Inverse modeling of global and regional CH₄ emissions using SCIAMACHY satellite retrievals

3.1.1. Implementation of Nonlinear 4DVAR System to Avoid Negative a Posteriori Emissions

[21] The previous system assumed a Gaussian probability density function (PDF) for the a priori emissions. In case of uncertainties that are of the same order of magnitude as the emissions themselves (which is typically assumed for CH₄ emissions), this implies a nonnegligible probability that emissions become negative. In fact, negative a posteriori emissions were sometimes obtained with the previous TM5-4DVAR system in some regions, in particular when strongly constraining observational data sets were applied, expressing a compensation for observational and model errors. In some cases, this artifact in the derived emissions also created significant artifacts in the simulated CH₄ mixing ratios, namely considerable CH₄ depletions close to regions of negative emissions.

[22] To enforce that a posteriori emissions remain positive, we apply a ‘semiexponential’ description of the PDF:

where the a priori emissions e_apri0 are used as a constant, and the emission parameter x is optimized instead. x is set a priori to zero, and assumed to have a Gaussian PDF. We chose this ‘semiexponential’ approach in contrast to a regular exponential function as applied e.g., by Müller and Stavrakou [2005], in order to avoid an increase in the PDF for emissions higher than the a priori emissions. Furthermore, test inversions showed somewhat better convergence of the minimization algorithm for the ‘semiexponential’ function compared to a regular exponential function.

[23] This ‘semiexponential’ approach introduces a nonlinearity of the forward model operator. While the tangent linear model corresponding to equation (4) and its adjoint can be readily obtained, the nonlinearity of the whole model operator H made a major update of the optimization procedure necessary (since the conjugate gradient algorithm can handle only linear systems). Therefore, a system with an outer loop for evaluation of the nonlinear model and an inner loop for incremental optimization of the linearized model has been implemented, similar to the ECMWF operational 4DVAR system [Tremolet, 2007] (but with same model resolution in the outer and inner loop). For the incremental optimization in the inner loop the ECMWF conjugate gradient algorithm [Fisher and Courtier, 1995] is used. After each inner loop cycle, the state vector for the outer loop evaluation of the nonlinear model, x_NL(i+1), is updated by the increment derived in the inner loop, δx_LI:

A disadvantage of the new system is that currently no a posteriori uncertainty estimates can be provided, since the number of iterations applied in the inner loop (as detailed in section 3.2) is not sufficient to reach satisfactory convergence of the approximation of the a posteriori uncertainties based on the leading Eigenvectors [Meirink et al., 2008b].

3.1.2. Improved Representation of Measurements and Their Errors

[24] As in our previous studies [Bergamaschi et al., 2007; Meirink et al., 2008a] we generally include high-accuracy surface measurements in the inversions of the SCIAMACHY data to derive and correct potential biases of the SCIAMACHY retrievals. This bias correction is modeled as a second order polynomial as function of latitude and month. To minimize the impact of potential systematic errors in simulated stratospheric CH₄ mixing ratios at higher latitudes, we limit the use of the SCIAMACHY data to the latitude region between 50°S and 50°N, where comparison of modeled CH₄ in the stratosphere with observations indicates generally good agreement (see section 4.2.3).

[25] Individual SCIAMACHY pixels have an extension of 30 km (along track) times 60 km (across track), and the column-averaged CH₄ mixing ratios, XCH₄, retrieved for these pixels are averaged on a regular 1° × 1° (longitude × latitude) grid. Modeled CH₄ fields are interpolated from the model resolution to the same 1° × 1° grid, and vertically integrated using the averaging kernels specified for the SCIAMACHY retrievals to obtain the column-averaged mixing ratio, XCH_4(TM5) [see also Bergamaschi et al., 2007; Frankenberg et al., 2006].

[26] Systematic errors may arise from the difference in horizontal resolution between observations and model, particularly in the case of complex topography. We therefore introduce a surface elevation filter to ensure that the atmospheric columns seen by SCIAMACHY are well represented by the model columns. This filter requests that the difference between surface elevation of the SCIAMACHY pixel, h_{surface(SCIAMACHY)}, and that of the model grid cell, h_surface(TM5), is within a specified limit

We set dh_MAX to 250 m as default value, leading typically to the rejection of 17% of the SCIAMACHY data over land.

