Evaluation of near-surface atmospheric composition reanalysis data in the metropolis of São Paulo, Brazil

2.1. Study area

The Metropolitan Area of São Paulo (MASP), the largest megacity in Latin America, covers ~8100 km², enclosing 39 municipalities, and hosts a population of 21.9 million inhabitants (Instituto Brasileiro de Geografia e Estatística 2023). MASP is located in the state of Sao Paulo, 60 km from the coast. With a demographic density of ~2710 inhabitants km^–2, MASP exemplifies the complexities faced by urban regions experiencing rapid expansion, which strains natural resources and causes deterioration of living conditions. Vehicles are the predominant source of air pollutants in the MASP, powered by a mixture of fossil and biofuels (de Fatima Andrade et al. 2017). According to the Sao Paulo state environmental agency, light and heavy-duty vehicles are responsible for 60% of NO_x (nitrogen oxide radicals), 71% of hydrocarbons, 40% of PM₁₀ and 37% of PM_2.5 emissions in the MASP (Companhia Ambiental do Estado de São Paulo 2024). The widespread use of biofuels results in particular atmospheric chemistry conditions in the MASP, favouring the formation of secondary pollutants like ozone and secondary organic aerosols (Salvo and Geiger 2014; Vara-Vela et al. 2016). Secondary aerosols comprise 25% of PM₁₀ and 51% of PM_2.5 in the MASP (Companhia Ambiental do Estado de São Paulo 2024). Other relevant emission sources include industrial (40% of NO_x and 10% of PM₁₀), soil dust (25% of PM₁₀) and biomass burning (7% of PM_2.5) according to Companhia Ambiental do Estado de São Paulo (2024). The climate in the region is subtropical, marked by a wet season from December to February and a dry season from June to August. Air quality is monitored in the MASP by the Sao Paulo State Environmental Agency, Companhia Ambiental do Estado de São Paulo (CETESB).

2.2. Ground-based data

Eight ground-based stations from the CETESB network were selected to represent the average air quality monitoring conditions in the MASP (Table 1). The criteria for choosing the stations included data coverage during the study period (2015–19) and the spatial representativeness of the stations (Companhia Ambiental do Estado de São Paulo 2016), which refers to the region for which a point measurement is representative, showing a consistent temporal evolution. Priority was given to stations with urban (>4 km), neighbourhood (0.5–4 km) and medium (101–500 m) spatial representativeness, avoiding stations strongly affected by local emissions. Additionally, the selected stations were strategically distributed across the metropolitan region to ensure the representation of diverse areas. Daily and monthly concentration averages were calculated considering data from eight monitoring stations, representing the average conditions in the MASP. Supplementary Fig. S1 shows the variability of concentrations within the eight stations. The period from 2015 to 2019 was selected based on the following criteria: (i) it is relatively recent, being representative of current atmospheric conditions and atmospheric pollutant emission patterns; (ii) it excludes the years affected by the COVID-19 pandemic, which exhibited atypical emission patterns in certain months.

2.3. CAMS reanalysis data

The CAMS global reanalysis (EAC4) provides reanalysis datasets for atmospheric composition, including aerosols and reactive trace gases. These datasets are produced by the ECMWF. Combining numerical weather prediction, aerosol chemistry models and satellite-based observations, CAMS uses the four-dimensional variational (4D-VAR) data assimilation technique within the CY42R1 version of the ECMWF Integrated Forecasting System (IFS) to deliver complete and physically consistent information on aerosol and trace gas concentrations (Inness et al. 2019). Atmospheric data are available across 60 hybrid sigma and pressure levels, extending up to 0.1 hPa, interpolated to 25 pressure levels, 10 potential temperature levels and 1 potential vorticity level. This study used the CAMS reanalysis data on O₃, NO₂, PM₁₀ and PM_2.5 concentrations near the surface (2 m), with a spatial resolution of 0.75 × 0.75° and a temporal resolution of 3 h (Table 2) from 2015 to 2019. The data were aggregated as daily and monthly averages of air pollutant concentrations. Six grid points intersecting the metropolitan area were selected, labelled C1–6 in Fig. 1: C1 (−23.25, −47.25°); C2 (−23.25, −46.50°); C3 (−23.25, −45.75°); C4 (−24.00, −45.75°); C5 (−24.00, −46.50°); C6 (−24.00, −45.75°).

