The evolution of the convective boundary layer is important for pollutant mixing, advection and plume injection and is idealised in Fig. 2. Surface heating initiates buoyant turbulent motions which deepen throughout the day and comprise the convective mixing layer. In the morning these thermals slowly destabilise the nocturnal stable layer before rapidly penetrating into the residual layer, i.e. the previous days' boundary layers. Growth of the mixed layer is inhibited by the statically stable entrainment zone, a region of overshooting thermals and downward entrainment from above. The lifting condensation level is typically within the entrainment zone but above the mixed layer top, leading to patchy cumulus fields. These clouds are either forced, i.e. fair weather cumulus which do not vent mixing layer air into the free troposphere, or active, where the level of free convection is reached and mixing layer air is vented through the tops of towering cumulus that may grow up to the limit of convection and detrain in the upper troposphere (Stull, 1985). Typically, a maximum in horizontal wind speed of variable magnitude and extent is present above the entrainment zone and referred to as the trade wind inversion.

The layer altitudes in Fig. 2 are based on a semi-automated analysis of individual thermodynamic profiles, the methodology for which is detailed in Sect. S2 in the Supplement. This analysis is summarised for an example profile in Fig. 3a–c. From the wind speed and direction profiles (panel a), regions of local positive wind shear (vertical brown line) and locations of wind speed maxima (open purple circles) are identified. Profiles of relative humidity, specific humidity, potential temperature and equivalent potential temperature (panel b) act as the basis for identification of a mixed layer and where this is the case, its depth. This involved a manual approach based on the spread of a number of automated methods. Profiles of temperature and dew point temperature (panel c) are presented in a skew-T log-P format, allowing calculation of the lifting condensation level (LCL), level of free convection (LFC) and limit of convection (LOC). The entrainment zone was not derived due to difficulties in determining its base with the information available. The format of Fig. 3 is replicated for each profile and displayed in Sect. S5.

In general, observed profiles of horizontal wind speed reflect those expected based on our understanding of synoptic flows over TSA (Campetella and Vera, 2002). In the lower troposphere (850 hPa) the mean synoptic circulation is an anticyclonic flow centred around south-eastern TSA: easterly trade winds turn south-eastward toward the extra tropics, parallel to the Andes (Fig. 4c). Horizontal wind speed is greatest in eastern coastal regions, northern Amazonia and Bolivia. Slacker wind speeds are prevalent in south-western TSA. At higher levels the influence of the Andes is reduced, resulting in a more zonal flow.

There is considerable structure in the observed profiles which is not captured by large-scale wind fields, such as those in the ECMWF ERA-Interim reanalysis. This includes multiple filament-like maxima in wind speed, often without accompanying changes in wind direction. In western TSA the expected sharp change in wind direction and increase in horizontal wind shear associated with a trade wind inversion are often not present, and instead a more gradual change in speed and direction was observed. In the mixed layer, horizontal wind speed is typically lower and of a more uniform direction and magnitude. There are a number of exceptions to this generalisation, whereby positive wind shear regions or wind speed maxima are present through the mixed layer. A wind speed maximum is generally observed above or collocated with the mixed layer top (e.g. Fig. 3). At times this was the trade wind inversion, but typically it was not. If this feature is associated with a region of positive wind shear below, mixed layer growth is likely inhibited, and in turn the magnitude of the jet and gradient of wind shear reduce.

Variations of the Fig. 2 idealised structure were also found to depend on the sample region, owing to the west–east asymmetries in land use. Compared to the Cerrado regions, the northern and western tropical forest regions experience higher levels of precipitation, soil moisture and cloud fraction, resulting in lower solar insolation and surface temperatures (Fig. S2 in the Supplement). On average these resulted in a deeper, faster developing mixing layer in the east than west and north (average of 1.8 vs. 1.1 km). As such the lifting condensation level (2.6 vs. 1.4 km) and level of free convection (3.1 vs. 2.4 km) were also greater. Only in 25 % of profiles was the limit of convection identified; it was higher in the east than the west and north (4.4 vs. 3.5 km).

