Methane retrieved from TROPOMI: improvement of the data product and validation of the first two years of measurements

The TROPOspheric Monitoring Instrument (TROPOMI) aboard of the Sentinel 5 Precursor (S5-P) satellite provides methane (CH4) measurements with high accuracy and exceptional temporal and spatial resolution. TROPOMI CH4 measurements are highly valuable to constrain emissions inventories and for trend analysis, with strict requirements on the data quality. This study describes the improvements that we have implemented to retrieve CH4 from TROPOMI using the RemoTeC full-physics algorithm. The updated TROPOMI CH4 product features a constant regularization scheme of the inversion that 5 stabilizes the retrieval and yields less scatter in the data, and includes a higher resolution surface altitude database. We have tested the impact of three state-of-the-art molecular spectroscopic databases (HITRAN 2008, HITRAN 2016 and Scientific Exploitation of Operational Missions – Improved Atmospheric Spectroscopy Databases SEOM-IAS) and found that SEOMIAS provides the best fitting results. The most relevant update in the TROPOMI XCH4 data product is the implementation of a posteriori correction fully independent of any reference data that is more accurate and corrects for the underestimation 10 at low surface albedo scenes and the overestimation at high surface albedo scenes. After applying the correction, the albedo dependence is removed to a large extent in the TROPOMI versus satellite (Greenhouse gases Observing SATellite – GOSAT) and TROPOMI versus ground-based observations (Total Carbon Column Observing Network – TCCON) comparison, which is an independent verification of the correction scheme. We validate two years of TROPOMI CH4 data that shows the good agreement of the updated TROPOMI CH4 with TCCON (−3.4 ± 5.6 ppb) and GOSAT (−10.3 ± 16.8 pbb) (mean bias and 15 standard deviation). Low and high albedo scenes as well as snow covered scenes are the most challenging for the CH4 re1 https://doi.org/10.5194/amt-2020-281 Preprint. Discussion started: 5 August 2020 c © Author(s) 2020. CC BY 4.0 License.

In this study we present the improvements that we have developed to retrieve CH 4 from TROPOMI measurements using the full-physics approach, and we validate the TROPOMI CH 4 product with satellite and ground-based measurements. Section 2 describes the data and analysis methods that we use and Sect. 3 focuses on the main improvements related to the regularization scheme of the inversion, the choice of the spectroscopic database for the absorption cross sections, the surface elevation database and a posteriori bias correction derived using only TROPOMI data. Finally, Sect. 4 and Sect. 5 show a detailed 5 validation of the improved TROPOMI CH 4 data. The study concludes in Sect. 6 with an outlook for future steps regarding CH 4 data retrieved from TROPOMI.
2 Retrieval algorithm and validation data sets 2.1 TROPOMI CH 4 retrieval algorithm The methane total column-average dry-air mole fraction (XCH 4 ) is retrieved from TROPOMI measurements of sunlight 10 backscattered by Earth's surface and atmosphere in the near-infrared (NIR) and shortwave-infrared (SWIR) spectral bands with the retrieval algorithm RemoTeC. This algorithm has been extensively used to retrieve both CO 2 and CH 4 from measurements of OCO-2 and GOSAT (e.g. Wu et al. (2018); Butz et al. (2011)) and it is the Sentinel 5-P and Sentinel 5 operational algorithm for CH 4 ; Landgraf et al. (2019)).
The S5P RemoTeC algorithm uses the full-physics approach that simultaneously retrieves the amount of atmospheric CH 4 15 and the physical scattering properties of the atmosphere. The algorithm aims at inferring the state vector x that contains all the parameters to be retrieved from the radiance measurements y in the SWIR (2305-2385 nm) and NIR (757-774 nm) spectral bands, where the forward model F simulates the TROPOMI measurements, Here, y and F are the measurement noise error and the forward model error respectively. The forward model employs the 20 LINTRAN V2.0 radiative transfer model in its scalar approximation to simulate atmospheric light scattering and absorption in a plane parallel atmosphere (Schepers et al. (2014); Landgraf et al. (2001)). Accurate modelling of absorption by molecules relies on spectroscopic databases, which provide the absorption cross-section of the target absorber CH 4 as well as of the interfering gases CO, H 2 O and O 2 .
The inversion to estimate the state vector x requires the use of regularization methods, as measurements typically do not 25 contain sufficient information to retrieve every state vector element independently. The RemoTeC retrieval algorithm uses the Philips-Tikhonov regularization scheme, which aims to find the state vector by solving the minimization problem where || · || describes the Euclidian norm, S y is the measurement error covariance matrix that contains the noise estimate, γ is the regularization parameter, W is a diagonal weighting matrix that renders the side constraint dimensionless and ensures that only the target absorber CH 4 and the scattering parameters contribute to its norm (Hu et al., 2016), and x a is the a priori state vector.
The retrieval state vector contains CH 4 partial sub-column number densities at 12 equidistant pressure layers. The total 5 column of the interfering non-target absorbers CO and H 2 O are also retrieved, together with the effective aerosol total column, size and height parameter of the aerosol power law distribution. A Lambertian surface albedo in both NIR and SWIR spectral range together with its first order spectral dependence is also retrieved, as well as spectral shift and fluorescence in the NIR band.
The TROPOMI CH 4 data product is given in the form of total column-averaged dry-air mole fraction, XCH 4 . It is calcu-10 lated from the methane vertical subcolumn elements x i and the dry air column V air,dry calculated with meteorology input from ECMWF (European Centre for Medium-Range Weather Forecasts) analysis product and surface topography from a high resolution database: The precision σ XCH4 is given as the standard deviation of the retrieval noise, which follows from the error covariance matrix 15 S x that describes the effect of the measurement noise on the retrieval (Hu et al., 2016):

