The two main instruments used for this study were the Condensational Particle Counter (CPC, TSI model 3025) and the Fast Cloud Droplet Probe (SPEC Inc., FCDP). The CPC instrument measures number concentration of aerosols between 3 nm and 3 µm using an optical detector after a supersaturated vapour condenses onto the particles, growing them into larger droplets. Particle concentrations can be detected between 0 and 105 cm-3 with an accuracy of ±10 %. Coincidence is less than 2 % at 104 cm-3 concentration and corrections are automatically applied for concentrations between 104 and 105 cm-3. The CPC was mounted in a rack inside the cabin and connected to an isokinetic inlet and an aerosol flow diluter and was operated using an external pump. The isokinetic inlet has an upper limit of 5 µm for particle diameter, with penetration efficiency higher than 96 %. A 1.5 L min-1 flow rate was maintained using a critical orifice. The dilution factor varied between one and five.

The FCDP measures particle size and concentration by using focused laser light that scatters off particles into collection lens optics and is split and redirected toward two detectors. The FCDP bins particles into 20 bins ranging between 1 and 50 µm, with an accuracy of approximately 3 µm in the diameters. Bin sizes were calibrated using glass beads at several sizes in the total range. The FCDP was mounted on the right wing of the G-1 aircraft. Shattering effects were filtered from the FCDP-measured DSDs (droplet size distributions), which is a built-in feature of the provider software. Additionally, measurements with low number concentration (< 0.3 cm-3) and low water contents (< 0.02 g m-1) were excluded.

The quality flag of the CPC instrument was used to correct the concentration measurements. Whenever an observation was flagged as “bad”, it was substituted by an interpolation between the closest measurements before and after it that were either “questionable” or “good”. For “good” measurements, which represent 59 % of all the measurements, the uncertainty is less than 10 %. The interpolation weights decayed exponentially with the time difference between the current observation and the reference ones. If the reference observations were more than 10 s apart, these data were excluded. Sixteen percent of the data were interpolated in that manner, while only 0.02 % had to be excluded. This process was required not only to smooth out the bad measurements but was also important for maintaining significant sample sizes (instead of simply excluding “bad” measurements). No averaging was applied to the 1 Hz CPC data. However, tests were made in order to check the impact that the sample frequency had on the results. The results were not sensible to moving averages of up to 10 s, which corresponds to roughly 1 km displacement given that the G-1 flew around 100 m s-1 in speed. Given this observation, the analyses are based on the 1 Hz CPC measurements.

Complementary measurements of meteorological conditions were obtained from the Aventech Research Inc. AIMMS-20 instrument (Aircraft-Integrated Meteorological Measurement System, Beswick et al., 2008). This instrument combines temperature, humidity, pressure, and aircraft-relative flow sensors in order to provide the atmospheric conditions during the measurements. From the aircraft measurements of relative flow, the vertical wind speed was obtained and was used herein to compare cloud properties in the up- and downdraught regions. The precision of vertical wind speeds is 0.75 at 75 m s-1 true airspeed.

In order to compare two different populations of clouds, namely those formed under background conditions compared to those affected by pollution, a classification scheme was developed. The most discernible and readily observable difference between a polluted and background atmosphere is the number concentration of aerosol particles per unit volume. Urban activities such as traffic emit large quantities of particles to the atmosphere, which are then transported by atmospheric motions and can participate in cloud formation, especially when they grow, age, and become more effective droplet activators. Their number concentration and sizes primarily determine their role on the initial condensational growth of cloud droplets through the aerosol activation mechanism. Even though the urban aerosols have a lower efficiency when becoming CCN (cloud condensation nuclei), their number concentrations are high enough to potentially produce a higher number of cloud droplets (see, for example, Kuhn et al., 2010). By affecting the initial formation of the droplets, increased aerosol concentrations due to urban activities can alter the cloud microphysical properties throughout its whole life cycle. It will be considered here that a simple, yet effective, classification scheme should consider primarily aerosol number concentration to discriminate polluted and background conditions with respect to cloud formation environments. The intent of the classification scheme is not to quantify specifically the aerosols concentrations available for cloud formation under background and polluted conditions. Rather, it is a way to identify atmospheric sections that presented urban or natural aerosol characteristics.