Superimposed on the regional variation was the impact of two distinct meteorological phases in north and west TSA. Between 14 and 22 September conditions were characteristic of the dry season, whilst between 23 September and 5 October conditions more closely resembled those of monsoonal transition (Brito et al., 2014). This second phase was characterised by an increased rate and area of rainfall, accompanied by reductions in solar insolation and increases in cloud fraction and soil moisture (Fig. 4). In ∼30 % of profiles no mixing layer could be identified, and instead the lower atmosphere was conditionally unstable throughout (and hence Fig. 2 is unrepresentative). This typically occurred in the west of TSA and toward the end of the campaign as wet season type conditions became more prevalent. Meteorological conditions during the 2012 SAMBBA period were typical of the climatology – there were no climatic extremes (Fig. S2).

An automated analysis of interactions between atmospheric structure and pollutant abundance was undertaken for each individual profile. Results from this analysis are summarised in Table S1 in the Supplement and illustrated for the example profile in Fig. 3d–g. The full analysis methodology is provided in Sect. S2.

When present, the convective mixing layer was investigated to determine if each pollutant was correspondingly well mixed throughout the layer. This is the case in approximately half of all profiles featuring a mixed layer. The profiles for which this was not the case are typically perturbed by fresh plumes and/or wind shear and jet interactions. The number of CO and CO2 profiles taken during the morning that were classified as well mixed was approximately half that of the afternoon. For example, in Fig. 3 – as indicated by the vertical green line and accompanying yes/no indicator – rBC and σsp_dry are considered well mixed, but CO and CO2 are not, due to concentrations increasing near the surface.

If pollutant abundance in the region above the mixing layer (or if absent, the near surface) and 4 km exceeded those of unpolluted background conditions, a pollutant residual layer was identified. Dry season background conditions were defined as 0.1 µgm-3 for rBC (Artaxo et al., 2013), 15 Mm-1 for σsp (Rizzo et al., 2013) and 140 ppb for CO (Andreae et al., 2012). Since CO2 has significant global trends and large biogenic sources/sinks, its background is harder to quantify and is not considered here. Over 70 % of profiles included a pollutant residual layer of rBC, σsp and CO, even those in remote regions away from fresh emissions (Table S1). Typically, the shape of these residual layers is similar to that from the previous day's (days') boundary layer. However there is substantial diversity in shape and magnitude, especially between 2 and 4 km, that results from variations in plume injection, long-range transport, cloud detrainment or entrainment of clean air from aloft. This is seen in the profiles in Fig. 3 in which the residual layer is indicated by the vertical brown line and accompanying yes/no indicator.

Two impacts of horizontal wind shear on pollutant vertical structure were explored – firstly, capping effects, indicated by a decline in pollutant abundance over regions of positive horizontal wind shear, present in 40 %–60 % of instances, and, secondly, the curtailment by wind shear on plume rise via increased drag and lateral entrainment, as previously explored by Freitas et al. (2010). This was indicated by a co-location of pollutant abundance maxima with wind shear maxima. This was present in 20 %–30 % of instances. Figure 3 contains such an example at ∼650 hPa, as illustrated by the purple circle in panel (f).

Plumes were identified by a significant enhancement in pollutant abundance (dotted-dashed grey lines, Fig. 3d–g) over a local background (dashed grey lines, Fig. 3d–g). The majority of profiles contain a plume of one or more pollutant. Only in 10 % of these plumes did all pollutant abundances exceed the identification threshold, despite often being well correlated, indicating variability in plume composition. The plumes observed in the mixing layer are likely more fresh and locally injected. Their frequency increases in the afternoon (Table S2), which may be associated with increased fire incidence (Giglio et al., 2006). Plumes were encountered between the mixing layer and lifting condensation level in 15 %–20 % of profiles, indicating a significant number of fire plumes have sufficient energy to overcome the stable entrainment zone. If the pyro-convection is such that the plume reaches above the lifting condensation level, as in 10 %–25 % of profiles, pyrocumulus form and detrain pollutants. Plumes encountered at these levels likely include those from local sources and those advected from regional upwind sources. The frequency of these plumes generally declines in the afternoon, indicating pollutants have been re-entrained into the mixing layer or, in western regions, transported following deep moist convection above our flight ceiling.

