Evaluation of Wetland CH4 in the JULES Land Surface Model Using Satellite Observations

Wetlands are the largest natural source of methane. The ability to model the emissions of methane from natural wetlands accurately is critical to our understanding of the global methane budget and how it may change under future climate scenarios. The simulation of wetland methane emissions involves a complicated system of meteorological drivers coupled to hydrological and biogeochemical processes. The Joint UK Land Environment Simulator (JULES) is a process-based land surface model that 5 underpins the UK Earth System Model and is capable of generating estimates of wetland methane emissions. In this study we use GOSAT satellite observations of atmospheric methane along with the TOMCAT global 3-D chemistry transport model to evaluate the performance of JULES in reproducing the seasonal cycle of methane over a wide range of tropical wetlands. By using an ensemble of JULES simulations with differing input data and process configurations, we investigate the relative importance of the meteorological driving data, the vegetation, the temperature dependency of wetland methane 10 production and the wetland extent. We find that JULES typically performs well in replicating the observed methane seasonal cycle. We calculate correlation coefficients to the observed seasonal cycle of between 0.58 to 0.88 for most regions, however the seasonal cycle amplitude is typically underestimated (by between 1.8 ppb and 19.5 ppb). This level of performance is comparable to that typically provided by state-of-the-art data-driven wetland CH4 emission inventories. The meteorological driving data is found to be the most significant factor in determining the ensemble performance, with temperature dependency 15 and vegetation having moderate effects. We find that neither wetland extent configuration out-performs the other but this does lead to poor performance in some regions. 1 https://doi.org/10.5194/bg-2022-2 Preprint. Discussion started: 12 January 2022 c © Author(s) 2022. CC BY 4.0 License.

each grid box the statistical distribution of topographic index (Marthews et al., 2015) is combined with the mean water table depth. This enables the simulation of a sub-grid water table distribution and therefore the extent of wetland in the grid box.

JULES Wetland CH 4 Emissions
The JULES land surface model calculates methane wetland emissions F CH4 , from three key factors, namely the amount of available substrate carbon, the temperature and the inundated area below the water table (Gedney et al., 2004;Clark et al., 2011): e −γz · C si,z · Q 10 (T soil ) 0.1(T soil −T0) (1) k CH4 is a dimensionless scaling constant (7.41x10 −12 ) for wetland CH 4 emissions when soil carbon is taken as the substrate for CH 4 emissions. The wetland fraction (i.e. the proportion of a grid cell where the water table is at/above the surface, and below a threshold indicative of significant flow (Gedney et al., 2004)) is denoted by f w . z is the depth of soil column (in m), i is the soil carbon pool, κ i (s -1 ) is the specific respiration rate of each pool (Table 8 of  (2x3x2x2). In order to identify ensemble members, we assign to each member a 4-digit ID as shown in Figure 1. Thus the ensemble member using WFDEI meteorology data (2), using dynamic vegetation (3), with the lower temperature dependency (1) and with the original JULES wetland extent (1) is ensemble member 2311.
In a post-processing step, the time series of annual wetland emissions of each ensemble member is separately scaled to give annual emissions of 180 Tg CH 4 yr −1 for the year 2000 (Saunois et al., 2016), as described in Comyn-Platt et al. (2018). 110 Maps of the CH 4 emissions for each ensemble member are presented in Figure 2 for August 2011. Clear differences are observed relating to the different ensemble configurations, including: substantial differences between ERA-Interim and WFDEIbased ensemble members with the magnitude of the emissions in the WFDEI members visibly smaller; and large spatial differences based on the Default vs SWAMPS wetland extent masking, with SWAMPS significantly reducing the wetland areas and concentrating the emissions, particularly removing the widespread but low emissions found more generally in the 115 Default members.

Driving Data: ERA-Interim vs WFDEI
Meteorological forcing data is used to drive the JULES land surface model. The meteorological parameters used in this study are: air temperature, surface pressure, precipitation, short and long-wave radiation, relative humidity and wind speed. In the ensemble we use two sources for the meteorological data, ERA-Interim and WFDEI.

