The approach used to retrieve aerosol and cloud properties from satellite spectral radiance measurements relies on the colour-ratio effect (Jethva et al., 2013). The signal backscattered by a liquid cloud is almost spectrally neutral from the UV to the near-infrared (NIR) ranges. Conversely, the absorption from biomass burning aerosols is typically larger at shorter wavelengths. Therefore, the presence of absorbing aerosols above clouds modifies the apparent colour of clouds. This enhancement of the spectral contrast can be detected by any passive remote-sensing instrument with two channels with enough separation in the UV–NIR region. The SEVIRI instrument, aboard the MSG satellite (Aminou et al., 1997), has channels centred at 0.64, in the visible, and at 0.81 µm, in the NIR ranges. Figure 1 plots the 0.81 µm radiance (R0.81) against the ratio of the 0.64 to 0.81 µm radiances (R0.64/R0.81), for absorbing aerosols above clouds over an ocean surface for several aerosol and cloud optical thicknesses. Throughout this paper, the radiances R refer to the normalized quantity as defined by Herman et al. (2005) and the optical thicknesses (i.e. AOT, COT) are given at 0.55 µm. The simulations have been performed with the adding–doubling method (De Haan et al., 1987), considering a viewing geometry of 20∘ for the solar zenith angle, 50∘ for the viewing zenith angle and 140∘ for the relative azimuth. The cloud is located between 0 and 1 km and the aerosol layer is between 2 and 3 km. Aerosols have a refractive index of 1.54–0.025i and the size distribution follows a lognormal with a geometric mean radius of 0.1 µm. The cloud droplets have an effective radius of 10 µm. Rayleigh scattering has been accounted for but the simulations do not include the absorption from atmospheric gases. A Lambertian surface with an albedo of 0.05 is assumed. For AOT = 0, the radiance ratio is around 1 and weakly depends on the COT. As the AOT increases, the radiance at 0.81 µm as well as the radiance ratio decreases, indicating that the attenuation from the aerosol layer is larger at 0.64 µm. This attenuation is mainly due to the absorption from the aerosol layer, which means that it is primarily correlated to the absorption AOT (AAOT).

As in the Nakajima and King technique (1990), the sensitivity of the retrieval to the CER comes from the measurements of the shortwave infrared (SWIR) channel of SEVIRI centred at 1.64 µm. Figure 2 shows the radiances at 0.81 and 1.64 µm for several COTs and CERs as well as the impact of overlying absorbing aerosols. The simulations without aerosol are plotted in blue and represent the signal typically used by cloud property retrievals that do not include light absorption from overlying aerosols. The orange and red grids are associated with an AOT of 0.5 and 1.5 at 0.55 µm. Compared to the no-aerosol case, these grids are shifted towards the upper left, which means that the presence of aerosols decreases the NIR radiance and increases the SWIR signal. As highlighted by Haywood et al. (2004), not taking into account the aerosol absorption above clouds leads to low biases in both the COT and the CER. These biases depend on the aerosol loading as well as on the brightness of the underlying cloud.

Although the aerosol microphysical properties have some influence on the signal measured by satellites, this kind of approach requires us to assume an aerosol model. Fundamentally, the algorithm developed here aims to retrieve the above-cloud AOT, the COT and the CER from the magnitude and the gradient of the radiances measured by SEVIRI at 0.64, 0.81 and 1.64 µm using a basic look-up table (LUT) approach and appropriate assumptions about the aerosol model for the region (Haywood et al., 2003) that have been refined based on measurements from the CLARIFY-2017 observational campaign (Zuidema et al., 2016).