It is possible that the layers of enhanced pollutants above the lifting condensation level are not direct plume injections but detrainment from active cumuli following shallow moist convection. However, mixing layer pollution transported in this manner will be of similar concentrations to that of the residual layer and as such may not be significant enough to trigger the plume thresholds defined here. Furthermore it would be expected that relative concentrations of CO would be conserved, but those of aerosol would be reduced following activation and wet scavenging if precipitation occurred. However, analysis of the plume composition (Sect. 3.4) indicates the opposite to be true. Together, this indicates the pollutant enhancements observed via this analysis are likely plumes rather than detrained layers. There is evidence of moist convection delivering CO to altitudes above ∼4 km but with significant wet scavenging of aerosol. In 81 % of profiles with sufficient vertical coverage, CO loadings increased by 40 ppb from a minimum at ∼4 km to the top of the profile. Unlike the discrete signal from plumes, the enhancement was often 1–2 km deep. Of the rBC and σsp profiles, only 8 % and 3 % had a similar increase in signal co-incident with CO enhancements. This indicates significant removal of aerosol from deeper convection to altitudes above ∼4 km.

During the dry season, fires are prevalent across TSA but especially between 0 and 25∘ S (Fig. 5a). Pollutants emitted from these fires accumulate in south-western TSA as the Andes act as a barrier to smoke advected westward via the trade winds (Fig. 5c). This regional smoke plume is optically thick, and its shape and magnitude are dependent on fire activity in a given year. Whilst the September 2012 dry season is generally representative of years since 2008 (Fig. 5a–c.ii), fire count and radiative power were greater in the eastern states. This likely explains an enhancement in AOD of ∼0.1 in these regions and southern Mato Grosso state. The shift to wet season conditions corresponded to a reduction in fire count (Fig. 6a) and emissions (Pereira et al., 2016). We speculate conditions were not optimal for human ignited fires, as a rise in relative humidity and increase in wind speed make ignition and control more difficult, whilst decreasing the fuel consumption. The increase in cloud cover may have also reduced fire detection efficiency. Together with increased wet removal, these effects acted to decrease the regional AOD by ∼0.2 and shift the peak AOD from central to north-eastern regions (Fig. 6c).

Based on the variability between the individual profiles and the regional differences in fire activity, land use and meteorology, we define four regimes which characterise the key synoptic differences in pollutant abundance and vertical distribution across TSA during SAMBBA:

deforestation-impacted western Amazonia around Rondônia and Mato Grosso states (68.5–54∘ W, 12–5.5∘ S) during dry season conditions (14 to 22 September), henceforth referred to as W1;

The average aerosol vertical profile shapes of W1, W2 and N1 are similar (Fig. 8a, b) – similar loadings occur from the near surface to 3 km with a small maximum between 1 and 2 km. Above 4 km, loadings decline to near-background values. Comparatively, E0 is characterised by a near-surface maximum, which may be because multiple fresh fires were sampled, followed by similar loadings up to 4 km, above which concentrations declined sharply to near zero. The shape of the aerosol profiles is not significantly affected if plumes are removed from the analysis.

There are also contrasts in the aerosol abundance between the regimes. There is a 2-fold reduction throughout the column between W1 and W2, consistent with the satellite AOD fields (Fig. 6). Within the boundary layer this represents a reduction from 0.85 to 0.4 µgm-3 for rBC and 115 to 60 Mm-1 for σsp_dry. rBC loadings of 0.85 µgm-3 in N1 are similar to those in W1, yet σsp_dry of 70 Mm-1 are more akin to W2. E0 features a similar ratio between rBC and σsp_dry as N1, although the abundance is much larger at 1.7 µgm-3 and 125 Mm-1.