Aerosol particle number concentrations (CN) measured by the CPC-3025 instrument were used to identify the plume location. The first procedure required is the elimination of possible artefacts related to measurements while the aircraft was inside a cloud. For that purpose, a cloud mask must be considered. The data are considered to be in-cloud by examining particle concentrations detected by several aircraft probes. The aircraft probes used to determine the presence of cloud are the Passive Cavity Aerosol Spectrometer (PCASP, SPEC Inc.), the 2D-Stero Probe (2D-S), and the Cloud Droplet Probe (CDP, Droplet Measurement Technologies). The thresholds for detection of cloud are when either the PCASP bins larger than 2.8 µm have a total concentration larger than 80 cm-3, the 2D-S total concentration is larger than 0.05 cm-3, or the CDP total concentration is larger than 0.3 cm-3. Thresholds were determined by examining the sensitivity of each instrument. Assuming that the presence of clouds can affect the CN measurements, the concentrations inside clouds were related to those in clear air. Whenever an in-cloud observation is detected, the CN concentration is substituted by the closest cloud-free measurement (given that they are not more than 15 s apart, in which case the data are excluded from the analysis). In this way, possible cloud and rain effects on aerosols concentrations, such as rainout or washout, can be mitigated on the analysis.

A simple and fixed threshold to separate the background and polluted observations is not enough because the altitude of the measurements should also be taken into account. For that purpose, all CPC data were used to compute vertical profiles of particle number concentrations in 800 m altitude bins. This resolution was chosen in order to result in significant amounts of data in each vertical bin. A background volume is identified whenever the measured particle concentration is below the 25 % quartile profile. The polluted ones are considered to be the ones above the 90 % profile. Additionally, it is required that the measurement is located in the general direction of the urban pollution dispersion in order to be considered a plume volume. Similarly, the background measurements are limited to the section outside the plume location only. It is important to note that, while the CPC data are available for the whole duration of the flights, in-cloud observations are limited to the times of actual penetrations. The choice of asymmetric 25 and 90 % profiles result in similar sample sizes for the classified polluted and background in-clouds measurements (305 and 424 s, respectively), while maximizing the differences between the populations.

Given the daily variations of meteorological characteristics, the plume direction, width, and overall particle concentrations may vary. For that reason, the plume angular section must be obtained for each day individually. Figure 3 shows an example of plume classification for the flight on 10 March 2014. The CN concentrations are shown as a function of the azimuth angle with respect to Manaus airport (0∘ is east, grows anticlockwise), irrespective of altitude. The colour represents the horizontal distance (km) from the airport. Note that there is an angular section where the concentrations are high not only close to the city but also as far as 70 km. This section is defined to be affected by Manaus pollution plume (delimited by grey dashed lines in Fig. 3). Note that the coordinate system is centred on Manaus' airport, where the G-1 took off, and not on the centre of the city or other point of interest. For this reason, it is also possible to observe relatively high CN concentrations close to the origin and to the north-east and south-east directions. This corresponds to high CN concentrations over the city. By keeping those directions outside the plume angular section, these data are not considered as plume. This is intentional because other aspects occur over the city that may contribute to the cloud formation. For instance, the heat island effect may contribute to the convection, changing the thermodynamic conditions compared to those over the forest. By keeping the origin point as the airport, which is located on the west section of the city, this problem is avoided.