The SEVIRI channels chosen for the retrieval are fairly standard in atmospheric science and have been widely used for aerosol and cloud analysis (e.g. Brindley and Ignatov, 2006; Thieuleux et al., 2005; Watts et al., 1998). However, the SEVIRI bandwidths are much larger than other state-of-the-art instruments such as MODIS. Hence, SEVIRI radiances are significantly more impacted by the absorption from various atmospheric gases. The spectral response functions for the 0.64, 0.81 and 1.64 µm SEVIRI channels are plotted in Fig. 3 together with the equivalent MODIS bands. The main absorbing gases in these spectral bands are ozone, water vapour, methane and carbon dioxide, gases which are typically produced and transported within biomass burning plumes (Browell et al., 1996; Koppmann et al., 2005). The contributions of each gas to the atmospheric absorption are also shown in Fig. 3 and the two-way transmittances (i.e. from the top of the atmosphere to the cloud top and from the cloud top to the top of the atmosphere) weighted by the spectral response function have been calculated. For the sake of simplicity, the two-way transmittances will be referred to as transmittances. Although the MODIS bandwidths are narrower than the SEVIRI ones, the weighted transmittances are similar for the 0.64 and 1.64 µm channels. In the NIR, the MODIS central wavelength (0.86 µm) is slightly larger than for SEVIRI (0.81 µm) and the spectral band is only weakly impacted by the humidity, with a weighted transmittance of 0.989. Within the SEVIRI band, water vapour absorption is much higher, with a transmittance of 0.931. As a result, humidity has an impact on the spectral contrast between the VIS and the NIR, and therefore on the above-cloud AOT retrieval. The atmospheric correction, especially for the water vapour, is essential to accurately retrieve the aerosol and cloud properties from SEVIRI.

In order to correct the SEVIRI measurements for atmospheric absorption, the transmittances Tatm,λ are calculated for each spectral band λ from the cloud top height to the top of the atmosphere using the fast radiative transfer model RTTOV (Matricardi et al., 2004; Hocking et al., 2014). The cloud top height is derived from the Met Office cloud property algorithm, which uses the 10.8, 12.0 and 13.4 µm channels of SEVIRI (Francis et al., 2008; Hamann et al., 2014). Water vapour profiles come from the operational forecast configuration of the global Met Office Unified Model (Brown et al., 2012). This forecast is assimilated according to the scheme described by Clayton et al. (2013) that uses humidity data from various sources, including radiosondes and remote-sensing sounding data from many meteorological satellites. The forecast is run every 6 h and the humidity profile used for the atmospheric correction comes from the latest time-appropriate forecast field available. The profiles of the remaining gases – including ozone, carbon dioxide and methane – are those implicitly assumed by the RTTOV calculations (Matricardi, 2008). The radiance measured by SEVIRI Ratm,λ is finally corrected using 1Ratm,λ=Tatm,λRλ, where Rλ is the radiance corrected from the gaseous absorption.

The choice of the aerosol microphysical properties to use for the retrieval is similar to that of Haywood et al. (2003), but based on more comprehensive in situ measurements acquired during the CLARIFY-2017 field campaign. The Facility for Airborne Atmospheric Measurements (FAAM) BAe 146 aircraft was deployed in August–September 2017 operating from Ascension Island, with a main objective of studying biomass burning aerosol interactions with both radiation and clouds over the SEAO. This analysis focuses on flight C050, performed on 4 September, 2017. A profile descent from 7.3 to 1.9 km altitude was performed in order to sample the aerosol layer above clouds.

The aerosol dry extinction and absorption were measured with the EXSCALABAR instrument (EXtinction, SCattering and Absorption of Light for AirBorne Aerosol Research), which consists of a series of cavity ring-down and photoacoustic absorption cells operating at different wavelengths (Davies et al., 2018). From these in situ measurements, the single-scattering albedo (SSA) has been calculated at the instrument wavelengths of 405 and 658 nm. The uncertainty in SSA calculations is related to the corresponding uncertainties in the extinction and absorption coefficients measured by EXSCALABAR. This error analysis has been performed previously and the reader is directed to Davies et al. (2019). Briefly, the measured extinction has an accuracy of ∼2 %, and we use a 2 % extinction uncertainty in the analysis here. The errors in absorption measurements using photoacoustic spectroscopy depend on uncertainties in the ozone calibration, microphone pressure dependence and the background response from laser scattering/absorption on the windows of the photoacoustic cell. We have shown in recent publications that our calibration uncertainties are ∼5 % (Cotterell et al., 2019; Davies et al., 2018), and the uncertainty in the pressure-dependent microphone response is 1.2 % (Davies et al., 2019). The background response from laser-window interactions is from 0.27 to 0.54 Mm-1. Thus, the total absorption uncertainty, propagating all the above uncertainties, is absorption-dependent and ranges from 29.0 % to 55.0 % (dependent on PAS measurement wavelength) at 1 Mm-1 and 8.1 % at 100 Mm-1 (independent of PAS measurement wavelength). We propagated these total measurement uncertainties for both extinction and absorption measurements to derive the standard deviation σ in our calculated SSA values. We find that the mean SSA uncertainties are 0.013 and 0.018 at the measurement wavelengths of 405 and 658 nm respectively.