120
The ERA-Interim Reanalysis (Dee et al., 2011) is a widely used global atmospheric reanalysis product produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). The WATCH Forcing Data ERA-Interim (WFDEI) is based on the ERA-Interim Reanalysis data but includes the modifications as outlined in (Weedon et al., 2014). Namely, interpolation to a 0.5 • x 0.5 • resolution, a sequential elevation correction and a monthly bias correction based on observations. As indicated in Equation 1, the CH 4 emission is strongly dependent on the temperature of the soil. This temperature dependency of methanogenesis is generally parameterised using a Q 10 value that approximates the Arrhenius equation. As discussed in Gedney et al. (2004), the approach that JULES takes due to applying this approximation globally over a wide temperature range is to use an effective or generalised Q 10 that fits the form of the Arrhenius equation exactly (Equation 2).
Q 10 (T ) = Q 10 (T 0 ) T0/T (2) 145 2.3.4 Wetland Extent: JULES vs JULES with SWAMPS mask JULES generates wetland extent following the TOPMODEL approach as outlined in Section 2.1. As accurate wetland extent is one of the largest challenges in relation to modelling wetland emissions of methane (Saunois et al., 2020), the ensemble also provides an alternative observationally-constrained wetland extent. In this instance, the JULES wetland area is simply masked by the SWAMPS dataset (Schroeder et al., 2015), meaning that any wetland extent that is inconsistent with the SWAMPS 150 observations is disregarded.

GOSAT CH 4 Observations
The primary observational dataset that we use for evaluation of the JULES CH 4 is the University of Leicester GOSAT Proxy XCH 4 (Parker et al., 2011(Parker et al., , 2020a. The GOSAT satellite, launched in 2009 by the Japanese Space Agency, was the first 155 dedicated greenhouse gas observing satellite (Kuze et al., 2009). This data was recently used (Parker et al., 2020b) to evaluate the WetCHARTs CH 4 emission database (Bloom et al., 2017a) and has previously been used for many wetland-related studies measurements over land and also over the ocean in cases where sun-glint reflection allows. The GOSAT Proxy XCH 4 retrieval provides around 15k-25k observations over land each month and, after changes to the sun-glint sampling in 2015, a comparable number over the ocean. For a full description of the data, including evaluation and validation, see Parker et al. (2020a).

TOMCAT Atmospheric CH 4 Simulations
In order to link surface CH 4 emissions as generated by JULES with atmospheric observations as measured by GOSAT, it is 165 necessary to run the emissions through a global chemistry transport model.
In this study, we use the TOMCAT 3-D model (Chipperfield, 2006) For the wetland CH 4 fluxes, the emissions generated for each of the 24 JULES ensemble members (Section 2.3) are assigned to individual tracers. These tracers each contain the wetland and non-wetland CH 4 fluxes and therefore an additional tracer containing no wetland emissions is used as a reference to remove the non-wetland effects.

Evaluation of JULES Wetland CH Seasonal Cycle
In this section we evaluate the seasonal cycle of the wetland CH 4 emissions generated from the ensemble of JULES simulations against atmospheric satellite observations. We perform the same analysis on the JULES wetland emission datasets as was used for the evaluation of the WetCHARTs emission dataset Parker et al. (2020b), thereby enabling comparison of results and conclusions.

