The modelled aerosol optical properties are compared to observations using level 2 AERONET (AErosol RObotic NETwork; ) photometer data, namely (i) AOD at a wavelength of λ=550 nm and (ii) the Ångström coefficient calculated using the AOD measured at λ=470 and 870 nm. The stations used for the model validation are listed in Table  and their location is shown in Fig. . Note that, except for Lope, most of the AERONET stations are located in the northern part of the studied region, mainly under the influence of mineral dust emissions. Comparisons are performed using statistical scores calculated with an hourly time step and are presented for a given AERONET station only if data are acquired on a regular basis over a period of 3 months (i.e. 2280 h) and if more than 30 values can be used to compute them (only scores for which at least 1.5 % of data are available are shown).

Three different satellite datasets are used in this study: (i) the Moderate Resolution Imaging Spectroradiometer (MODIS) for AOD, (ii) the Infrared Atmospheric Sounding Interferometer (IASI) for CO, and (iii) the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) for aerosol classification. The first two correspond to vertically integrated data when CALIOP provides vertical profiles.

The MODIS AOD product at 550 nm (from the MODIS–Terra aerosol 5 min L2 swath 10 km data collection 5.2) is used to quantify the increase in aerosol due to biomass burning . The model outputs and observations are collocated in space and time in order to exactly compare the model to the available observations.

The IASI CO total column retrievals by the FORLI algorithm are used. CO is a product of incomplete combustion with a lifetime of several weeks. It can be used here as a tracer of biomass burning long-range transport. These observations are thus used to check if biomass burning aerosol plumes in the model are realistically represented and transported. The comparison between the model and the IASI observations consists of 3-day-averaged column-integrated CO concentrations. The model outputs are collocated in space and time with the satellite observations when they are available. They are also vertically corrected using the satellite averaging kernels before the vertical integration. For comparison to the model results, six sub-domains are defined to represent several regions as follows.

SW: the South-West domain is the only region entirely over the sea and may be under the plume of biomass burning aerosols coming from Central Africa.

The CALIOP lidar measurements, on-board the Cloud-Aerosol Lidar Pathfinder Satellite Observation (CALIPSO) satellite , are analysed to obtain an aerosol sub-type classification (CALIOP v4.10 product), as proposed in and . In addition, the product provides information on the vertical extent of aerosol layers as shown by . The aerosol sub-type classification is built on thresholds of lidar-derived optical characteristics and is not error free, as mentioned by and . Limitations associated with this aerosol classification are described in . A specific development was carried out for the comparison between CALIOP and the model results. It is described in detail in Appendix .

The tracer release experiments are aimed at addressing the following question: what are the regions of Central Africa for which biomass burning aerosols can reach the coastal cities of the Gulf of Guinea?

The passive tracers are released from two locations in the western and eastern part of the biomass burning area in Central Africa for May–July 2014: trcW in Gabon at 12∘ E and -5∘ N, and trcE in the Democratic Republic of Congo at 25∘ E and -5∘ N (see Fig. ). The corresponding experiments are named “trcW” and “trcE”. For each location, two vertical profiles of injection are used: experiments for which aerosols are injected between the surface and 3000 m a.g.l. are labelled “1” and experiments for which aerosols are injected from 3000 to 6000 m a.g.l. are tagged “2”. These two altitude intervals enable the estimation of the sensitivity of the biomass burning transport to different regimes of the injection height (Hp) values. The tracers behave as aerosols, with a density and a size distribution, and are thus subject to deposition during transport. The tracers are continuously released from 15 June to 30 July. There is no diurnal cycle, with the emissions flux being constant during the whole period. The released amount is arbitrary and has no unit (but for realism, the emitted fluxes are of the same order of magnitude as anthropogenic emissions in SWA).

The fires in Central Africa generally start in April and peak in July . A lot of parameters are involved in the calculation of these emissions, making the wildfire fluxes one of the most uncertain sources in chemistry-transport models . This flux calculation may be divided into three parts: (i) the emissions fluxes, (ii) the injection height Hp, and (iii) the shape of the injection height profile. The first two items have already been developed in CHIMERE and are now considered as validated schemes. They are detailed in Appendix .

For this study, a specific development has been carried out on the shape of the vertical injection profile. This quantity is difficult to estimate but is often considered as a very sensitive parameter because the way emitted particles are vertically distributed in the lower troposphere will likely impact the long-range transport of biomass burning aerosols.