[27] While in the previous studies the uncertainty of the SCIAMACHY retrievals was assumed to be constant, we take now explicitly the contributions from the estimated random and systematic errors into account:

where ΔXCH_4(retrieval) is the statistical fit error (mainly related to the signal-to-noise ratio of the recorded spectra [Frankenberg et al., 2006]) of the individual SCIAMACHY pixels, averaged over the 3h assimilation time slot and 1° × 1° grid, and ΔXCH_4(STD) is the standard deviation of the pixels over this averaging period and domain. ΔXCH_{4(systematic error)} is meant to represent systematic errors not covered by our bias correction, such as e.g., remaining systematic errors due to aerosols or surface albedo [Frankenberg et al., 2006], and is set to an assumed constant value of 1%.

[28] Furthermore, we apply a new scheme to estimate the model representativeness error for surface stations. This new scheme includes estimates of the impact of the subgrid-scale variability of emissions on simulated mixing ratios for stations in the boundary layer. In the present study, however, the effect of this new scheme on retrieved emissions is relatively small, mainly because we use here only remote sites, for which emission of the local model grid cells play a minor role only. The details of the new scheme will be discussed in more detail elsewhere.

4.1.1. Reference Inversion S1

[39] Figure 2 shows column-averaged CH₄ mixing ratios, XCH₄, for year 2004 (3 month composite averages): IMAP V5.0 retrievals (Figure 2, left), and assimilated values (Figure 2, right). The small bias correction derived by TM5-4DVAR has been subtracted from the retrievals (see below). In general, the major regional patterns and their seasonal variation apparent in the retrievals are well captured in the assimilated XCH₄ fields, demonstrating the major progress of the 4DVAR system compared to the previously applied synthesis inversion [Bergamaschi et al., 2007]. Some smaller-scale enhancements in the retrievals (e.g., over Venezuela and Columbia), however, are somewhat less pronounced in the assimilated fields, partly due to the coarse model resolution applied here (6° × 4°).

[40] In Figure 3 we show the spatial and seasonal distribution of emissions and the partitioning among the 3 principal source categories which have been optimized (Figure 3, left: a priori; Figure 3, right: a posteriori). The a priori ‘remaining emissions’ vary very little with season (only the included soil sink is varying with season), and have their maximum between ∼20°N and ∼60°N. These principal patterns remain also in the a posteriori emissions. However, the ‘remaining emissions’ attributed to the Southern hemisphere are somewhat higher than in the a priori inventory. Emissions from ‘wetlands and rice’ and ‘biomass burning’ show a large seasonal variability. In particular, wetland emissions in tropical wet-dry climate regions largely follow the respective wet seasons, while wetland emissions in the continuously humid inner tropical zones persist during the whole year. This general seasonal behavior is clearly visible also in the a posteriori emissions, but a posteriori values of tropical emissions are generally higher compared to a priori values (see also Tables 6 and 7). Emissions from NH extratropical wetland regions and from rice paddies in India and South East Asia exhibit a very pronounced seasonality, however with significant differences between a priori and a posteriori distributions. The inversion leads to a significant reduction of NH extratropical wetland emissions in particular during the second quarter of the year, and an earlier end-of-season decline of rice emissions (with much lower values compared to the a priori inventories during the last quarter of the year), consistent with our previous results based on the IMAP V1.1 retrievals [Bergamaschi et al., 2007]. Emissions from biomass burning represent only a small part of the annual total emissions, but constitute a significant source in the tropics during the respective NH and SH dry seasons, opposite in phase with the tropical wetland emissions. The inversion leads to a moderate enhancement of biomass burning emissions mainly in the SH dry season (July–September).

[41] In addition to changes in the seasonality, the inversion also leads to significant changes in the spatial emission patterns (see also Figure 4 for annual mean emissions). For example, in India a posteriori emissions are most pronounced over the Ganges valley, while a priori emissions are much more homogeneously distributed over the Indian subcontinent. In Africa, a posteriori emissions are extended over a large part of the tropics (while a priori emission are dominated by very strong wetland emissions from the Congo basin), and in South America the inversion leads to larger increments in particular over Venezuela and Columbia, in the region of the Amazon delta, and along the east coast between 0 and 30°S.

