For the definition of seasons, we follow the description already given in Sect. 1: FRS (15 May–14 July), LDS (15 July–31 August), and SRS (1 September–31 October). Figure 1 shows a map of the region of interest in SWA. Longitudinally we concentrate on the region 8∘ W–6∘ E, which avoids higher topographic features and contains all major cities along the Guinea coast such as Abidjan, Accra, and Lagos as well as Kumasi in inland Ghana. Our main region of study reaches 3.5∘ of latitude inland from the coastline (red border in Fig. 1), a distance for which we suppose to find aerosol effects on precipitation due to the fast northward transport of pollutants from the cities with the monsoon flow during the LDS. As during the SRS the monsoon flow weakens considerably, implying that pollutants will remain in the densely populated coastal plains, we concentrate on a smaller region, referred to as “coastal strip”, during this season (blue border in Fig. 1). The coastal strip only reaches 0.75∘ inland and is restricted to 8∘ W–4∘ E due to the strong curvature of the coastline to the east of Lagos.

Two different rainfall data sets are used in this study: first, the Climate Hazards Group InfraRed Precipitation with Stations CHIRPS, data set. It utilizes both high-resolution satellite imagery and in situ station data and provides, amongst other things, daily rainfall estimates over land on a 0.25∘ × 0.25∘ horizontal grid from 1981 onwards. CHIRPS uses an underlying static rainfall climatology CHPclim;, which contains input from historical station precipitation averages, historical thermal infrared satellite estimation averages, and a global topographic grid. The analyses are performed every 5 d (pentad). First, the pentad estimates based on thermal infrared are corrected through local regression between cold-cloud-top duration and Tropical Rainfall Measuring Mission (TRMM) 3B42 precipitation . In the second step, the product of the infrared estimates and CHPClim is bias-corrected using 5 d accumulated gauge observations. The disaggregation to daily values is done using infrared-based cold-cloud duration. Despite the product being relatively insensitive against changes in the station network (Chris Funk, personal communication, 6 July 2018) and a successful use in trend studies in SWA e.g., large changes in surface station availability (e.g. https://data.chc.ucsb.edu/products/CHIRPS-2.0/diagnostics/stations-perMonth-byCountry/pngs/Benin.003.station.count.CHIRPS-v2.0.png, last access: 8 May 2020) may induce inhomogeneities in long-term trends. This has been noticed by for western Uganda, who stress the necessity to validate satellite-derived trends by ground-based measurements. Furthermore, the algorithm tends to smooth spatial inhomogeneities.

Therefore as an additional source of precipitation data in the region, the Karlsruhe African Surface Station Database (KASS-D), containing daily, quality-controlled rain gauge measurements from manned weather stations operated by national weather services e.g., has been used in this study. Only stations with at least 50 % data coverage are considered to allow a meaningful trend analysis. As this criterion is hardly fulfilled before 1983 and after 2015, the main analysis is restricted to this period. In addition, the trend analysis is repeated for the more recent, shorter time span 2001–2017, for which surface observations of incoming solar radiation and more satellite data are available (see Sect. 2.4). Unfortunately, the availability of KASS-D data deteriorates during this period, mostly in Nigeria, where our database has few data after 2015. Despite this, we decided not to end this recent period in 2015 because trends become less meaningful for shorter time spans. Note that in the 1990s to 2010s, KASS-D contains daily data from many stations in SWA that have not been used in CHIRPS.