TCCON reference dataset
To validate XCH 4 retrieved from TROPOMI we use independent ground-based XCH 4 measurements from the Total Carbon Column Observing Network (TCCON) (Wunch et al., 2011a) as a reference (data version GGG2014). Table 1  which minimizes the errors due to assumptions on the vertical CH 4 distribution used to correct for differences between the surface elevation of TROPOMI particular pixels and the ground altitude at the TCCON sites. paired collocations at each station, we compute the mean bias defined as the mean of the difference of individual collocations (∆CH 4 ) and its standard deviation (σ) as a measure of the spread in the data. We then compute the average of the station biases (b) and its standard deviation (σ(b)) as a measure of the station-to-station variability. The station-to-station variability is an important diagnostic parameter as it indicates regional biases in our data, and it might be used as an overall uncertainty estimate. We use the GOSAT proxy XCH 4 data product produced at SRON in the context of the ESA GreenHouse Gas Climate Change Initiative (GHG CCI) project (Buchwitz et al., 2019(Buchwitz et al., , 2017. This XCH 4 product is retrieved using the RemoTeC/proxy retrieval algorithm. The proxy approach (Frankenberg et al., 2005) infers a CO 2 and CH 4 total column from observations at 1.6 µm ignoring any atmospheric scattering in the retrieval. Substantially, the XCH 4 product is derived by where the column-average dry-air mole fraction XCO 2 mod is taken from the Carbon Tracker data assimilation system, and V CH4 and V CO2 are the vertical column densities. This approach assumes that light path modifications due to scattering in the atmosphere are the same for the target absorber (i.e. CH 4 ) and the proxy absorber CO 2 , whose prior is assumed to be known with high accuracy.
The proxy approach cannot be applied to retrieve XCH 4 from TROPOMI since it does not cover the 1.6 µm CH 4 and CO 2 15 absorption bands. Schepers et al. (2012) compared both the physics and proxy retrievals applied to GOSAT measurements to retrieve XCH 4 and concluded that both retrievals performed similarly when validating the retrieved XCH 4 with ground-based TCCON measurements.

