Two aerosol datasets based on shipborne observations are considered here as ground-based reference: the Microtops Sun photometer (Microtops II manufactured by Solar Light Inc.) and the GUVis shadowband radiometer (GUVis-3511 plus BioSHADE accessory manufactured by Biospherical Instruments Inc.). Both instruments are well suited for operation on moving platforms such as ships. Their measurement principles however are rather different. The technical specifications of the Microtops and GUVis instruments are summarized in Table . The configurations of both instruments allow a direct comparison of all spectral channels of the Microtops versus corresponding GUVis observations.

The Microtops is a hand-held Sun photometer, which has to be pointed manually at the Sun. To minimize uncertainties arise from manual pointing, more than five consecutive scans are averaged to form one measurement . The Microtops instrument measures the incident direct normal solar irradiance with a field of view of 2.5∘ . The MAN Microtops Sun photometers are calibrated against an AERONET master Cimel Sun photometer, which in turn is calibrated using the Langley technique.

The GUVis shadowband radiometer utilizes an entrance optic with a global field of view combined with a shadowband that performs a 180∘ sweep, while the global irradiance is measured at a high temporal frequency of 15 Hz. Several corrections are applied as post-processing to correct the influence of the ship motion, and to retrieve the direct spectral irradiance for later AOD calculation, as is described later. The measurement principle of the shadowband radiometer can be described as follows. While the global irradiance is observed with the shadowband in its lowest position between sweeps, the shadowband blocks a fraction of the incoming diffuse irradiance during its rotation and will occlude the direct irradiance at a specific angle determined by instrument orientation and Sun position. From the irradiance time series measured during the sweep, the global, diffuse and direct irradiance components can be inferred . Prior to the processing of the GUVis sweeps, the measured irradiance data have to be corrected to compensate for the motion of the ship and the imperfect cosine response of the instrument. The actual cosine response of the entrance optic is measured by the manufacturer during lab calibrations, and can be corrected by applying correction factors depending on the spectral channel and Sun elevation if the orientation angles of the ship are known. The motion correction utilizes the method of based on the ship motion angles to correct the direct and diffuse irradiance components. The GUVis instrument has been calibrated in a laboratory at regular 2-year intervals using a 1000 W FEL standard calibration lamp as absolute reference. The correction and processing of GUVis irradiance data as well as the calculation of AOD is described in detail in . The concept of the “field of view” of a Sun photometer is not directly applicable to a shadowband radiometer. Instead, there is the “shading angle” as described in , which is the minimum angle between the edges of the shadowband as viewed from the centre of the global entrance optic. For the GUVis, the shading angle is about 15∘ (depending on shadowband position), and thus relatively large in comparison to the Microtops field of view. The wide angle of the shadowband of the GUVis causes an underestimation of AOD caused by the influence of the forward scattering of the aerosol . The GUVis processing algorithm has received a substantial update (see Sect. ) to compensate at least partially for this effect. The reduction of measured irradiance during the shadowband sweeps is stronger in situations with increased aerosol forward scattering. Besides some other refinements, an offset was introduced for estimation of the blocked diffuse irradiance as part of the processing algorithm update in order to compensate for this effect (see Sect. ).

Given the direct normal irradiance obtained from both instruments and a given spectral band, the AOD can be calculated using the well-known Beer–Lambert law, and by subtracting optical depth contributions from Rayleigh scattering and gas absorption. In the following, an overview of the shipborne datasets based on both instruments is presented:

(i) As the first dataset, all observations conducted during numerous cruises with the Microtops II Sun photometer in the framework of AERONET MAN since 2004 to 2018 in the area of the Atlantic Ocean are used. The uncertainty of Microtops AOD is estimated to be within ±0.02 . The datasets include a total number of 19 250 valid data points. This dataset (referred to as MIC in the following text) also provides the diversity needed to investigate aerosol-type-related effects for the evaluation of satellite products, and for the comparison with CAMS RA.

(ii) The second reference dataset (GUVis) is based on the GUVis shadowband radiometer. Observations with the GUVis were conducted within the framework of OCEANET during Atlantic transect cruises with the German research vessel Polarstern operated by the Alfred Wegener Institute since 2014. Until now, five cruises including the shadowband radiometer observations have been performed, namely PS83, PS95, PS98, PS102, and PS113. A Microtops instrument from MAN has also been operated in parallel on all these cruises. This offers the opportunity to directly compare both datasets. A direct comparison of Microtops and GUVis AOD product has already been presented for PS83 in . The AOD uncertainty is estimated to be within ±0.02 . Following the same procedure, this comparison is extended to all available cruises, with some minor changes to obtain more meaningful results. The total number of valid GUVis observations is 10412.

In order to improve the agreement of the aerosol products of both instruments to acceptable limits, it has been found necessary to introduce a cross calibration to the MIC instrument, and an empirical correction for aerosol forward scattering, to account for differences arising from the limited accuracy of lab-based instrumental calibration, and the broad shadowband of the GUVis instrument. The correction is done fitting a linear regression curve (Eq. ) to the GUVis AOD (see Sect. ), similar to the approaches adopted by and . This enhanced dataset is denoted as GUVisE in this study.

