Table  summarizes the location, type of instrument and observation periods. We have used different types of sun photometers, automatic and handheld. The automatic CIMEL sun photometer is the reference instrument used in the AERONET network for measuring the AOD and retrieving columnar aerosol optical properties and size distribution. We have used the level 2 quality-assured daily averages processed with the version 3 of the aerosol-optical-depth algorithm . We used the data for Ghana (station named Koforidua_ANUC located at 6∘6′ N, 0∘6′ W), Nigeria (station named Ilorin located at 8∘29′ N, 4∘40′ E) and Côte d'Ivoire (station named Lamto located at 6∘13′ N, 5∘2′ W). The geophysical station of Lamto was equipped early on, in 1997–1998, then the automatic sun photometer was restored back in 2017.

Handheld sun photometers are a well-known scientific instrumentation for measuring atmospheric transmission . The first type of handheld photometer we used is the one manufactured by CIMEL, hereinafter called the HHC. The HHC was operated during 2 years between April 2006 and March 2008 at Lamto geophysical station. The operating wavelengths are 440, 670 and 870 nm. The second handheld sun photometer is a new lightweight instrument manufactured by TENUM (http://www.calitoo.fr, last access: 5 February 2021) and named CALITOO . CALITOO operating wavelengths are 465, 540 and 619 nm. The sun photometer measures the sun irradiance at the three wavelengths, so no additional check on the AOD curvature can be applied; however the spectral consistency between the AODs (observed at 540 nm and computed using the Ångström exponent) is checked. The atmospheric optical depth is then retrieved following the Beer–Lambert law, knowing the calibration constant for each instrument and the relative air mass. The AOD is then retrieved after subtracting the Rayleigh and trace gas optical depth.

For the HHC, observations were acquired twice a day at around 09:00 and 15:00 UTC. For the CALITOO sun photometer, the observations were acquired at around 13:00 LT. The operators were asked to make measurements only when the sun was not obscured by clouds and have proceeded with a sequence of five measurements within about 15 min. The presence of sub-visible cirrus or broken clouds within the field of view induces spurious variation in the atmospheric transmission that can be easily detected by looking at the standard deviation of the 15 min series of AOD measurements. An arbitrary threshold of 0.2 on the standard deviation has been selected to remove the cloud-contaminated observations. The diurnal variability range is expected to be less than 10 % for our site conditions . The sun photometer observations are reported as daily averages.

The total uncertainty in AOD for the AERONET instruments is ±0.01 for λ>440nm and ±0.02 for shorter wavelengths . CALITOO sun photometers were calibrated prior to the site deployment using the Langley plot method at the Izaña high-altitude observatory . CALITOO observations were compared to coincident AERONET observations before and after the field experiment. The total uncertainty in AOD is estimated to ±0.02 for all the wavelengths. Post-field measurements indicate a change of about 1 % yr-1 in the calibration.

AOD measurements are all reported at 550 nm because this wavelength is a reference for visibility calculation and satellite missions e.g., . The Ångström exponent (AE) is computed between wavelengths 465 and 619 nm for the CALITOO, 440 and 670 nm for the HHC, and 440 and 675 nm for the AERONET.

The Moderate Resolution Imaging Spectroradiometer (MODIS) aerosol products have been widely used by the scientific community for assessing the impact of aerosols on global climate or air quality . The MODIS AOD is also used in operational data assimilation for weather forecast . The MODIS level 2 product has a spatial resolution of 10km×10km at nadir, which increases to 20km×40km at the edges of the swath. The MODIS level 3 is a regular gridded product having a spatial resolution of 1∘×1∘. Most of the time, the validation exercise of MODIS-derived aerosol parameters consists of a comparison between sun photometer observations and MODIS level 2 pixels co-located in space and time . A box of 5×5 MODIS pixels and a time slot of ±30 min around the satellite overpass is considered to be a good compromise; however the window-size dependence is small, and this compromise is more dictated by statistics rather than physics .