180
The evaluation is performed over 7 large-scale areas (Global, Northern Hemisphere, Southern Hemisphere, 60 • S-60 • N, Tropics, North Tropics, South Tropics) as well as 16 specific wetland areas as indicated in Figure 3.
To calculate the XCH 4 seasonal cycle, we apply the NOAA CurveFitting routine (Thoning et al., 1989; NOAA) to the GOSAT CH 4 observations as well as the TOMCAT model simulations for each of the JULES wetland emission ensemble members. To determine the wetland-specific signal, we apply the same technique to the TOMCAT tracer that contained no 185 wetland emissions and subtract that signal. This results ( Figure 4) in a wetland XCH 4 seasonal cycle for each region from GOSAT and from each of the model ensemble members. The observed (GOSAT) seasonal cycle magnitude varies significantly between regions (e.g. contrast the Pantanal to East Amazon) and can also be seen to vary strongly between years for the same region (e.g. contrast S.E. Asia for 2010 to 2017). Qualitatively, the ensemble of JULES-based simulations are not dissimilar to the observations, however the simulated seasonal cycles are typically weaker in magnitude than the observations. Although the 190 ensemble spread can be large in some regions (e.g. Indo-Gangetic), the regions with a strong observed seasonal cycle typically produce a box-and-whisker plot for the distribution of the model-GOSAT wetland XCH 4 seasonal cycle amplitude differences (∆A), combining all ensemble members and all years for each region. Further, the box is coloured according to the mean value of the correlation coefficient (R cycle ) between the GOSAT and model seasonal cycles.
Globally we find that the JULES ensembles underestimate the XCH 4 wetland seasonal cycle amplitude by approximately 6.6 ppb (quartiles: 5.6 ppb -7.9 ppb) with a correlation coefficient of 0.85. When considering the northern and southern When focusing on specific wetland regions we find that the evaluation is varied and performance is very region-dependent.
For example, although R cycle = 0.83 for the Pantanal region suggesting that the phase of the seasonal cycle is reasonably well-205 captured, the seasonal cycle amplitude is significantly underestimated (∆A = -19.5 ppb) and furthermore, this underestimation has a very large spread between ensemble members and years (ranging from -42.4 ppb to -5.7 ppb). In contrast, the Paraná region has a slightly poorer R cycle (0.70) and slightly better ∆A (-15.3 ppb) but with significantly smaller spread between ensemble members (-21.8 ppb to -7.9 ppb). values: 0.23, 0.31, 0.01 respectively). We revisit these regions in Section 5 and perform a more detailed evaluation in order to explain the poor performance here.
Despite these few poorly-performing regions, JULES shows reasonable-to-good performance overall in representing the observed seasonal cycle. It is informative here to judge the performance of JULES against the current state-of-the-art wetland emission dataset, WetCHARTs. In Parker et al. (2020b) we evaluated the performance of WetCHARTs in the same way as we 220 evaluate JULES here so a direct comparison of the ability to model the observed seasonal cycle can be made. We reproduce analysis and our previous WetCHARTs analysis are that: for WetCHARTs the ensemble spread (σ A ) in the Congo is far larger than for JULES, while R cycle is reasonable compared to poor for JULES; although the biases for Southern African are very similar, R cycle for WetCHARTs is reasonable while again, it is poor for JULES. The above all suggests that the performance of JULES is very comparable to that of the observation-driven WetCHARTs emissions, albeit with some differences in key regions.

Attribution of Performance to Specific Configuration Choices
A significant feature apparent in the analysis so far is that the spread in ∆A across the ensemble members is typically large, often in excess of 20 ppb between the minimum/maximum ∆A values. Understanding which ensemble members perform well (and poorly) is an important step towards identifying which parameters and processes are driving the discrepancies to observations. To investigate this, we calculate the change in two metrics, the correlation coefficient between the GOSAT and 235 modelled wetland seasonal cycle (R cycle ) and the standard deviation of the seasonal cycle amplitude(σ A ), above the minimum value for that metric. We denote these changes as ∆R cycle and ∆σ A . We do this for the different ensemble parameter groupings (meteorological driving data, vegetation, temperature dependency, wetland extent) individually and hold the other parameters constant. To elaborate, out of the 24 ensemble members, the ensemble is split into (2x3x2x2) groupings (see Section 2.3 and Figure 1). Using the meteorological driving data as an example, there are 12 different configurations that use ERA-Interim 240 and 12 configurations that use WFDEI. We compare the statistics for the performance of these configurations for pairs of configurations where the only difference is which meteorological driving data is used and calculate the change in the metric between the highest and lowest values. We then do likewise for the other parameters (vegetation, temperature dependency and wetland extent). Note that for vegetation there are 3 configuration possibilities (phenology, fixed-TRIFFID and dynamic-TRIFFID) and this results in triplets rather than pairs of members that are compared.