(iii) The enhanced GUVis dataset (GUVisE) has been combined with the Microtops dataset to obtain a merged surface product, to test whether the combination can lead to further improvements in accuracy. This combined surface dataset (COMB) serves as third reference dataset for the evaluation of the satellite products. COMB consists of the mean of the collocated GUVisE and MIC AOD for this purpose. As shown in Table , the total amount of data points decreases to 1033 due to the collocation requirement.

Satellite-based aerosol datasets over the ocean considered here are obtained from both the MODIS and SEVIRI satellite instruments. The MODIS Collection 6.1 (C6.1) level-2 aerosol products MxD04_L2 and MxD04_3K are used from both Terra and Aqua satellites. This dataset includes the AOD at 470, 550, 660, 860, 1240, 1630, and 2130 nm. The AOD(500 nm) and AE obtained from the SEVIRI instrument aboard the MSG satellite provided in the SEV_AER-OC-L2 product of the ICARE Data and Services Center is also considered. In the following text, unless otherwise stated, the terms “MODIS aerosol products” or “MODIS retrieval” refer to the MxD04_L2 and MxD04_3K products, and similarly, the term “SEVIRI aerosol product” refers to the ICARE SEV_AER-OC-L2 aerosol product.

Both aerosol retrievals are based on the inversion of the measured reflectance at top of atmosphere to estimate the AOD at the instrumental spectral channels using lookup tables of radiative transfer calculations. The accuracy of these estimates critically depends on realistic assumptions about the optical properties of aerosols assumed in the calculations. A larger number of channels enables a more accurate choice of aerosol type used by the retrieval, and is thus expected to increase the overall accuracy. In addition, factors such as the spatial resolution of the sensor, the viewing geometry, and sensor calibration, as well as the accuracy of cloud screening will influence the overall accuracy. While the SEVIRI retrieval is based on only two wavelengths (630 and 810 nm) , the MODIS retrieval utilizes seven spectral channels. In addition, it is continuously monitored with ground-based observations at AERONET stations . A degraded accuracy for aerosol properties in the presence of desert dust in both satellite products is expected, since dust particles are non-spherical. This leads to an increased side-scatter effect compared to spherical particles which are assumed in both retrievals.

Besides the retrieval differences, MODIS and SEVIRI products are also different due to their satellite platform characteristics. MODIS is operated on both the Terra and Aqua satellites, which fly in a polar orbit. For studies targeting aerosol properties at a specific location, MODIS observations are only available for the two overpasses during daylight, compared to SEVIRI with a time resolution of 15 min. On the other hand, the geostationary orbit of MSG leads to lower spatial resolution of nadir 3 km for SEVIRI versus a 1 km nadir resolution of MODIS. In order to avoid cloud contamination in the aerosol product, the MODIS retrievals consider multiple pixels together with a strict cloud mask, leading to a decrease of the spatial resolution to 3 km for the high-resolution aerosol product (MxD04_3K) and 10 km for the standard aerosol product (MxD04_L2).

CAMS RA is the latest global reanalysis dataset of atmospheric conditions produced by ECMWF . Amongst other atmospheric constituents, it contains the spectral AOD at a temporal resolution of 3 h on a global grid of 0.7∘ (corresponding to a T255 spatial resolution). The advantage of utilizing CAMS RA over satellite observations is the availability of aerosol properties independent of factors such as cloud coverage or satellite orbit, albeit the accuracy of AOD under cloudy sky conditions in the model might be questionable.

CAMS RA was developed based on the experiences gained with the former Monitoring Atmospheric Composition and Climate (MACC) reanalysis and the CAMS interim analysis . It relies on the assimilation of global observational datasets into the Integrated Forecast System (IFS) from various satellites to provide a global picture. In terms of aerosol properties, the AODs from the products of the MODIS C6 from both Terra and Aqua are assimilated, while the composition mixture is maintained as given from the IFS. Before its failure in March 2012, retrievals from the Advanced Along-Track Scanning Radiometer (AATSR; ) flown aboard the Envisat mission were also being assimilated. The influence of this additional source of information for data assimilation on the accuracy is investigated in Sect. . Currently, the dataset covers the period 2003–2016 and will be extended in the following years. For the evaluation of the CAMS RA aerosol dataset, an accuracy close to the MODIS aerosol product is expected.

A first validation presented within emphasizes the high quality of AOD in the CAMS RA system, as judged by a comparison to AERONET stations around the world. However, an overestimate of AE was shown during desert dust events and was attributed to the fixed component mixture (e.g. less dust in CAMS RA) in the forecast model. Further evaluation with a focus on individual aerosol components as well as aerosol properties over the ocean has been recommended .