Fig. 1.

Location of the study area in Brazil showing the São Paulo state as an insert (left). MASP is depicted on the right, showing the municipal divisions and the coastline. Selected air quality monitoring stations are marked as red stars, CAMS grid centroids in purple, and MERRA-2 grid centroids in yellow. All CAMS grid points are over land areas, except for points C6 and M6, which are over the Atlantic Ocean. Point C5 is in the Baixada Santista area. The points that showed the best comparison against in situ data are circled in red.

CAMS uses an updated version of the MACCity inventory for anthropogenic emissions of chemical species (Granier et al. 2011) and the Global Fire Assimilation System (GFAS 1.4) for wildfire and biomass burning emissions. The MACCity inventory covers the period 1960–2010, and it has been updated for subsequent years based on the Intergovernmental Panel on Climate Change Representative Concentration Pathways (IPCC-RCP 8.5) scenario, which corresponds to the ‘business as usual’ scenario. In CAMS, the MACCity inventory provides anthropogenic emissions for the following species: CO, NO, SO₂, black carbon, organic carbon and anthropogenic secondary aerosols (Inness et al. 2019). Although the total anthropogenic emissions may be similar in local and global emission inventories, the literature points out large discrepancies in the distribution of pollutants among different sectors like transport, industry and agriculture emissions (Castesana et al. 2022; Paton-Walsh et al. 2022). Dust and sea salt aerosol emissions are simulated online, relying on meteorological conditions (European Centre for Medium-Range Weather Forecasts 2024). The chemical mechanism in the IFS-COMPO includes 123 tracers for reactive gases and 16 tracers for various aerosol species and size ranges, 157 gas-phase reactions, 3 heterogeneous reactions and 2 aqueous phase reactions, including the production of secondary organic and inorganic aerosols. The following processes are simulated, affecting the mass mixing ratio (kg kg⁻¹) of each tracer: transport (advection and turbulent diffusion), emission, dry and wet deposition, chemical and microphysical processes (European Centre for Medium-Range Weather Forecasts 2024).

The IFS aerosol module (AER) includes the following prognostic species: sea salt (SS), dust, black carbon (BC), organic matter (OM), sulfate (SU), nitrate (NI), ammonium (AM) and secondary organic aerosols (SOAs). Sea salt, dust and nitrate are represented in up to three size bins, from 0.03 to 20 µm. Black carbon and OM are divided into hydrophobic and hydrophilic components. SOA is divided into biogenic (having isoprene and terpenes as gaseous precursors) and anthropogenic (having aromatics as gaseous precursors). The concentrations of PM_2.5 and PM₁₀ are assessed using the following equations (European Centre for Medium-Range Weather Forecasts 2024):

where ρ represents air density (kg m⁻³); DUST_k, SS_k and NI_k refer to soil dust, sea sal, and nitrate components in different size fractions. All aerosol prognostic species are represented as dry mass mixing ratios (kg kg⁻¹). From the different aerosol components, the model calculates the AOD, using prescribed optical properties for each aerosol species and size range, and integrates them in the atmospheric column, assuming that aerosols are externally mixed.