Visibility and low- and medium-cloud-cover data from the Met Office Integrated Data Archive System (MIDAS) are used from those stations in SWA where data are available for at least 50 % of the days in each season. Horizontal-visibility data are categorized into ranges of “below 10 km”, “10–20 km”, and “above 20 km”, for which time series and trends are calculated. Observations of surface downwelling shortwave radiation (SDSR) are available at Parakou (October 2001–June 2017) and Lamto (January 2001–May 2018). The instrument in Parakou was replaced in March 2009 and again in March 2014, potentially influencing trend calculations due to inconsistencies in the measurements. However, a shorter time series (2002–2015) from Djougou, about 100 km north-east of Parakou, indicates that the observations in Parakou are consistent (not shown). No visibility data are available for Lamto, but human observer estimates of total cloud area fraction (TCAF) are available for January 2000–July 2016. From these (sub-)daily observations seasonal and monthly averages were calculated, again with a 50 % data coverage criterion. For a more complete look at radiation, monthly data of SDSR and effective cloud albedo (ECA) of the SARAH-2 data set from the Satellite Application Facility on Climate Monitoring CM SAF; are analysed. The comparison of normalized trends (see Sect. 2.4) of SDSR and ECA allows us to estimate a residual potentially related to aerosol. This technique of normalization is also applied to observations from the surface stations in Parakou and Lamto.

Monthly satellite measurements of aerosol optical depth (AOD) on a 1∘ × 1∘ horizontal grid from the Moderate Resolution Imaging Spectroradiometer (MODIS) , i.e. the “combined dark target and deep blue AOD at 0.55 µm for land and ocean: mean of daily mean”, are used to calculate seasonal trends between July 2002 (beginning of data set) and October 2018. As these are monthly data, they are weighted when calculating season-averaged values. AOD for the LDS (15 July–31 August), for instance, is computed as AOD(LDS) = (0.5 × AOD(July) +1.0 × AOD(August))/1.5. Especially over land, clouds often inhibit AOD measurements from space, leading to many missing values in the MODIS monthly products. To cover the diurnal cycle, only months are used when AOD data from both the Aqua and Terra platforms are available, which are then averaged to obtain one single monthly value. For every year a seasonal mean value is computed if data are available for all months of the respective season. Again a sufficiently complete record with data for at least 50 % of the seasons between 2002 and 2018 is required before calculating a trend.

To give climatological context to the trend analysis presented here, relative humidity, cloud cover, and meridional wind speed data from the fifth generation of European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalyses of the global climate ERA5; are used. The influence of ocean temperatures on rainfall is analysed using the five most important SST-based climate indices for the region, namely Atlantic 3 (3∘ S–3∘ N, 0–20∘ W), Niño3.4 (5∘ S–5∘ N, 120–170∘ W), the coupled ocean–atmosphere Atlantic Meridional Mode, Indian Ocean (10∘ S–30∘ N, 50–90∘ E), and the Atlantic Multidecadal Oscillation index (Atlantic Ocean from 0–70∘ N), all as used by . The Atlantic Meridional Mode is based on the National Centers for Atmospheric Prediction/National Center for Atmospheric Research (NCEP/NCAR) reanalysis and the Atlantic Multidecadal Oscillation index on the Kaplan SST data set. The other three indices are computed from SST values from the Hadley Centre Sea Ice and Sea Surface Temperature (HadISST) data set. The same weighting as for the AOD data is used for the conversion from monthly climate indices to seasonal averages.

Long-term trends for seasonal rainfall totals are computed for every CHIRPS grid point and KASS-D station, applying Sen's slope method . It calculates slopes for each pair of 2 consecutive years in a time series, and the final trend is the median from all slopes. These trends are tested on statistical significance using the Mann–Kendall test . Generally, trends are considered statistically different from 0 if the two-sided p value is smaller than the tested significance level α. All our tests are performed for α values of 5 % and 20 %, the latter being a relatively weak criterion for statistical significance. As discussed in the context of “climate change attribution” by and , the choice of significance levels depends on how one intends to interpret the results. Small α values reveal high confidence that a trend found in the data has some other reason than natural variability. However, a less strict significance level (α=20 % in our case) can still be useful for a study like ours that is dealing with the challenge of a relatively short and incomplete data record subject to multiple influence factors and possibly non-linearities. In such a situation, choosing a large α implies reducing so-called “type II errors”, i.e. retaining the null hypothesis of no trend beyond natural variability, although such a trend actually exists but is hard to detect with the information at hand. Taking into account additional factors such as geographical distribution and seasonal behaviour can help in the evaluation of a trend with weak statistical significance. Ultimately, it is the balance of all available evidence that does or does not suggest that an identified trend has other-than-natural causes , and this is the philosophy we are following in this study. We feel that such an approach is particularly justified in the given situation as a risk assessment for a potential local human influence on rainfall in SWA is urgently needed.