The final result of the classification scheme for 10 March is shown in Fig. 4. A visual inspection of radiosonde (released from the Ponta Pelada Airport located in southern Manaus) trajectory plots confirmed the overall direction of the plume for each flight. Given the nature of the meteorology in the Amazonian wet season, i.e. its similarities with oceanic conditions concerning horizontal homogeneity, there should be no significant difference between the thermodynamic conditions inside and outside the plume region for the G-1 flights. In this way, differences observed in pollution-affected clouds are primarily due to the urban aerosol effects. It should be noted that even though the plume classification is defined from the CN measurements, there are also observable differences regarding CCN concentrations. The in-plume CCN concentrations (for altitudes lower than 1000 m) averages at 257 cm-3 for a 0.23 % supersaturation, while the respective background concentration is 107 cm-3 (Fig. 5). Note the overall low concentrations representative of the wet season. In that case, the plume increases the CCN concentrations by more than a factor of 2. For higher supersaturation conditions (which can be achieved in strong updraughts), the differences are even more pronounced. At 0.5 % supersaturation, the average CCN concentration inside the plume is 564 cm-3, while outside it is 148 cm-3. This shows that the plume increases the concentration of aerosol particles that are able to form cloud droplets under reasonable supersaturation conditions, even though they are less efficient than the particles in the background air.

In addition to the plume, the river breeze also plays a role in the convection characteristics over the region and the respective microphysics. The clouds directly above the rivers are usually suppressed by the subsidence from the breeze circulation. This was addressed by comparing the DSDs under plume and background conditions only for measurements over land, and it showed a similar picture to what will be shown in the next section. In this way it is possible to confirm that the results presented here reflect the effect of the Manaus pollution plume and not the river breeze, even though the clouds over land were indeed more vigorous. The results shown in the next section consist of the data probed both above rivers and above land.

Given that the aerosol population directly affects cloud formation during the CCN activation process, bulk DSD properties under polluted and background conditions may differ. Figure 6 shows the frequency distribution of the droplet number concentrations (DNC), liquid water content (LWC), and effective diameter (De) for all measurements inside the plume and under background conditions, irrespective of altitude. Those bulk properties were obtained from the FCDP-measured DSDs. The background clouds presented droplet number concentrations below 200 cm-3 for most cases, while being more dispersed for the polluted DSDs. It shows that higher DNC is much more likely to be found under polluted conditions than on background air. This observation may be tentatively justified as an increase in the water vapour competition, which leads to the formation of a higher number of droplets with smaller diameters. However, the water vapour competition is usually discussed for a fixed LWC, which is not the case for the statistics shown here. The measured background clouds presented lower water contents overall, which could also partly justify the lower concentrations observed.

The effective diameter histograms show distinct droplet sizes distributions for both populations. While around 50 % of droplets in the polluted clouds have De between 8 and 12 µm, the frequency distribution for the background clouds shows more frequent occurrence of De > 12 µm, even though they peak at similar diameters. This factor shows that, despite condensing lower amounts of total liquid water, the background clouds are able to produce bigger droplets than their polluted counterparts. Overall, Fig. 6 shows a picture consistent with the water vapour competition concept. However, DSD formation under a water vapour competition scenario depends on two factors. One is commonly cited in the literature (e.g. Albrecht, 1989) and is related to the impacts on effective droplet size as a function of aerosol number concentration. The other factor is how much bulk water the systems are able to condense while the vapour competition is ongoing. Figure 6 suggests that the Manaus pollution plume affects both mechanisms, which are more complex than the water vapour competition process.

An interesting question to address is why LWC is lower for background clouds, i.e. why this type of cloud is relatively inefficient for converting water vapour to liquid droplets. One possible answer is related to total particle surface area in a given volume. Considering a constant aerosol size distribution, when their total number concentration is increased, the total particle surface area per unit volume also increases. In this way there is a wider area for the condensation to occur, leading to higher liquid water contents. Additionally, if there is higher competition for the water vapour, the more numerous and smaller droplets formed under polluted conditions will grow faster by condensation than their background counterparts (because the condensation rate is inversely proportional to droplet size) and will readily reach the threshold for detection by the FCDP (around 1 µm). One point to remember is the high amount of water vapour available during the wet season. Those differences in the bulk condensational growth under polluted or background conditions may explain in part the differences observed in Fig. 6c–d, even if the aerosol size distribution changes from the background to the polluted sections. If the bulk condensation is more effective in a polluted environment, it should also lead to increased latent heat release and stronger updraughts. In a stronger updraught the supersaturations tend to be higher, which feeds back into an even higher condensation rate.