The aerosol size distribution was characterized between 0.05 and 1.50 µm radius using a wing-mounted passive cavity aerosol spectrometer probe (PCASP). Before and after the campaign, the bin sizes of the PCASP were calibrated using aerosolized diethyhexyl sebacate and polystyrene latex of known size and refractive index (Rosenberg et al., 2012). Further calculations based on Mie-scattering theory are performed in order to determine the bin sizes at the refractive index of the biomass burning aerosol sample. Partial evaporation of water is expected in the PCASP due to the heating of the probe, which may decrease the aerosol size. However, the sonde dropped during the flight indicates an average relative humidity above clouds of 29.2 % with a maximum of 38.6 %. According to Magi and Hobbs (2003), the light scattering coefficient of an aged African biomass burning plume only increases by a factor of 1.01 for a relative humidity of 40 %. For this reason, the impact of humidity on the PCASP and EXSCALABAR measurements is neglected. Three sources of errors have been taken into account on the PCASP measurements: the error on the bin concentration is calculated according to Poisson counting statistics, the sample flow rate error is assumed to be 10 % and a bin edge calibration error of half a bin has been considered.

The aerosol properties needed for the SEVIRI retrieval include the size distribution and the complex refractive index. The normalized number size distribution (dN/dln⁡r) is commonly represented by a combination of lognormal modes: 2dNdln⁡r=∑iNi2π1ln⁡σiexp⁡-(ln⁡ri-ln⁡r)22(ln⁡σi)2, where Ni, ri and σi are the number fraction, the geometric mean radii and the standard deviation of the mode i respectively. As in most remote-sensing applications, it has been chosen to represent the particle size distribution for the aerosol during CLARIFY-2017 with fine- and coarse-mode contributions. The aerosol optical properties are calculated using the Mie theory, as the spherical approximation is expected to be valid for biomass burning particles from 1 h after being released in the atmosphere (Martins et al., 1998). The aerosol model is selected by iteratively adjusting the refractive index and fitting the PCASP measurements (Fig. 4a) until the aerosol model matches the SSA from EXSCALABAR (Fig. 4b). In order to obtain the most suitable aerosol optical parameters for the retrieval, it is important to accurately fit the PCASP measurements where the aerosols contribute the most to the SEVIRI signal. Each bin of the PCASP has been assigned a weight for the fit of the bimodal distribution. The weights have been calculated in a similar way to Haywood et al. (2003), which means that they are proportional to the contribution of each bin to the total aerosol extinction in the 0.6 µm band. The bins corresponding to the 0.15 to 0.25 µm radius range contribute about 77 % of the extinction. Consequently, these bins have been assigned appropriate larger weights during the fitting process of the size distribution. Due to the small fraction of coarse-mode aerosols, the standard deviation of this mode σcoarse could not be reliably fitted and has been set to a value of 2.23, which is within the same order of magnitude as the one assumed for absorbing aerosol (∼2.12) in the MODIS Dark Target operational algorithm (Levy et al., 2009).