To obtain improved estimates of trace gases and aerosol gridded concentrations, CAMS uses a data assimilation system, minimising a cost function that measures the differences between the model fields and the corresponding observations, resulting in the best possible estimate by adjusting the model initial conditions. CAMS assimilates various atmospheric composition satellite observations (refer to European Centre for Medium-Range Weather Forecasts 2024 for the complete list) for the variables CO, O₃, NO₂, SO₂ and AOD. The data assimilation for species like O₃ is rather simple because the observations can be directly compared with a variable included in the model. However, in the case of aerosols, AOD observations do not provide sufficient information to constrain each of the individual aerosol species and bins, so the total aerosol mass mixing ratio, defined as the sum of all aerosol species, is used as the control variable in the assimilation process (European Centre for Medium-Range Weather Forecasts 2024).

2.4. MERRA-2 reanalysis data

The Modern-Era Retrospective Analysis for Research and Applications, version 2, MERRA-2, is NASA’s latest atmospheric reanalysis of the modern satellite era, produced by the Global Modeling and Assimilation Office (GMAO) (Gelaro et al. 2017). MERRA-2 includes aerosol data assimilation, creating a multidecadal reanalysis where both aerosol and meteorological observations are integrated within a global data assimilation system. MERRA-2 utilises the Goddard Earth Observing System (GEOS 5.12.4) atmospheric data assimilation system, which comprises the GEOS atmospheric model and the Gridpoint Statistical Interpolation (GSI) analysis scheme. The model features a finite-volume dynamical core with cubed-sphere horizontal discretisation at ~0.5 × 0.625° resolution (Table 2) and 72 hybrid-eta levels up to 0.01 hPa. The analysis is performed on a latitude–longitude grid with the same spatial resolution using a three-dimensional variational assimilation algorithm (3DVAR) and a 6-h update cycle (Gelaro et al. 2017). Temporal resolution ranges from 1 to 3 h, depending on the product. In this study, MERRA-2 data were aggregated into daily and monthly averages of air pollutant concentrations. Six grid points intersecting the metropolitan area were selected, identified M1 through M6 in Fig. 1: M1 (−23.000, −46.875°); M2 (−23.000, −46.250°); M3 (−23.500, −46.875°); M4 (−23.500, −46.250°); M5 (−24.000, −46.875°); M6 (−24.000, −46.250°).

For anthropogenic emissions of PM, MERRA-2 uses the AeroCom Phase II HCA0 v1, including the sectors: residential, biofuel, industry, power, land and inland waterway transport. The AeroCom Phase II dataset varies annually in the period 1979–2006. Unlike CAMS, MERRA-2 has not updated emissions, repeating the 2006 emissions for the subsequent years. Biomass burning emissions are obtained from the Quick-Fire Emissions Dataset (QFED v2.4-r6), based on daily satellite retrievals. Dust and SS aerosols have wind speed-dependent emissions (Randles et al. 2017). The aerosol module, Goddard Chemistry Aerosol Radiation and Transport (GOCART), considers sources, sinks and chemistry of the following aerosol species: dust, sulfate, SS, BC and organic carbon (OC). Loss processes for aerosols include dry and wet deposition, transport and convective scavenging. The carbonaceous aerosols are divided into hydrophobic and hydrophilic components. All aerosol species are treated in bulk, except SS and dust, which contain five non-interacting size bins each (Randles et al. 2017). Compared with CAMS, the MERRA-2 aerosol chemistry module is less sophisticated because it does not include nitrate aerosols, gas phase photochemistry and anthropogenic SOAs.

Surface O₃ concentrations (M2I3NVCHM product) and aerosol components (M2T1NXAER product) from the MERRA-2 reanalysis were utilised in this study. NO₂ concentrations were not available in the MERRA-2 product. The concentrations of PM_2.5 and PM₁₀ were calculated as follows (Buchard et al. 2017):

where ρ represents air density (kg m⁻³); DUST refers to the soil dust component in different size fractions; SS represents the sea salt component in different size fractions; BC and OC stand for black carbon and organic carbon as the hydrophilic and hydrophobic fractions; SO₄ is the sulfate ion, assumed to be neutralised by ammonium (NH₄⁺). All aerosol prognostic species are represented as dry mass mixing ratios (kg kg⁻¹).