For all regimes, CO mixing ratios are greatest at the near surface: 340 ppb in W1, 310 ppb in E0, 220 ppb in W2 and 150 ppb in N1. CO mixing ratios reduce with altitude to a minimum in the free troposphere: 125 ppb at an unknown altitude in W1, 95 ppb at 4–5 km in E0, 150 ppb at 4–5 km in W2 and 100 ppb at 3–4 km in N1. The reduction of CO mixing ratios in the boundary layer from W1 to W2 corresponds to that of aerosol abundance. CO2 mixing ratios are also greatest at the near surface: 402 ppm in E0, 393 ppm in N1, 394 ppm in W1 and 397 ppm in W2 representing a particularly prominent enhancement. CO2 mixing ratios also reduce with altitude with minima of 391 ppm in E0 and N1 or 392 ppm in W1 and W2 at the same altitudes as those for CO.

Profiles of pollutant ratios further illustrate the contrasting aerosol and gas-phase vertical distributions (Fig. 9a, b). Compared with CO and CO2 aerosol abundance is enhanced at higher altitudes in the boundary layer, even accounting for the removal of plumes (Fig. S3). The peak enhancement is at ∼2 km in W2, ∼1.5 km in N1, ∼4 km in E0 and, in W1, ∼2.5 km for σsp_dry and ∼1.5 km for rBC. These distributions are also seen when examining the pollutant ratios within plumes – as altitude increases so does the plume Δaerosol:ΔCO ratio (Fig. 9c, d). We use the term Δaerosol:ΔCO to describe both ΔrBC:ΔCO and Δσsp_dry:ΔCO since these aerosol properties co-vary. As these enhancement ratios are typically within the boundary layer, share a common source and in most cases are likely relatively fresh, it is unlikely they are biased by sudden changes in the composition of background air driving the observed gradient, as warned against in Yokelson et al. (2013).

There is limited evidence for a diurnal variation in aerosol abundance and vertical profile shape. CO, however, is enhanced near the surface in the morning (Fig. 8). In W2 and N1, CO2 is also enhanced at the near surface in the morning but, given the poor correlation with other pollutants (Sect. S5), likely arises via biogenic activity. This may also help explain the greater CO2 mixing ratios in W2 vs. W1, as the increased cloud cover reduces photosynthesis and hence CO2 uptake (Graham et al., 2003).

Within the boundary layer there is a significant contrast in the coating thickness of rBC between E0 (55 nm) and W1 and W2 and N1 (80–90 nm; Figs. 10, S4). Coating thickness increases with altitude in the boundary layers of W2, N1 and E0. This may be associated with partitioning of organic material into the particle phase as temperatures decrease and/or the greater prevalence of fresh emissions near the surface. In W1, W2 and most markedly N1, the rBC coating thickness declines above the boundary layer. Comparatively, rBC coating thickness increases above the boundary layer in E0.

Fire class determines the relative mix of pollutants emitted. In E0 there are relatively greater loadings of rBC owing to emissions from more efficient flaming combustion in Cerrado fires, as observed by Hodgson et al. (2018). In W1 and W2 there are relatively greater magnitudes of σsp_dry and CO as the higher biomass density and moisture content result in more inefficient smouldering combustion. Significantly, the shift in meteorology between these two phases does not substantially impact the relative abundances of rBC, σsp_dry and CO to each other.

Meteorological conditions influence the absolute pollutant concentrations via their impact on removal rates/processes, advection and fire incidence. The driest conditions and greatest fire incidence meant the greatest pollutant concentrations were observed in E0. The reduction in aerosol loadings from W1 to W2 is attributable to a corresponding decline in fire count and increase in wet removal associated with the more widespread and heavy precipitation. Aside from wet removal, the ubiquity of the pollutant residual layer indicates pollutants are not rapidly removed. As such, advection becomes a key modulator of pollutant concentrations local to, and distant from, the source region. The latter is the case in N1 – back trajectory analysis (Fig. S6) shows that pollutants are advected from the fire hotspot around Maranhão and northern Pará states (Fig. 5). The pollutant mix – relatively greater concentrations of rBC and CO2 – is consistent with the Cerrado burns at these sources. Whilst CO2 mixing ratios are also greater in the east (relative to CO), as would be expected from more flaming combustion, the role of fire combustion processes cannot be evaluated given the confounding influence from biogenic sources/sinks in both regions.