4.1.2. Comparison of Reference Inversion S1 With Inversions Based on Previous SCIAMACHY IMAP V1.1 Retrievals (S2) and Based on Surface Stations Only (S3)

[42] Figure 4 shows annual total emissions for the reference inversion S1 and the sensitivity inversions S2 and S3 (including the a priori emissions which are identical for all 3 cases), and the differences between these inversions. The largest impact of the revision of the SCIAMACHY retrievals is apparent in the tropics, with significantly lower a posteriori emissions derived for the new retrievals, compared to the old IMAP V1.1 (sensitivity inversion S2). This is a direct consequence of the generally lower tropical XCH₄ values in the IMAP V5.0 retrievals (see Frankenberg et al. [2008a] and section 2.1). Summing up emissions from all tropical regions ‘TR_sam’, ‘TR_afr’, and ‘TR_asi’ (see Figure 1), yield annual total emission of 244 Tg CH₄/yr for sensitivity inversion S2, compared to 203 Tg CH₄/yr for the reference inversion S1 (see Table 7), i.e., a difference of 20%. The inversion based on NOAA stations only (sensitivity inversion S3) yields total tropical emissions close to those of reference inversion S1 (193 Tg CH₄/yr). An important difference between S1 and S3, however, is that the spatial patterns of the inversion increments in S3 largely follow those of the a priori emissions, because the remote surface stations mainly constrain the large-scale (continental scale) emissions. In contrast, the SCIAMACHY based inversions put significant constraints on the regional emission distributions over the continents. The total tropical emissions derived for inversions S1−S3 (Table 6) are very close to the first estimates presented by Frankenberg et al. [2008a], which were still based on the previous TM5-4DVAR system.

[43] While the observed XCH₄ can be reproduced rather well for the new retrievals (as shown in Figure 2), significant differences are apparent for the IMAP V1.1 (sensitivity inversion S2; see auxiliary material Figure S1). In particular the very pronounced tropical enhancements of the IMAP V1.1 data set cannot be fully reproduced in the simulations despite higher tropical emissions derived in sensitivity inversion S2. This is clearly evident in the latitudinal XCH₄ averages with differences of up to 20 ppb between IMAP V1.1 and TM5 in the tropics (Figure S1).

[44] Allowing for potential latitudinal and seasonal biases of the retrievals, both IMAP V1.1 and IMAP V5.0 can be reconciled with the surface measurements. However, using the new retrievals (reference inversion S1), the calculated bias correction is significantly smaller than that calculated for the IMAP V1.1 based inversion (sensitivity inversion S2), as illustrated in Figure 5: The latitudinal dependence of the bias corrections for S1 is typically only one third of that calculated for S2. Furthermore, a posteriori simulations of reference inversion S1 agree better with measurements at the tropical sites (RPB, CHR, SEY, ASC and SMO) than those of sensitivity inversion S2 (see auxiliary material Figure S3).

[45] This altogether clearly demonstrates the major progress achieved with the new retrievals, leading to an overall much better consistency with the high-accuracy surface measurements. This is also further supported by independent validation data (see section 4.2).

[46] The large tropical XCH₄ derived in the previous IMAP V1.1 retrievals seemed to support the hypothesis of significant CH₄ emissions from plants under aerobic conditions proposed by Keppler et al. [2006], as further evaluated by Houweling et al. [2008]. However, the study of Keppler et al. [2006] was questioned by Dueck et al. [2007], who did not find any evidence for substantial CH₄ emissions of plants under aerobic conditions. Nevertheless, later studies suggest that breakdown of detached plant material under high UV radiation may release some CH₄ [Vigano et al., 2008], but the importance of this process for the global CH₄ budget is probably rather small [Nisbet et al., 2009]. Nisbet et al. [2009] discuss also the role of transport of CH₄ dissolved in soil water through plants, which may partly explain the experiments of Keppler et al. [2006]. Such processes, however, are mainly important for wetland plants (for which they are usually covered by the emission inventories), and should play only a small role for plants living on dry land [Nisbet et al., 2009]. Our study suggests that wetlands play a dominant role for the tropical enhancements observed by SCIAMACHY, but clearly does not rule out additional sources which were not considered (e.g., additional plant emissions). Furthermore we note that the inversions do not provide direct information on the source categories. Only applying some a priori knowledge on the likely spatial and temporal distribution of the different source categories allows the attribution of the derived emissions to these categories in the inversion.