The data assimilation of atmospheric composition in MERRA-2 follows a similar approach to CAMS. MERRA-2 assimilates satellite data for variables like AOD and O₃, adjusting the model’s initial conditions to minimise the difference between observation and model fields. For the complete set of datasets assimilated in MERRA-2, refer to Gelaro et al. (2017). The three-dimensional profiles of aerosol mass mixing ratios simulated by the model are converted into AOD using prescribed aerosol optical properties and assuming external mixing. Just like CAMS, AOD is used as the control variable in the assimilation process. After the AOD data assimilation, apportionment of aerosol mass among the aerosol species is determined by a convolution of aerosol parameterisations, meteorology and assumed aerosol emissions (Randles et al. 2017).

2.5. Evaluation methods for the reanalysis products

Correlation analyses were performed between the MASP in situ data and the concentrations at each reanalysis grid point for the four air pollutants under analysis to select the most representative CAMS and MERRA-2 grid points. Metrics to evaluate the performance of reanalysis data in representing in situ observations included the Pearson correlation coefficient (r), relative mean bias (RMB) and mean absolute error (MAE). The Pearson coefficient is defined as

where n is the number of observations, x_i and y_i respectively represent the reference (ground-based observations) and reanalysis data, and mean values are represented by bar symbol (x̅ and y̅). The coefficient r indicates the strength and direction of the linear relationship between the two variables. A value r = 1 signifies a perfect positive correlation, r = −1 indicates a perfect negative correlation and r = 0 suggests no linear correlation between the variables.

Least squares linear regression was applied to the scatter plots of reanalysis (y) v. observations (x) to retrieve linear correction curves for the reanalysis products:

where a represents the slope and b is the intercept. A 95% level of significance was adopted for this study (P < 0.05). The P-value associated with the Pearson coefficient was used to evaluate whether there was a significant linear correlation between reanalysis and observations (null hypothesis: there is no linear relationship between the variables). Also, the P-values associated with the linear regression coefficients were used to evaluate whether the coefficients differed significantly from zero (null hypothesis: the coefficient equals zero).

MAE is defined as the mean absolute difference between the reanalysis data (y_i) and the reference ground-based observations (x_i):

MAE provides the average error without considering the direction of the errors (whether they are positive or negative), making it a useful metric for understanding the overall prediction accuracy. The lower the MAE, the more accurate the model.

The RMB was used to obtain the percentage difference between reanalysis and reference observations:

A positive RMB indicates that the reanalysis estimates predictions are, on average, overestimated compared with the reference ground-based observations.

Taylor diagrams were used to support the choice of the best CAMS and MERRA-2 grid points to represent air quality conditions in the MASP, allowing for simultaneous comparison of different metrics (Taylor 2001). In this work, the Taylor diagrams have two dimensions: the Pearson correlation coefficient (r) and the ratio (σ_r) between the standard deviation of reanalysis (σ_y) and ground-based (σ_x) observations:

3.1. Comparison of different reanalysis grid points against in situ air quality time series

As depicted in Fig. 1, a few grid points represent the whole MASP in the reanalysis products. Nearest neighbour and interpolation are the most common methods for collocation of in situ and model gridded atmospheric data (Albuquerque de Almeida et al. 2018; Ali et al. 2022). A different collocation approach was taken by comparing in situ observations against data from six reanalysis grid points intersecting the metropolitan area. The average air quality conditions in the MASP were represented by the mean of observations in eight monitoring stations (Section 2.2). Fig. 2 shows Taylor diagrams for each air pollutant to compare statistical metrics for the different reanalysis grid points against in situ observational data. The objective is to determine which reanalysis grid points better reproduce observations in the MASP.

Fig. 2.

