) height anomalies in East Africa from the GRACE-FO mission (Landerer et al 2020) (figure 1(a)), between 2018–07 and 2019–12. The data show a minimum reached in 2019–04, before an increase in the latter part of 2019. Previous studies used LWE anomalies to isolate wetland emissions of CH₄ using satellite column measurements of CH₄ (Bloom et al 2010, Bloom et al 2012), implying likely changes in wetland emission anomalies as a result of the LWE changes during this period.

Corresponding seasonal mean XCH₄ enhancements over East Africa (figure 2) are calculated by subtracting a model background component from the XCH₄ column. The model background component is generated by including only that part of the modelled CH₄ field that originated from outside the regional domain in the GEOS-Chem simulation. We assume that the dominant factor affecting XCH₄ enhancements is surface emissions, acknowledging that any errors in generating the modelled background component will affect this quantity. Gaps in the XCH₄ data are due to cloud coverage, surface albedo or aerosol interference preventing successful retrievals; some seasonal mean values may be indicative of only a few successful retrievals in that season (figure S2).

Zoom In Zoom Out Reset image size

We find the largest XCH₄ enhancements over the Sudd wetlands in South Sudan and the Sobat basin to its east, which are present in OND 2018 and 2019. October represents the end of the wet season in South Sudan, but it is usually the time of peak flooding of the Sudd, due to inflow from the region to the south known as the torrents (Sutcliffe and Parks 1999, Lunt et al 2019). In MAM 2018 large XCH₄ enhancements (mean 17 ppb; max 93 ppb) are observed in South Sudan and regions south and east of Lake Victoria. In contrast, in MAM 2019 these regions display smaller enhancements (mean 12 ppb; max 52 ppb) consistent with reduced precipitation and emissions. Similarly, while enhancements (mean 17 ppb; max 65 ppb) are seen during OND 2018, consistent with wetland emissions from South Sudan, even larger enhancements (mean 20 ppb; max 86 ppb) are present in OND 2019, particularly over South Sudan (figure 2). The mean XCH₄ enhancements in MAM 2018 and OND 2019 across East Africa are 33% and 18% larger than the equivalent season in the opposite year, consistent with the large positive seasonal precipitation anomalies (figure 1).

The enhancements in the satellite data provide a useful indication of potential emission anomalies. However, to fully understand the behaviour of the fluxes a more formal Bayesian inversion is required. XCH₄ measurements from TROPOMI and GOSAT were assimilated into the LETKF inversion to provide monthly CH₄ emission estimates. The a posteriori CH₄ emission estimates corresponding to each set of data are generally consistent within their uncertainties (figure 3). Both sets of estimates show prominent emission peaks during MAM 2018 and particularly OND 2019. Our a posteriori emissions estimate from TROPOMI XCH₄ data in MAM 2018 is 6.2 ± 0.3 Tg CH₄ and in OND 2019 is 8.6 ± 0.3 Tg CH₄. These represent a 0.5 ± 0.2 Tg CH₄ (10%) and 2.3 ± 0.2 Tg CH₄ (37%) increase from MAM 2019 and OND 2018, respectively. We calculate similar a posteriori seasonal emissions inferred from GOSAT XCH₄ data of 6.4 ± 0.5 Tg CH₄ and 8.1 ± 0.5 Tg CH₄ in MAM 2018 and OND 2019, respectively, with an emissions difference in OND of 2.1 ± 0.3 Tg CH₄ and a larger MAM difference of 1.2 ± 0.3 Tg CH₄ (table S1).

Zoom In Zoom Out Reset image size

The additional seasonal pulses of CH₄ emissions in MAM 2018 and OND 2019 coincide with large growth rates of global atmospheric CH₄ of 8.5 ppb and 10.4 ppb in 2018 and 2019 respectively (Dlugokencky 2020). By incorporating these global mean CH₄ data into a one box-model calculation, global emissions growth is estimated to be 5 and 7 Tg CH₄ in the respective years, assuming a constant atmospheric lifetime (see appendix and figure S3). Therefore, the additional seasonal emissions pulses we report from East Africa, based on TROPOMI (GOSAT) data, represent 10 (24)% and 32 (30)% of the increase in global CH₄ emissions in 2018 and 2019 respectively. Although our calculation is subject to assumptions regarding the atmospheric loss rate of CH₄ it demonstrates that the seasonal East African emission pulses could account for a considerable fraction of the global emissions growth in 2018 and 2019.

The distribution of OND emissions in 2018 and 2019 from both TROPOMI and GOSAT derived estimates is shown in figure 4. Both estimates show increases in emissions in the region surrounding Lake Victoria. The TROPOMI estimates indicate an increase of 0.7 ± 0.1 Tg CH₄ from this region and the GOSAT inversions 0.4 ± 0.3 Tg CH₄. The major regions of change from the GOSAT estimates are seen in S Sudan and Ethiopia with two distinct hotspots around 8^∘ N. The increase in emissions from the S Sudan regions is 0.4 ± 0.1 Tg CH₄, and from Ethiopia 0.8 ± 0.2 Tg CH₄. The TROPOMI results indicate a smaller increase of 0.2 ± 0.1 Tg CH₄ from both regions, but a change in the emissions distribution, with greater emissions in 2019 around the South Sudan/Ethiopia border in the Sobat basin. Greater emissions from this wetland region to the east of the Sudd, are consistent with significant positive soil moisture anomalies in OND 2019 (figure S4).