TROPOMI CH 4 retrieval updates
Our updated TROPOMI XCH 4 product corresponds to the S5P-RemoTeC algorithm version 1.3.0 that will be suggested for 20 use in the operational processing (Hu et al. (2016), data product 1.2.0) in the next processor update. The updates to the S5P-RemoTeC retrieval algorithm relate to the regularization scheme, the selection of the spectroscopic database, the implementation of a higher resolution digital elevation map (DEM) for surface altitude and a more sophisticated a posteriori correction for the albedo dependence. In this section we present the updates and quantify the improvements, and we use the comparison with TCCON and GOSAT as a benchmark to test the performance of the retrieval after implementing the updates. 3.1 Regularization scheme Hu et al. (2016) determined the regularization parameter γ in the inversion (Eq. 2) using the L-curve criterion (Hansen (1998), Hu et al. (2016)) in each iteration of the TROPOMI measurement inversion. As TROPOMI has been measuring for more than two years, it is possible to select a constant regularization optimized for real observations. This includes a dedicated 6 https://doi.org/10.5194/amt-2020-281 Preprint. Discussion started: 5 August 2020 c Author(s) 2020. CC BY 4.0 License.
regularization parameter for the target absorber CH 4 and one for each of the aerosol parameters (aerosol distribution height and size parameter, and aerosol column). The advantage of the constant regularization is a more stable performance compared to the L-curve method in which the regularization strength changes at each iteration for every scene. The regularization parameters are selected such that the degrees of freedom for CH 4 are between 1 and 1.5 and that retrieved aerosol parameters have realistic distributions.

5
The main improvement of the constant regularization is that the dispersion in the retrieved XCH 4 is significantly reduced. This is noticeable in the XCH 4 distribution over small regions where we do not expect large gradients of XCH 4 . At regions with relatively low albedo, the decrease in the spread of the data can reach 10-20 % (e.g. from 18 ppb to 14 ppb over Canada and 11 to 9 over Australia). Furthermore, the validation with TCCON shows a decrease in the station-to-station variability of 4 ppb (25 % decrease of the 15 ppb station-to-station variability using the L-curve approach) when analysing one year of 10 data. The comparison with GOSAT shows that the new regularization scheme reduces the standard deviation of the difference between collocated GOSAT and TROPOMI XCH 4 observations by 9 % (19.7 ppb to 24.5 ppb).

Spectroscopy database
The TROPOMI CH 4 retrieval uses external spectroscopic information to simulate the molecular absorption lines of the target absorber CH 4 as well as of CO and H 2 O. The baseline retrieval algorithm employs the HITRAN 2008 spectroscopic database The TCCON validation shows that after substituting HITRAN 2008 by HITRAN 2016 and SEOM-IAS for all the molecules in the CH 4 retrieval, the station-to-station variability does not change significantly (less than 1 ppb, see Table 2). The change in the mean bias shows that the different spectroscopy databases introduce an overall bias in the retrieved XCH 4 with respect to HITRAN 2008 (+15.5 pbb for HITRAN 2016 and −14.7 ppb for SEOM-IAS), but the correlation of the bias with other 25 retrieved parameters (surface albedo, H 2 O) does not improve or worsen with any of the spectroscopic database. The spectral fitting quality parameters (e.g. the root mean square of the spectral fit residuals (RMS) and the corresponding χ 2 ) show a slight improvement over TCCON stations when using the SEOM-IAS spectroscopic database, similar to what was found for the CO retrieval from TROPOMI (Borsdorff et al., 2019). The comparison with XCH 4 measured by GOSAT also shows that different spectroscopic database introduce an overall bias but the standard deviation of the bias does not change significantly (Table 2).

30
On a global scale, we see that both the RMS and χ 2 improve significantly when using the SEOM-IAS database, with HI-TRAN 2008 giving the worst fitting results. Figure 1 shows the latitudinal distribution of XCH 4 retrieved with HITRAN 2008, SEOM-IAS and HITRAN 2016, referenced to the value at 0 • latitude. XCH 4 retrieved with HITRAN 2016 shows the least latitudinal variation at latitudes higher than 55 • where differences between the datasets are largest, however the global distri- The results of the sensitivity tests do not point to an improved data quality when HITRAN 2016, SEOM-IAS or HITRAN 2008 spectroscopic database is used. Each of them introduces an overall bias that cannot be used as an independent argument 5 to favour a specific database. In view of the better spectral fitting results in the retrieved XCH 4 we have decided to use the SEOM-IAS spectroscopy database.