Gridded daily or monthly mean MODIS level 3 AODs have also been demonstrated to fit the AERONET retrievals . As the objective of the paper is to address the ability of MODIS data to reflect aerosol changes in a specific area rather than the validation of the retrieval algorithm, hereinafter we rely on the gridded level 3 MODIS product of the AQUA satellite (namely MYD08_D3) from 2003 to 2019. MODIS AQUA has been selected following the recommendations of for long-term trend analysis. This product provides several values for the AOD depending on the underlying surface and the algorithm used. For the sake of consistency between the different sites, we use the product named AOD_550_Dark_Target_Deep_Blue_Combined_Mean from version 6.1 of the MODIS processing algorithm, which is a combination of the “Dark Target” and “Deep Blue” methods. For the coastal sites, both AOD over land (namely Aerosol_Optical_Depth_Land_Mean or Deep_Blue_Aerosol_Optical_Depth_550_Land_Mean) and over ocean (Aerosol_Optical_Depth_Average_Ocean_Mean) are also provided. We use the “Deep Blue” AE (namely Deep_Blue_Angstrom_Exponent_Land_Mean) and compute an Ångström exponent from the spectral AODs over land and ocean, respectively. For the purpose of satellite validation, the satellite AOD and AE of the nearest cell to the photometer location are extracted. We have adopted the evaluation metrics proposed by , including the linear correlation coefficient, the median bias, the root mean square error, the mean absolute percentage error and the fraction of data falling within the MODIS expected error (EE) given by EE=±(0.05+0.15×AOD).

From February 2015 to March 2017, Abidjan and Cotonou were equipped with PM2.5 monitoring stations . Particles were collected on 47 mm diameter filters (quartz and PTFE filter types) at a flow rate of 5 Lmin-1. Samplers were equipped with a PM2.5 mini Partisol impactor. PTFE filters were weighted before and after the sampling with a Sartorius MC21S microbalance.

The total volume of filtered air is measured by a GALLUS-type G4 gas meter. Mass concentrations of PM2.5 are estimated from the mass load of particles on the filters and the total volume of air. The exposure duration of the filters is 1 week. A period of 1 week is sufficient to capture the main temporal pattern of atmospheric aerosols over a long period of time and has been selected as a trade-off between logistics and observations.

A total of 2323 handheld-sun-photometer observations (including data collected during the 2006 campaign) have been acquired. Starting and ending dates are reported in Table  along with the number of observations, median and interquartile range (IQR) of AOD and AE distributions. We select the AERONET data until the end of the CALITOO observation period, i.e., March 2017, for a total number of 1248 daily observations. There is an excellent time coverage for the stations of Lamto and Cotonou by the CALITOO observations. Observations were performed for 66 % of the time in Cotonou and 68 % in Lamto. As a comparison, this rate is 68 % for the automatic sun photometer in Koforidua, indicating that handheld measurements can be as representative as the automatic ones. This rate drops to 24 % in Abidjan and 39 % in Savè due to operating issues leading to gaps in the time series.

Considering all the stations, the AOD ranges between a minimum of 0.07 and a maximum of 3.76. The highest AOD acquired by the CALITOO instrument is 3.50 in Cotonou in March 2015, and the highest AOD recorded by the AERONET sun photometer is 3.76 in Ilorin in December 2016. The median AOD ranges between a minimum of 0.47 at Lamto and a maximum of 0.66 at Comoé. Considering all the daily measurements for all the sites, the median AOD is 0.52, and the IQR is (0.33, 0.82). AOD observations at Cotonou and Abidjan are rather similar, with a median AOD = 0.55 (0.38, 075) and 0.58 (0.35, 0.86), respectively. The observations with the automatic sun photometer in Koforidua show AOD in the same range as the two aforementioned stations, with a median AOD = 0.49 (0.33, 0.84). The observations performed at Lamto with the three kinds of sun photometers show similar AOD: 0.47 (0.3, 0.72), 0.55 (0.35, 0.80) and 0.56 (0.38, 0.85) for CALITOO, the HHC and AERONET, respectively. The difference in the statistics for the three instruments at Lamto is due to different sampling periods (see Table ), although the coincident observations between CALITOO and AERONET in 2017 show an excellent agreement (see below).