245
The results of this analysis are presented in Figure 6 with all regions collated into a single set of results. The ensemble members driven by WFDEI consistently out-perform the ERA-Interim based members with both a significantly higher ∆R cycle (a median increase of 0.12 with quartile values of 0.02 and 0.24) and significantly lower ∆σ A (a median decrease of 0.53 ppb).
For the vegetation configurations, the results are more mixed without any single configuration being substantially better than the rest but the phenology-based configurations do exhibit a slightly higher ∆R cycle (0.03, 0.01 and 0.01 for Phenology,

250
TRIFFID-Fixed and TRIFFID-Dynamic) and lower ∆σ A than the TRIFFID configurations suggesting that overall it performs slightly better. However the significant overlapping spread here suggests that these results are much more region-dependent.
For temperature dependency the lower Q 10 value (3.7) performs better than the higher Q 10 value (5.0) but again, the spread in both ∆R cycle (e.g. 75th-percentile values of 0.15 and 0.09 for a Q 10 of 3.7 and 5.0 respectively) and ∆σ A (75th-percentile values of 0.53 ppb and 0.72 ppb) are high, suggesting a large region-to-region variability (consistent with Turetsky et al.

255
(2014) who measured a wide range of Q 10 values across different wetland types). Finally the choice of wetland extent between JULES and SWAMPS is found to make little difference with SWAMPS very slightly increasing the correlation and decreasing the standard deviation over the original JULES. We discuss this aspect in more detail below.
Overall we can conclude that the source of the meteorological driving data (ERA-Interim vs WFDEI) is the most significant factor in how well JULES is able to reproduce the wetland seasonal cycle with WFDEI performing (almost) unanimously 260 better than ERA-Interim over the 16 wetland regions that we consider. This highlights the importance of the meteorological data and the value in the interpolation and bias-correction that is performed as part of the WFDEI methodology (Weedon  et al., 2014). The choice of the vegetation and temperature dependency configurations were found to improve (or worsen) the representation of the seasonal cycle depending on their choice but this was found to be much more region-dependent with a greater spread. Perhaps surprisingly, the choice of wetland extent configuration was found to have less of an effect when 265 collating results across all regions. However, an important point to make here is that we are solely comparing the performance between two extent configurations and find that neither is significantly better than the other. This does not preclude extent itself from being important. It should also be remembered here that for the majority of regions, R cycle already shows a good correlation to observations for the majority of ensemble members (see Figure 5) implying that the extent is already sufficiently well-reproduced in these regions. In the following section we focus on case studies over the 3 poorly-performing African 270 wetland regions and demonstrate the significance of poorly reproducing wetland extent in these regions. 14 https://doi.org/10.5194/bg-2022-2 Preprint. Discussion started: 12 January 2022 c Author(s) 2022. CC BY 4.0 License.
We now investigate three significant African wetland regions (the Sudd, the Congo and Southern Africa) in detail and evaluate the performance of the JULES wetland methane emission estimates in these regions. Figure 7 presents the same analysis as performed in Figure 6 but broken down individually for the three African wetland 275 regions. Overall, the same general pattern that we find for all regions persists individually for these regions but with some interesting exceptions.
For the meteorological data, the WFDEI ensemble members show improved ∆R cycle (0.26, 0.12 and 0.46 medians for Sudd, Congo and Southern Africa respectively) with ERA-Interim worsening the ∆σ A value (by 0.60 ppb, 0.50 ppb and 1.35 ppb respectively). As a reminder here, a value of 0 (as is the case for the change in ERA-Interim), indicates that the selection 280 consistently performs the same (be that the lowest correlation coefficient or the smallest standard deviation) in relation to the other possible selection(s).
