Deriving Aerosol Absorption Properties from Solar Ultraviolet Radiation Spectral Measurements at Thessaloniki, Greece

: The gap in knowledge regarding the radiative e ﬀ ects of aerosols in the UV region of the solar spectrum is large, mainly due to the lack of systematic measurements of the aerosol single scattering albedo (SSA) and absorption optical depth (AAOD). In the present study, spectral UV measurements performed in Thessaloniki, Greece by a double monochromator Brewer spectrophotometer in the period 1998–2017 are used for the calculation of the aforementioned optical properties. The main uncertainty factors have been described and there is an e ﬀ ort to quantify the overall uncertainties in SSA and AAOD. Analysis of the results suggests that the absorption by aerosols is much stronger in the UV relative to the visible. SSA follows a clear annual pattern ranging from ~0.7 in winter to ~0.85 in summer at wavelengths 320–360 nm, while AAOD peaks in summer and winter. The average AAOD for 2009–2011 is ~50% above the 2003–2006 average, possibly due to increased emissions of absorbing aerosols related to the economic crisis and the metro-railway construction works in the city center.


Introduction
Atmospheric aerosols can be of natural or anthropogenic origin and their impact on human health is well established [1][2][3]. Furthermore, they play an important role in determining the levels of solar radiation reaching the Earth surface [4][5][6][7]. Decreasing aerosols over Europe and North America due to improvement of the air quality in the last two decades led to increased solar radiation at the Earth surface, while increasing emissions of anthropogenic aerosols over South-East Asia affected the levels of solar radiation in the opposite direction [8,9]. Although changes in aerosol load and properties may induce important changes in climate [10,11], their interactions with solar radiation, as well as with different atmospheric constituents have not been fully understood yet [12][13][14]. Despite progress in the description of aerosols in global climate models in the recent past, uncertainties related to their efficiency to absorb solar radiation still constitute one of the largest sources of uncertainty in current modeling estimates of aerosol climate forcing [15].
The absorption efficiency of an aerosol layer is usually quantified either using the mean columnar single scattering albedo of the layer (in many studies simply called single scattering albedo, and from now on is referred as SSA) or the absorption optical depth (AAOD). Particular aerosol types, such as sulfate, mainly scatter solar radiation in the ultraviolet (UV) and visible (VIS) wavelengths, while correspond to cloudy skies which were not detected by the cloud filtering, or problematic measurements not detected by the quality control.
For each of the six wavelengths (310-360 nm with a 10 nm step), the DG irradiance ratio is simulated for SSA ranging from 0.3 to 1 with a step of 0.01. Then, each of the measured DG ratios is matched to the closest value of the modeled ratio. At the lowest and highest SSA limits (0.3 and 1) the calculated SSA values are excluded from the analysis when the measured DG ratios differ from the closest modeled ratios by more than 1.5%. The AAOD is finally calculated by multiplying (1-SSA) for each wavelength and each spectral scan with the corresponding value of the AOD. Only SSA and AAOD for SZAs smaller than 70 • have been used in the analysis since uncertainties increase fast for higher SZAs. Climatological analysis of the SSA and AAOD is performed only for wavelengths from 320 to 360 nm. Measurements at 310 nm are only used to assess the spectral dependence-not for climatological analysis-and only for SZAs smaller than 55 • and AOD (at 340 nm) above 0.3, because of the very high uncertainties at larger SZAs and lower AOD.
The level 2, version 3 AOD at 340 nm from a co-located CIMEL sun-photometer [60] for the common period of measurements with Brewer#086 has been used for two purposes: (1) For the validation of the AOD from Brewer#086, and (2) as RTM input (instead of the AOD from Brewer#086) for the re-calculation of the SSA in order to further evaluate the reliability of the particular product from Brewer#086. The closest measurements performed by CIMEL within ± 15 mins from the corresponding measurements of Brewer#086 have been used. The SSA and AAOD at VIS and NIR wavelengths from CIMEL are used for comparison with the corresponding Brewer#086 products in the UV. Level 2 SSA and AAOD from CIMEL are available only for AOD (at 440 nm) above 0.4 and SZA above 50 • . Thus, in addition to the level 2 data, we also used the level 1.5 products whenever the level 2 size distribution was available in order to enrich the dataset as also suggested by Bond et al. [71].
Different limitations were applied to the Brewer products used for different parts of the analysis depending on the uncertainties described in the following section and the availability of the AERONET products. More specifically: • In order to investigate the spectral dependence of the SSA and AAOD over Thessaloniki (Section 4), close-in-time (within ± 15 min) SSA and AAOD measurements from Brewer#086 and CIMEL have been analyzed for the period 2005-2017. Only measurements at a narrow range of SZAs (45 • -55 • ) were used, because measurements from CIMEL are available only for SZA larger than 45 • , even for the enriched dataset, while the Brewer measurements at 310 nm can be highly uncertain at SZAs larger than 55 • . Measurements for AOD below 0.3 were also excluded from the analysis to further reduce the uncertainties. • All available measurements of Brewer#086 for SZAs smaller than 70 • have been used to study the diurnal variability of the SSA and the AAOD (Section 5.2) in the UV. Measurements before 2003 have not been used in this part of the study, since data for the period 1998-2002 are reliable only for SZAs 50-70 • as discussed above. Measurements at 310 nm have also not been used because they are highly uncertain. In order to study the diurnal variability hourly averages were calculated.