Interannual variations in precipitation in SWA are, depending on the exact region and season, influenced by (remote) climate indices. In order to distinguish between these climatic signals and other parameters such as (local) air pollution, we calculate so-called “residual” time series through subtraction of a linear-regression model based on five climate indices defined in the previous section from the original time series. For each time series, 31 different linear models are constructed that consist of either single indices (i.e. five linear models) or any combination of the five indices (i.e. 10 combinations with two indices, 10 combinations with three indices, 5 combinations with four indices, and the one linear model with all five indices). For each model the leave-one-out cross-validation procedure for example is performed and evaluated calculating the residual sum of squares. Aiming to keep the model as simple as possible, the linear regression was constructed to consist of as few indices as possible to avoid (near) multicollinearity. However, if a combination of two indices yields a model for which the residual sum of squares is at least 3 % lower than that of a single-index linear model, the two-index model is chosen. A third, fourth, or fifth index is again only added if the performance of such a linear model increases by more than 3 % compared to a model consisting of fewer indices. Thresholds between 1 % and 5 % were tested for the performance improvement but all result in the same choice for the linear model at the majority of CHIRPS grid points and stations.

The optimal linear model determined this way and its correlation coefficient with the original rainfall time series are shown in Fig. 2 for the LDS and SRS during the 1983–2015 period. During the LDS a simple linear model built on the Atlantic 3 index alone yields best results for a large majority of CHIRPS grid points analysed (yellow in Fig. 2a). This is consistent with previous results for a slightly larger region along the Guinea coast and a longer season from June to September . Exceptions are (a) the Kwahu Plateau in south-western Ghana (combination of Atlantic 3 with the coupled ocean–atmosphere Atlantic Meridional Mode; green), (b) central Ghana and the far south-west of Ivory Coast (combination of all five climate indices; pink), and (c) northern Benin and adjacent Nigeria (Atlantic Multidecadal Oscillation index; orange), but this region consists of only five grid points. The agreement between CHIRPS and the stations is largely good, with the exceptions of Kara in Togo (Atlantic Multidecadal Oscillation index; orange), Sokodé in Togo (Niño3.4 index; turquoise), and Warri in Nigeria (Indian Ocean; blue). We are not aware of any convincing physical reasons for these deviations and therefore mostly consider them to be statistical fluctuations. This is consistent with Fig. 2b showing correlations with the best models to be above 0.5 through most of the region, largely irrespective of model choice. Only the orange points from Fig. 2a at the far north-eastern fringe of the study region show a lower correlation, indicating problems with data or stronger terrestrial forcings farther away from the ocean. The simple picture that emerges from this analysis is that a considerable part of the season-to-season variability is linearly controlled by the SST over the nearby Atlantic Ocean, with warmer waters leading to more rainfall. This is related to a weaker temperature and pressure gradient towards the Sahel, which leads to stronger convergence and higher moisture content in SWA .

The picture during the SRS is also dominated by the Atlantic 3 index, but in the northern and especially eastern parts of the region, other climate indices yield better linear models (Fig. 2c). Given the weaker monsoon flow during this season, this may indicate a less close link to the nearby Atlantic Ocean and more room for teleconnections to influence rainfall. Correlation coefficients are generally lower than during the LDS but still reach 0.5 across considerable parts of the region, particularly along the coastal strip (marked by a black line in Fig. 2d). It is striking that the regions that do not show the simplest model based on the Atlantic 3 index also show the lowest correlations, which again points towards data problems and/or more local influences such as topographic forcing by the Oshogbo Hills in Nigeria.

