The CAMS model and data assimilation system is a specific configuration based on ECMWF's Integrated Forecasting System (IFS). The chemical mechanism of the IFS versions used between 2018 and 2021 is an extended version of the Carbon Bond Mechanism 5  (CB05; Huijnen et al., 2010), as implemented in Chemical Transport Model (CTM) Transport Model 5 (TM5). It is documented in Flemming et al. (2015, 2017), and more recent updates are given in Inness et al. (2019b) and Huijnen et al. (2019). CB05 is a tropospheric chemistry scheme with 57 species and 131 reactions. The horizontal resolution of the CAMS model is approximately 40 km (T511 linear spectral truncation and 0.35∘×0.35∘ grid), i.e. coarser than the resolution of the TROPOMI data.

The NRT global CAMS system has used CAMS_GLOB_ANT anthropogenic emissions (Granier et al., 2019) since July 2019 and used MACCity anthropogenic emissions (Granier et al., 2011) before then. Biomass burning emissions have come from the Global Fire Assimilation System V1.4 (GFAS; Kaiser et al., 2012) since October 2020 and were GFAS V1.2 before. Biogenic emissions are from CAMS_GLOB_BIO from July 2019 onwards, based on the Model of Emissions of Gases and Aerosols from Nature (MEGAN; Guenther et al., 2006), and from MEGAN_MACC, a different application of the MEGAN model, before then (see also Table S1 in the Supplement for CAMS model changes).

The IFS uses an incremental four-dimensional variational (4D-Var) data assimilation system going back to Courtier et al. (1994). In the current CAMS 4D-Var setup, a cost function that measures the differences between the model fields and the observations is minimized to obtain the best possible forecast through the length of the assimilation window by adjusting the initial conditions. Several atmospheric composition fields, including CO, are included in the control vector and minimized together with the meteorological control variables. The data assimilation methodology for the atmospheric composition fields remains unchanged compared to the one described in Inness et al. (2015b), but the background errors for CO were updated in July 2019 (Fig. 1). CAMS uses 12 h assimilation windows from 03:00 to 15:00 and 15:00 to 03:00 UTC and two minimizations at spectral truncations T95 (∼210 km) and T159 (∼110 km).

TCCO products from IASI and MOPITT instruments are routinely assimilated in the CAMS NRT system (see Table 1). The MOPITT data used by CAMS are the NRT thermal infrared (TIR) MOPITT retrievals from the 4.7 µm band (Deeter et al., 2017, 2019, 2022). IASI TCCO is retrieved from the same band (spectral range 2143–2181.25 cm-1), and the data used during the period covered in this paper are produced by Eumetsat using LATMOS/ULB's Fast Optimal Retrievals on Layers for IASI (FORLI, v20151001) algorithm, documented in George et al. (2009) and Clerbaux et al. (2009). The TIR retrievals have the largest sensitivity to CO in the mid-troposphere (Deeter et al., 2013; George et al., 2015). In the CAMS system, IASI and MOPITT TCCO data are thinned to a horizontal resolution of 0.5∘×0.5∘ by randomly selecting one observation in each grid cell.

The observation operator for TCCO in the CAMS system applies the averaging kernels of the observations to the model fields and calculates the model equivalent at the observation location and time, giving departures between the observations and the model as 1d=y-H^xm. Here, d is the departure between observation and model equivalent in observation space, y the TCCO retrieval, and H^ the observation operator to calculate the model equivalent of the observations from the 137 model level fields (xm). The TCCO retrievals can be written using the averaging kernel A, that relates the true vertical CO profile xt to the retrieved total columns 2y=xap+Axt-xap+ϵ, where xap is an a priori profile used in the retrieval of y and ε an error term for measurement errors and errors in the forward model. In the observation operator we apply the averaging kernels A to the model profiles, xm, to smooth the model profiles according to the sensitivity of the retrievals. This means Eq. (1) can be written as 3d=xap+Axt-xap-xap+AH(xm)-xap+ε=Axt-H(xm)+ε, where H is an operator to calculate CO layers from the model profiles on the vertical grid of the a priori profile. Using this observation operator, we remove the explicit influence of the a priori profile in the calculation of the departures, but knowledge of the a priori profile is still needed in the observation operator calculations for IASI and MOPITT TCCO retrievals. Also, the impact of the a priori profile remains implicitly through the dependence of the retrieval y and the retrieval error on the a priori profile, since a badly chosen a priori profile will generally lead to larger retrieval errors and larger departures.