Our study aims to compare shipborne and satellite AOD products also with respect to the role of aerosol types. A satellite-independent aerosol classification is applied, which is based on the empirical method presented in for Cimel instruments from AERONET. This method is also applicable to the Microtops AOD product, as it contains all required parameters. The aerosol classification is done by comparing the AOD at λ=440 nm with the AE calculated based on the 440 and 870 nm channels (Eq. ). The pair of AOD and AE values is checked against empirical thresholds to identify the dominant aerosol type of the current situation as being one of maritime background (AOD <0.15), mineral dust transport (AE <0.5, AOD >0.15), continental transport (AE >1, AOD >0.15), or mixed (0.5<AE <1, AOD >0.15) types. It should be noted that all categories are expected to cover mixed aerosol types to some extent. Therefore, the mixed category consists of a mixture of aerosol without a dominant type.

In the following text, we shorten the aerosol type description from maritime background (maritime), mineral dust transport (desert dust, since over the Atlantic Ocean most dust cases originate from the Sahara desert), and continental transport (continental). All results shown in this study separated by aerosol type (maritime, desert dust, continental, mixed) are based on this aerosol classification method. It should be noted that the shipborne observations at wavelengths of 440 and 870 nm are utilized for classification, even if figures and tables present AOD and AE for different channels (e.g. Fig. ).

As common practice for spatiotemporal collocation with MODIS, a window size of 50×50 km and a time window of 1 h are recommended by for Sun photometer observations. For the MODIS C6 validation by , a spatial radius of 25 km and a temporal window of ±30 min have been used. Both the MxD04_L2 and MxD04_3K products have been validated using a window of 5×5 pixel by , resulting in different window sizes of 50 and 15 km2, respectively. The following collocation technique is utilized here to find the appropriate pixel of the satellite dataset and to compare it to the shipborne data obtained at a certain position. First, eligible satellite images are selected using a time frame of ±30 min around observations, and checking if the ship position is located within the field of view of the satellite image. Then, the distance angles of all pixel coordinates to the ship position have been calculated. The satellite AOD is finally calculated as the median of all non-cloudy pixel values with a distance angle less than 0.2∘.

Choosing a distance angle threshold of 0.2∘ for the collocation of all satellite and model datasets to the shipborne observation assures that the same area around the reference observation is chosen regardless of satellite or model product, spatial resolution, and projection, and ensures comparability of results. This threshold results in a spatial radius of about 22 km.

Applying the collocation strategy introduced above to the 19 250 MIC data points results in a total number of remaining 1517 data pairs for MxD04_L2, 1448 for MxD04_3K, 10 061 for SEVIRI, and 2474 for CAMS RA, as shown in Table .

After collocation with the GUVis dataset, consisting of a total number of 10 412 data points, the resulting number of data pairs is 147 for MxD04_L2, 210 for MxD04_3K, and 1126 for SEVIRI. The collocation with CAMS RA results in 141 data pairs. The number of collocated data pairs is rather small, limiting the statistical significance of the comparison results.

The number of data pairs per aerosol type classified based on the shipborne reference data as described in Sect.  is given in Table . Since the observations are performed across the Atlantic Ocean, the dominant aerosol conditions are mainly maritime or desert dust originating from the Sahara desert (see Fig.  and Table ).

To assess the agreement of two measures (X, Y) of the same quantity such as AOD from Microtops versus GUVis, or the shipborne dataset versus satellite products, linear regression statistics and the Pearson product-moment correlation coefficient R (referred to simply as the correlation in the following text) are calculated. Further, the analyses are extended with the so-called “limits of agreement” (LOA) method first introduced by . This method considers the mean of the differences of both quantities X–Y (i.e. the bias), and the LOA is defined as the 95 % confidence interval for those differences as additional parameters. As not stated otherwise, Y denotes the reference dataset for comparisons presented in this study.

For the evaluation of the uncertainty estimates for the shipborne observations, the method of is adopted, weighting the difference between X and Y (D) with their uncertainty estimate (σX, σY): 2(X-Y)/σX2+σY2. Thus, utilizing the LOA method together with the weighted difference, the uncertainty estimate can be confirmed if the uncertainty-weighted difference lies within the range of ±1.96 for the 95 % confidence interval (see Fig. ). The percentage of outliers exceeding the limits of ±1.96 is used as quantitative measure for the validation.

For the evaluation of the satellite products and CAMS RA, the bias and LOA (95 % confidence interval) are used as a measure for the agreement to the shipborne reference datasets. Additionally, Gfrac, defined as the percentage of data lying within expected error (EE) limits, is calculated in order to be consistent with other validation studies e.g.. Expectations of the error are met if 67 % of data points of the satellite or model product fall into the EE range compared to the shipborne reference . Two EE limits are chosen here, originally presented for the MODIS aerosol product based on former validation studies, e.g. by and : 3EE1=±(0.03+0.05AOD), and more recently in : 4EE2=+(0.04+0.1AOD),-(0.02+0.1AOD). EE1 is a general measure of agreement, since the boundaries are equally distributed around the reference dataset. EE2 has been specialized for the MODIS aerosol product, since a known overestimation is considered via different intercepts.