During the FRS, the Atlantic 3 index is also dominating, but regions where other indices yield better linear models are larger than during the other two seasons (Fig. S1a). The correlation of the best models with rainfall is low during the FRS, rarely exceeding values of 0.5 (Fig. S1b) and showing a rather weak effect of ocean temperatures on rainfall during this season. As before, the agreement between station and CHIRPS data in both the optimum index combination and the correlation is very good.

The optimal linear models described above are now used to remove the dominant influences of SST fluctuations from the original time series. Trends in these residual time series will differ from those computed from the original time series, and we refer to them as “residual” trends and “full” trends, respectively. Trends and their statistical significance are also calculated for all other observed parameters described in Sect. 2.3, namely visibility, clouds, radiation, and AOD, but no residual time series are calculated in these cases.

Given the issues with aerosol observations in SWA described above, we have to rely on indirect indicators to determine changes in air pollution. These are observations of cloud cover, radiation, and horizontal visibility. As there is not sufficient information for even the simplest radiative transfer estimates, we chose to estimate a potential aerosol effect on radiation through considering normalized trends. In this context, normalizing means subtracting the mean and dividing by the standard deviation of the respective time series. The dimensionless trends in such normalized time series can be directly compared to each other assuming no significant changes in variance during the considered time period (stationarity). If surface radiation was entirely determined by cloud cover, identical but opposite normalized trends would result. A non-zero sum of the normalized trends points to an additional aerosol effect and/or changes in cloud optical thickness, which could at least in part be related to aerosol. Positive sums (e.g. slight reduction in radiation but strong increase in cloud) would then indicate reduced aerosol, while negative sums (e.g. strong reduction in radiation but only slight increase in cloud) point to increased aerosol. This concept is applied to the time series of CM SAF satellite data and surface observations from Parakou and Lamto in Sect. 3.3.

Another data set we use as an indirect aerosol indicator are surface observations of horizontal visibility as estimated by human observers. Trends in this parameter are most likely an effect of changes in the aerosol burden, but changes in low-level humidity and clouds may also have an effect as they change incoming light and the split between direct and diffuse radiation.

Figure 3 shows the time mean seasonal evolution of key variables based on ERA5 and CHIRPS rainfall averaged over the main study region shown in Fig. 1. The main seasons as used in this paper are delineated by vertical black lines. During the main dry season from December to February, rainfall drops to values well below 1 mm d-1 followed by a gradual ramp-up towards the FRS starting in mid-May. Peaks during the two rainy seasons FRS and SRS reach about 8 mm d-1, while rainfall drops to below 4 mm d-1 during the LDS. The monsoonal retreat in October and November occurs much more rapidly than the onset.

recently developed a seasonal climatology of rainfall types over the region based on satellite imagery and reanalysis data. They found that the location of the African easterly jet close to the Guinea coast during the FRS creates enhanced vertical wind shear, leading to the highest degree of convective organization into long-lived mesoscale clusters during the year (marked with ++ in Fig. 3). In contrast, during the SRS, the vertical profiles of shear and instability lead to a more local convective triggering, for example by the land–sea breeze, and therefore smaller and shorter-lived but often intense systems (+ in Fig. 3). Conditions during the LDS are significantly different. During this period, SWA is subject to a strong southerly monsoon flow, leading to fast advection of thermodynamic and chemical properties from the Gulf of Guinea, typically along a low-level jet that forms at night. Rainfall during the LDS is often due to relatively weak, isolated, short-lived, and weakly organized convection ∘ in Fig. 3;, mostly occurring in the afternoon. Land–sea breeze days have a seasonal minimum , and stability is often too high to allow the transition from cumulus clouds in the coastal hinterland into deep cold convective clouds in the course of the afternoon.