A variational bias correction (VarBC) scheme (Dee and Uppala, 2009), where biases are estimated during the analysis by including bias parameters in the control vector, is used for the TCCO IASI data. In this scheme, the bias corrections are continuously adjusted to optimize the consistency with all information used in the analysis. VarBC is applied to the IASI TCCO data, with three predictors: (1) the thickness of the 1000–300 hPa layer, (2) the thermal contrast between the surface temperature and the temperature of the lowest model level over land, and (3) a global constant. The thickness parameter was chosen to allow for latitudinal variations of the bias correction. MOPITT TCCO data are used to “anchor” the bias correction, i.e. are assimilated without bias correction. Experience has shown that it is important to anchor the variational bias correction scheme, to avoid drifts in the model fields (Inness et al., 2013). MOPITT is used as the anchor because it was the first instrument assimilated in the early CAMS system, and the assimilation of IASI CO was added later. To prepare the CAMS system for the eventual loss of the MOPITT instrument, and because IASI and CAMS CO show good agreement (see Figs. 7 and 8 below) in the next CAMS model upgrade (planned for Q2/2023), the VarBC settings will be changed to use IASI-C as anchor and to bias-correct MOPITT and IASI-B CO.

The CAMS NRT model and data assimilation system is continually improved as the NWP model and the chemical scheme are further developed, and new satellite data are added. Table S1 lists the model upgrades that were implemented since the monitoring of TROPOMI CO data began in the CAMS NRT system in November 2018. The change to model CY46R1 on 9 July 2019 had a big impact on the CO analysis field and led to increased negative biases (see Fig. 3 below), likely to be linked to a change of the emission inventories.

In the CAMS 4D-Var data assimilation system, the background error covariance matrix is given in a wavelet formulation (Fisher, 2004, 2006). This allows for both spatial and spectral variations of the background error covariances. The background errors for CO used in model cycles CY46R1-CY47R2 (OLDBGE) were calculated using the National Meteorological Center (NMC) method (Parrish and Derber, 1992). For this, 150 d of 2 d IFS forecasts (using a 137 L configuration of CY45R1) were run, and differences between pairs of 24 and 48 h forecast fields were calculated whose statistical characteristics serve as a proxy for the background errors. For model CY47R3, new CO background errors were calculated (NEWBGE), again with the NMC method but with a newer model configuration (CY47R1), again using 150 d of 2 d forecasts. By updating the background errors, we remain consistent with the model as it evolves. The 150 d covered the end of NH winter, spring, and the beginning of NH summer and thus accounts for some seasonal differences in CO distributions. For future recalculation of the background errors, it is planned to use a sample composed of a NH summer and winter period to better account for the seasonal differences in CO, including its lifetime and changes in CO emissions, that might affect the background errors.

Figure 1 shows the globally averaged vertical CO background error standard deviation profiles and the horizontal correlation profiles from NEWBGE and OLDBGE. NEWBGE standard deviation values are smaller than OLDBGE ones in the troposphere below model level 80 (about 260 hPa) and larger between model levels 80 and 65 (about 260–130 hPa). This means that tropospheric CO increments will be smaller with NEWBGE (assuming the data and all other model settings are the same), as the background is penalized less in the 4D-Var analysis with NEWBGE than with OLDBGE. The global mean horizontal correlations are longer for NEWBGE than OLDBGE in the boundary layer (below model level 120, about 920 hPa) but shorter above. The horizontal correlations are longest over the continents at the surface, in the tropics in the free troposphere, and in mid-latitudes around 200 hPa (not shown). The CO background errors are univariate; i.e. the error covariance matrix between CO and other chemical species or dynamical fields is diagonal. They are also constant in time, and globally averaged vertical background error standard deviation profiles are used in the 4D-Var procedure. Work is ongoing to allow for the use of seasonally and geographically varying background standard deviation profiles for the CAMS atmospheric composition fields. This should improve the characterization of background errors that are likely to vary with season and region due to factors such as CO emissions, lifetime, and tropopause height.