The median AE is between a minimum of 0.33 (0.13, 0.55) at Comé and a maximum of 0.78 (0.56, 1.09) at Koforidua. Observations performed with CALITOO at Lamto show a slightly lower range of AE values than the ones performed with the HHC. The difference between CALITOO median AE and HHC median AE is -0.09. However it should be noted that the statistical distribution of AOD values has an impact on the corresponding AE distribution . AOD observed at Lamto during the HHC period (2006–2008) being higher than the ones observed during the CALITOO period (2014–2016), it is not expected to have the same AE values. Nonetheless, the difference is within the expected accuracy.

Coincident AERONET and CALITOO observations (N=31) were acquired between January and March 2017 at Lamto (see Fig.  and Table ). Figure  shows the scatterplot for the corresponding daily AOD and AE There is an excellent agreement for both AOD (regression coefficient R=0.93) and AE (R=0.87) between the two instruments.

The daily AODs and AEs for each site and each instrument between 2015 and 2017 (between 2006 and 2008 for HHC) are presented in Figs.  and , respectively. A similar seasonal pattern is observed in the different time series. There is an increase in AOD during the main dry season (December to March) and a decrease during the rainy season (April–July). The 2-week smoothing average reveals a high degree of correlation between time series. The correlation coefficient between Cotonou and Lamto time series is R=0.82, with R=0.90 between Cotonou and Koforidua. During the short overlap period in March 2017, the CALITOO and AERONET instruments show similar AOD values at Lamto station. The Comoé time series is the weakest one, with only 82 data points. The 2006–2008 HHC Lamto time series has the same pattern as the one recorded by CALITOO in 2015–2017 and shows two maxima in the dry season, one in December and another one in January–February.

The seasonal pattern for AEs (Fig. ) shows an opposite cycle, with lower values in the dry season and higher values during the rainy season. AE seasonal cycle is clearly affected by the winter dry period, with dust-laden air masses that decrease the AE values. The median AE value during the first half of the year (all sites except Comoé) is 0.36 (0.23, 0.62) and 0.69 (0.43, 1.00) during the second half of the year. AODs in the last quartile (AOD above 0.82) are mostly (72 %) observed during the months of December, January and February and are associated with a median AE of 0.44 (0.24, 0.64), while low AODs (first quartile, i.e., below 0.33) are associated with a median AE of 0.89 (0.61, 1.12) and are observed between August and October (51 % of the observations). The difference in AOD between the inland and the coastal sites is less than 0.05, with differences up to 0.1 between April and June, the AOD at the coastal stations being higher than inland. AEs are higher at the coastal stations than at the inland stations by 0.15 on average, reflecting the influence of urban air pollution at the coastal stations.

Table  gives the statistics of the regressions for each site and per instrument presented in Fig. . We have then adopted a log–log representation on the scatterplots presented in Fig.  as the AOD distribution has a significant right skewness . Figure  also presents the MODIS expected error (EE; blue lines). Whatever the site is, there is a significant correlation between the MODIS and sun photometer observations. The Pearson correlation coefficient R ranges between 0.75 (Comoé) and 0.94 (Koforidua). For the CALITOO observations, R is between 0.75 and 0.90 (Cotonou). The lowest RMSE values are found for the measurements operated using the CALITOO at the coastal sites of Abidjan and Cotonou. The mean absolute percentage error (MAPE) is on average 30 %. The sites in Cotonou and Abdijan are not biased and present a fraction of data falling within the MODIS EE above 60 %. All the inland sites are biased, and it results in a rather low fraction of data falling within the MODIS expected error.

The bias has a seasonal behavior that is highest during the dry season, between December and March. An underestimation of the MODIS AOD is then observed at maximum in January, with an absolute bias of -0.33 (relative bias of 39 %) at the inland sites. have already pointed out the possible differences in the “Dark Target” and “Deep Blue” algorithms. It appears from Fig. 6 in that the dry-to-humid-savanna transition zone in SWA is an area where large differences exist in both retrieval techniques during the dry season. Those differences can explain that the “merge” product used in this study has a large bias during the dry season in the northern part for the inland sites, so the north–south AOD gradient in this area remains difficult to assess based on satellite products.