A lot of global models simply inject the emitted mass in an homogeneous way in the boundary layer or from the surface to a prescribed height Hp (see references in , among others). distribute the flux homogeneously between Hp/3 and Hp. Other models use more complex parameterizations based on thermal convective approaches primarily developed for boundary layer convection in dynamical models and adapted to the specific problematic of pyroconvection . However, this “thermal” approach is numerically cost consuming and difficult to use, being very sensitive to the chosen input parameters. Finally, some vertical profiles are close to the vertical diffusivity profile (Kz) shape with the maximum of injection at the height Hp/2, such as in Raffuse et al. (2012) and Veira et al. (2015).

In this study, and in order to reduce the uncertainty of our results, three simulations are performed.

NoFIRE. This simulation takes into account all processes (dynamic and chemistry) available in the CHIMERE model. All emissions are taken into account except the biomass burning emissions.

The differences between the PR1 and PR2 injection profiles are as follows.

PR1: 80 % of emissions are injected in the model layers included in the interval 0.9×Hp<z<1.1×Hp. The rest, 20 %, are injected between the surface and 0.9×Hp. This profile was selected to (i) estimate the long-range transport of biomass burning plumes and (ii) determine whether fires mainly injected in the mid-troposphere may have an impact on remote surface concentrations. This profile represents an idealized shape similar to that generally used for “thermal” parameterization under convective conditions.

The goal is to estimate whether the biomass burning emissions which occurred in Central Africa reach the Gulf of Guinea.

Results are presented in Fig.  for the four emitted tracers and for the three sites Lope, Lagos, and Abidjan. The four emitted tracers provide non-zero surface concentrations on the three sites. This means that the meteorological conditions are favourable to the transport of biomass burning from Central Africa to the Gulf of Guinea.

Lope is close to the most important biomass burning observed during the modelled period. The tracers are first emitted on 15 June and the first non-zero tracer concentrations in Lope are modelled on 17 June. As expected, the most important tracer concentrations are modelled for the trcW1 experiment (i.e. trcW experiment with the PR1 injection profile), the site being very close to the source. The values are important (up to 500 in arbitrary units). For the same release source but emitted at altitude, the tracer concentrations from the trcW2 experiment are lower but not negligible. This shows that even if a tracer is emitted between 3000 and 6000 m a.g.l., the daily dry convection in the lower troposphere is strong enough to mix significant concentrations down to the surface layer. The tracer experiments further east (i.e. trcE1 and trcE2) also have non-negligible concentrations in the surface layer in Lope. The first non-zero tracer concentration values are modelled on 23 June, 8 days after the initial tracer emissions. This means that even though the emissions are far to the east, the mixing and long-range transport brings biomass burning aerosols to the west coast of Central Africa in 1 week.

Even if Lagos and Abidjan are far from the tracer sources (≈ 1000 km), the biomass burning proxies also exhibit significant concentrations at the surface in the area of those cities. The most important concentrations are modelled for the tracer emissions in the western domain. For this location, the peak values are not completely correlated in time and depend on the altitude of injection. This shows that the main biomass burning plume follows the same transport in the troposphere, but also that vertical mixing coupled with differential advection may change the transport pathways to the surface layer of the studied cities. Finally, note that in Abidjan, the highest impact is related to the injection of particles at altitude (i.e. in the trcW2 experiment) and not to injection in the ABL (i.e. in the trcW1 experiment). The tracer concentrations in Abidjan and Lagos for the trcE1 and trcE2 experiments are 4 to 5 times lower than for the trcW1 and trcW2 experiments, indicating that most of the tracers are transported away from the Gulf of Guinea northern coast.

To increase our understanding of the complex transport pathways of biomass burning aerosols, we analyse the regional distribution of tracer concentration at the end of the period covered by the tracer simulations, during which long-range transport pathways from Central Africa to the Gulf of Guinea cities are best established.

Figure  presents surface concentrations for 27 July 2014 at 12:00 UTC and for each of the four tracer experiments. This day was selected as an example because it corresponds to (i) the end of the modelled period when the biomass burning transport is the highest, and (ii) the availability of CALIOP data with biomass burning plumes. As previously discussed, the tracers in all experiments reach the Gulf of Guinea cities of Lagos and Abidjan. The most important transport from the fire region to these cities is associated with the western tracer experiment trcW. For trcW1, the main transport from the emission region is to the south and the north-east. Up to latitude ϕ=+5∘ N, the direction of the tracer transport changes and follows the Harmattan (a dry and dusty north-easterly wind) towards the west. The most important contribution comes from the tracer emitted at altitude, i.e. in trcW2. A large part is observed in the southern part of the emission region, while another contribution follows the coastline towards the north and Nigeria before veering to the west upon reaching West Africa and being advected over Lagos and Abidjan.

For trcW2, the main part of the tracer plume is transported to the east over the continent. Upon reaching longitudes higher than λ>20∘ N, a part of this plume is redirected towards the west as a result of interactions with the Harmattan. Even though less important than for trcW1, a non-negligible part of trcW2 tracers is observed in Lagos and Abidjan.