CAMS NRT data are routinely validated, and validation reports are produced every 3 months (https://atmosphere.copernicus.eu/global-services, last access: 28 October 2022, for example, Errera et al., 2021). These reports show that the seasonality of the CO field is reproduced well by the global CAMS NRT system, i.e. currently still without the assimilation of TROPOMI data, when compared with independent data. However, the CAMS CO data generally have a negative bias. Compared to IAGOS (In-service Aircraft for a Global Observing System; https://www.iagos.org/, last access: 28 October 2022) aircraft CO data (Nedelec et al., 2015) the NRT CAMS data show the largest underestimations in the lower troposphere, while upper layers show a better agreement. On average, the modified normalized mean biases (MNMBs) with respect to IAGOS CO range between -10 % and 5 %, while correlations are mostly between 40 % and 60 % (e.g. Errera et al., 2021). Figure 2 shows profiles of seasonal mean relative differences between CAMS profiles and IAGOS (calibrated level 2) CO profiles at Frankfurt Airport and averaged over airports in the eastern United States. These airports were chosen because they had the best data coverage for the period discussed here. Figure 2 shows that the largest relative errors are found in the lower troposphere, with negative biases between -15 % and -25 % in all seasons. Differences above 700 hPa vary between -5 % and -15 %. Comparisons with IAGOS cruise level data given in the CAMS validation reports showed mostly negative biases for CAMS data in September, October, and November (SON) and March, April, and May (MAM) 2021 and no systematic biases in June, July, and August (JJA) 2021 and December, January, and February (DJF) 2020/2021 when positive and negative biases within ±20 % were found in most regions (Errera et al., 2021).

Errera et al. (2021) document that biases against GAW CO surface observations are within -8 % for European GAW stations and Asian stations and around -16 % for stations located in the Southern Hemisphere. Their comparisons with EEA AirBase surface observations in Europe show high temporal correlations, small biases over Belgium, Germany, Austria, and Switzerland, and larger negative biases over Spain (-30 %), Estonia (-30 %), Poland (-50 %), the Czech Republic (-60 %), and Bulgaria (-65 %).

Compared to NDACC FTIR tropospheric column CO data (Fig. 3) the CAMS NRT system underestimates the tropospheric CO column in the Northern Hemisphere (NH). This negative bias increased in July 2019 after the CAMS model upgrade to CY46R1 to between -3 % to -15 % for most stations and is larger than the reported 3 % measurement uncertainty. Compared with MOPITT and IASI TCCO data, the CAMS validation reports (e.g. Errera et al., 2021) find negative biases of up to -20 % and -30 % respectively. These also increased after the CAMS model upgrade to CY46R1 in July 2019. The negative CO bias of the CAMS model is even more pronounced in experiments without CO data assimilation (Errera et al., 2021).

Figure 6a shows a map of the relative differences between TCCO fields from ASSIM and CTRL averaged over the period 1 August to 30 December 2021 and illustrates that the assimilation of TROPOMI TCCO data leads to increased CO columns in the analysis. The largest relative increase is found at high latitudes north of 70∘ N (25 %–30 %), over North America (10 %–20 %), and over oceans (10 %–20 %) in the tropics. Such increases are expected if the CAMS analysis is influenced by the TROPOMI data, considering the positive TROPOMI analysis departures seen in the monitoring plot (Fig. S2) and the generally negative bias of the CAMS CO fields reported in comparison with independent observations. The large differences north of 70∘ N are in the area where no IASI or MOPITT TCCO data are assimilated; i.e. no TCCO data at all are assimilated here in CTRL. Figure 6b shows a cross section of the zonal mean absolute differences between ASSIM and CTRL for the same period and illustrates that CO is increased most at the surface and in the lower troposphere in the NH, with differences north of 60∘ N extending throughout the troposphere, as well as in the upper troposphere in the tropics. Cross sections and maps of analysis increments (not shown) illustrate that the increased CO in the upper troposphere is a direct result of the TROPOMI assimilation, rather than a result of convective transport redistributing CO in the vertical, with the largest impact over Africa, South America, and the Maritime Continent. Figure 6c shows the time evolution of the zonal mean relative differences between ASSIM and CTRL and illustrates that the largest differences are found in July and August, i.e. at a time when large boreal wildfires burned in Siberia and North America (see also Sect. 3.2.4 below).

Next, we look at the fit of the CO analysis to the assimilated CO data in ASSIM and CTRL. Figure 7 shows box plots of averaged global mean analysis departures from CTRL and ASSIM. In CTRL, TROPOMI analysis departures vary between 7.3 % to 9.1 %, for the different data types. Analysis departures for MOPITT (0.6  %), IASI-B (1.5 %) and IASI-C (1.3 %) are much smaller because these data are assimilated, and the analysis is influenced by the data. MOPITT has the smallest analysis departures because the data are assimilated without bias correction. In ASSIM, the analysis is influenced by the TROPOMI data in addition to MOPITT and IASI-BC, so that TROPOMI analysis departures are much reduced and now lie between 0.3 % and 0.8 % for the various data types. IASI-BC analysis departures are also reduced slightly, illustrating that the assimilation of TROPOMI CO improves the fit of the CAMS analysis to the IASI-BC data globally. MOPITT departures are increased in ASSIM because MOPITT TCCO values are generally lower than TROPOMI and IASI values, and by assimilating TROPOMI (without bias correction) in addition to the other data products, the TCCO analysis values are increased (Fig. 6), and the fit to MOPITT is degraded.