Surface elevation
Satellite remote sensing of XCH 4 requires accurate knowledge of surface pressure and thus of surface elevation, which is specially relevant for the spatially highly resolved measurements of TROPOMI. The effect is two-fold: (1) through the pressure 10 dependence of the absorption cross sections and (2) through the dry air column used to calculate dry air mixing ratio from the retrieved column (Eq. 3).
In a first pre-processing step of the retrieval, the elevation data from a digital elevation map (DEM) is interpolated in space to the ground pixel. Then a correction is applied to the atmospheric variables (i.e. surface pressure and model pressure levels) based on the difference between the coarse resolution ECMWF altitude and the surface elevation from the DEM. To minimize 15 errors, a filter is applied on terrain roughness, which excludes scenes with a standard deviation of the surface elevation higher than 80 m within the observed area. The default source for surface elevation information for all TROPOMI products is the  perform a spatial sampling of 0.5 km and compute the mean altitude and its standard deviation for each scene. Figure 2 (upper 5 panel) shows altitude differences between S5P-DEM and SRTM collocated to TROPOMI pixels (before altitude correction) on 5 May 2019 over the United States. In this specific area, 5 % of the pixels have differences in altitude greater than 45 m, with the highest differences over mountain regions. For these scenes the differences in retrieved XCH 4 are up to 7 pbb. On a yearly average (and after correction and quality filtering), 1 % of the retrievals present altitude differences greater than 50 m, which result in surface pressure differences above 5 hPa and XCH 4 differences above 10 ppb ( Fig. 2 lower panels). The 10 terrain roughness within TROPOMI pixels has not changed significantly with the SRTM DEM, so it does not affect the data yield due to the 80 m threshold. Although globally the average altitude difference is small, the analysis of small scale XCH 4 enhancements will benefit from this update. Due to its higher resolution the SRTM DEM is a better representation of elevation not only over mountains, but also close to coastlines and over rough terrain (e.g. Greenland, Sahara desert).

Posteriori correction 15
Greenhouse gas concentrations retrieved from satellite instruments like TROPOMI generally show systematic biases with different instrumental or geophysical parameters. Retrieved CO 2 and CH 4 from GOSAT and OCO-2 are typically corrected for dependencies on goodness of fit, surface albedo or aerosol parameters (e.g. Guerlet et al. (2013), Inoue et al. (2016), Wu et al. (2018)). In the approach that O'Dell et al. (2018) derived for OCO-2 CO 2 retrievals, such parametric bias is part of a more complex correction that also accounts for footprint-level and global biases using a set of four "truth proxies" as a reference.  The comparison of TROPOMI and TCCON XCH 4 measurements shows a dependence of the bias (i.e. difference between TROPOMI and TCCON) on surface albedo, while for the other retrieved parameters the dependence is negligible (compared to that of the surface albedo, see Fig. 3). Figure 3a shows that for low albedo values, TROPOMI XCH 4 strongly underestimates TCCON measurements, while for relatively high albedo values TROPOMI overestimates TCCON measurements. The comparison of TROPOMI XCH 4 with XCH 4 retrieved from measurements of GOSAT shows the same dependence of the bias 5 with the retrieved surface albedo. For scenes with low albedo values, generally the retrieval's sensitivity is low due to the large measurement noise. therefore errors from unaccounted light path modification due to scattering processes can be more significant than for scenes with a relatively higher albedo. For low albedo scenes, this effect leads to an underestimation in the retrieved trace gas (Guerlet et al. (2013); Aben et al. (2007)), resembling the TROPOMI XCH 4 underestimation in Fig. 3a.
To account for the albedo dependence, we apply an a posteriori bias correction to the retrieved XCH 4 . In the baseline algo-10 rithm, we applied a correction based on the comparison of TROPOMI XCH 4 with GOSAT retrievals .
After more than two years of measurements, we have sufficient data to derive the correction using only TROPOMI XCH 4 measurements. We use a similar approach to the "small area approximation" applied to OCO-2 (O'Dell et al., 2018), assuming The new correction is derived as follows: 1. We select areas at several latitudes and longitudes throughout the globe, small enough so we can assume that XCH 4 does 5 not vary, but large enough to cover scenes with a wide range of albedo values. Figure B1 shows the different regions.
2. For each region we estimate a XCH 4 reference value for a surface albedo around 0.2 and then we calculate the ratio of the retrieved XCH 4 to the reference value to obtain the albedo dependence. The specific value for surface albedo is selected because XCH 4 retrieval errors are lower in the SWIR for that albedo range: errors because of unaccounted light path modifications due to scattering and surface albedo are minimal around a surface albedo of 0.2 (e.g. Guerlet et al.
3. We combine the albedo dependence for all the areas, we fit the curve using B-spline interpolation and least squares fitting.
The B-spline method fits piece-wise polynomials that are continuous at the pre-selected knots. The knots and the order of the polynomials are chosen such that the residual RMS of fit residuals is minimum and that the shape of the fit at the edges of 15 the surface albedo range does not vary sharply. Figure 4 shows the distribution of the reference to TROPOMI XCH 4 ratio for all the areas and the result of the B-spline fit. We observe two distinct features: (1) the strong underestimation for low albedo values (already shown in the TCCON comparison in Fig. 3a), for which the B-spline fit corrects more strongly than the regular polynomial fit that was previously used and (2)   The correction applied to the retrieved XCH 4 can be expressed as: The correction function f depends on the retrieved surface albedo A s at each pixel i. Figure 5 shows the global distribution of XCH 4 before and after applying the correction. Distinctive features that correspond with low and high surface albedo areas are visible in the difference map. After correction, for example, the XCH 4 underestima-5 tion for low albedo values (e.g. over high latitudes over Canada and Russia) is corrected. Similarly, the XCH 4 overestimation for high albedo values over desert areas like Sahara is accounted for in the correction. The change in XCH 4 induced by the bias correction is in the range of 2 %, in agreement with the errors observed in the TCCON comparison.
As the correction is derived using only TROPOMI XCH 4 data, the comparison with TCCON and GOSAT is an independent verification of the approach. The validation with TCCON shows a reduction of 5.9 ppb (50%) in the station-to-station variability 10 and of 13.6 ppb in the bias due to the albedo correction. The dependence of the bias on surface albedo is removed (Fig. 3a vs.   Fig. 6a) and the dependence on other parameters remains negligible (not shown). The comparison with GOSAT measurements shows that bias dependence on albedo is removed after applying the correction (Fig. 6b), which reduces by 4 ppb the scatter of the differences in XCH 4 measured by the two satellites. In the remainder of the paper the corrected XCH 4 product will be used.