Other possible physical explanations for the higher LWC in polluted clouds include processes associated with precipitation-sized droplets (i.e. outside the FCDP size range) and aerosol characteristics. If the aerosol-rich plume is able to reduce the effective sizes of the liquid droplets, it will also be able to delay the drizzle formation. In this way, the liquid water would remain inside the cloud instead of precipitating. On the other hand, the fast-growing droplets in the background clouds may grow past the FCDP upper threshold, effectively removing water from the instrument size range. However, the penetrated clouds were predominantly non-precipitating cumulus at the early stages of their life cycles. Therefore, the warm-phase was not completely developed and the condensational growth plays a major role in determining the overall DSD properties. The second process identified (i.e. suppressed precipitation staying longer inside the clouds) probably has a lower impact. The average ratio between the second moment of the polluted and background DSDs is around two, which shows that the former has around twice the total area for condensation than their background counterparts. In this way, the increase in the bulk condensation efficiency is probably significant. Further studies are encouraged in order to detail and quantify the processes that lead to the observed LWC amount. However, based on Koren et al. (2014), the most determinant factor contributing for the high amount of cloud water under polluted conditions seems to be related to the condensation process. In the referred paper, it is shown that the amount of total condensed water tends to grow with aerosol concentration in a pristine atmosphere.

In order to detail the pollution effects on the total condensation rate and on the DSD properties, averaged properties for different water content and updraught speeds are analysed. Firstly, given that the LWC is a measure of the total amount of water condensed onto the aerosol population, its correlation with the updraughts should be assessed. The updraught speed at cloud base can be understood as a proxy for the thermodynamic conditions, as it is a result of the meteorological properties profiles in lower levels. In this way, it is possible to disentangle the aerosol and thermodynamic effects by averaging the LWC data at different updraught speed levels. Figure 7a shows the result of this calculation for only the lower 1000 m of the clouds, while also differentiating between polluted and background clouds. The 1000 m limit is chosen both for maximizing statistics and capturing the layer in which the aerosol activation takes place. That layer is possibly thicker under polluted conditions, given the higher availability of nuclei. For similar updraught conditions, i.e. similar thermodynamics, the averaged total liquid water is always higher for polluted clouds. By eliminating the dependence on the thermodynamic conditions, it is possible to conclude that the LWC values are significantly influenced by the aerosol population. This figure shows that, on average, not only are the polluted clouds more efficient at the bulk water condensation but the resulting LWC scales with updraught speed (linear coefficients, considering the error bars, are 0.13 g s m-2 for plume measurements and 0.033 g s m-2 for background clouds). In a background atmosphere, most of the aerosols have been activated, and increasing updraught strength does not result in further condensation. On the other hand, the higher availability of aerosols inside the plume allows for more condensational growth as long as enough supersaturation is generated, especially considering that the critical dry diameter for activation is inversely proportional to supersaturation and, consequently, to the updraught speed. However, a deeper analysis in a bigger data set would be required to assess the statistical significance. The enhanced condensation efficiency and the possible LWC scaling with updraught strength at least partly explain the higher liquid water contents in the plume-affected clouds. The standard deviation bars in Fig. 7a indicate that while there is high variability for the LWC in polluted clouds, the clean ones are rather consistent regarding the condensation efficiency.

The water vapour competition effect can be observed by examining droplet effective diameter and number concentrations at a certain LWC interval, as shown in Fig. 7b and c. In this way, the polluted and background DSD properties can be evaluated irrespective of the bulk efficiency of the cloud to convert water vapour into liquid water. It is clear that, even with the dispersion observed, the two DSD populations present consistently different average behaviours for all LWC intervals. For similar LWC, the averaged effective diameter is always larger on background clouds, with lower droplet number concentrations on average. Those results show a picture clearly consistent with enhanced water vapour competition in polluted clouds. It shows that, given a bulk water content value, droplet growth is more efficient in background clouds. This process should make background clouds more efficient at producing rain from the warm-phase mechanisms because of the early initiation of the collision–coalescence growth.