Zoom In Zoom Out Reset image size

The simulated background component of the data is consistent between the two observational datasets. Therefore, differences in estimated emissions reflect differences in the spatial coverage and density of observations as well as any offsets in observed XCH₄ values. The sparser GOSAT coverage, relative to TROPOMI, can explain why emissions are generally confined to regions where a priori emissions are most prominent (figure S5).

We solve for total CH₄ emissions, and cannot directly determine the underlying source sectors behind the seasonal emissions pulses we find. However, we can make the crude assumption that the fractional contribution of each source sector in each grid cell is the same in the a posteriori distribution as in the a priori. Given this assumption, we find that the largest component of the seasonal OND emissions peak in both years is from wetlands emissions (60%) followed by agriculture (15%), based on the difference between a posteriori emissions in OND and June. The division into sector emissions is limited because the a priori source sectors are not spatially distinct (figures S5–S6). However, at a qualitative level it suggests wetlands are the major driver of the a posteriori seasonal OND peak.

To further investigate the impact of the a priori distribution on the estimated short rains emission pulse of 2019 we conducted a sensitivity test, redistributing the wetland component of emissions to follow the soil moisture distribution (figure S7), since errors in the wetland inventory extent have been shown to cause a mismatch to satellite observations (Parker et al 2020b). Using our revised a priori emissions distribution we estimate a posteriori emissions of 8.4 ± 0.4 Tg CH₄ in OND 2019 that are close to the estimate (8.6 ± 0.3 Tg CH₄) corresponding to the main results, although with a different distribution of emissions (figure S7). We find the East Africa seasonal emission pulses are also relatively insensitive to changes in lateral boundary conditions and various assumptions made in the inversion method (figure S8).

Annual mean total CH₄ emission estimates from East Africa for 2018 and 2019 are 25 ± 1 and 27 ± 1 Tg CH₄ yr⁻¹ respectively. Our GOSAT derived estimates for both years are 24 ± 1 Tg CH₄ yr⁻¹. Results from both satellites are substantially larger than the a priori fluxes, the mean of which is 15 Tg CH₄ yr⁻¹ in both years. Through scaling the relative a priori emissions of each source sector in each grid cell by the a posteriori ratio we find that, in June (when emissions are at a minimum) agricultural emissions from livestock account for 60% of this difference between a priori and a posteriori emissions. A sensitivity test using a priori fluxes that are twice as large in all grid boxes returned the same annual mean emissions for East Africa, indicating the minimal impact of a priori emissions on our results at this regional scale (figure S9).

Our East African emissions estimates are slightly smaller than the estimate from previously published work for the 2011–2016 mean of 27 Tg CH₄ yr⁻¹ (Lunt et al 2019). We attribute this to differences in emissions from South Sudan. Our annual mean emission estimates for total CH₄ from S Sudan are 3.4 ± 0.2 Tg CH₄ yr⁻¹ in 2018–2019 from TROPOMI XCH₄ data and 3.4 ± 0.3 Tg CH₄ yr⁻¹ from GOSAT XCH₄ data, compared to the recently published estimates of 6.3 Tg CH₄ yr⁻¹ for 2015–2016 (Lunt et al 2019). Assuming a dependency on water table height, these estimates can be reconciled by GRACE LWE height anomalies which dropped from +2.6 cm in 2015–2016 to −1.5 cm in 2018, returning to levels closer to values in 2011–2012 when estimates from this previous study were 3.5 Tg CH₄ yr⁻¹ (Lunt et al 2019).

However, our estimates for the sum of all sources from S Sudan are inconsistent with another recent study which estimated emissions from only wetlands of 7.2 ± 3.2 Tg CH₄ yr⁻¹ in 2018–2019 using TROPOMI data and a mass balance method (Pandey et al 2020). A bias in the definition of the background concentration used by either approach will lead to larger or smaller XCH₄ enhancements and subsequent emission estimates. For instance, our sensitivity tests show that a 5 ppb positive bias in the background results in East African emissions that are larger by 6 Tg CH₄ yr⁻¹. However, we find that this bias translates into a S Sudan estimate that is larger by only 0.5 Tg CH₄ yr⁻¹, suggesting that the background contribution to the total XCH₄ column would have to be substantially smaller to reconcile the two estimates. When we relax our TROPOMI observation selection criteria to closely follow Pandey et al (2020), allowing greater aerosol loading and surface albedo thresholds, our a posteriori estimate for S Sudan increases to 4.6 Tg CH₄ yr⁻¹. While this partly helps to reconcile the emission difference the resulting emission distribution is forced to be unrealistic, and inconsistent with results using GOSAT data (supplementary material). A further explanation is likely to lie in the use of different meteorology fields (GEOS-FP vs ERA-5) and the use of a mass-balance method in Pandey et al (2020) versus the 3D modelling and EnKF approach used here.