TCCON validation
We perform a detailed comparison of the TROPOMI XCH 4 corrected with XCH 4 measured at 13 TCCON stations selected for the validation (Table 1). TROPOMI is able to capture the temporal XCH 4 variability, both the seasonal cycle and the year- to-year increase. This is clearly visible in the time series (e.g. Pasadena or Lamont) in Fig. 7, which shows the time series of 5 13 https://doi.org/10.5194/amt-2020-281 Preprint. Discussion started: 5 August 2020 c Author(s) 2020. CC BY 4.0 License. The mean bias is below 1 % for all stations; the validation results are summarized in Table 3. The average bias for all stations is -0.2 % (−3.4 pbb) and the station to station variability is 0.3% (5.6 pbb), both parameters below the mission requirements for TROPOMI XCH 4 retrievals. Compared to the uncorrected TROPOMI XCH 4 , the mean bias is reduced significantly (from 5 −3.4 % to 0.2 %) even though the correction approach does not include any term to correct a global bias. As the overall negative bias is driven by the strong XCH 4 underestimation for low albedo values (Fig. 3a), correcting for the albedo bias partly accounts for the overall bias. Figure 8a shows the mean bias and the standard deviation for each of the stations and Fig. 8b shows the correlation plot. For a more strict collocation criterion of 100 km radius instead of 300 km, the number of points is reduced significantly but the 10 results of the validation do not change.