•
For the study of the annual variability (Section 5.1), the hourly averages calculated for the study of the diurnal variability, but only for the period 2005-2017 and for hours 10-14 (LT) were used. The reason for not taking into account measurements for the years 2003 and 2004 is that continuous AERONET data, which were also used, are available only since 2005. The reason for using only measurements between 10 and 14 LT is that throughout the year the SZA cut-off limit of 70 • corresponds to different times of the day (in the morning and the evening). Thus, using all measurements for SZAs smaller than 70 • would introduce biases because SSA and AAOD change in the day as a function of the local time (see Section 5.2). Daily averages were calculated from the hourly averages, and then, monthly means were calculated when at least ten days of measurements were available. Monthly means were averaged over the whole period of study to get monthly climatological values.

Uncertainties in the Retrieval of the SSA and the AAOD
The methodology for the quantification of the uncertainties in SSA and AAOD from Brewer#086, as well as the results of the analysis, are discussed below. The uncertainties in the final products are mainly due to uncertainties in the measurements of Brewer#086, as well as uncertainties in other atmospheric parameters used as inputs for the RTM.

Uncertainties in the DG Ratio
A major uncertainty factor for the calculation of the DG ratios is the accuracy in the calculation of the CTF used for the transfer of the absolute calibration from global irradiance to direct irradiance measurements. In the context of the present study, measurements for the calculation of the CTF were repeated in a number of days in 2017, at different SZAs and under different atmospheric conditions. As long as the measurements for the calculation of the CTF are performed at SZAs smaller than 50 • and for AOD (at 340 nm) below 0.2, the results are consistent to within 4% for wavelengths between 310 and 360 nm. Measurements performed in the past under similar conditions (small SZA, low AOD) within periods of a few months confirm the above findings. The corresponding standard uncertainty is thus estimated to 1.7%, which is lower than the uncertainty (of 3%) reported by Bais [57]. This difference is mainly due to the improved characterization of the Brewer#086 for the effects of dead-time and temperature, and the use of the corresponding updated correction factors.
Accurate correction for the effects of temperature and dead time is of exceptional significance when the temperature and the recorded signal are very high (as usually happens with measurements for the determination of the CTF). Instrumental temperatures of~40 • C are not unusual in summer, inducing changes of~4% in the response of the particular instrument relative to the response at the reference temperature of 25 • C [72,73]. In the past, direct sun measurements were not corrected for the effect of temperature since it is not easy to accurately characterize the instrument [74], while the used correction factors for the global measurements were not accurate, especially for temperatures below~20 • C [72]. Updated methods were used for the characterization of both, the global [72] and the direct sun measurements [73], and the new correction factors were then applied to the dataset. Dead time is important when the recorded signal is very high. Thus, neutral density filters are used to keep the signal within the desired levels, wherein errors in the correction for the effect of dead time are not significant. However, before the re-evaluation of the dataset, there were periods during which the error in the dead time constant used for the correction of the measurements was of the order of 10 ns, leading to errors up to~5% in the irradiance when the recorded signal was very high. Detailed discussion regarding the dead time effect, and the improvements in the characterization and correction procedures can be found in Fountoulakis et al. [63]. Circumsolar irradiance (e.g., Blanc et al. [75]) is a major factor of uncertainty in the direct sun measurements, especially