The shapes of pollutant vertical distributions are primarily controlled by meteorological conditions, in particular vertical convective motions and horizontal wind shear (Fig. 7). The former acts to mix pollutants released near the surface toward the mixing layer top, the altitude of which can be modulated by the latter, soil moisture and solar insolation. The difference in profile shape from west to east to north is primarily driven by contrasting mixed layer depths. Pollutant loadings remained relatively high above the mixing layer in residual layers, indicating wet removal is not significant at these altitudes. Large unmixed plumes perturbed the mixed and residual layers, although they contribute only 15 % (E0), 11 % (W1), 8 % (W2) and 1 % (N1) to the scattering-only column AOD (calculated from the nephelometer; Sect. S2). Such plumes were seldom seen above 4 km, in contrast to previous observations in approximately the same sample region (Supplement in Andreae et al., 2004), indicating the mass flux from large pyrocumulus detrainment into the upper troposphere (within the aircraft range) was not significant. The observed increases in CO concentrations above ∼4 km indicate vertical transport of mixing layer pollution into the free troposphere. The presence of co-incident increases in rBC or σsp_dry in less than 10 % of these plumes indicates moist deep convection transports CO and presumably other gaseous and non-soluble components to altitudes greater than 4 km but efficiently removes aerosol from the air by wet scavenging. Such efficient aerosol removal and transport of CO to the upper troposphere was also observed in TSA during the 2014 dry season by Andreae et al. (2018). This is consistent with the decrease in rBC coating thickness at these altitudes in W1, W2 and N1 and also is similar to previous observations in boreal Canada that showed preferential wet deposition of the largest and most coated particles (Taylor et al., 2014). As deep moist convection is not common in eastern regions (e.g. using TRMM rainfall as a proxy; Fig. 4a), the source of elevated and enhanced CO is unlikely to arise from the mixing layer in the east. It is possible the source is from CO aloft in the west which is recirculated in the persistent anti-cyclonic flow at 500 hPa (Fig. S2f), as has previously been observed from satellite (MOPPIT) CO observations by Deeter et al. (2018) and characterised by trajectory analysis in Andreae et al. (2018). A long ageing time is consistent with the larger rBC coatings observed aloft in E0.

A transect flight from east to west (Fig. 11) captures and summarises the meteorological drivers of the regional contrast in pollutant vertical distribution. A declining mixing layer depth from east to west is evident from the reduction in altitude of the sharp gradient (i.e. the entrainment zone) of θe from ∼3 km in the east to ∼1.5 km in the west. Above the mixed layer top, relative humidity increases, especially so above the lifting condensation level where the high humidity distinguishes the cloud convective layer. This layer is deeper in the west, associated with deeper moist convection. A wind speed maximum is present at 5–6 km, coincident with the entrainment zone. Together, this structure can explain the lidar-derived extinction coefficient distribution (first published by Marenco et al., 2016). Aerosol is capped below the first wind speed jet, is well mixed within the mixed layer and features a maximum at ∼1.5–2 km. Visible plumes at ∼59 and ∼52.5∘ W lie at injection heights typical of those observed in the in situ profile data. The similarity of the in situ σsp_amb and lidar extinction coefficient profile shapes at the regional (Fig. 11) and local (Fig. S7) scales engenders confidence in the representativity of both datasets. Disparity in the absolute magnitudes is primarily explainable by differences in the sampling coverage (Fig. S8).