Figure 8 shows time series of daily mean analysis departures from the four instruments averaged over the area between 60∘ N and 60∘ S, i.e. excluding the polar regions where the NRT MOPITT product does not provide data. The analysis in ASSIM is influenced strongly by the TROPOMI data (Fig. 8a), and TROPOMI analysis and first-guess departures are much reduced in ASSIM compared to CTRL. IASI-BC analysis departures are also reduced in ASSIM during July and August and of similar size to the ones in CTRL for the rest of the time series. MOPITT departures in ASSIM are more negative than in CTRL (as already seen in Fig. 7b).

The increased MOPITT departures in ASSIM (Figs. 7 and 8) do not imply that there are problems with the MOPITT dataset. MOPITT data have been extensively validated (e.g. Deeter et al., 2019, show that MOPITT V8 products generally have biases of less than about 5 %) and are used with good results in the CAMS reanalysis, where their assimilation leads to a CO analysis that agrees well with independent observations (Inness et al., 2019b; Wagner et al., 2021). The underlying problem is the current negative CO bias of the CAMS model, which increased with the implementation of model cycle CY46R1 in July 2019. The CAMS CO analysis in ASSIM is improved when TROPOMI data are assimilated as TROPOMI has a small positive bias with respect to observations (see Sect. 2.2), which compensates for the negative bias of the underlying CAMS model and gives an analysis that agrees better with independent observations than if MOPITT is used as the main instrument. There are differences between the CO retrievals from TROPOMI, MOPITT, and IASI, whose investigation is beyond the scope of this paper. A study by George et al. (2015) using older versions of the MOPITT and IASI retrievals showed that using the same a priori profile in the MOPITT and IASI retrievals led to better agreement between the datasets in source regions and during periods of low sensitivity but that differences between the retrievals remained. They attributed these to differences in time and location of the observations, differences in the vertical sensitivity of the instruments, and differences in auxiliary parameters used in the retrievals (such as temperature, humidity, and cloud cover). MOPITT is likely still to be the instrument of choice for a future CAMS2 reanalysis because of the long dataset going back to 2000 and the long-term stability (Deeter et al., 2019).

Figure 9 takes a closer look at the analysis departures in the Arctic from IASI-B and IASI-C. Here we see a much improved fit in July and August when the high CO columns retrieved by IASI-BC, which are the result of strong boreal wildfires in Siberia and North America, are not well captured in CTRL (see also Sect. 3.2.4). The improved fit to IASI-BC in ASSIM continues throughout September and October, i.e. while TROPOMI data are available for assimilation north of 60∘ N. The departures are more similar in both experiments after November, when no TROPOMI data are available for assimilation in this latitude band. Examples of the impact of boreal wildfires in August 2021 on the CAMS TCCO analysis are discussed in Sect. 3.2.4 below.

To assess the quality of the TROPOMI analysis, we compare the TCCO fields from ASSIM and CTRL with independent observations, i.e. observations that were not used in the analysis. Figure 10 shows comparison against NDACC FTIR data at all available stations. The negative CO total-column and tropospheric column biases seen in CTRL are much reduced in ASSIM everywhere, except at the Antarctic station of Arrivals Heights, where ASSIM shows a larger positive bias which increases with time. As no TROPOMI CO data are assimilated south of 60∘ S in ASSIM, but the assimilation leads to increased CO values elsewhere, this increase over Antarctica must be the result of transport into the Antarctic region. Stratospheric CO in ASSIM has a reduced bias in the SH but larger positive biases in the NH and tropics than in CTRL. As the CAMS system only uses a tropospheric chemistry scheme, we do not assess the changes to stratospheric CO any further. Figure 11 shows the mean biases and standard deviation values at each NDACC station for the period 6 July to 31 December 2021 and confirms the strong reduction of the total and tropospheric column CO biases in ASSIM. This suggests that large parts of the TROPOMI TCCO analysis departures shown in Figs. S3 and S4 are the result of a low bias of the CAMS model, rather than a high bias of the TROPOMI TCCO product.