4.1.3. Sensitivity Experiments

[47] Further experiments have been performed to investigate the sensitivity of the inversion results to critical assumptions and settings of the 4DVAR system (Figure 6 and Tables 5, 6, and 7). The reduction of the horizontal correlation length from 500 to 100 km (sensitivity inversion S4) results in slightly more pronounced emission hot spots, but the overall impact on the derived spatial patterns is rather small, and the totals for the major global regions listed in Tables 6 and 7 remain very close to our reference inversion S1.

[48] In sensitivity inversion S5, only a monthly constant bias is allowed instead of the standard second order polynomial as function of latitude. The resulting tropical emissions are only slightly higher (∼5%) compared to our reference inversion, and the impact on the derived spatial emission patterns is relatively small. Furthermore, both the SCIAMACHY data and the surface observations can still be reproduced relatively well (see auxiliary material Figures S2 and S3). The calculated bias is relatively constant throughout the year (−13.2 ± 2.5 ppb). While the correction of the large biases of the previous retrievals was considered an issue of major concern with a potential significant impact on the derived emissions [Meirink et al., 2008a], the sensitivity inversion S5 suggests that the bias correction is now much less critical for the new retrievals.

[49] The use of the OH fields from Spivakovsky et al. [2000] in sensitivity inversion S6 results in significantly higher derived global total emissions (∼7%), mainly attributed to the tropics. These higher emissions are a direct consequence of the higher OH values (compared to the standard fields from TM5 used in all other inversions), which need to be balanced in the inversion. However, the total photochemical sinks in S6 are only 4% higher than in the reference inversion i.e., sources and sinks are almost in balance in this case resulting in somewhat better agreement with observations in the extratropical SH toward the end of 2004 (see auxiliary material Figure S3). The difference of ∼4% for the total photochemical sinks is smaller than the currently assumed uncertainty and potential interannual variability of global OH (∼10%) [Bousquet et al., 2005; IPCC, 2007; Krol and Lelieveld, 2003; Rigby et al., 2008]. Hence, the real uncertainties of derived emissions due to OH are likely to be even larger than the difference between sensitivity inversion S6 and the reference inversion.

[50] In sensitivity inversion S7 we also included the SCIAMACHY pixels over the ocean. The largest difference in derived emissions is apparent in tropical South America, with higher emissions at the east coast and lower emissions at the west coast. At the same time, the XCH₄ residuals (i.e., XCH_{4 TM5} − XCH_{4 SCIAMACHY}) show some tendency to positive values over the ocean, and to negative values over land close to the coastline (not shown), pointing to a potential small land-ocean bias of the SCIAMACHY retrievals, which could lead to systematic errors in the inversion. For marine air masses arriving at the east coast (trade wind zone), a bias toward lower XCH₄ values for the measurements over the ocean is compensated in the inversion by enhancing emissions along the coast, while for air masses leaving the continent at the east coast the opposite effect occurs. Further investigation of this systematic effect and its dependence on the retrieval selection criteria will be necessary to make better use also of the ocean pixels. In this study the ocean pixels are generally excluded (except for this sensitivity experiment S7).