Consistent values of the scattering Ångström exponent at 700/550 nm within the boundary layer (Fig. S5) indicate a similar aerosol type throughout. These values were all above 1.5, typical of submicron biomass burning aerosol and indicating no significant regional role of super-micron dust or primary biological aerosols which have values closer to zero (Clarke et al., 2007; Russell et al., 2010). This is consistent with size distributions reported and from the aircraft (Darbyshire et al., 2019) and literature values of the Ångström exponent during the dry season (Rizzo et al., 2013; Saturno et al., 2018). This also indicates that model scaling factors to match remotely sensed AOD are not significantly biased by non-biomass burning aerosol. The observed increase in humidity with altitude in the boundary layer (Fig. 7c) will likely have a significant impact on AOD and therefore the required scaling factor as it remains an uncertain model process (Johnson et al., 2016; Reddington et al., 2018). We note that findings from SAMBBA suggest that omission of burned area from small undetected fires is the most significant source of inventory under-representation of aerosol emissions (Hodgson et al., 2018; Reddington et al., 2016), as in other burning regions (Nowell et al., 2018).

The average vertical distribution of pollutants in this study are similar to those from previous in situ and remote sensing observations in the region, despite temporal, geographic and instrumental sampling differences. This indicates driving processes are common throughout. The aerosol vertical profile is approximately vertical from the surface to a height of: 2.25 km (Huang et al., 2015; Table 2 therein), ∼1.5 km (Bourgeois et al., 2015; Fig. 6 therein), ∼2 km (Baars et al., 2012; Fig. 12 therein), ∼1.5 km (Andreae et al., 2012; Fig. 6 therein), ∼2 km (Marenco et al., 2016; Fig. 9 therein) or 1.5–3.5 km (this work, Fig. 8). Profiles of rBC in Andreae et al. (2018) were similar to those presented here – showing greater concentrations in Cerrado regions where lofted layers of greater concentrations were also observed between 2 and 4 km. Loadings reduce to near baseline values above ∼4 km in all. Only in Andreae et al. (2012) have similar regionally averaged profiles of CO been presented. These exhibit a similar profile shape as here – greatest near the surface and reducing thereafter to a minima at ∼4 km. Andreae et al. (2012) therefore also demonstrated the same relationship of Δaerosol:ΔCO with altitude as observed here. Whilst studies displaying only a few profiles (e.g. Andreae et al., 1988, 2004; Chand et al., 2006; Pereira et al., 1996; Yokelson et al., 2007) may not mirror these average distributions, comparable profiles are present in our library of individual profiles (Sect. S5), indicating that our more extensive dataset is consistent with previous reports and captures variability between different studies.

To try to understand the relationship between Δaerosol:ΔCO and altitude, it is necessary to examine multiple potential drivers. Substantial dry deposition could remove near-surface aerosol and reduce the observed Δaerosol:ΔCO. However, dry deposition fluxes reported in the literature (Ahlm et al., 2010) are small compared to the number concentrations observed (Darbyshire et al., 2019), so this is an unlikely cause. Likewise we discount a significant anthropogenic non-biomass burning source of CO near the surface, as emissions hotspots are over 1000 km from the flight region and are approximately an order of magnitude lower than from fires (Fig. S9). Increasing aerosol concentrations with altitude are typically attributed to partitioning of organic and inorganic species from lower-volatility gas-phase species into the particulate phase as temperatures decline and relative humidity increases (Heald et al., 2011; Morgan et al., 2010) . A significant fraction of boundary layer organic aerosol in the regional haze will be secondary in nature following oxygenation and condensation of semi-volatile vapours during smoke plume evolution. During SAMBBA no net addition of organic mass was observed in the near or far field (Morgan et al., 2019), unlike previous studies (e.g. Akagi et al., 2012). However, the contribution of further biomass burning secondary organic aerosol owing to increasing altitude is difficult to determine with the data available. There is an increase with altitude of the contribution from secondary inorganic aerosol with altitude in E0 but not W1, W2 or N1 (Darbyshire et al., 2019). This may account for the increase in rBC coating thickness with altitude (Fig. 10), but the decrease with altitude in W1, W2 and N1 indicates no significant addition of secondary organic or inorganic aerosol. Furthermore, as the approximate 3-fold increase of involatile ΔrBC:ΔCO with altitude is of the same magnitude to that of Δσsp:ΔCO, these secondary aerosol processes cannot explain our observations.