Taylor diagrams comparing the performance of reanalysis products for different grid points against in situ observational data. Pearson correlation coefficients (r), standard deviation ratio (σ_r) and relative mean bias (RMB) are depicted for the air pollutants O₃ (a), NO₂ (b), PM_2.5 (c) and PM₁₀ (d). Circles represent the MERRA-2 reanalysis grid points M1–6. Squares represent the CAMS reanalysis grid points C1–6. The star symbol references the statistical metrics that would result in a perfect comparison against in situ data, with r = 1.0 and σ_r = 1.0.

The Pearson correlation coefficients for the CAMS grid points were consistently higher than MERRA-2 for all air pollutants and grid points. O₃, NO₂ and PM_2.5 concentrations estimated by CAMS showed the highest correlation coefficients (r > 0.6) with the reference observations, especially at points C2 and C4 (Fig. 2a–c). CAMS data dispersion was typically greater than the reference observations (σ > 1.0). Data from the CAMS C4 grid point showed a data dispersion closer to the reference measurements when compared with the C2 grid point. The CAMS C4 grid point was selected for the analysis below. Concerning the MERRA-2 dataset, grid point M3 showed the best combination of relatively high correlation coefficients (r≃0.4) and standard deviation ratio close to 1.0 for O₃ (Fig. 2a). The PM MERRA-2 dataset was uncorrelated with the ground-based observations, with r < 0.2 in all of the grid points (Fig. 2c, d). The reasons behind the poor representation of PM in MERRA-2 is discussed in Section 3.3. It is noteworthy that the selected grid points (C4 and M3) are typically downwind of the MASP considering the prevailing eastern winds in the region (Silva et al. 2017).

3.2. Overall performance of the reanalysis products at the monthly temporal scale

After selecting grid points C4 and M3 as the most suitable to represent the reference observations, statistical metrics were calculated to characterise the overall performance of the reanalysis products for each air pollutant at the monthly temporal scale. CAMS reanalysis data showed significant correlations with the reference observations, with r values in the range 0.74–0.89 (Table 3). MERRA-2 data showed a significant correlation with observations only for O₃.

O₃ concentrations were overestimated by CAMS and MERRA-2 reanalysis, with RMB values above 100% and MAE values in the range 52–66 µg m⁻³ (Table 3). Despite the overestimation, O₃ data from CAMS was linearly correlated with observations, reaching an r of 0.76 (Fig. 3a). By contrast, O₃ data from MERRA-2 showed inconsistent behaviour along the annual cycle, showing a weaker linear correlation with reference observations, especially in the first semester, which corresponds to the austral summer and autumn periods (Fig. 3b). CAMS overestimation of O₃ concentrations is consistent with the underestimation of NO₂ concentrations (Table 3), given that O₃ photochemical production is limited by hydrocarbons in the MASP (Alvim et al. 2017). Under this production regime, a decrease in NO₂ concentrations typically results in increased O₃ concentrations. NO₂ concentrations were not available in the MERRA-2 dataset, hindering the interpretation of the O₃ product.

Fig. 3.

Comparison between reanalysis and observational data for O₃ (a, b), and NO₂ (c) concentrations, considering the CAMS and MERRA-2 datasets. The red dashed line represents the 1:1 ratio.

PM_2.5 and PM₁₀ concentrations estimated by CAMS were strongly correlated with observations, with r values above 0.85 (Fig. 4 and Table 3). The relative mean bias was 50% for PM_2.5 and 16% for PM₁₀, indicating that the CAMS product is suitable to represent PM concentrations in the MASP at the monthly scale, after the correction of biases. However, MERRA-2 estimates for PM were not correlated with the reference observations, failing to reproduce the main features of PM seasonal variability in the MASP.

Fig. 4.

Comparison between reanalysis and observational data for PM_2.5 (a, b), and PM₁₀ (c, d) concentrations, considering the CAMS and MERRA-2 datasets. The red dashed line represents the 1:1 ratio.