For all the sites considered in this study and whatever the sun photometer is, the correlation between MODIS AE and sun photometer AE is non-significant. This finding is coherent with the results of and . The histograms of the sun photometer and MODIS-derived AE are presented in Fig. . The default values AE = 1.5 and AE = 1.8 in the MODIS AE Deep Blue product have been removed . The MODIS median AE is biased by 0.32, and the distribution of AE values does not follow a normal distribution, whereas the distribution of AE values for the sun photometers does. At the coastal sites, the MODIS AE Ocean algorithm reproduces well the left side of the histogram (lowest AE), while it significantly underestimates AE above 0.8. MODIS AE Land algorithm at 0.5 indicates that a dust-like aerosol model has been selected ; however there is no systematic association with low sun photometer AE. However it could be possible to adapt the interval bounds (lower and upper limits of AE values) for each category; the statistical distribution of MODIS AE values does not fit the sun photometer ones and could lead to misclassification of daily observations.

AE is an intensive aerosol optical parameter and depends on the spectral aerosol extinction coefficient . AE is influenced by the aerosol size distribution and is commonly used to identify aerosol types . Aerosol types having a dominant fraction of their size distribution in the coarse mode, like dust and sea salt particles, are associated with a lower value of AE than aerosol types with a size distribution dominated by the accumulation mode, like secondary and combustion aerosols. The concurrent changes in AOD and AE help to distinguish generic aerosol types in sun photometer time series . Mineral dust tends to increase atmospheric AOD and decrease AE , while biomass-burning events tend to increase both AE and AOD .

The AOD vs. AE scatterplot can be used to cluster the observations by aerosol broad categories corresponding to a main source, like coarse mineral dust of biomass-burning aerosols. The thresholding in AOD and AE for aerosol type identification varies from one site to another and also depends on the distance from aerosol sources upwind of the site . In particular, the classification based on AOD vs. AE values is incapable of determining aerosol absorption properties . Figure  presents the scatterplots of AODs (log-scale) vs. AEs for each site and split by seasons. We have considered four seasons corresponding to the long dry season (December–March), the long wet season (April–June), the short dry (August–September) and short wet season (October–November). For the sites with the most comprehensive data set over the different seasons (Lamto, Cotonou, Koforidua, Ilorin) the AOD vs. AE plots show a similar pattern, with AODs decreasing almost linearly as AE increases. The lowest AEs are observed during the long dry season and are associated with the largest AODs, indicating the presence of coarse mineral dust. The presence of dust can be also observed in the long wet season. During the short dry season, all the sites excepted Comoé show larger AEs and lower AODs than for the other seasons.

It can be noticed from Fig.  that there is no clear definition of AOD and AE thresholds for each aerosol category, and the scatterplots of Fig.  reflect the high mixing of different aerosol types. The absolute error in AE is a function of the relative error in AODs and depends on the spectral range investigated . Typical error in AE is ±0.3 for an AOD of 0.2, and there is a risk of overinterpreting AE variations.

In this paper we classify the daily observations according to the AE values using a simple statistical analysis and a threshold on AOD. The whole sun photometer data set is divided into three quantiles. The first third corresponds to AE≤0.45, and observations having an AOD≥0.8 are labeled “coarse dust”, while observations having an AOD < 0.8 are labeled as “mixed”. The threshold on AOD corresponds to the third quantile of AOD distribution and is used to better identify dust events. The last third corresponds to AE≥0.80 and is labeled “urban-like”. The data having 0.45 < AE < 0.80 fall into a “mixed” category, being more populated than the two others. This rather crude classification enables us to identify the main aerosol influence with a significant number of observations in each category.