[51] Finally, we performed a further sensitivity experiment to explore the impact of the applied a priori emission inventories. In the standard setup the spatial patterns of the a priori inventories strongly constrain the solution space, since uncertainties of the grid cell emissions are specified as percentage of the emissions. This implies that model grid cells with low a priori emissions can usually not turn into grid cells with very high emissions, while for emission hot spots in the a priori inventories a relatively small fractional increase or decrease may have a large regional impact, leading to a clear tendency of the inverse modeling system to “optimize” such hot spots. At the same time it is recognized that existing bottom-up inventories have clear deficiencies regarding their spatial and temporal emission patterns. This is in particular the case for source categories with larger spatial and temporal variability, such as wetlands. On the other hand, the almost global coverage of the SCIAMACHY observations provides strong constraints, which may allow at least for some regions to derive emissions without the use of a priori emission inventories with detailed spatial and temporal disaggregation. For sensitivity inversion S8 we constructed an alternative a priori emission inventory in which emissions are distributed homogeneously over land (except Antarctica), using an annual total of 500 Tg CH₄/yr. In addition we account for emissions over the ocean and distribute an assumed total of 17 Tg CH₄/yr over the ocean. The a priori uncertainty of these homogeneous emissions have been set to a very large value (300%) to give the inverse modeling system a large degree of freedom to optimize the spatial emission patterns. Figure 6 shows that the derived emission patterns for this sensitivity inversion agree remarkably well with those from our reference inversion S1 for several regions. This is particular the case for emissions in India, China, and South East Asia, which are well constrained by the SCIAMACHY retrievals due to the large observed XCH₄ gradients. Also for the United States a very good agreement in the spatial patterns is visible. Major differences to the reference inversion S1 are apparent in particular for Western tropical Africa (with significantly lower emissions especially over the Congo basin for sensitivity inversion S8) and over tropical South America, where higher emissions are attributed along the East coast in sensitivity inversion S8, but less in the inner part of the Amazon basin. These discrepancies point to some potential deficiencies in the spatiotemporal patterns of the applied bottom-up inventories. At the same time, however, it is clear that this ‘free inversion’ of sensitivity inversion S8 is more sensitive to potential systematic errors in the retrievals (and modeling errors), as it is not guided by the a priori information about the expected spatial and temporal patterns. The free inversion appears to be successful in deriving very low emissions over the Sahara, Australia and Greenland, consistent with our standard a priori inventory (Figure 4).

4.2.1. Ocean Transects

[53] The oceanic transects are depicted in Figure 7. The observed latitudinal gradient and its seasonal variation are reproduced very well by the model simulations, demonstrating the overall very good simulation of the marine background mixing ratios. This good agreement was expected, since all inversions use the measurements from the global NOAA monitoring sites. Therefore, the surface mixing ratios over the remote ocean are well constrained by the observations (observed and simulated mixing ratios at the NOAA sites are shown in auxiliary material Figure S3). In general, model simulations of the oceanic transects are very close to each other for the different inversions. A striking difference, however, is the clear tendency to higher values (∼10 ppb) for sensitivity inversion S2 in the Atlantic Ocean (AOC) and West Pacific (WPC) transects close to the equator, while the model simulations of the other inversions match the observations much better. These higher mixing ratios are the result of the ∼20% higher tropical emissions of sensitivity inversion S2, resulting in a less consistent picture with the oceanic transect validation data.

4.2.2. Aircraft Profiles

[54] In contrast to surface mixing ratios, the simulated vertical gradients are much less constrained by the observational data, and are largely determined by the vertical mixing of the transport model. Validation of the simulated vertical gradients is essential to investigate the model's capability to simulate the column-averaged mixing ratios observed from SCIAMACHY. Furthermore, validation of vertical mixing is very important to evaluate potential systematic errors in the derived emissions [Peters et al., 2007; Stephens et al., 2007]. In Figure 8 we present the comparison of model simulations with aircraft profiles, showing overall good agreement. Most of the profiles are over the United States, covering the altitude range between ∼1 and ∼8 km. Typical gradients are on the order of 50–100 ppb between the boundary layer and the upper free troposphere which is generally well captured by the model simulations. Model simulations for the different inversions are very close to each other, except for small differences in the boundary layer for some profiles, reflecting small differences in the derived emissions. Integrated over the whole altitude range of the measurements, the difference between model simulations and measurements translates into a maximum difference in column-averaged mixing ratios of ∼4 ppb, but is in most cases much smaller (see Figure 8).

[55] For the subtropical site HAA (21.2°N) model simulations are somewhat higher (∼10 ppb) than observations in the free troposphere (between ∼3 and ∼5 km), while they are lower at RTA (21.3°S) between ∼2 and 6.5 km. Hence the simulated latitudinal gradient across the equator in the free troposphere is smaller than observed. Despite the relatively small integrated effect (ΔXCH₄ = 1.4 ppb (HAA) and ΔXCH₄ = −3.0 ppb (RTA)) this discrepancy could be an indication that interhemispheric exchange might be too weak in TM5, consistent with the observation that simulated SF₆ mixing ratios in the SH are slightly smaller than measurements [Bergamaschi et al., 2006; Peters et al., 2004]. However, the 3D interhemispheric mixing time derived from SF₆ simulations of 10.4 months [Bergamaschi et al., 2006] is well within the range derived in the TransCom2 intercomparison [Denning et al., 1999].