On a daily scale, there is a greater dispersion in the comparison between reanalysis and observations, possibly associated with flaws in the models’ representation of atmospheric mesoscale and air pollution removal processes. Section S3 in the Supplementary material depicts an example of a daily concentration time series with contrasting conditions of atmospheric dispersion, showing that CAMS was able to reproduce the main features of PM concentration variability whereas MERRA-2 showed weak variability, disconnected from the ground-based observations.

3.3. Performance of the reanalysis products in reproducing air quality seasonal cycles

To assess the performance of the reanalysis products concerning air quality seasonal variations, mean annual cycles were calculated for the study period (2015–19), depicted in Fig. 5. The annual cycles of NO₂ and PM in the MASP are mostly driven by climate conditions, because air pollutant emissions, dominated by vehicular emissions, are rather constant throughout the year (Companhia Ambiental do Estado de São Paulo 2024). The winter (June–August) in this subtropical metropolis is mild, cold and dry, hindering atmospheric dispersion and favouring the accumulation of primary air pollutants (Carvalho et al. 2015; Oliveira et al. 2022). Frequent rain showers in the summer and unfavourable atmospheric dispersion conditions in the winter result in higher concentrations of NO₂, PM_2.5 and PM₁₀ during the winter in the MASP, as depicted in Fig. 5b–d. Secondary air pollutants like O₃ follow a different pattern, with higher concentrations in the austral spring and summer (Fig. 5a), when the increase in solar radiation input and reduced cloud cover favour the photochemical production of secondary pollutants in the MASP (Schuch et al. 2019).

Fig. 5.

Mean annual cycles of air pollutant concentrations from in situ observations, MERRA-2 and CAMS reanalysis products, considering the period 2015–19: O₃ (a), NO₂ (b), PM_2.5 (c), and PM₁₀ (d). Shadows represent standard deviations.

Concerning O₃, both reanalysis products, CAMS and MERRA-2, approximately reproduced the observed seasonality (Fig. 5a). Both reanalysis products estimated an O₃ concentration decrease from the austral spring to summer, which was not observed in the ground-based measurements. Also, both reanalysis products overestimated the in situ O₃ concentrations in all seasons by as much as 50 µg m⁻³, with RMB > 100%.

The CAMS reanalysis was also able to reproduce the seasonal variability of NO₂, which, like other primary pollutants, shows higher concentrations during the MASP austral winter (Carvalho et al. 2015) (Fig. 5b). CAMS consistently underestimated NO₂ concentrations along the year, with RMB = −45%. According to the product’s documentation, CAMS NO₂ concentrations are significantly affected by the prescribed anthropogenic emissions. It is possible that NO₂ emissions in the MASP are underestimated in CAMS’ emission inventory. Moreover, the overestimation of O₃ concentrations by CAMS may partially explain the underestimation of NO₂ concentrations under an O₃ photochemical production regime limited by hydrocarbons (Alvim et al. 2017).

Concerning PM concentrations, CAMS reanalysis correctly reproduced the increased concentrations in the austral winter (Fig. 5c, d), which was also reflected by r values above 0.85 (Table 3). Atmospheric dispersion is hindered during the mild cold and dry winter period in the MASP, favouring the accumulation of PM (Oliveira et al. 2022). CAMS overestimated PM concentrations, especially PM_2.5. However, MERRA-2 did not reproduce the austral winter increase in PM concentrations. Rather, MERRA-2 PM showed an increase in September, when the MASP is typically influenced by the long-range and regional transport of biomass burning aerosols above the planetary boundary layer (Yamasoe et al. 2017). Occasionally, the biomass-burning smoke trapped in the residual layer can affect PM_2.5 near-surface concentrations in the MASP (de Arruda Moreira et al. 2021; Souto-Oliveira et al. 2023; Vieira et al. 2023), which could partially explain the secondary peak observed in September for PM concentrations (Fig. 5c, d). Supplementary Fig. S2 shows that exceptionally high PM concentrations were observed at ground level in 2017, which may have contributed to this secondary peak in the mean annual cycle. Despite the contribution of the transport of biomass burning emissions, the main peak in the ground-based PM concentrations occurs in July, demonstrating that the unfavourable meteorological conditions in the winter are the main driver behind the PM seasonal variability in the MASP, which was not captured by the MERRA-2 product.