Another noteworthy point shown in Fig. 7 is the difference between the relationships of De and LWC, and of DNC and LWC. While the average effective diameter varies linearly with LWC (R2=0.95 for plume and R2=0.92 for background DSDs), there seems to be a capping on DNC. This means that for low LWC (< 0.4 g m-3), increases in the total water content are reflected in increased droplet concentrations. For higher LWC values, the averaged DNC remains relatively constant, while the effective diameter grows with the water content. This suggests that at low water content levels, i.e. at the early stages of cloud formation, the formation of new droplets has a relatively higher impact on the overall LWC. As the cloud develops, the LWC is tied to the effective diameter of the droplets, as the impact of new droplet formation is weaker at this point. This effect is clearer in background clouds given the limited aerosol availability.

The analysis of bulk DSD properties indicates a clear difference between the polluted and background cloud microphysics. However, it is desirable now to further detail those differences. As most of the aerosol activation takes place close to cloud base (Hoffmann et al., 2015), the direct effects of enhancements in particle concentrations should be limited to this region. However, the aerosol effect can carry over to later stages of the cloud life cycle given that it will develop under perturbed initial conditions. One proxy for the cloud DSD evolution in time is to analyse its vertical distribution. For a statistical comparison, a relative altitude for all flights is defined. This relative altitude is calculated as follows: firstly, the closest radiosonde is used in order to obtain the cloud base altitude (as the lifting condensation level) and the freezing level. In case the aeroplane reached high enough altitudes, its data are instead used to obtain the altitude of the 0 ∘C isotherm. From those two levels, the relative altitude is calculated as percentages where 0 % represents the cloud base and 100 % is the freezing level. The altitudes of the cloud base and freezing levels range, respectively, from 100 to 1200 m and from 4670 to 5300 m approximately. Three layers are then defined: (1) bottom layer in which relative altitudes vary between 0 and 20 %; (2) middle layer for 20 to 50 %; and (3) top layer, where the altitude is above 50 %. Those specific relative altitude intervals were chosen in order to capture the physics of the cloud vertical structure and to minimize the differences in sample sizes for each layer, as there are more measurements for lower levels. Despite probing individual clouds, the DSD measurements can be combined into the three layers defined and interpreted as representative of a single system. It is conceptually similar to satellite retrievals of vertical profiles of effective droplet radii (e.g. Rosenfeld and Lensky, 1998), where the cloud top radius is measured for different clouds with distinct depths and combined into one profile. This approach was validated by in situ measurements for the Amazon region by Freud et al. (2008).

Figure 8 shows statistical results for the DSDs in the three defined warm layers, while Table 2 shows the respective mean bulk properties. The altitude-averaged values show that the polluted clouds present higher number concentrations and water contents but lower diameters for all layers. Additionally, DNC decays much more slowly with altitude, and droplet growth is significantly suppressed. Those observations point to enhanced collisional growth in the background clouds.

The overall picture of cloud DSD vertical evolution can be seen in Fig. 8a. The most discerning feature between the DSDs at different altitudes is related to the concentrations of droplets greater than 25 µm. The concentrations in this size range grow with altitude on average. On the other hand, the concentrations of droplets smaller than 15 µm tend to diminish from the bottom to the top layer. Considering that the vertical dispersion of the DSDs represents at least in part its temporal evolution, this feature is associated with droplet growth, where the bigger droplets grow in detriment of the smaller ones. This growth mechanism is the collision–coalescence process, where the bigger droplets collect the smaller ones and acquire their mass. The shaded areas on the figure show that this is not only an average feature, but is also visible in the quantiles.