Plume Δaerosol:ΔCO enhancement ratios increase with altitude within the boundary layer, indicating that these plumes may be the source of the observed Δaerosol:ΔCO profiles in the regional haze. Past measurements (Fig. S10; Ferek et al., 1998; Yokelson et al., 2007) and those from SAMBBA (Fig. S11, Hodgson et al., 2018) indicate Δaerosol:ΔCO increases with the modified combustion efficiency (MCE). We speculate the distribution may be driven by the relationship between MCE and fire intensity – and thus by extension, when normalising for meteorology, plume height (Lavoué et al., 2000). The peak heat flux (i.e. intensity) of an open landscape fire occurs when there is both a flame front, where flaming combustion occurs (high MCE), and a zone of smouldering combustion (moderate MCE) in its wake. Once the flame front is extinct, the fuel bed remains smouldering (moderate MCE), and although reduced, there may still be sufficient heat flux to generate a convective column to loft these emissions, albeit to intermediate altitudes. Once the heat flux is insufficient to generate any convective plume emissions from residual smouldering (low MCE) are released at the surface. Together one would thus expect a continuum from high-altitude release of high and moderate MCE emissions to surface release of low MCE emissions, which could explain our Δaerosol:ΔCO observations. The evolution of plume height as a function of heat flux has previously been observed for savannah fires in South Africa (Stocks et al., 1996). Alternatively, there may be multiple convective cores for a given burn as seen elsewhere (Achtemeier et al., 2011). Where these are away from the primary flame front, the MCE will be lower along with heat flux and therefore injection height. Another complicating factor may be from very intense fires which experience oxygen deficiency in the flaming zone and combustion efficiency drops (Ward and Hardy, 1991) – and therefore aerosol:CO reduces. We recommend further fire-scale plume dynamics observations in tropical regions to test these speculations and acknowledge that many case studies would be required to account for the other impacts on “real-world” fires' plume rise (including stability, vertical profile of wind speed/direction, latent heat release, flame front annihilation and topography). In addition to being related at the individual fire level, regionally there is a connection between MCE, average fire intensity and injection height. Fires in eastern Cerrado regions have a higher MCE (Hodgson et al., 2018), are more intense as illustrated by the FRP distributions in Fig. 5b and Gonzalez-Alonso et al. (2019) and have higher plume injection heights in spite of more stable conditions (Fig. S12, Gonzalez-Alonso et al., 2019; Marenco et al., 2016).

Long-range transport of emissions from biomass burning in tropical regions of West Africa may provide an alternative explanation for, or provide an additional contribution to, the increase in aerosol:CO with altitude. Such a transport pathway has long been observed (Andreae et al., 1994; Baars et al., 2011; Das et al., 2017) and typically peaks between August and October (Saturno et al., 2018). A basic investigation of CALIPSO lidar retrievals in the air mass history footprints of E0 (Fig. S13) shows frequent – but not persistent – offshore layers of enhanced backscattering between 2 and 5 km, identified as “elevated smoke” by the aerosol subtype algorithm. These aerosol extinction coefficient of these layers would account for approximately half of the burden observed in E0. This is consistent with satellite AOD observations on and offshore from September 2012 (Saturno et al., 2018). The elevated smoke layers can be tracked back across the Atlantic to West Africa, where the elevated layers are typically deeper, more backscattering and geographically wider. If the aerosol:CO and BC:CO ratios in these layers are higher than that of emissions in TSA, then introduction of this pollution could well generate the aerosol:CO gradient observed. As current literature estimates of the combustion efficiency and ratio between emission factors of rBC and CO are similar for both regions (Hodgson et al., 2018, and references therein) one may discount this theory. However, observations from recent (and as yet unpublished) field projects by the research group behind this study show greater rBC:CO enhancement ratios in lofted outflow from West Africa than over TSA. Furthermore, during the 2014 ACRIDICON-CHUVA aircraft campaign in TSA (Wendisch et al., 2016), offshore pollutant layers between 3 and 4 km with air mass histories from West Africa had rBC:CO enhancement ratios of approximately 20 and rBC loadings of up to 2 µgm-3 (Meinrat Andreae, personal communication, 2019). It remains unclear if these lofted layers with high rBC:CO enhancement ratios are driven by a difference in fire intensity or perhaps result from the MCE–plume dynamics mechanism speculated above, which would be consistent with past literature values.