The relatively large differences originally observed in the comparison between Microtops and GUVis (Sect. ), and their changes from one cruise to another, lead to the hypothesis that calibration of the GUVis instrument might introduce significant uncertainties and be responsible for the differences, given the importance of calibration for the AOD accuracy see.

Despite the fact that the deviation of AOD between Microtops and GUVis due to forward-scattering effects of aerosol is partially compensated by the processing update of GUVis (see Sect. ), a remaining linear dependence of the bias has been observed (Sect. ), which can most likely be attributed to the wide shadowband of the GUVis instrument, and the resulting difference in the field of view of both instruments. If AOD increases, this effect increases due to the enhanced circumsolar radiation. Although this effect does not have a major impact on the correlation in the direct comparison of Microtops and GUVis AOD datasets of this study (see Sect. ), it introduces a substantial relative bias and needs to be compensated to ensure the consistency of both shipborne datasets and for the comparison to the satellite and model datasets. The compensation is done using a linear scaling factor for measured AOD (S), as is explained later in this section.

To improve the consistency of the GUVis and MIC datasets, the following approach has been adopted to both transfer the calibration from the MIC instrument to the GUVis instrument, and to empirically correct for the effects of forward scattering. The first correction is accomplished following the method introduced by for the Multi-Filter Rotating Shadowband Radiometer (MFRSR). The spectral direct irradiance measured by the GUVis can be represented by the following equation: 5Ii=CiIi0exp⁡-τiμ0, where Ii0 and Ii are the spectral direct irradiance at top of atmosphere and surface, respectively, for a spectral channel i. The inverse of the air mass is denoted by μ0, the cosine of the solar zenith angle. τi is the atmospheric column extinction optical depth for a spectral channel i. Following , a correction factor Ci for the calibration is introduced.

The absolute calibration of GUVis spectral channels is carried out in the laboratory to obtain the channel-specific calibration factors (ki; VW-1m2nm1) for the conversion of the measured voltage (Vi) to spectral irradiance (Ii): 6Ii=Viki. The relation of the calibration factor ki and the correction Ci can be obtained from Eq. () as 7ksi=kiCi, where ksi denotes a corrected calibration factor.

τi can be expressed as the sum of AOD τA,i and the remaining contributions to the atmospheric optical depth (τ̃i) from Rayleigh scattering and gaseous absorption as 8τi=τA,i+τ̃i. The AOD can now be obtained from Eqs. () and () as 9τA,i=-μ0ln⁡IiIi0+μ0ln⁡Ci-τ̃i. This equation shows that the calibration correction factor Ci introduces a change in AOD which is proportional to the product of the cosine of the solar zenith angle and the logarithm of the correction factor. Introducing also a linear scaling factor Si for the AOD to account for the effects of aerosol forward scattering see, the following correction equation is used here in a bilinear fit, using μ0 and the GUVis-based AOD τGUV,A,i as dependent variables and the MIC-based AOD τMIC,A,i as an independent variable: 10τMIC,A,i=μ0ci+SiτGUV,A,i. Thus, the scaling factor Si and the calibration correction factor Ci=exp⁡ci can be obtained simultaneously from this bilinear fit.

In the approach adopted for this study, the factor Ci has been determined independently for each of the five Polarstern cruises (PS83, PS95, PS98, PS102, PS113), in order to account for potential temporal changes in calibration between the different ship cruises, while a single constant value is assumed for Si. The correction factors obtained by multi-linear regression based on Eq. () (Cij, Si) are listed in Table . Excluding individual cruises from the regression has been found to cause only negligible influence on the remaining coefficients, confirming the stability of this correction approach. In addition, adding either a constant or quadratic correction term such as that used by does not lead to a significantly improved fit quality and has thus not been used.

The final procedure adopted here for the correction of GUVis AOD is done in the following steps: i.

First, the closest GUVis and MIC data points regarding time of measurement are selected for comparison within a time frame of 30 min.

An evaluation of the AOD product of the GUVis shadowband radiometer compared to the Microtops Sun photometer as reference was previously described by , considering one cruise of the research vessel (RV) Polarstern (PS83). This study extends the comparison to include four additional cruises with the RV Polarstern (comprising PS83, PS95, PS98, PS102, and PS113) (see Fig. ). Regarding the comparability of both datasets, certain shortcomings are expected, as already mentioned. (i) Since the radiometers of both instruments utilize different calibration methods, and the spectral response of comparable channels might slightly differ, a deviation due to calibration is expected. (ii) Due to the different measurement methods of the Sun photometer and shadowband radiometer, an underestimation of AOD is expected for the GUVis instrument.