The aerosol model that best represents the PCASP and EXSCALABAR measurements is shown in blue in Fig. 4a, b. A refractive index of 1.51–0.029i has been obtained, associated with an SSA of 0.85 at 0.55 µm, which is within the range of SSA measured over the SEAO during the SAFARI and the DABEX campaigns (Johnson et al., 2008) and on the upper end of the values from Ascension Island reported by Zuidema et al. (2018). Regarding the refractive index, it should be noted that the SSA is not very sensitive to the real part, suggesting that the value of 1.51 is not particularly well constrained. However, a real part of 1.51 is consistent with the AERONET retrievals for African biomass burning particles (Sayer et al., 2014) and is adopted here. The best-fit size distribution is characterized by [Rfine, σfine, Nfine; Rcoarse, σcoarse, Ncoarse] = [0.12 µm, 1.42, 0.9996; 0.62 µm, 2.23, 0.0004]. By way of comparison, the three-mode lognormal distribution obtained for aged biomass burning aerosols during the SAFARI 2000 campaign (Haywood et al., 2003), defined by [R1, σ1, N1; R2, σ2, N2; R3, σ3, N3] = [0.12 µm, 1.30, 0.996; 0.26 µm, 1.50, 0.0033; 0.80 µm, 1.90, 0.0007], is plotted in orange in Fig. 4a. The radius associated with the first mode is consistent with the CLARIFY-2017 model. The absence of the second fine mode in this study is compensated for by a larger standard deviation for the fine mode. Finally, the radius of the CLARIFY-2017 coarse mode is slightly smaller than the SAFARI-2000 one but the coarse-mode fractions of the two models are close to each other. The uncertainties on the aerosol properties have been estimated using the errors on the PCASP and EXSCALABAR measurements. The uncertainty on the imaginary part of the refractive index is 0.02 for the real part and 0.004 for the imaginary part. For the size distribution, the uncertainty is 0.016 µm, 0.09 and 0.00045 for the radius, standard deviation and number fraction of the fine mode respectively.

The algorithm relies on the comparison of the corrected SEVIRI signal at 0.64, 0.81 and 1.64 µm with precomputed radiances. The simulations have been performed using an adding–doubling radiative transfer code (De Haan et al., 1987). The surface is assumed to be Lambertian with an albedo of 0.05 at all wavelengths, which is typical of the sea surface albedo under diffuse radiation conditions. The aerosol and cloud properties assumed for the LUT are summarized in Table 1. The truncation of the cloud droplet phase function has been carried out using the delta-M method (Wiscombe, 1977) and the TMS correction (Nakajima and Tanaka, 1988) has been applied. The cloud layer is assumed to be located between 0 and 1 km and the aerosol layer between 2 and 3 km. The sensitivity of the algorithm to the altitudes of the aerosol and cloud layers is expected to be negligible due to the small contribution of the Rayleigh scattering to the signal at the SEVIRI wavelengths. We have evaluated the error due to the fixed aerosol and cloud altitudes to be lower than 2.5 % on the AOT and 0.3 % on the cloud properties. The cloud droplets are assumed to follow a gamma law distribution characterized by an effective variance of 0.06. When the cloud is optically thin and/or the cloud droplets are too small, it is not possible to separate the contribution to the optical signal arising from aerosols from that of clouds. Therefore, the minimum values for the CER and the COT in the LUT are 4 and 3 µm respectively. This also justifies the assumption of a relatively simple sea surface reflectance parameterization as, at COTs exceeding 3, the sea surface has little impact on the upwelling radiances above clouds. Clouds associated with lower COT and/or CER are rejected. The aerosol model corresponds to the CLARIFY-2017 model mentioned above, assuming the same refractive index at the three SEVIRI wavelengths.

The retrieval of the above-cloud AOT, COT and CER is performed simultaneously. The result corresponds to the parameters that minimize the difference ε between the simulated radiances Rsim and the corrected satellite signal Rλ: 3ε=∑λRλ-Rsim,λRλ2. When the simulated signal is not close enough to the satellite measurements (i.e. ε>0.0006), the result is rejected. The retrieval of the above-cloud AOT is highly uncertain at the cloud edges and for inhomogeneous clouds. In order to remove these results, the products are aggregated onto a 0.1∘×0.1∘ grid and the standard deviation of the AOT and the CER are calculated. Note that each grid cell represents approximately 12 SEVIRI pixels. The inhomogeneity parameter ρ is defined by the ratio of the standard deviation of a parameter to the average value of this parameter. The results corresponding to a standard deviation of the AOT larger than 0.7 and/or ρCER>0.2 as well as grid cells associated with fewer than nine successful retrievals are rejected.