The seasonal pattern for each PM component was assessed to further investigate the reasons behind the poor performance of MERRA-2 for PM. OC and SS components dominated the MERRA-2 PM_2.5 concentrations (Fig. 6b), with concentration peaks in September and November respectively (Fig. 6a). SS aerosols dominated the MERRA-2 PM₁₀ concentrations, representing 58% of the mass (Fig. 6d), and also showing a peak in November (Fig. 6c). Only the MERRA-2 BC showed the expected seasonality for PM in the MASP, with higher concentrations in the austral winter. CAMS reanalysis, however, estimated a dominance of OM, accounting for more than 80% of PM_2.5 and PM₁₀ concentrations (Fig. 7). Inorganics were the second most important PM component according to CAMS (~8%), followed by SS and BC. CAMS’ OM, BC and sulfate components peaked in July, consistent with the expected seasonality of PM concentrations in the MASP.

Fig. 6.

Mean annual cycle of MERRA-2 near-surface PM_2.5 (a), and PM₁₀ (c) components as mixing ratios (kg kg⁻¹). The relative contribution of each aerosol component is also shown for PM_2.5 (b), and PM₁₀ (d). MERRA-2 accounted for the following aerosol components: dust and sea salt (SS) in different size ranges, organic (OC) and black carbon (BC) in different size ranges and sulfate (SO₄).

Fig. 7.

Mean annual cycle of CAMS near-surface PM_2.5 (a), and PM₁₀ (c) components as mixing ratios (kg kg⁻¹). The relative contribution of each aerosol component is also shown for PM_2.5 (b), and PM₁₀ (d). CAMS accounted for the following aerosol components: dust and sea salt (SS) in different size ranges, organic matter (OM), black carbon (BC), sulfate (SU), ammonium (AM) and nitrate (NI).

3.4. Temporal variability of concentrations at the monthly scale

Fig. 8 shows the monthly time series of air pollutant concentrations estimated by the CAMS and MERRA-2 reanalysis and the reference observations in the MASP. There were no significant concentration trends during the study period (2015–19), although some interannual variability is present. For example, the relatively low O₃ concentrations observed in the austral spring of 2018 (September–November) were correctly reproduced by the CAMS reanalysis product (Fig. 8a and Supplementary Fig S3). The MERRA-2 data showed a rather constant O₃ annual cycle. In some years, the MERRA-2 O₃ concentration peak was not in phase with the reference observations, such as in the austral spring of 2018.

Fig. 8.

Monthly time series of concentrations from the reanalysis CAMS and MERRA-2 and from ground-based observations for the following air pollutants: O₃ (a), NO₂ (b), PM_2.5 (c), PM₁₀ (d).

CAMS was able to reproduce the relatively high PM_2.5 concentrations in the austral winters of 2017 and 2018 but overestimated PM_2.5 in the winter of 2019 (Fig. 8c). Oscillations in the PM_2.5 and PM₁₀ concentrations observed in the austral summers of 2016 and 2017 were remarkably well reproduced by the CAMS reanalysis (Fig. 8c, d), consistent with the high correlation coefficients presented in Section 3.2. The CAMS dataset properly represented PM monthly concentration anomalies (refer to Section S2 in the Supplementary material). The correct reproduction of the PM variability indicates that the CAMS reanalysis can represent the variability of aerosol emissions and atmospheric removal processes in the MASP, at least at the monthly temporal scale. As discussed in Section 3.3, the MERRA-2 product failed to reproduce the basic features of PM seasonal variability. Likewise, the MERRA-2 PM time series showed a monthly variability clearly disconnected from the reference observations (Fig. 8c, d, and Supplementary Fig. S3).