Table  presents the typology of the sites according to the aforementioned classification. Comoé is the most influenced by coarse dust aerosol (20 %), followed by Ilorin (12.7 %), Savè (12.2 %), Cotonou (10.2 %) and Lamto (8.6 %). The southwestern sites, Abidjan (3.7 %) and Koforidua (4.9 %), are less influenced by dust events than the eastern sites. As expected from Fig. , all the sites show the “urban-like” category that also corresponds to low AODs. Two sites are less influenced by urban-like aerosols than the others, namely Savè (20 %) and Comoé (8.5 %). For the other sites, the “urban-like” category ranges between 47.7 % (Koforidua) and 32.2 % (Cotonou).

As SWA lacks dedicated studies on aerosol characterization, there are few other data to compare with. have proposed a sophisticated aerosol classification for Africa based on AERONET observations; however none of the sites are located in our area. have classified Djougou (northern Benin, located at 9∘42′ N) as a dust site that is seldom affected by biomass burning. Savè and Ilorin are located around 200 km south of Djougou, and the influence of dust is still significant compared to the coastal sites. Comoé is located 600 km westward of Djougou and probably less influenced by the dust transport from the Bodélé area in Chad; however measurements acquired at Comoé do not cover a full season, and the exact frequency of dust or biomass-burning events remains uncertain.

The changeover between the monsoon and the harmattan results in a change in the vertical distribution of aerosol layers and in the type of aerosols . The harmattan flow carries continental aerosols in the lowest part of the atmosphere during the dry winter season (December to March). During the dry winter season the days with high AOD are often associated with an increase in the PM2.5 surface concentration, leading to a high correlation coefficient between AOD and PM2.5.

The correlation coefficient between weekly mean AOD and PM2.5 measured in Cotonou and Abidjan is R=0.75 (N=105) when considering the whole observation period. The correlation coefficient can reach R=0.96 (N=6) during specific aerosol events observed from December 2015 to January 2016 in the heart of the dry season. During other periods of the year, the correlation remains weak because the concentrations are less fluctuating than during the winter period.

PM2.5/AOD ratios are estimated using the daily AOD observations and the weekly PM2.5. The PM2.5/AOD is basically the amount of PM2.5 that is expected per unit of AOD. It was first promoted by as a conversion factor . The PM2.5/AOD ratio reflects how a change in the AOD affects the ground surface concentrations; however there is no evidence of a unique relationship between both quantities. The PM2.5/AOD ratio depends on the vertical stratification of the aerosol layers in the atmosphere due to mixing processes in the boundary layer or large-scale advection . The ratio depends also on the aerosol size distribution and chemical properties that are changing during the transport and the aging of the aerosols.

In the specific case of the coastal cities of the Gulf of Guinea, we are interested in evaluating how the change in aerosol type during the season, and in particular the seasonal advection of mineral dust from the desert area, may affect the PM2.5 surface concentrations. For this purpose, we estimate a PM2.5/AOD ratio per aerosol type and per season.

Each daily AOD observation is associated with an aerosol type (coarse dust, mixed or urban-like) depending on the corresponding daily AE value. The daily AODs are associated with the corresponding PM2.5 observation using Eq. (). 1PM2.5weekly=1n∑i=1n∑t=13βt,sτi,t,s The corresponding PM2.5/AOD coefficients – βt,s in Eq. (), where t represents the aerosol type and s the season – are evaluated using a multilinear regression on the observations collected in Cotonou and Abidjan for each season independently. Cotonou and Abidjan samples are pooled together to increase the statistical significance and to retrieve average coefficients at the regional level. As the season are not equal in length, and the number of observations differs in Abidjan and Cotonou, the number of samples differs, ranging between 71 samples during the long dry season and 24 samples during the short dry season. The significance of the regression and standard error in the coefficients depends on the number of samples. None of the weeks in the short dry period are affected by dusty days, so the coefficient for coarse dust is not retrieved for this period. During the short wet season, only 2 weeks over 26 have a dust contribution, and the coefficient is not significant. The PM2.5/AOD ratios by aerosol category are presented in Fig.  as a function of the season and with their respective uncertainties. The average PM2.5/AOD ratio without accounting for the aerosol category for a given season is also reported. The uncertainties correspond to the standard error in the coefficients found by regression. The standard error depends on the occurrence of an aerosol category and its relative weight in a given season. For all the seasons the coefficients for each aerosol type are significant (p<0.05), except for the coarse dust category during the short dry (no data) and the short wet season. The resulting adjusted coefficient of determination for the regression is between 0.76 (long dry season) and 0.83 (short wet season).