Given the importance of calibration for the AOD accuracy of the GUVis see, only the calibration difference of both instruments is corrected for first, based on the method presented in (see Sect. ). The correction factor Cij for each spectral channel i and each cruise of RV Polarstern j is given in Table . The Microtops calibration is considered as consistent and trustworthy, due to traceability to the mature AERONET retrieval and calibration process. Therefore, it serves as the calibration reference. In the following, all versions of the GUVis datasets are calibration corrected towards the Microtops by the method presented in Sect. . The correction of AOD with the linear scaling factor S is only applied to GUVisE.

The extended comparison is presented in the top part of Table . The GUVis irradiance data are first processed with the original algorithm used in . The correlation (R>0.95) found for all spectral channels comparing GUVis (old processing) and MIC generally confirms the findings of . However, the goal of an outlier ratio below 5 % (see Table ) as well as the weighted LOA within ±1.96 to verify the uncertainty estimate of ±0.02 for GUVis is missed with the old processing algorithm. As expected, an underestimation of AOD measured by the GUVis is reflected in the negative bias of -0.03. Since the observations are performed over the ocean, the dominant aerosol conditions are maritime or desert dust (mineral dust transport) from the Sahara desert (see Fig.  and Table ), which significantly differ in their forward-scattering behaviour. Comparing Sun photometers with a narrow field of view to measurements with shadowband radiometers with a wide shading angle the influence of the forward scattering of the aerosol causes an underestimation of AOD of the shadowband radiometer . This has previously been confirmed by for the MFRSR as well as for the autonomous marine hyperspectral radiometers presented by .

The GUVis processing algorithm has received a substantial update to improve the data quality and to compensate for the underestimation of aerosol forward scattering. This update is described in detail in the Appendix (Sect. ). The GUVis AOD data of all cruises with RV Polarstern have been reprocessed with the new algorithm, and the resulting improvement of the measured AOD compared to the Microtops is also shown in Table . The correlation of GUVis AOD compared to MIC increases from >0.954 to >0.988 for all channels, indicating that any non-linear deviations due to aerosol forward scattering and other effects have been substantially reduced. The underestimation of AOD is still present, indicated by a negative bias of -0.02. The uncertainty estimate of ±0.02 can be verified for spectral channels with wavelengths larger than 500 nm, since the statistics show a weighted LOA within ±1.96 (see Sect. ). The uncertainty of GUVis AOD increases with decreasing wavelengths, indicated by the increase of outlier percentage and LOA.

The difference (D) of GUVis and MIC AOD shows a higher linear correlation as |R(D)| increases from >0.4 to >0.6 going from the old to new processing. As also shown by Fig. a, the underestimation of GUVis AOD increases linearly with increasing AOD. This linear dependence is here attributed to the difference of field of view of both instruments. If AOD increases, the effect of the circumsolar radiation due to differences in the field of view will increase. Since GUVis utilizes a broad shadowband resulting in a shading angle of 12 to 15∘ compared to the field of view of Microtops of 2.5∘, this effect results in an underestimation of AOD for the GUVis radiometer.

The lower part of Table  shows the results of the comparison of the enhanced GUVis dataset (GUVisE) and MIC, which includes both the calibration and forward-scattering corrections. The expected uncertainty of ±0.02 is verified again, as the values of the weighted LOA are all within ±1.96 (see Sect. ). This is also shown in Fig. b, where the LOA falls within the uncertainty limits. In addition, the outlier percentage is close to zero, indicating a close agreement of MIC and GUVisE. The correlation increases to >0.992 for all comparable channels. While the corrections introduced here need to be reconfirmed based on future observations, they are able to reconcile the observed differences between the MIC and the GUV products for the currently available observational data. Hence, we consider the MIC and GUVisE datasets consistent due to their strong linear correlation and their agreement within the individual uncertainty limits. Therefore, the datasets are used as reliable ground-based reference datasets in the following.

In the following, the comparison of the two MODIS and the SEVIRI aerosol products to collocated shipborne observations is shown.

(i) First, the satellite AOD at 550 nm is validated. The wavelength of 550 nm is chosen as this channel is mainly used in previous validation studies e.g. and EE limits are defined for it . Since the SEVIRI dataset does not provide AOD at 550 nm, it was calculated with Eq. () using the AOD of 630 nm and the AE from the SEVIRI dataset.

Figure  shows the comparison of MxD04_L2 and the SEVIRI AOD at 550 nm to the shipborne aerosol datasets within AOD bins of 0.1. The validation with respect to the EE limits shows that the MODIS aerosol product meets the goals of both EE1 and EE2 compared to the Microtops dataset. As expected, the SEVIRI aerosol product shows a higher deviation versus Microtops than MODIS and only meets the goal of 67 % for EE2, since it accounts for a general overestimation of satellite AOD. The SEVIRI retrieval shows an even stronger tendency to overestimate AOD in comparison to the MIC reference dataset. The bias of satellite AOD also shows a dependence on the magnitude of the AOD. A positive bias (overestimation) is mostly found in situations with AOD values below 0.5, and decreases for larger AOD. This behaviour is most evident in Fig. a and b, as the reference datasets are GUVisE and COMB. A similar behaviour also appears in the comparison to Microtops (Fig. c), although it is far less pronounced. Since the satellite instruments measure reflected radiance, the reflecting properties of the ground used in the retrievals influence the retrieved AOD. Especially for clean atmosphere, e.g. low AOD, the influence of such parameters (e.g. surface albedo) is strong, since the values measured reflectance at the top of the atmosphere (TOA) are close to the values of surface reflectance. For larger AOD values, the uncertainty of those characterizations shrinks; therefore, the overestimation of AOD decreases. Since the GUVisE and COMB datasets contain more maritime and desert dust cases than the MIC, this behaviour is strongly visible.