It is important to realize that the uncertainties that we quantify here are structural and parametric uncertainties related to assumptions made in the retrieval algorithm. When using a fixed aerosol model, no account is made for natural variability in the aerosol optical parameters and the associated uncertainty; this is dealt with in the uncertainty analysis that follows.

The algorithm has been applied to an event of biomass burning aerosols above clouds captured by SEVIRI on 28 August 2017 at 10:12 UTC. The RGB composite and the retrieved above-cloud AOT, COT and CER over the SEAO region are shown in Fig. 5. The largest AOTs are observed off the coast of Angola, with a local average value of 1.0 and a maximum of 1.6 at 0.55 µm. The AERONET site of Lubango (14.96∘ S–13.45∘ E) measured an average AOT of 0.75 that day with an Ångström exponent of 1.83, indicating the expected domination of fine-mode biomass burning aerosols. A gradient of AOT is observed towards the south-west, as we move away from the source as might be expected from a pre-campaign analysis of satellite retrievals (Zuidema et al., 2016). Absorbing aerosols above clouds are also detected in the north-west part of the region. Around Ascension Island (7.98∘ S–14.42∘ W), the above-cloud AOT from SEVIRI is around 0.37 while the AERONET site indicates a value of 0.48 associated with an Ångström exponent of 1.271. This suggests that coarse-mode aerosols, such as sea salt within the boundary layer but generally below cloud, are contributing to the total column aerosol load. The cloud properties retrieved are within the range of values typically observed for marine stratocumulus (Szczodrak et al., 2001) with more than 90 % of the COT lower than 25 and 99 % of the CER between 4 and 20 µm. As a comparison, Fig. 6 shows the equivalent aerosol and cloud properties retrieved from MODIS Terra with the MOD06ACAERO algorithm (Meyer et al., 2015) for the 10:00 and 11:30 UTC overpasses. The MODIS above-cloud AOT pixels associated with an uncertainty larger than 100 % have been removed. A good spatial agreement is observed between the two satellite products. The above-cloud AOT from MODIS is also 1.0 on average close to the coast. On average over the area, the MODIS above-cloud AOT is larger by 0.05 compared to SEVIRI. Considering that MODIS is less sensitive to the atmospheric absorption and that the two algorithms are based on the same principle, the small differences observed between the two above-cloud AOT tend to validate the atmospheric correction applied to the SEVIRI measurements for that case. There is a good consistency between the MODIS and the SEVIRI COT. Finally, the CER retrieved with the MOD06ACAERO algorithm is larger by 2.2 µm compared to the SEVIRI CER. This almost systematic difference is mainly due to differences in the satellite instruments, and especially the difference in the channels used for the retrieval (Platnick, 2000). A fully statistical analysis against the MODIS algorithm, and against airborne remote-sensing and in situ measurements will be presented in a companion paper.

The atmospheric transmittances above clouds used to correct the SEVIRI measurements from the gas absorption are calculated based on forecast water vapour profiles. In order to assess the sensitivity of the retrieval to the atmospheric correction, new transmittances have been calculated for the event studied here, modifying the specific humidity by ±10 %. The aerosol and cloud properties retrieved with the modified atmospheric corrections are aggregated on a 0.1∘×0.1∘ grid. Figure 7 compares the retrieved aerosol and cloud properties from SEVIRI-measured radiances using the original specific humidity forecast with the perturbed specific humidity (+10 % in orange and -10 % in blue). The uncertainty on the water vapour content impacts mainly the retrieval of the above-cloud AOT, and then the COT, because of its effect on the radiance ratio. A +10 %/-10 % bias on the humidity leads to an overestimation/underestimation of the AOT and COT respectively. On average, errors of 18.5 %, 5.5 % and 2.3 % have been calculated for the AOT, COT and CER respectively, based on biases of 10 % in the specific humidity forecast. These errors are likely upper estimates because forecast errors in specific humidity are unlikely to reach these values owing to the extensive assimilation of satellite data and sonde profiles by the data assimilation process used in the Met Office forecast model as previously mentioned. However, the differences between forecast model specific humidities and those of simple standard atmosphere climatological values (e.g. those of McClatchey et al., 1972) frequently exceed 10 %, indicating that accurate retrievals of aerosol and cloud need synergistic retrievals or data-assimilated forecasts of specific humidity.