The statistical results of the vertical evolution of the DSDs are discriminated for the measurements inside the plume and in background regions in Fig. 8b–c. At first glance, it is quite clear that the two DSD populations present different behaviours with altitude, meaning that the droplets grow differently depending on the aerosol loading. The plume DSDs present a high concentration on the bottom layer and shows weak growth with altitude. The concentration of small droplets (< 15 µm) does not change much with altitude and the top layer DSD is relatively similar to the middle one. On the other hand, the DSDs in the background clouds show a stronger growth with altitude (Fig. 8c). The bottom layer DSD presents lower concentrations of small droplets but higher concentrations of bigger droplets than its polluted counterpart does. This coexistence of relatively big and small droplets readily activates the collision–coalescence process, accelerating droplet growth. After comparing both polluted and background DSDs with the overall averages (Fig. 8a), it is clear that enhanced aerosol loading leads to less than average growth rates, and the opposite is true for background clouds. The average growth rate for De is 2.90 and 5.59 µm km-1 for polluted and background clouds, respectively.

The vertical speed inside the cloud is a critical factor as it helps determine the supersaturation and consequently the condensation rates in the updraughts. The interaction between the updraught speeds and aerosol loadings ultimately determines the initial DSD formations at cloud base. As mentioned before, the characteristics of the initial DSD may have impacts on the whole cloud life cycle, making the study of the vertical velocities critical for understanding the system development. Figure 9 shows averaged DSDs for different cloud layers and vertical velocities conditions, discriminating between the plume and background cases. The first row shows results for the bottom layer under (a) plume and (b) background conditions. The middle and top layer results are shown together in the second row, for (c) plume and (d) background conditions. “Strong” and “Mod” are references to the up- or downdraught speed (strong or moderate). The middle and top layers are considered in conjunction in order to increase the sample size.

For the bottom layer, the vertical velocity has an impact mainly on the concentration of small droplets on polluted DSDs in the range D < 5 µm. The regions that presented updraughts are associated with higher concentrations of such droplets because of new droplets nucleated under supersaturation. The downdraught regions mainly contain droplets that already suffered some processing in the cloud system and have relatively lower concentrations of small droplets that were probably collected by bigger ones. Additionally, small droplets ascend readily with the updraughts due to their low mass, which is also a factor that can contribute to the differences between up- and downdraught DSDs. However, the dispersion shown in the shaded areas shows that the populations of DSDs in up- and downdraughts are relatively similar, suggesting a homogeneous layer with respect to DSD types. The DSDs shown on Fig. 9a indicate single-mode distributions, which hamper collection processes and explain the similarities between the different vertical velocity regions. On the other hand, the background clouds have a second mode, especially in the downdraughts, given the additional cloud processing which favours the collision–coalescence process. The particles associated with background air in the Amazon are not only less numerous but also bigger overall compared to the urban pollution, and both of those features favour faster growth by condensation because of less vapour competition and larger initial sizes. It is interesting to note that the background DSDs in the strong updraught regions are narrower when compared to their polluted counterpart. In a polluted environment, there is not only the natural background aerosol population but also the urban particles emitted from Manaus. The mixture of the two, with the consequent physicochemical interactions, permits the formation of droplets over a wider size range, with a prolonged tail towards the lower diameters. The shaded areas show that the differences between the DSDs in the up- and downdraught regions are statistically relevant for the background clouds and are not a mere averaging feature.

Cloud droplets keep growing as they move to higher altitudes, but the way in which it occurs is rather different in a background or plume-affected environment. For polluted DSDs, there are two modes at the higher altitudes: one reminiscent of the lower levels and the other probably mainly a result of additional condensational growth. In those systems, the additional processing does not seem to be effective at producing bigger droplets, as shown by the blue line and shaded area in Fig. 9c. For the background clouds, DSDs in the updraught regions show similar modes to their polluted counterparts, one close to 10 µm and the other at around 18 µm. However, there are appearances of droplets bigger than 30 µm that contribute to the formation of a third mode in the middle and top layers. This mode appears on the strong downdraught regions, which suggests it appears after in-cloud processing.