Figure  presents the same comparison as Fig. , but simultaneous availability of data from all datasets (SEVIRI, MODIS, MIC) is required to preclude differences arising from a different sampling of cases. Therefore, the accuracy of SEVIRI and MODIS is directly comparable with respect to the MIC reference. This comparison shows that both AODs retrieved from SEVIRI and MODIS agree well with the shipborne reference, although the non-linear behaviour of over- and underestimation is more pronounced for the SEVIRI retrieval. Since 550 nm is not a native spectral channel of SEVIRI, increased deviations in AOD are expected, since the uncertainty of AE calculated from SEVIRI native channels is high, as shown later. Therefore, a strong improvement of agreement is found comparing the 630 nm AOD to the shipborne reference as the SEVIRI AE is not used for calculation in Fig. b. At lower AOD values, SEVIRI AOD is close to the MODIS AOD.

Table  summarizes the results of the AOD validation at 550 nm. For the statistics presented in this table, the SEVIRI AOD is calculated using MIC AE since using the native AE leads to high uncertainties. Generally, the satellite-based AOD is higher compared to the shipborne reference datasets, as is reflected in the bias >0 for all selections, except for larger AOD values, where the SEVIRI bias turns negative. The MODIS aerosol products show the highest linear correlation (0.93 for MxD04_L2, 0.95 for MxD04_3K) and lowest values for LOA. LOA values are even slightly lower than those for the MxD04_L2 product. Thus, this finding does not confirm the expectation of higher noise in the 3 km versus the 10 km product of MODIS expressed in . The analysis has also been repeated separately for the MODIS datasets based on the Terra and Aqua satellites (not shown), but only minor differences in the evaluation statistics for the individual satellites were found. Thus, only the combined MODIS dataset from Terra and Aqua is presented here. It also has to be stressed that the considered dataset is still relatively small compared to other validation studies and should be repeated if more reference data become available. Nevertheless, the correlations found here agree well with the findings of considering the MODIS C6.1 aerosol product (0.937) and the 550 nm channel. A smaller dataset of Microtops observations was compared to MODIS aerosol products by , where a general overestimation of AOD and a high correlation were found, similar to our results. For SEVIRI, the findings exceed the values found in the study of at 630 nm over the ocean (0.795 versus 0.88 in this study), indicating a significantly better performance over the ocean than over land. The results for SEVIRI and MODIS show a similar agreement of the AOD compared to the reference data but with a larger scatter of ±0.15 LOA for SEVIRI versus ±0.13 LOA for MODIS AOD. Also, SEVIRI AOD shows higher bias values for AOD <0.4 and negative bias for AOD values ≥0.4.

(ii) Second, the AE calculated from the satellite products is validated. We chose to calculate the AE for each aerosol product from the channels matching closest to the wavelengths 440 and 870 nm with Eq. (). This obviously leads to increased uncertainties for the SEVIRI product but also demonstrates its limitations.

Figure  shows the comparison of the difference of AE versus the different shipborne reference datasets as a scatter plot, indicating the bias, as well as the LOA and EE limits. The EE for AE is estimated to be ±0.4 for the MODIS products , and the same EE is applied for the SEVIRI AE product. In general, the MODIS AE agrees with this estimate of the EE limits but shows a tendency to overestimate the shipborne AE, as reflected by the positive bias. The bimodal behaviour of AE of the MODIS products found for C5 in is not reproduced here, which agrees with the findings for C6.1 presented in . Also, MODIS AE meets the expected Gfrac of >67 % for an EE of ±0.4, as was already found in . The results for the SEVIRI AE show a general overestimation versus MIC, indicated by the positive bias. Furthermore, a bimodal behaviour of AE is found, similar to that reported for the C5 MODIS products in . The SEVIRI-based AE mostly lies close to two values: AE close to zero is associated with the models of oceanic and maritime aerosol used by the retrieval (O99, M99; ). Another large fraction of the dataset is related to purely tropospheric aerosol models (types T99, T90, T50; ), covering AE from 1.29 to 1.61. Another frequent assignment of aerosol model is that for very small particles which cover the AE range from 1.8 to 2.4 . Therefore, it can be concluded that SEVIRI retrieval of AE cannot realistically capture the variability in the AE which is observed from shipborne products, which is expected given the limitations of the product. It should be noted that the results comparing AE from satellite to MIC can be reproduced with the COMB dataset (Fig. a and b), although the number of collocated measurements is small. Thus, this study should be extended with data from additional cruises in the future.