For the vegetation configuration, as found across all regions combined, there is not a distinctly better configuration. The phenology-based ensemble members perform best for the Sudd, with the highest ∆R cycle (0.12) and lowest ∆σ A (0.0 ppb, indicating that it consistently out-performs the other configurations). However, for the Congo region there seems to be very 285 little improvement, or indeed variability, between the 3 different vegetation options. This is largely expected due to low variability/seasonality in the tropical broadleaf vegetation. For the Southern Africa region, the dynamic TRIFFID configuration performs slightly worse than the others (∆σ A increasing by 0.28 ppb) but the performance of phenology and fixed-TRIFFID is hard to differentiate.
The temperature dependency exhibits very strong regional behaviour. For example, for Southern Africa the temperature 290 dependency can improve ∆R cycle by 0.75 for the lower Q 10 value versus the higher value and at the same time, the higher Q 10 value can worsen the ∆σ A by over 1.6 ppb. In contrast, for the Congo the higher Q 10 value improves ∆R cycle by 0.32 with the lower Q 10 value worsening ∆σ A by over 0.6 ppb. While this does not leave us with a clear indication that one Q 10 value is universally better than the other, it does highlight the potential for significantly improving the ∆R cycle by selection of appropriate region-specific values. It should be noted that while some studies (e.g. Turetsky et al. (2014)  as summarised in Figure 5 and the reason for the poor R cycle value is that JULES appears to be out of phase with observations. This all suggests a fundamental lack of variability is being generated by JULES, with wetland extent an obvious parameter to evaluate in greater detail.
We compare in Figure 9 the JULES wetland fraction for these three regions against that generated using JULES-CaMa-Flood simulations which are capable of explicitly representing river and floodplain water dynamics and hence incorporate fluvial   (Figure 9). In Figure 10 we present CH 4 emission maps over the Sudd from two of the JULES ensemble members 365 (one with the default wetland extent (Fig. 10a) and one with the additional SWAMPS mask (Fig. 10b)). Furthermore, we also show the CH 4 emissions derived from a GEOS-Chem flux inversion (Fig. 10e) and from the WetCHARTs ensemble (Fig. 10f).
In addition to the CH 4 , we show the JULES wetland fraction (Fig. 10c), MODIS imagery (Fig. 10d), the JULES-CaMa-Flood wetland fraction (Fig. 10g) and the WAD2M wetland fraction (Fig. 10h). By using this wide range of information we are able to more confidently assess and evaluate the performance of JULES in this region and determine whether wetland area (and with additional wetlands in the Machar marshes on the border with Ethiopia (e.g. Fig. 10h). When using the SWAMPS masking of the JULES wetland extent, slightly more emissions are generated in the correct location due to the removal of the majority of the spurious Ethiopian emissions but emissions remain significantly too small in both area and magnitude.
As further confirmation for where CH 4 emissions should be present in this region, CH 4 observations from TROPOMI are used, allowing us to map CH 4 in the region. Figure 11 (bottom) shows the enhancement in the TROPOMI data over the Sudd The reason that JULES fails to produce these wetlands is largely due to the topography in this region. Rainfall here occurs 385 in the Ethiopian Highlands, flowing downhill to maintain the Sudd wetlands. Without the addition of a river routing and inundation mechanism within the JULES simulations, wetlands are instead created erroneously in the Ethiopian Highlands (as indicated in Figure 10a).
It is important to highlight here that the JULES-CaMa-Flood simulations (Fig. 10g) are capable of producing wetlands in the correct location and as such, future developments within JULES that incorporate some of the CaMa-Flood capabilities for 390 river routing and fluvial inundation would be expected to significantly improve the ability of JULES to successful reproduce the correct temporal and spatial distribution of wetlands, and ultimately CH 4 emissions, over the Sudd region.