We have observed a marked contrast in pollutant composition and vertical distribution from west to east across tropical South America which has not previously been identified. Given predicted “savannisation” of the southern and eastern edges of the Amazon rainforest owing to climate change, deforestation and fire–climate feedbacks (Nobre et al., 2016), we recommend model analyses are carried out to assess the magnitude of these effects which may impact convective motions, cloud formation and regional dynamics.

The linkages between pollutants and atmospheric structure presented here reinforce the first-order requirement for representation of realistic thermodynamics within models. This is currently a challenge for regional and global models, especially sub-grid convective mixing, which drives the aerosol profile shape (Archer-Nicholls et al., 2016; Hong and Dudhia, 2012). For instance ECMWF ERA-Interim equivalent potential temperature profiles compare poorly to our measurements (Fig. 11) and those of Beck et al. (2013). These model fields are often used as boundary conditions for regional models and as such may introduce unrealistic thermodynamic structures and thus pollutant vertical distributions. Whilst the broad shape of horizontal wind shear from model output is captured on the individual profile basis, the fine structure present in observations is missing. Comparisons with ECMWF ERA-Interim and WRF output indicate this is the case for both global models and finer-scale regional models. This may be problematic in accurately representing sub-grid processes, such as plume rise and convective mixing.

We speculate the Δaerosol:ΔCO profile may be driven by a coupling between combustion efficiency and plume dynamics. If confirmed by further enquiry, then future modelling studies will have to consider how best to represent the phenomenon. Replicating the gradient Δaerosol:ΔCO may be particularly important for model simulations which draw results based on a realistic vertical distribution, for example, aerosol–cloud and aerosol–radiation interactions or surface air quality simulations. Emissions factors for residual smouldering combustion in the Amazon have been collected (Bertschi et al., 2003; Christian et al., 2007) and could be used to test predictions from novel (and non-trivial) model set-ups against surface and satellite observations. Past satellite observations have suggested residual smouldering is a large source of CO and not fully captured by emissions inventories (Deeter et al., 2016; Pechony et al., 2013). If emissions from higher MCE combustion are released at altitude, further exploration of the processes and variability in TSA is required. Archer-Nicholls et al. (2015) demonstrated how the parameterised Freitas et al. (2007) plume rise model calculated injection heights which were too high in TSA. Climatological distributions of plume injection heights, which can prescribe model injection heights, show substantial diversity in TSA. The distributions derived by Marenco et al. (2016), consistent with plume heights here (Fig. S10), show injection heights at greater altitudes than those from CALIOP (Sofiev et al., 2013) and in turn MISR (Gonzalez-Alonso et al., 2019).

Although the magnitude and bearing may differ, the fundamental drivers of the pollutant vertical distribution identified here will remain so in drought years, which may be increasing in frequency (Jiménez-Muñoz et al., 2016). Dry convection may be more vigorous and the atmosphere more stable, deep convection less vigorous and aerosol scavenging reduced, fires more intense and fire hotspots located in different regions, but as long as model simulations well represent the fundamental drivers identified in this work, then they ought to be able to replicate the resultant vertical distribution. This is a promising avenue for future research to predict the impacts in future years, following on from the study of Thornhill et al. (2018).