The arrows in Fig. 3 show average wind profiles for the three seasons of interest. Clearly the monsoon flow is strongest in the LDS, reaching meridional wind speeds exceeding 4.5 m s-1 at 950 hPa with southerly winds up to 875 hPa, marking the depth of the monsoon layer. The FRS is also characterized by marked southerlies and an even deeper monsoon layer, while the SRS shows weaker flow reduced to a shallower layer. This is consistent with , who investigated upper air data at the coastal station of Cotonou. As already discussed earlier, we anticipate that this should restrict the largest aerosol effects to the coastal strip, which contains the main pollution sources. The higher rainfall in SRS as compared to the LDS should also lead to more wet deposition, which would further support the concentration to the vicinity of sources. Wet deposition will of course also be enhanced during the FRS.

Another parameter of interest is the frequency of relative humidity of 95 % and higher (light-blue curves in Fig. 3) as this determines the potential for wet growth of aerosol particles, which strongly enhances radiative effects . Here large differences are seen between the deep layer of very moist conditions in LDS as compared to the much shallower layers in the rainy seasons, with maxima at 950 hPa (FRS) and even 975 hPa (SRS). Qualitatively, these results suggest that during the LDS the potential for wet aerosol growth is higher and that considerable aerosol effects could in fact be spread over a much larger region through the combination of faster transport and less wet removal than in the other seasons. For the LDS and the SRS, time series of the frequency of relative humidity of 95 % and higher (Fig. 4) reveal slight decreases over 1983–2015 of about 0.1 % yr-1, which are significant only on the α=20 % level. Given the highly non-linear dependence of wet growth of aerosol particles on relative humidity e.g., this relatively small change implies that the radiative effect per amount of dry aerosol will be smaller in the later part of the study period. As our estimates for aerosol trends rely on horizontal visibility and surface radiation, the change in dry aerosol may be even larger than deduced from these parameters alone. The trend of relative humidity exceeding 95 % during the FRS is indistinguishable from 0 (Fig. S2 top).

Finally the red curve in Fig. 3 shows the mean vertical profile of cloud cover at 06:00 UTC (corresponds to local time in the study region), the analysis time closest to the diurnal maximum of low clouds . These clouds are crucial for the radiative energy budget and, thus, for vertical mixing and the triggering of convection in the afternoon . We anticipate this factor to influence the significance of aerosol changes to the surface energy balance when incoming radiation is strongly reduced by cloud cover. Peak coverage at about 950 hPa differs only little between the three seasons, but in the LDS the cloud layer is somewhat thicker.

For the remainder of the paper, the analysis concentrates on the LDS and the SRS, for which we found the largest indications for aerosol effects. The corresponding results for the FRS are provided in the Supplement and are referred to along the way. Figure 3 offers relatively little in terms of potential reasons of fundamentally different behaviour in the FRS, so we hypothesize that the much larger degree of convective organization during this season found by is key in reducing sensitivity to local aerosol effects. In addition, spatial patterns of positive and negative rainfall trends during FRS change depending on the time period analysed (cf. Fig. S3a, b with S3d, e), and there are more pronounced discrepancies between CHIRPS and station data. Regionally averaged trends are weak and statistically not significant on the 20 % level (Fig. S3c, f).

In the previous section, we have demonstrated that, once influences of SST changes are corrected for, large parts of SWA have undergone an accelerating drying over recent decades. Seasonal and geographical patterns, together with the acceleration, are consistent with the hypothesis of a human influence through rapidly growing emissions of pollutants. What other evidence do we have to support this idea? As we discuss in more detail in the next subsection, usable aerosol measurements are largely restricted to the ocean adjacent to SWA, impeding the establishment of a direct link to rainfall. Therefore we turn here to indirect indicators such as horizontal visibility (Sect. 3.3.2) and SDSR (Sect. 3.3.3), which need to be regarded in concert with cloud cover that influences both quantities.