The Congo
The second region that we focus on is the Congo. The Congo Basin contains flooded forests and peatlands, known as the  Furthermore, the observed seasonality exhibits more complex behaviour with double-peaks in some (but not all) years, highlighting the complex hydrology in this region.
405 Figure 9 (middle) shows that the seasonality produced by JULES-CaMa-Flood is in good agreement with that from JULES but with significantly lower average inundation ∼ 0.02 vs ∼ 0.10. When applying the SWAMPS masking to JULES, the average inundation is reduced (to ∼ 0.05) with the seasonality is largely lost.  By comparing against the additional datasets we see why the Congo remains a difficult area to model. The default JULES simulations lead to groundwater inundation of the entire Congo Basin (Fig. 12c), leading to fairly low widespread emissions 410 whereas the JULES simulations with the extent masked by SWAMPS produce significantly more emissions (Fig. 12b), more tightly constrained to the area in the vicinity of the river system, albeit still very widespread. These latter emissions with the SWAMPS mask do appear to be in more reasonable agreement spatially with the CH 4 emissions from both the GEOS-Chem inversion (Fig. 12e) and from WetCHARTs (Fig. 12f). Some care needs to be taken here as WetCHARTs itself is used as the prior for the GEOS-Chem flux inversion so the two should not be considered fully independent, and the major difference between ( Fig. 12g) produce wetland extent close to the river which is largely missing from the standard JULES simulations. MODIS imagery (Fig. 12d) agrees with the JULES-CaMa-Flood simulations and does not show clear signs of inundation over this area except directly at the rivers. However, this may be misleading due to the dense tree canopy in this area. Indeed, wetlands (i.e. swamps and flooded forest) in the Congo can exist in relatively hilly areas, not directly fed by river flooding, but more due to 420 local precipitation or groundwater. The pattern of wetland fraction from WAD2M (Fig. 12h), employing microwave observations that can partially penetrate the canopy layer, does suggest that there is a combination of both groundwater inundation and fluvial inundation. This does highlight the challenge in simulating such flooded forests where evaluation can be challenging and observations lacking. Additionally, dense cloud-cover in this region results in very few successful CH 4 retrievals from satellites (both GOSAT and TROPOMI), again reducing our capability to accurately evaluate model performance in this region.

425
The Congo remains one of the most significant global wetland regions but equally remains one of the most challenging to simulate and evaluate, with a significant uncertainty in the CH 4 emissions. Ongoing model development (Gedney et al., 2019) related to inclusion of methane emissions from trees in flooded areas (Pangala et al., 2017;Gauci et al., 2022) as well as improvements in the soil ancillary data to represent oxisol and ultisol soils in this area are expected to improve our ability to more accurately model the CH 4 emissions from the Congo in future work.

Southern Africa
The final region that we evaluate is Southern Africa, primarily focusing on the Zambezi River Basin in Zambia and Angola but also including parts of Namibia, Botswana, Zimbabwe, Mozambique and the Democratic Republic of Congo. Wetlands in this region are primarily swampland and seasonally inundated savannah/grasslands (Zimba et al., 2018;Lowman et al., 2018). The region also encompasses the Okavango Delta in northern Botswana (McCarthy, 2006;Wolski et al., 2012).

435
The values of R cycle for this region are found to vary significantly, ranging from reasonable positive correlations (R cycle = 0.67) to similar large negative correlations (R cycle = -0.68). This region is one in particular where the WFDEI-based ensemble members perform much better than the ERA-Interim members as shown in Figure 7. across the ensemble explains why the average correlation is found to be very poor (Fig. 5).
When comparing the wetland extent from the best performing ensemble members to that produced by JULES-CaMa-Flood ( Fig. 9 (bottom)) we find a good agreement in the seasonality between all three. However, in terms of the magnitude, the 445 average groundwater inundation for the default JULES configuration is augmented by approximately 50% in the simulation with JULES-CaMa-Flood, with the SWAMPS-masked inundation in contrast being far too low. Figure 13 clarifies that although the seasonality is reasonable, the spatial distribution is again, incorrect. The default JULES wetland extent for this region places wetlands in northern Zambia and southern Democratic Republic of Congo. In contrast, the SWAMPS masking places the wetlands primarily along the Zambezi and Bangweulu wetlands in the west and north-east of Zambia respectively. The