High latitude stations
Measurements at high latitude stations such as East Trout Lake (54.36 • N) and Sodankylä (67.37 • N) show the highest variability and the highest bias in the validation before correction, which is partially reduced by the albedo correction (see validation results in Table 3). There is a seasonality in the bias which is positive during February -April period and changes to a negative bias 15 around May that then increases to reach the highest (negative) values in fall. This seasonality can be attributed to the fact that during the winter there is snow in these regions at high latitudes as a result of cold, dry air, influencing XCH 4 measurements by TROPOMI that affect the validation with TCCON measurements. Figure 9 shows the time series of the bias between TROPOMI and TCCON XCH 4 together with the surface albedo retrieved in both the SWIR and NIR spectral range over East Trout Lake, Sodankylä and Lamont, the latter included as a mid-latitude  reference. Low surface albedo in the SWIR together with high surface albedo in the NIR indicates the presence of snow which is highly correlated with the seasonality in the bias in East Trout Lake and Sodankylä, seasonality that is more pronounced in 2019 than 2018. The seasonal bias is also correlated with high hydrogen fluoride (HF) and low H 2 O concentrations (not shown). High HF concentrations are an indication of the influence of the vortex in a specific location, as HF is mostly found in the stratosphere; HF together with the contrast between surface albedo retrieved in the SWIR and NIR spectral ranges can be 5 used as a proxy to identify the presence of snow and dry air from dynamic meteorological situations at high latitudes.
The presence of snow at high latitude stations shifts the focus to retrieval errors as the most probable cause of the seasonal bias between TCCON and TROPOMI, rather than errors due to collocation or influence of the different priors. Scenes covered by snow are characterized by low spectrum intensity in the SWIR, so signal-to-noise ratio is a limiting factor. On the other hand, the high TROPOMI signal in the NIR suggests that the weighting of each band might not be optimal in the inversion. 10 As the optical properties are different in the NIR and SWIR bands, errors in the quantification of light path modifications over snow covered scenes can lead to an overestimation of retrieved XCH 4 . Furthermore, if H 2 O may compensate for any radiometric offset in the strong CH 4 absorption bands, then in such dry conditions this would not be as effective in winter as in spring-fall, causing the seasonality on the bias. A high bias in high latitudes correlated with H 2 O columns was also found in H 2 O/HDO retrievals from TROPOMI by Schneider et al. (2020). Note that the seasonal bias is also present when XCH 4 is 15 retrieved using the spectroscopic databases discussed in Sect. 3.2.  To filter for scenes covered with snow or ice, Wunch et al. (2011b) introduced the so-called "blended-albedo", which combines the surface albedo in the NIR and SWIR to be used as a filter. By applying it to Sodankylä and East Trout Lake, we found that a threshold value of 0.85 is optimal to remove these scenes. The influence of snow needs to be further investigated from the retrieval algorithm perspective, and it should be considered when interpreting the validation results and when analysing TROPOMI XCH 4 data over snow-covered scenes, most prevalent at high latitudes.

Comparison with GOSAT satellite
We compare XCH 4 retrieved from TROPOMI and GOSAT measurements for a period of two years (Jan 2018 -Dec 2019). The comparison yields a mean bias of −10.3 ± 16.8 ppb (−0.6 ± 0.9 %) and a Pearson's correlation coefficient of 0.85. Figure 10 shows TROPOMI and GOSAT XCH 4 and their ratio averaged to a 2 • x 2 • grid. Overall compared to GOSAT, TROPOMI underestimates XCH 4 , specially in the regions around the tropics in South America (−0.6 ± 0.8 %) and in the African continent (−0.9 ± 0.8 %). In Asia there is higher variability (up to 1 %) compared to other regions, with areas of underestimation as well as overestimation. The overall underestimation is stronger by about 2 % in the non-corrected XCH 4 , reflecting that the albedo correction improves the too low TROPOMI XCH 4 in areas where the surface albedo is low (e.g. forests around the Equator).
For higher latitudes, the underestimation is less strong, and in some areas TROPOMI overestimates XCH 4 compared to GOSAT 5 (e.g. Greenland and Antarctica), in agreement with the high bias in XCH 4 reported in the TCCON validation at East Trout Lake and Sodankylä.
The latitudinal distribution of XCH 4 from TROPOMI, GOSAT and TROPOMI collocated with TCCON stations is shown in Fig. 11, summarising the validation of TROPOMI XCH 4 and showing the good agreement between the three datasets. Similar to Fig. 10, it shows that TROPOMI underestimates GOSAT at most latitudes but both overlap within the XCH 4 variability. 10 It also shows the shift to an overestimation at high latitudes where TROPOMI retrieves higher XCH 4 . This agrees with the conclusion that over snow TROPOMI XCH 4 is too high and although this distribution resembles the latitudinal distribution of XCH 4 shown in Fig. 1, it cannot be attributed to the selection of the spectroscopic database.