This tracer experiment allows us to better understand the complex transport pathways of the biomass burning aerosols from Central Africa to the cities of Lagos and Abidjan. This can be summarized as follows.

Over continental Central Africa, the main transport pathway for biomass burning aerosols is towards the north-east. For fires, in the western part of the emissions region, the aerosol plume may follow the coastline.

The main conclusion regarding the tracer experiments is that the whole area of biomass burning in Central Africa is impacting the surface concentrations in the Gulf of Guinea coastal cities. The second main conclusion is that wildfire particle injection profiles PR1 and PR2 (peaking in the lower and the mid-troposphere, respectively) lead to different biomass burning transport pathways. Nevertheless, after a few weeks, the fire emissions injected in the mid-troposphere have an impact on the Gulf of Guinea cities of the same order of magnitude as those emitted in the boundary layer.

Results are presented in Fig.  for the month of July 2014, when biomass burning intensity is at its maximum for the studied period. The satellite observations are compared to the three model configurations NoFIRE, FIRE PR1, and FIRE PR2.

Over Africa, the MODIS data show two large areas of AOD > 0.5: in Central Africa (corresponding to fire emissions) and up to latitude 10∘ N (corresponding to mineral dust emissions). Without fire emissions, the NoFIRE simulation enables the validation of the mineral dust modelling and shows that the model tends to underestimate the AOD over the Sahel between 10 and 15∘ N. On the other hand, the plume transported to the Atlantic is slightly overestimated. Over the Gulf of Guinea, the modelled AOD is overestimated (0.5 when MODIS shows 0.3). The AOD due to mineral dust is mostly underestimated and many factors may explain this. As already discussed in , the modelled size distribution may be inaccurate, while it is very sensitive for the AOD estimation. It was also shown that a bias in AOD calculation may exist but is not necessarily related to erroneous modelled surface concentrations of particulate matter (PM). Over this region and during this period, an additional explanation for this bias could be related to the way the model handles precipitation events. The results presented in Appendix C, dedicated to the analysis of the precipitation, show that the modelled precipitation patterns correspond to what was observed with the Met Office MIDAS land surface stations. However, as discussed in , , and , these processes remain highly variable, uncertain, and difficult to validate, and it is possible that the scavenging was not modelled correctly, leading to these differences between the model and observations.

When including the calculation of biomass burning emissions and their transport, a general increase is observed in the FIRE simulations. While AODs are less than 0.05 in the NoFIRE simulation, AOD values can reach 1 over Cameroon in the FIRE simulations. The westerly winds transport these biomass burning plumes over the Gulf of Guinea and the model results shows that the whole coast is under these dense plumes, from Nigeria to the Ivory Coast. With MODIS, two high AOD regions related to fires are observed: one in Central Africa and the other along the coast. With the model, the increase in AOD is located more to the north and less intense. Finally, it is worth noting that there are no significant differences between the results of the two FIRE (using PR1 and PR2) simulations.

The conclusion is that the model reproduces the two large areas of high AOD due to mineral dust and biomass burning emissions, but that the intensities are not correctly modelled. Over Central Africa, the modelled AODs due to biomass burning are underestimated. This may be due to fire intensity or the size distribution of the modelled aerosol. This will be further discussed in Sect.  with the comparison between observed and modelled CO.

Results are presented as statistical scores in Table . For the time series, the two FIRE simulations using PR1 and PR2 are displayed. But for the scores, only the results for FIRE PR2 are presented, with the differences between the two FIRE simulations being negligible.

Except for the station of Lope, differences between the simulations NoFIRE and FIRE are very small (Table ). The correlation values range between -0.08 (Ascension) and 0.77 (Lope). The low score in Ascension is related to the offshore location of the site and the fact that long-range transport over the sea is difficult to reproduce: being less turbulent, there is less horizontal diffusion and vertical mixing. The plumes are thinner and more concentrated, and the results are more sensitive to a possible model error in the wind direction. The comparison with observations located at one single point over the sea thus often exhibits a lower correlation than for comparisons conducted over land. For other sites, the correlations are larger and show that the mineral dust variability is well modelled. The only site with differences between the NoFIRE and FIRE simulations is Lope, close to the biomass burning areas. The correlation increases from 0.46 to 0.77 when biomass burning emissions are added. This shows that the timing of the fire emissions and the transport is precise enough to clearly improve the simulation.