4.2.3. Validation of Stratospheric CH₄ Mixing Ratios

[56] In addition to vertical profiles in the troposphere, stratospheric mixing ratios also contribute significantly to the XCH₄ derived in the SCIAMACHY retrievals, although the averaging kernel is typically smaller in the stratosphere [Frankenberg et al., 2006]. In our previous analysis of the IMAP V1.1 retrievals [Bergamaschi et al., 2007] we compared simulated CH₄ mixing ratios in the stratosphere with measurements from HALOE/CLAES [Randel et al., 1998], concluding that potential deficiencies of the modeled stratosphere are likely to have only a very small impact on simulated XCH₄ (except some potential constant offset, and except the polar vortices). We repeated this comparison with the model fields of this study, basically confirming the earlier conclusions. However, it cannot be ruled out that the HALOE/CLAES data themselves may have some bias, and an intensive intercomparison including various further remote sensing data (including ACE-FTS, MIPAS, and ground-based FTIR) identified significant biases among these different data sets [De Mazière et al., 2008]. Therefore, we use high-accuracy measurements of air samples taken from balloon soundings for further validation of our model fields. These measurements can be assumed to be unbiased, but are sparser in space and time than the remote sensing data. For that reason we use all measurements made between 1999 and 2005 for comparison with the model simulations of year 2004. Model values are sampled at the same day of the year (and time) as the measurements, but no correction was made for interannual variations or trends. Figure 9 depicts this comparison for 3 month 5° latitude bins. Over the tropics, model simulations match the observations almost perfectly. Integrated over the altitude range of the measurements the difference between model simulations and observations translate into a difference of ΔXCH_4eff = 0.0 ppb (taking into account also the SCIAMACHY averaging kernels). In the midlatitudes (40–45°N) some differences are apparent in the exact shape of the profile (in particular for the 3rd and 4th quarter of the year), but the integrated difference remains relatively small with derived ΔXCH_4eff values between 2.9 and 7.8 ppb. In contrast, large differences are visible at high latitudes (65–70°N), particularly in the first quarter of the year due to the polar vortex which is not well reproduced in the model, resulting in a ΔXCH_4eff value of 26.0 ppb. Since no SCIAMACHY measurements are available at high latitudes during this period, this mismatch would not be relevant. However, in the 2nd quarter the high latitude profile also shows a significant CH₄ depletion around 20 km with a ΔXCH_4eff value of 12.6 ppb. This feature is caused by remnants of the polar vortex, and could be reproduced by the Chemical Lagrangian Model of the Stratosphere (CLaMS) [Grooß et al., 2008]. Since the measurements are from a single profile only, taken at a different year than the TM5-4DVAR model simulations (B41, 9 June 2003; see Table 3) they represent a specific synoptic situation, which might not be representative for other years. Nevertheless, they demonstrate the potential systematic error arising from such structures related to the polar vortex. In order to minimize the impact of model deficiencies in the stratosphere at high latitudes, we therefore chose to generally exclude the high-latitude SCIAMACHY measurements in this study, and use only the data between 50°S and 50°N. An additional advantage of this restriction is that the data coverage is more continuous in time, reducing a potential seasonal bias in the inversion due to seasonally varying availability of observations. The difference in ΔXCH_4eff between midlatitudes and tropics (2.9…7.8 ppb) is close to the variation of the derived bias correction between these latitudes (see Figure 5), suggesting that significant parts of the derived bias corrections might be attributable to the deficiencies of the model stratosphere. We note, however, that our analysis is limited by the fact that we use measurements from different years than the model simulations. Hence interannual variations and trends in stratospheric CH₄ (which are not considered in our analysis) may also affect the derived ΔXCH_4eff values [Rohs et al., 2006].

4.3.1. South America

[58] Figure 10 depicts observed and simulated XCH₄, along with a posteriori emissions for inversion S1Z. Overall, the spatial patterns in observed XCH₄ and their seasonal variation are well reproduced by the model. Large emissions are attributed to the major wetland areas, in particular to the Amazon basin, but also to the Orinoco River plain (Venezuela), the Llanos de Mojos (Bolivia; influx area of Madeira river), the Pantanal region, and the Paraná River floodplain. The wetlands show a pronounced seasonality with largest emissions during the wet season (January−March). During the SH dry season (July to September) significant emissions are attributed to biomass burning between 5 and 15°S. Furthermore, anthropogenic emissions play an important role, in particular over Venezuela, Columbia, and South Eastern Brazil.