Conclusions
Overall we find that existing configurations of JULES can simulate wetland CH 4 emissions comparable in performance to those generated via state-of-the-art data-driven emission inventories such as WetCHARTs. are the configurations that use the WFDEI meteorological driving, the lower Q10 value and phenological vegetation as these were shown to provide the best result over this region (Fig: boxplot).
The wetland methane seasonal cycle amplitude from JULES is typically underestimated compared to observations by between 1.8 ppb and 19.5 ppb across the different wetland regions examined. However, the correlation coefficient to the observed 465 seasonal cycle is typically reasonable-to-good for most wetland regions (r = 0.58 to 0.88) although several regions do exhibit a poor correlation (r < 0.31) and these are explored in more detail.
Across the JULES ensemble, there are significant differences between ensemble members with the WFDEI driving data giving universally better performance than ERA-Interim. This highlights the vital role that the meteorological driving input data has on determining the wetland response within the model and emphasises the benefits of bias-correcting to observations 470 as done in the generation of the WFDEI data.
We find that the specific vegetation configuration of the ensemble member has a small effect on the performance (with Phenology typically performing better than either TRIFFID configuration) suggesting that there are potential improvements to consider when using a dynamic vegetation model such as TRIFFID. The effect of the temperature dependency is moderate, with the lower value (Q 10 = 3.7) generally performing best but there are some important regional differences where the effect is 475 much larger. We recommend further investigation into the variability in Q 10 across different ecosystems and the consequences that has for CH 4 emissions.
Neither choice of wetland extent, either the original JULES as is or masked with SWAMPS data, tends to perform better and both clearly have significant deficiencies. We find that a simple masking of the JULES wetland extent with the observed SWAMPS wetland mask is not sufficient to reproduce the wetland seasonal cycle in key areas and instead, fundamental changes 480 to the way the inundation is modelled are necessary in some regions, particularly those regions where fluvial inundation plays a significant role in the hydrology. This is demonstrated by the significant improvement in the agreement to multiple observation-based wetland and CH 4 datasets when using the JULES-CaMa-Flood wetland extent which incorporates fluvial inundation compared to the original (interfluvial) JULES data over key African wetland regions. Incorporating such fluvial inundation changes into JULES is expected to significantly improve the ability of JULES to better represent the wetland extent 485 and subsequently, produce more accurate CH 4 emissions.
Despite our analysis pointing towards the potential for significant improvements in key regions, the Congo wetland region in particular remains both challenging to model and to evaluate, highlighting the need for further study and additional groundbased observations that are less affected by the extensive cloud coverage of the region. Improved mapping of the wetland extent (by both groundwater and fluvial inundation) as well as measurements of the temperature dependency of the CH 4 emissions 490 would help in further constraining the CH 4 emissions from this region.
Finally, ongoing developments within JULES, such as the chimney venting of CH 4 by vegetation and the improved representation of soil properties, are expected to lead to additional improvements in the model. With these additions coupled to an improved representation of wetland extent and variability through more advanced hydrological modelling, we greatly improve our capability to model the emission of CH 4 from tropical wetlands both historically and under a changing future climate.

495
Sentinel-5 Precursor ground-segment development has been funded by the ESA and with national contributions from the Netherlands, Germany, and Belgium. The generation of the TROPOMI methane product by University of Bremen has been funded by ESA (GHG-CCI+ project) and by the State and the University of Bremen We thank the Japanese Aerospace Exploration Agency, National Institute for Environmental Studies, and the Ministry of Environment for 530 the GOSAT data and their continuous support as part of the Joint Research Agreement.

Appendix A
In the main text (Section 4) we make reference to previous work (Parker et al., 2020b) we have undertaken to evaluate the WetCHARTs data-driven emission inventory (Bloom et al., 2017a) using a similar methodology as used in this study. That allows a direct comparison between the performance of the JULES wetland CH 4 emissions for these regions to the WetCHARTs 535 performance. Figure A1 reproduces Figure Figure A2 shows the correlation coefficient between the different ensemble members and the observed wetland CH 4 seasonal cycle for the Sudd region. The majority of the ensemble members correlate strongly to each other (r > 0.9) but poorly to the observed seasonal cycle (r < 0.2). The set of ensemble members that correlate best to observations (members 2121 and 2122 -WFDEI meteorology, Phenology vegetation and high Q 10 value) correlates the least to the remaining ensemble members, 540 suggesting a significant difference in the characteristics of these few ensemble members. This is discussed in the main text in Section 5.2.