Conclusions
We have presented several improvements that have been implemented in the retrieval of XCH 4 from TROPOMI measurements in the NIR and SWIR spectral range. Now that TROPOMI has been measuring for more than two years, the amount of data allows the implementation of a series of updates that were not previously possible without the use of any reference data (i.e. regularization scheme and a posteriori correction derived using only TROPOMI XCH 4 data).

5
The regularization scheme with constant regularization parameters stabilizes the retrieval and yields less scatter in the TROPOMI XCH 4 data compared to the operational data product (version 1.2.0 Hu et al. (2016)). We have investigated the effect of the horizontal resolution of the surface elevation database by replacing GMTED2010 S5P with the SRTM 15" database, relevant in the XCH 4 retrieval for which accurate knowledge of surface pressure is necessary. The higher resolution database results in a more realistic representation of surface altitude, particularly for mountainous regions and places with rough surfaces, 10 where differences in surface pressure above 5 hPa result in retrieved XCH 4 that varies up to 10 ppb for specific scenes.
We have tested three state-of-the-art spectroscopic databases (HITRAN 2008 with updates from Scheepmaker et al. (2013), HITRAN 2016 and SEOM-IAS). Using the SEOM-IAS database results in the best spectral fitting quality parameters in the retrieved XCH 4 . Each of the different spectroscopic database introduces a bias in the distribution of XCH 4 with respect to each other, but there is not any additional bias (e.g. latitudinal, albedo bias) that could point to the fitness for purpose of any of the databases. In view of the best fitting results, we decided to use the SEOM-IAS database, which was derived specifically for TROPOMI. However, there is a need for a thorough and detailed analysis of these databases focusing on the different absorbers that are relevant in the CH 4 absorption bands to learn about the underlying processes that are driving the overall bias.

5
One of the most relevant updates is the implementation of a posteriori correction that is fully independent of any reference data. We have derived a correction for the bias dependence on albedo using only TROPOMI XCH 4 data. This has been possible due to the high resolution of TROPOMI and its global coverage. We select regions around the globe which cover different albedo ranges and dependencies to estimate the albedo bias. The new correction is more accurate than the regular polynomial fit for the strong XCH 4 underestimation at low surface albedo scenes, and also corrects for the positive bias in scenes with 10 high surface albedo. After applying the correction, the albedo dependence in the TROPOMI-GOSAT and TROPOMI-TCCON comparison is removed, which is an independent verification of the correction scheme. The change in XCH 4 induced by the bias correction is in the range of 2 %, and although we attribute it mostly to unaccounted light path modification due to scattering processes, Butz et al. (2012) predicted residual scattering errors to be mostly below 1 % which suggests that other errors might exist that needs to be further investigated. 15 The good agreement of TROPOMI XCH 4 with TCCON (−3.4 ± 5.6 ppb) and GOSAT (−10.3 ± 16.8 pbb) highlights the high quality of the TROPOMI measurements. Low and high albedo scenes are the most challenging for the XCH 4 retrieval algorithm, and although the posteriori correction accounts for most of the bias, there is a need to further understand the 20 https://doi.org/10.5194/amt-2020-281 Preprint. Discussion started: 5 August 2020 c Author(s) 2020. CC BY 4.0 License.
underlying cause and whether it originates in the instrument or in the retrieval algorithm. Also the overestimation of XCH 4 over snow covered scenes requires further investigation from the retrieval algorithm perspective. With respect to the validation, additional sites would be beneficial to cover the under-sampled regions and conditions. The network is currently limited to relatively low albedo values, so there is a lack of reference data for high albedo scenes, particularly around the Equator.
Furthermore, there is a clear imbalance between the number of stations in the Northern and Southern hemisphere, as well as a 5 lack of stations below 45 • S. This is not only relevant for a complete validation of current and future satellite instruments, but also to have a complete global network to monitor concentrations of CH 4 in the atmosphere.

Appendix B: Regions selected for the posteriori correction
The regions selected to estimate the posteriori correction are shown in Fig. B1.