[59] Table 8 lists the emissions for major wetland areas in South America, derived for the 4 inversions run with zooming over South America (see Table 5). The use of the IMAP V5.0 retrievals (inversion S1Z) results in ∼20% lower emissions over the Amazon compared to the previous IMAP V1.1 data set (sensitivity inversion S2Z) and is rather close to the value derived for sensitivity inversion S3Z (using only the NOAA observations), hence largely reflecting the differences for the total tropical emissions as discussed in section 4.1.2 based on the global coarse resolution inversions. The ‘free inversion’, sensitivity inversion S8Z, yields only slightly lower values for the Amazon (47.5 Tg CH₄/yr) compared to S1Z (52.0 Tg CH₄/yr).

[60] Based on microwave remote sensing and an extensive set of flux measurements available from the literature, Melack et al. [2004] estimated the total emissions from wetlands in the Amazon basin to be 29.3 Tg CH₄/yr. We note, however, that the definition of the Amazon region is somewhat different in their study (and could not be exactly reproduced from the information given in their paper), adding some uncertainty to the comparison with the values derived in our study.

[61] Inversion S1Z yields 40.2 Tg CH₄/yr for the category ‘wetland and rice’ (with rice paddies playing only a minor role in this region according to our a priori inventories (<1% of ‘wetland and rice’)), i.e., ∼30% higher than the Melack et al. [2004] estimate, which, given the overall uncertainties of the bottom-up and top-down estimates (including the correct attribution to the different source categories), uncertainties from the exact area used for the comparison, and potential interannual variability of wetland emissions, is considered to be broadly consistent. Sensitivity inversion S2Z, on the other hand yields substantially higher emissions for the Amazon region (52.0 Tg CH₄/yr for category ‘wetland and rice’, i.e., 77% higher than the Melack et al. [2004] estimate).

[62] As already seen from the global coarse resolution inversions (section 4.1.2) the use of SCIAMACHY data results in significant changes in the smaller-scale emission patterns, compared to the inversion based on the NOAA data only. This is in particular noticeable for the Llanos de Mojos and the Orinoco River plain, for which inversion S1Z yields significantly higher emission than S3Z (∼40–50%) (see Table 8). The emissions attributed to ‘wetlands and rice’ in S1Z are ∼79% higher for the Llanos de Mojos and ∼36% higher for the Orinoco River plain than the estimates of Melack et al. [2004], while the corresponding emissions derived in S3Z are very close to the Melack et al. estimates. Beside the use of the observational data, the emissions derived in the inversions depend significantly on the use of the a priori inventory. This is in particular the case for the Llanos de Mojos, with the ‘free inversion’ S8Z yielding only ∼52% (2.9 Tg CH₄/yr) of the total emissions derived in inversion S1Z (5.6 Tg CH₄/yr). For the Pantanal region, the differences among the different inversions are generally rather small, and the emissions attributed to ‘wetland and rice’ are very close (∼10%) to the Melack et al. [2004] estimates. As for the Amazon, rice cultivation plays an only minor role (<1% according to our a priori inventories) in these other wetland regions.

[63] Finally, we compare simulated mixing ratios with measured aircraft profiles for the 3 South American sites Santarem, Manaus, and Fortaleza (Figure 11). The upper panel shows this comparison using all available observations over the period 2000–2005 (with model values from 2004, but taken at the same day of the year (and time), and interannual variations of the observations corrected using measurements at the NOAA site Ascension Islands (ASC) as reference [see also Meirink et al., 2008a], while the lower panel shows the comparison using only the observations of year 2004. Observed mixing ratios in the free troposphere are matched very well for all inversions which use either the new SCIAMACHY retrievals (S1Z and S8Z) or the NOAA surface measurements only (S3Z), while sensitivity inversion S2Z (based on the old IMAP V1.1 retrievals), results in higher mixing ratios. This, again, is a strong indication that sensitivity inversions S2 and S2Z overestimate the tropical emissions, leading to a significant enhancement of CH₄ mixing ratios in the free troposphere, which is not visible in the measurements. The comparison is more difficult to interpret in the boundary layer, since simulated mixing ratios strongly depend on the emissions of the local model grid cell and subgrid scale variability of emissions has to be taken into account. While simulated mixing ratios in the boundary layer are broadly consistent with observations at Santarem, they are generally higher at Manaus. Interestingly, S8Z yields the best agreement, since emissions for the Manaus grid cell are lower in this sensitivity inversion compared to inversions S1Z, S2Z, and S3Z which are ‘guided’ by large a priori emissions in this grid cell. At Fortaleza simulated mixing ratios in the boundary layer are generally higher than observed, most likely because measurements mainly sample air coming from the Atlantic Ocean, while the simulations are influenced by the emissions of the whole model grid cell, as frequently observed at sites at the land-sea border [Peters et al., 2004].