Figure 12 shows comparisons of tropospheric CO profiles from ASSIM and CTRL with IAGOS aircraft data at various airports averaged over the period July to December 2021, and Fig. 13 shows time series of the monthly mean MNMB against IAGOS CO for layers in the upper troposphere, mid-troposphere, and lower troposphere. We see a clear improvement in ASSIM with reduced biases, particularly in the lower and mid-troposphere at Frankfurt and North American and Middle Eastern airports. Here, the assimilation of the SWIR TROPOMI TCCO retrievals provides additional information to the CAMS system that is already constrained by the TIR MOPITT and IASI TCCO data in CTRL. At the West African airports, the improvement is largest above 800 hPa. Here, the number of clear data is smaller than in the other areas, reducing the sensitivity of TROPOMI to near-surface CO. The reduced bias at the West African airports in the upper troposphere confirms that the increase in CO seen in ASSIM in the upper tropical troposphere in Fig. 6b is a good result.

The comparisons against independent CO observations in this section show that the assimilation of NRT TROPOMI CO data in the CAMS global system reduces the negative bias of the CAMS CO analysis in the troposphere and has the largest impact in the lower and free troposphere. This is because the SWIR TROPOMI CO retrieval is sensitive to the CO column, including contributions from the boundary layer, especially for clear data. It therefore provides additional information for the CAMS CO analysis, which is already constrained by the assimilation of TIR MOPITT and IASI CO retrievals that are most sensitive to CO in the mid-troposphere.

Increased flammability and wildfire risk, due to high temperatures and drought conditions, were manifested by the development of large-scale and persistent wildfires between June and August 2021 in the boreal regions of North America, particularly in the western United States and Canada, and Eurasia, particularly in the Sakha Republic of Russia. To illustrate the increased emissions from these fires, Fig. 14 shows that daily total biomass burning emissions of CO for Canada, the United States, and Siberia in July and August 2021 from the Global Fire Assimilation System (GFAS; Kaiser et al., 2012) were persistently above the typical daily values based on the 2003–2020 data in July and August 2021. The annual total biomass burning CO emissions for 2021 of 46 Tg (Siberia), 17 Tg (Canada), and 9 Tg (United States) were considerably larger than the multi-year mean (2003–2020). GFAS uses fire radiative power observations from the two Moderate Resolution Imaging Spectrometer (MODIS) instruments on the NASA Terra and Aqua satellites to produce daily global estimates of emissions from biomass burning, with a spatial resolution of 10 km × 10 km, and is one of the emission input datasets for the CAMS system (see Sect. 2.1). GFAS emission estimates are likely to be an underestimation of the real biomass burning emissions, and uncertainties (which are also valid for other biomass burning emission datasets) arise from (1) the detection limit of the sensors for observing smaller fires or low temperature fires (such as in peatlands) and (2) the knowledge of fuel types (i.e. vegetation types and peat) and associated emission factors – particularly in Eurasia/Siberia.

The CAMS global NRT forecasts captured the transport of smoke from the persistent large-scale Siberian and North American wildfires. In July and August 2021, thick smoke from the North American wildfires was transported eastwards across North America in several episodes. It was observed at many Aeronet sites (not shown) and led at times to severely degraded air quality as far away as the north-eastern US coast, with haze clearly apparent in New York City. Satellite images, Aeronet observations, and CAMS forecasts showed that smoke from some of the North American fires was transported further across the North Atlantic Ocean, passing Greenland, and reached Europe. A concern with such high amounts of smoke crossing Greenland is the potential deposition of black carbon to the ice sheet, and previous studies have shown smoke particles from Canadian wildfires in the snowpack (Thomas et al., 2017). A very high number of wildfires also burned in eastern Russia through July and early August, with the worst affected area the Sakha Republic. In terms of the total estimated wildfire carbon emissions for June–August in the Sakha Republic, a new record in the GFAS period was set in early August 2021, when the cumulative daily total emissions for 2021 were already larger than annual total of the previous record year 2020. CO from these fires was transported north across the Arctic Ocean as far as the North Pole. Figure S5 shows maps of daily TCCO from ASSIM from 4 to 23 August 2021, depicting CO from the strong fires in eastern Russia, the North American fires, and several of the episodes of CO transport from North America eastwards over the Atlantic and from Siberia over the North Pole. Note that on 6 August 2021 some of the highest CO columns globally were found over the normally clean North Pole.