(iii) Third, the representation for different aerosol conditions within the satellite-based AOD and AE products is investigated. To examine the representation of AOD and AE with respect to aerosol type, the layout presented in is used for example in Figs.  and . Instead of the AOD at 440 nm (as chosen by ), the wavelength of 630 nm is chosen to match the native SEVIRI channel. Also we restrict the comparison choosing only data points where AOD at 810 nm >0.05 to avoid uncertainties from calculating AE from low AOD values. The aerosol type is classified based on the MIC data (see Sect. ). Points related to a certain aerosol type are combined in the form of a covariance ellipse which spans 67 % of the related data points.

Figure  shows that, overall, the AODs of the different products and instruments lie very close together, with a slight tendency to overestimate the AOD for desert dust (only MODIS products) or maritime aerosol types. In general, the satellite-based datasets overestimate the AE. These results confirm the statistics discussed before. The satellite ellipses are tilted compared to the MIC ellipses, as a result of the assumed relation of AOD and AE in the retrieval, which is determined by the choice of aerosol model. This effect is strongly visible for SEVIRI in situations other than maritime, because the AE is calculated from the native channels only. SEVIRI AOD in maritime conditions exhibits a stronger overestimation as also shown in Table . This effect might be related to the coarser spatial resolution of the satellite pixels or undetected cloud contamination (see Sect. ). The spatial-mean AOD inferred from satellite pixels can deviate from the AOD which is retrieved from slant transmission in the case of MIC, due to the mismatch of spatial scales. The most prominent feature of Fig.  is the deviation in AE for desert dust conditions and to a lesser extent for the mixed type aerosol. The MODIS product shows a more than 2 times larger and the SEVIRI product a more than 3 times larger AE for desert dust situations, when compared to the shipborne products. This relates to a lack of realistic mineral dust models in the satellite retrievals. This emphasizes that the Ångström behaviour is not applicable for desert dust conditions, at least with a limited set of spectral channels. However, the AE is still the method of choice for extrapolating the AOD at the desired wavelengths to validate or increase observation capabilities, as is done in several studies .

Figure  confirms the above findings. In this figure, each dataset has been collocated individually to the MIC reference to increase the diversity of conditions. No noteworthy discrepancies are found for continental aerosol types. With the exception of the positive bias in AOD especially at lower values of AOD, and the overestimation of AE in particular for desert dust and therefore to some extent also mixed aerosol, the satellite aerosol products are found to agree closely to MIC, in particular for continental aerosol.

The previous statistics confirm that the AOD retrieved from satellite agrees well with the shipborne reference but slightly overestimates AOD in general and especially at low AOD. AE is also overestimated for maritime and especially desert dust aerosol. Therefore, AOD is only represented well for the native spectral channels of the satellite instruments. The estimation of the spectral behaviour of AOD remains challenging, due to the lack of realism of the aerosol models (MODIS and SEVIRI) and the number of spectral channels available (SEVIRI). These findings are in particular applicable for conditions dominated by mineral dust.

Last, we investigate the value of increased temporal resolution within the SEVIRI aerosol product versus MODIS products. While the MODIS aerosol product is clearly the product of choice for many applications, i.e. for data assimilation and climate studies, due to its accuracy, availability, and global coverage, the SEVIRI aerosol product is still of scientific interest due to its high temporal resolution of 15 min . The high temporal resolution, however, adds information compared to products from polar-orbiting satellites if the temporal variations of aerosol properties since the last overpass of a polar-orbiting satellite exceed the error limits of the retrieval. Thus, it is not clear how much information can actually be gained from the higher temporal resolution of SEVIRI, as it is expected that AOD variations are generally small on the timescale of hours. To further investigate this point, MODIS collocations with the shipborne datasets are used to serve as random samples to study the AOD variability between successive overpasses. For each pixel of a MODIS image, the corresponding SEVIRI AOD for every available SEVIRI image between overlapping MODIS images of consecutive Terra and Aqua overpasses was acquired to calculate the AOD variation. Relative to the linear regression line of the MODIS AOD of each overpass, the standard deviation (SD) of SEVIRI AOD was calculated. Therefore, SD is a measure of the additional variation of AOD which cannot be seen in a MODIS-only AOD product. Figure  shows the SD calculated for different time intervals between consecutive MODIS images. The mean SD of AOD within 6 h is slightly larger than 0.02. The SD is compared to the mean EE2 calculated using Eq. () and mean SEVIRI AOD, indicated by the dashed green line in Fig. . If the SD is larger than EE2, it points out a situation where AOD variation cannot be captured by MODIS and can be called significant. SEVIRI aerosol measurements add information to the general AOD monitoring only if the AOD variation is significant. As Fig.  reveals, this is only true for slightly above 8 % of all situations, knowing that, in general, Terra and Aqua satellites pass over the same region every 3 h. This emphasizes that, in terms of climate studies or data assimilation, the significant higher temporal resolution of SEVIRI does not lead to improvements for the majority of situations, unless the accuracy of this product could be significantly improved. In fact, such an improvement will be possible in the future with the third generation of Meteosat (MTG). We are aware that the analyses presented here do not provide a complete picture of the AOD variability over the full diurnal cycle. It was only possible to analyse the variability between daytime overpasses of MODIS. Continuous evaluation of the daily cycle of AOD are only possible with geostationary satellites such as SEVIRI. With the high temporal resolution, the SEVIRI product is needed for many applications, such as extreme events such as dust or smoke plume development, where high variability of AOD is expected.