Examples of detailed comparisons between AERONET and the model are displayed in Fig. . In Cinzana, the AOD hourly variability is well represented and the majority of observed AOD peaks are modelled. The site being mainly under the influence of mineral dust emissions, there is no significant difference between NoFIRE and FIRE. This is very different in Lope. The addition of the biomass burning emissions increases AOD during the whole period. The modelled AOD remains lower than the observations, but the timing and the absolute value are more realistic.

As opposed to the comparison with MODIS, these time series and correlation values show that the AOD is not always overestimated by the model. This result shows the large variability obtained with different sets of data and also reflects the difficulty of modelling this parameter, which is strongly dependent on the optical properties of the modelled aerosols and the estimation of the extinction with the modelled size distribution (in our configuration, 10 bins may be considered as a correctly resolved size distribution for a CTM).

Complementary to the AOD, the Ångström exponent is also compared to the AERONET retrievals and for the same two stations of Cinzana and Lope. Results are presented in Fig.  (right column). This exponent expresses the ratio between the AODs at two different wavelengths and its value is inversely proportional to the aerosol size. Low values of the Ångström exponent will be representative of mineral dust (aerosols mainly in the coarse mode), while high values will be representative of biomass burning. In Cinzana, the Ångström exponent is low, with values between 0 and 0.3 (except some peaks). This means that the aerosol content is mainly mineral dust. On the other hand, in Lope, the Ångström exponent is higher and values range between 1 and 1.75, which is representative of finer particles and thus concentrations related to biomass burning emissions.

The CO comparison is presented in Fig.  as time series with the daytime IASI measurements and the corresponding model results. Each time series corresponds to the sub-domains described in Fig. .

The IASI data show the increase in vertically integrated CO concentrations over Central Africa and the eastern Atlantic from May to July (sub-domains SW and SE): under the influence of biomass burning emissions, the CO concentrations increase by 100 %, from ≈ 1.5 × 1018 to ≈ 3 × 1018 molecules cm-3.

For NoFIRE, the CO concentrations are quasi-constant. For FIRE, the observed CO increase is correctly reproduced. Even though this increase is slightly underestimated by the model in the southern part (SW and SE), the temporal variability and intensity are better modelled in the central part (CW and CE) where the studied cities are located. North of the studied region (NW and NE), the biomass burning emissions have a very low impact on the CO concentrations, with the outputs of NoFIRE and FIRE being close. At this latitude, the model tends to slightly underestimate CO concentrations (by ≈ 0.2 × 1018 molecules cm-3) with respect to IASI.

The differences between observations and the model may be due to several factors. First, the boundary conditions used for the simulations are global and “climatological” in model outputs. The transition from “mean” time-averaged values and this real test case may induce biases due to the lack of temporal variability in the climatologies. For long-lived species such as CO, these biases may be transported inside the model domain. Secondly, underestimated CO may be due to overestimated OH or to an underestimate of the production of CO from the oxidation of VOCs. showed that this last process results in a large variability in model results. However, without complementary observations it remains difficult to disentangle different contributions.

The vertical cross sections of aerosol types derived from CALIOP observations and CHIMERE simulations along the CALIPSO track for 26 and 27 July 2014 are displayed in Fig. . We focus on these two days because (i) CALIOP data are available above the studied region, and (ii) the long-range transport of biomass burning is at a maximum at the end of the studied period. The two CALIOP ground tracks are shown in Fig. .

The first result with this comparison is that the aerosol characteristics of the main air masses are well reproduced by the model: over land, the main aerosol is mineral dust, while over sea, sea-salt aerosols dominate the composition near the surface. Over sea at altitude, the main aerosol type is related to biomass burning (denoted as smoke). For the two days, the model is able to estimate the latitudinal extension of the smoke plume, from -15 to +10∘ N. Regarding the vertical extension of smoke, the model underestimates the altitude of the top of the plume on 26 July but represents it correctly for 27 July. For this latter day, the vertical structure of the plume (exhibiting two distinct features) is correctly reproduced by the model. The main difference between the model and the observations is that the smoke plume reaches the surface with the model but not in the observations. pointed out that in the case of an optically thick aerosol layer, the sensitivity of the CALIOP backscattered signal to the altitude of the base of the aerosol layer is strongly attenuated by the two-way transmission term. As a result, the operational algorithm may locate the base of the aerosol layer too high when it could actually be deeper and extend towards the surface.

However, the CALIOP data are only for “elevated smoke” (see Appendix ), meaning that this is not because the CALIOP aerosol typing algorithm did not detect and attribute a smoke value above the marine layer but rather that there is no smoke. In this sense, the model provides complementary insight about the plume vertical extent.

Finally, this comparison with “instantaneous” measurements in the whole troposphere proves that the model is able to correctly estimate the location, latitude, and altitude of the main studied aerosols. This improves our confidence in the model robustness.