The LUT used for the SEVIRI retrieval uses an assumed aerosol model based on in situ measurements from CLARIFY-2017. However, the absorption property and the size of biomass burning particles are expected to vary during the fire season and across the SEAO (e.g. Eck et al., 2003). Here, we analyse the impact of the aerosol assumptions on the retrieved aerosol and cloud properties.

In order to create a range of aerosol optical properties, a thousand aerosol models have been processed using the Mie theory. The radius and the standard deviation of the fine mode and the real and imaginary part of the refractive index of the models are random values following a normal distribution. Their mean corresponds to the CLARIFY model values provided in Table 1, with standard deviations of 0.01 µm and 0.1 for the radius and the standard deviation of the fine mode, 0.02 for the real part of the refractive index, and 0.008 for the imaginary part. Figure 8a and b show the histograms of the simulated SSA and asymmetry factor g at 0.55 µm in orange. As a comparison, histograms of the AERONET SSA and g are plotted in blue. The data correspond to the AERONET level 2.0 retrievals for August–September, from 1997 to 2018 and for inland sites of southern Africa (10–35∘ S, 10–40∘ E). Only data associated with an Ångström exponent larger than 1.0 have been used in order to remove measurements dominated by coarse-mode particles (such as dust and sea salt) that are less likely to be observed above clouds in the SEAO. The mean SSA (0.862) and the mean g (0.620) from AERONET are respectively slightly larger and smaller than the CLARIFY model. Small differences between above-cloud and full column aerosol properties could be explained by the contribution of aerosol within the boundary layer, such as pollution, desert dust and sea salt. The dashed lines in Fig. 8a and b represent the mean ± the standard deviation of SSA and g. The AERONET standard deviation is 0.023 for the SSA and 0.024 for g while the simulation produces a standard deviation of 0.036 for the SSA and 0.041 for g. The simulated range of both optical properties is larger than the range observed by AERONET. Therefore, the variation in the aerosol microphysical properties used for the simulations is wide enough to cover the range of observed aerosol optical properties.

From the simulated standard deviation σ of g and SSA, eight aerosol models have been defined and their properties are summarized in Table 2. The first four are used to test the sensitivity of the retrieval to g and SSA independently ([SSACLARIFY±σSSA, gCLARIFY] and [SSACLARIFY, gCLARIFY±σg]) and the sensitivity to both parameters will be assessed with the last four ([SSACLARIFY±σSSA, gCLARIFY±σg]). New LUTs have been processed with these modified aerosol models and used to re-process the case study from Sect. 3a. After aggregating the data on a 0.1∘×0.1∘ grid, the AOT as well as the absorption AOT (AAOT), the COT and the CER are compared against those obtained with the standard CLARIFY-2017 aerosol model. Results are shown in Figs. 9 and 10. For each aerosol and cloud property, a linear relationship is observed between the retrieval using the standard CLARIFY-2017 aerosol model and the modified one. The retrieval of cloud properties (Figs. 9c, d and 10c, d) appears to be weakly sensitive to the assumed aerosol model, with g having a slightly larger impact. On average, differences lower than 4.1 % are observed on the COT and lower than 2.4 % on the CER. As expected, the choice of the aerosol model has much more influence on the AOT retrieval. The uncertainty on the AOT is dominated by the SSA assumption. When aerosols are more absorbing than the CLARIFY model, the algorithm overestimates the AOT by 25.7 %. Conversely, the retrieved AOT is underestimated by 32.6 % when aerosols are less absorbing than the CLARIFY model. The impact of g alone on the retrieved AOT is far less significant and lower than 4.3 %. Figure 9a, which shows the impact of a perturbation on both the SSA and g, confirms that the SSA is the parameter with the strongest influence on the AOT retrieval. The largest overestimation (27.5 %) is observed when both the SSA and g are overestimated (Fig. 10a), while the largest underestimation (-33.3 %) is obtained when the SSA is underestimated and g is overestimated. The retrieval of the above-cloud AOT depends mostly on the aerosol absorption of the light reflected by the cloud. Therefore, it is expected that the retrieved AAOT is less sensitive to the absorbing property of the aerosol than the AOT. The sensitivity of the AAOT to the assumed aerosol properties is shown in Figs. 9b and 10b. The uncertainty in the AAOT due to an error in g is similar to the uncertainty in the AOT (<5 %). However, the influence of the SSA assumption alone on the AAOT is smaller than the influence on the AOT, with differences of 1.9 % and -8.7 %. This means that a perturbation of the SSA primarily impacts the scattering AOT. The largest overestimation of the AAOT (2.7 %) is obtained when the assumed aerosol model overestimates g. An underestimation of the SSA and an overestimation of g lead to the largest underestimation of the AAOT (-5.1 %).