Alongside the evaluation of satellite aerosol products described above, results for the CAMS RA AOD are presented in Table .

In comparison to MIC as a reference dataset, Table  shows that CAMS RA AOD agrees closely to MIC, since the correlation is 0.92 and the bias is about zero. The LOA of ±0.13 is similar to the one found for the products of SEVIRI and MODIS compared to MIC, with values ranging from ±0.12 to ±0.15. The CAMS RA outperforms the SEVIRI aerosol dataset in all presented statistical measures at least slightly (e.g. a correlation of 0.88 versus 0.92 or LOA of 0.13 versus 0.15). Further, the bias of AOD and its dependency on AOD is reduced for the CAMS RA product, as it shows low bias values for both low and high AOD values. This is expected since the MODIS AOD bias must be corrected before assimilation into the reanalysis product. This effect is clearly shown in Fig. , together with a tendency of CAMS RA towards an underestimation of AOD for larger values of AOD. For maritime aerosol, CAMS RA AOD has the lowest correlation (0.7). The values of LOA are lowest for maritime, which is expected since this measure favours lower AOD and maritime aerosol situations are generally connected to low AOD values. A slight overestimation of 0.01 is shown by the bias considering only maritime aerosol situations, which is lower than the bias found for the MODIS products. For desert dust conditions, the correlation of CAMS RA to MIC (0.85) is similar to the one found for MxD04_L2 (0.87) in Table , although the correlation of MxD04_3K is largest with 0.92. As for maritime aerosol, the overestimation of AOD is compensated in the CAMS RA aerosol product. This emphasizes that the CAMS RA aerosol product is comparable in accuracy to the MODIS products in maritime and desert dust situations.

In terms of assimilated aerosol observations, data from MODIS and AATSR are used by the IFS for CAMS RA starting from the year 2003 to March 2012, when the Envisat mission ended due to loss of contact to the satellite. After March 2012, only the MODIS AOD is used . Table  shows the evaluation of CAMS RA versus MIC for the different time periods to investigate potential differences in quality. Without AATSR, MODIS is the only contributor for data assimilation in terms of AOD. Comparing the results of CAMS RA with and without AATSR, the performance of CAMS RA with additional AATSR data is increased, indicated by increased correlation from 0.87 to 0.90 and lower LOA, dropping from 0.18 to 0.14. This shows that the AATSR observations lead to an improvement of the representation of aerosol in CAMS RA. suspected a slight increase of CAMS RA AOD without AATSR, which cannot be observed in this study. As the analysis presented here is based on a limited number of data points, it is unclear whether these findings are statistically significant, and the discussed tendencies should be considered with caution.

reported an overestimation of AE in CAMS RA compared to AERONET stations of about 5 %–20 %. From the comparison to MIC presented in Fig. f, the same conclusion can be drawn based on our dataset, showing a positive bias of 0.17. Compared to MODIS, similar values are found for Gfrac, but while the MODIS AE scatters more equally around the reference AE, CAMS RA AE is clearly distributed above zero. Also, the AE difference of CAMS RA and MIC shows a increased linear dependency indicated by increased correlation R(D). This indicates that, similar to the SEVIRI product, certain aerosol models are favoured in the processing. Nevertheless, the overall scatter of AE indicated by the values of LOA is lower for CAMS RA.

As stated by , the overestimation of AE results from a deficit in the handling of the coarse dust fraction in the model. The total AOD calculated in CAMS RA is composed of less dust than in its predecessor versions, which explains the higher overall AE. Nevertheless, the comparison of CAMS RA AE with respect to aerosol type in Fig.  reveals that the AE for desert dust agrees best with the MIC reference, compared to the satellite products. The slightly better representation compared even to MODIS indicates that the representation of the spectral dependence of AOD for dust is most realistic in CAMS RA. For maritime aerosol, CAMS RA AE shows a similar overestimation compared to the satellite products but with less scatter. This emphasizes a more consistent representation of maritime aerosol in CAMS RA as compared to satellite products. The CAMS RA AE representation versus MIC in Fig.  shows a close agreement for all aerosol types, except an overestimation of AE in maritime conditions, and a tendency for overestimation during dust conditions with low AOD. In general, the AE and AOD of CAMS RA are similar to or in some instances even exceed the accuracy of the satellite retrievals including MODIS in comparison to the reference data presented in this study.