We assess how far the TROPOMI CO assimilation can improve the representation of such plumes in the CAMS system. Total-column CO values are clearly improved in the Arctic in August 2021 in ASSIM, as can be seen in Fig. 15, which shows average TCCO fields from TROPOMI and IASI-C. The highest CO columns are found over Siberia as well as over North America, the North Pacific, the North Atlantic, and the Arctic Ocean. The same was seen for IASI-B (not shown). These high values are better captured in ASSIM than in CTRL, as seen by the smaller IASI-C analysis departures in ASSIM in Fig. 15, in the area north of 65∘ N where IASI data are not assimilated. This better agreement was also seen in Sect. 3.2.2 in the time series of departures averaged over the Arctic latitudes (Fig. 9).

The differences between CAMS CO fields in CTRL and the satellite observations can have many reasons and cannot purely be traced back to shortcomings in the biomass burning emissions. Underestimation of CO is a common problem with many atmospheric chemistry models (e.g. Gaubert et al., 2020) and not just the IFS. Other studies have related it to possible overestimation of the hydroxyl radical OH (Strode et al., 2015) as the reaction with OH is the main removal of CO, underestimation of anthropogenic emissions and of non-methane volatile organic compounds (NMVOCs) from traffic emissions (Stein et al., 2014), or underestimation of secondary CO sources from the oxidation of methane and NMVOCs (Gaubert et al., 2016). Problems with the deposition fluxes could also play a role. Further work is needed to investigate the reason for the negative CO bias in the CAMS model, not only for the Arctic in this case, but also in general.

Some IAGOS flights in August 2021 intersected parts of the CO plumes that were transported eastwards over North America and further over the Atlantic. Figure 16 shows CO profiles at Boston and New York airports on 6 August which crossed such a plume, visible in the CO values greater than 150 ppb above 500 hPa. The subsequent profile at New York (Fig. 16c) on 7 August still showed the remnants of the plume, with values greater than 100 ppb between 450 and 350 hPa. Both ASSIM and CTRL have elevated CO values above 500 hPa at both airports on 6 August but do not quite reach the values seen in the observations. The vertical structure of the CO profile is reproduced better in ASSIM than in CTRL, but the largest improvement is found in the lower troposphere in all three cases, and the actual impact at the plume altitude is smaller.

Smoke plumes from the North American wildfires reached Europe on 7, 9, and 11 August 2021 and again between 19 and 21 August and were visible in IAGOS aircraft CO profiles at Frankfurt Airport (Figs. S6 and S7). Figure 17 shows weekly averaged CO profiles at Frankfurt Airport between 26 July and 29 August 2021 and illustrates clearly the improved fit to the IAGOS profiles below about 500 hPa in ASSIM. The figure also shows that the high CO values seen in IAGOS profiles between 500 and 400 hPa in the week 2–8 August are not quite captured in either model run, while the weekly mean profile of ASSIM for the week 16–22 August agrees well with the IAGOS profiles at all altitudes. A closer look at the daily IASI CO profiles in Figs. S6 and S7 shows elevated CO values above 500 hPa on 7 and 9 August, with maximum values located around 400 hPa and exceeding 300 and 250 ppb, respectively. On 11 August the plume was located at lower altitude (between 600 and 500 hPa) and only had values of about 180 ppb. ASSIM and CTRL show elevated CO values above 500 hPa on 7 August, but maximum CO values remain below 200 ppb. On 9 August, the plume is located at higher altitude in ASSIM and CTRL than in the observations. The weaker plume on 11 August is captured to some extent in ASSIM but does again not reach the values seen in the observations. On 19–21 August 2021 ASSIM and CTRL both show elevated CO values in the upper troposphere, with larger values in ASSIM, but the maximum values of about 180 ppb seen in the observations above 300 hPa are not quite reached. On 20 August both ASSIM and CTRL still show the plume, while the observations do not show it anymore. This might be because of a horizontal or temporal mismatch of the plume location in the model and observations.

The comparison with IAGOS data for these case studies shows that the largest improvement from the assimilation of TROPOMI CO data in ASSIM is found in the lower and mid-troposphere, i.e. below the altitude of the smoke plumes. In the upper troposphere, the CAMS analysis is already constrained by the assimilated TIR MOPITT and IASI data, and the additional impact from the TROPOMI data is smaller. While TROPOMI is very successful in correcting the total CO column, it cannot completely correct such deficiencies in the CAMS vertical CO profiles, suggesting that modelling aspects such as transport and accurate emission estimates remain essential to capture such events in the absence of vertically higher-resolved satellite data for use in the assimilation.