The PM2.5/AOD ratio for coarse dust aerosols ranges between 54±8 µgm-3 per unit AOD in the long dry season and 20±4 µgm-3 per unit AOD in the long wet season. The seasonal changes in the PM2.5/AOD ratio for coarse dust reflects well the vertical shift in the dust layer between the dry and the wet season. During the wet (April–July) season, the air masses are uplifted by the monsoon flow. PM2.5 concentrations remain moderate (21 µgm-3 in April), while AODs are still significant (0.57 on average in April) due to the aloft transport. The impact of coarse dust on PM2.5 is higher during the dry season (higher ratio and high AODs), when the dusty air masses are advected close to the ground surface.

The PM2.5/AOD ratio for mixed aerosols ranges between 53±7 µgm-3 per unit AOD in the long dry season and 27±11 µgm-3 per unit AOD in the short dry season. During the short dry season, only 30 % of the weeks are affected by mixed-type aerosols, the remaining being classified as urban-like. The corresponding PM2.5/AOD ratio for mixed-type aerosol is close to the one found for dust during the previous season (long wet), indicating that the aloft dust transport can be still active but incorrectly classified to the mixed aerosol type due to a low intensity (small AOD).

The PM2.5/AOD ratio for urban-like aerosols ranges between 96±15 µgm-3 per unit AOD in the short wet season and 37±5 µgm-3 per unit AOD in the long wet season. The PM2.5/AOD ratio retrieved in the long dry season for the urban-like category is affected by a larger uncertainty due to a limited impact of urban-like aerosol during the long dry season compared to coarse dust. There is a shift in the PM2.5/AOD ratio toward higher values during the short wet season. The short wet season is a transition period during which the stagnation of air masses over land favors the accumulation of pollutants and also combustion by-products emitted over Nigeria .

Figure  presents the weekly average AODs and satellite-derived and in situ PM2.5 for both Abidjan and Cotonou. The label attributed to each week corresponds to the aerosol type having the largest mean AOD over the week. The period from March to May is dominated by the coarse dust type, and there is a clear shift to urban-like type in June–July. A second period of coarse dust is observed in December (2015 and 2016) and is associated with a significant increase in both AOD and PM2.5. PM2.5 during the dusty period of December rises over 100 µgm-3. Another sharp increase is observed in February and is associated with the mixed aerosol type. For both years, the two intense periods (December and February) are separated by an interim period showing moderate PM2.5 and AOD and classified as urban-like aerosols.

On average, satellite-derived PM2.5 agrees with the in situ PM2.5 observations. Indeed the mean difference between retrieved and observed PM2.5 during the 2015–2016 period is less than 1 µgm-3 (3 %). The MAE is 14 µgm-3, and the RMSE is 21 µgm-3. The RMSE found here is within the range of previous studies for other regions of the world and different algorithms. The very intense periods are underestimated; e.g., the mean difference between retrieved and observed PM2.5 is -51 µgm-3 (-70 %) in December 2015 in Cotonou and nearly a factor of 2 lower in December 2016. The satellite-derived concentrations in January and in March are overestimated. Despite introducing a characterization of the aerosol type, there is still a clear smoothing effect on the weekly concentrations that results from the adjustment of the regression coefficients on a seasonal basis. Using seasonally adjusted coefficients only rather than the seasonally adjusted and aerosol-type-adjusted coefficients has a limited impact on the comparison and decreases the RMSE by only 2 µgm-3 on average. The biggest impact is during the long wet season (20 % decrease in RMSE), when a lower PM2.5/AOD coefficient is selected for the identified dust cases.