The variation in the solar zenith angle, and therefore in the satellite observation geometry during the day, can impact the sensitivity of the retrieval to the aerosol assumptions. Therefore, the 15 min SEVIRI observations for 28 August have been processed using the eight aerosol models described above and compared to the aerosol and cloud properties retrieved with the CLARIFY aerosol model. The difference Δxi of a product x is defined as Δxi=(xCLARIFY-xi)/xi×100%, where xCLARIFY and xi is the mean product x retrieved over the SEVIRI slot with the aerosol CLARIFY model and the modified model i respectively. Figure 11 shows the time series of ΔAOT (a), ΔAAOT (b), ΔCOT (c) and ΔCER (d) obtained with the modified aerosol models. The sensitivity of the retrieved cloud properties to the aerosol model assumptions remains small (lower than 5.6 % for the COT and 2.6 % for the CER) and dominated by the sensitivity to g. Apart from a small decrease in ΔCOT at midday when g is overestimated (solid blue line) and an increase in ΔCOT in late afternoon when the SSA is underestimated (solid red line), no significant trend is observed in the cloud property sensitivities. As observed previously, the uncertainty on the AOT is led by the SSA assumption, with the AOT being overestimated (respectively underestimated) when the assumed SSA is overestimated (respectively underestimated). Until 15:00, ΔAOT stays within ±40 %, with the sensitivity to the SSA being slightly larger at midday. Then it increases up to 60 % when the SSA is overestimated and g is underestimated (dashed blue line). Similar trends are observed in ΔAAOT, with generally lower values than ΔAOT. An increase in the uncertainty is observed on the AAOT after 15:00, which reaches up to 27 % at 16:30. Before 15:00, there is a larger AAOT sensitivity to the SSA around midday (+8.9 % / -15.2 %), but there is no evident evolution of the sensitivity to g with time. The case that leads to the largest biases on the AAOT is when the SSA is underestimated and g overestimated (dashed green lines), with an underestimation of up to 23 %. However, it should be noted that 0 % of the AERONET observations used in Fig. 8 are associated with an SSA lower than SSACLARIFY-σSSA and a g larger than gCLARIFY-σg. Otherwise, the sensitivity of the AAOT to the aerosol property assumptions stays between -16.6 and +9 % before 15:00.

In conclusion, the retrieved AOT is less sensitive to the aerosol property assumption before 15:00, with an uncertainty of 40 %. This uncertainty is dominated by the sensitivity of the retrieval to the SSA. An overestimation (respectively underestimation) of the AOT is expected when the observed aerosols are more (respectively less) absorbing than the aerosol model assumed for the retrieval. A better accuracy is obtained on the retrieved AAOT, with an uncertainty generally lower than 17 % before 15:00. The sensitivity of the cloud properties to the aerosol model assumption remains small all day long, with an uncertainty of 5.6 % on the COT and 2.6 % on the CER.