4.3.2. Africa

[64] Similar to South America, tropical Africa is characterized by strong seasonal variations of emissions from wetlands and biomass burning following the wet and dry season (Figure 12). Overall, the spatial patterns of observed XCH₄ and their seasonal variation are well reproduced in the simulations. In the third quarter of the year, however, simulations show a stronger enhancement in XCH₄ over the Congo basin than the SCIAMACHY observations. Obviously, the a posteriori emissions have been strongly guided by the very pronounced wetland emissions from the Congo basin in the bottom-up wetland inventory, while the SCIAMACHY measurements suggest a more homogeneous distribution over a larger area.

[65] The inversion leads to a significant increase of emissions around Lake Victoria, a feature which is also clearly seen in the global coarse resolution inversions (Figure 3). Furthermore, large emissions are attributed to the swamp area of the Upper Nile in Sudan, the wetlands in Southern Chad and the Niger delta, especially during the wet season (mainly third quarter of the year). Considerable emissions over the Niger delta, however, persist during the remainder of the year, although at lower intensity, derived from significant XCH₄ enhancements over this region discernible throughout the year.

[66] A prominent emission hot spot is visible around Johannesburg, South Africa, attributed according to the EDGAR database mainly to coal mining in this area. It is interesting to note that this emission peak can also be seen in the coarse resolution free inversion (sensitivity inversion S8, Figure 6), i.e., it is identified from the SCIAMACHY observations even without a priori information (but attributed a lower intensity than in reference inversions S1/S1Z).

4.3.3. Asia

[67] Observed and simulated XCH₄ over Asia are shown in Figure 13, along with the derived emissions. This region is strongly influenced by emissions from rice cultivation with a very pronounced seasonality. In particular during the third quarter of the year a prominent enhancement in XCH₄ is visible over India, Bangladesh, South East China, and Mainland Southeast Asia, largely attributed to rice emissions in this area. A major difference from the applied a priori emission inventory is the earlier end-of-season decline of the rice emissions in inversion S1Z, resulting in significantly lower rice emissions in the last quarter of the year. This change in the seasonality is also clearly visible in the various coarse resolution inversions that use the SCIAMACHY observations (see e.g., Figure 3), and had been observed already in our previous studies based on the IMAP V1.1 retrievals [Bergamaschi et al., 2007; Meirink et al., 2008a].

[68] Beside rice paddies, further anthropogenic emissions play an important role in this densely populated area, in particular emissions from ruminants and waste, and coal mining especially in China. However, all inversions attribute somewhat lower emissions (∼22–27%) to our category ‘remaining emissions’ compared to the a priori estimates (Table 7). A further striking difference compared to the a priori inventory is a change in the spatial distribution of emissions over India, which is much more concentrated over the Ganges valley for all inversions which use the SCIAMACHY observations, while the a priori emissions and the emissions derived for sensitivity inversion S3 (using the NOAA data only) exhibit a more homogeneous distribution over the whole country (see Figure 4).

[69] A further interesting feature visible in Figure 13 is the significantly enhanced XCH₄ observed over the Sichuan Basin (also called Red Basin) in China, a region with intensive rice cultivation, but also coal mining and gas production. The enhancement is noticeable throughout the whole year, and shows during the 2nd quarter of the year even the highest XCH₄ values over the entire Asian zoom region.

[70] Overall, the spatial patterns in XCH₄ observed over Asia are well reproduced in the model assimilations. An exception, however, is the tendency to lower simulated XCH₄ values over the Himalaya and Tibetan Plateau. A potential explanation could be that over these regions at very high altitudes systematic errors related to topography and model deficiencies in the stratospheric CH₄ mixing ratios have a larger impact.