at low wavelengths and large SZAs, thus it affects the DG ratio. Measurements of the sky radiance near the solar disc revealed that, for AOD (at 340 nm) below 0.3, the circumsolar component contributes less than 1% in the direct-sun UV measurements above 310 nm, even at large SZAs, which is also confirmed by the results for visible wavelengths reported by Blanc et al. [75]. The effect of circumsolar irradiance is more important for higher AOD [76]. However, for AOD of the order of 0.5, or higher, even an overestimation of 3% of the direct component due to the contribution of circumsolar radiation has small impact relative to the overall uncertainties in the calculation of the SSA as discussed in the following.
Additional uncertainties in the DG ratio may come up from the technical characteristics of the instrument if proper corrections are not applied [77][78][79]. The wavelength shift [80] of both the direct and global measurements used for the calculation of the DG ratio is the same, thus, the effect on the DG ratios is negligible. Proper characterization of the entrance optics for the effects of temperature [72,74] and polarization [81] keeps the uncertainties related to these factors below 1%. Application of updated corrections for non-linearity due to the effect of dead-time [63] keeps the relative errors below 2%, even for the highest levels of the recorded signal.
The overall uncertainties in the calculation of the DG ratios are dominated by the uncertainties in the calculation of the CTF. Based on the above discussion, it is estimated that the standard uncertainty in the calculation of the DG ratios is~2%. The corresponding errors in SSA due to increased or decreased DG ratios by 2% were calculated for the entire dataset. The estimated errors for AOD > 0.3 are 0.03 or lower (relative errors up to 5%) while for AOD between 0.1 and 0.3 they can be up to 0.1 (relative errors up to 20%). Errors of~0.2 (relative errors up to 40%) were found for AOD below 0.1. The aforementioned errors in SSA induce absolute errors of up to 0.02 in the calculation of the AAOD. The corresponding relative errors are below 20% for AAOD larger than 0.05 and reach 100% for AAOD below 0.02. AOD below 0.1 represents however a very small fraction of the dataset. For example, at 340 nm, the daily average AOD is below 0.1 for~1.5% of the days in the period 1998-2017. For~27% of the days daily average AOD is between 0.1 and 0.3, while it is larger than 0.3 for~71.5% of the days.
At 310 nm, increased noise at larger SZAs [63,82] may introduce additional uncertainties. Furthermore, the contribution of diffuse light entering the field of view of the instrument in the measured solar beam can be important, also affecting the retrieval of the AOD. Since uncertainties in the DG ratios related to these effects cannot be easily quantified, DG ratios at 310 nm for SZAs larger than 55 • are not included in the analysis.

Uncertainties in AOD
Kazadzis et al. [36] report that the most important source of uncertainty in the retrieval of the SSA is the AOD used as input in the RTM. In another study Kazadzis et al. [59] estimated that standard (1-sigma) uncertainty in the retrieval of the AOD from Brewer#086 is 0.05 for the AOD at 320-360 nm and 0.07 for AOD at 310 nm. Higher uncertainty for 310 nm is mainly due to the uncertainty in the retrieval of total ozone [83]. However, after improvements in the characterization of the instrument and re-evaluation of Brewr#086 dataset these uncertainties are now estimated to be lower. Comparison of the AOD at 340 nm from Brewer#086 with close in time (within ±15 min) measurements of the CIMEL for the period 2005-2017 results in very good agreement between the two instruments (correlation coefficient = 0.99, mean offset = 0.009, R2 = 0.98, RMSE = 0.034). Given that the uncertainty in the AOD from CIMEL is of the order of 0.02 in the VIS and larger in the UV [60], these numbers confirm the good quality of the AOD derived from Brewer#086. The offset of 0.009 is due to the nearly perfect agreement of the two instruments in the summer and slightly higher differences during the rest of the year, reaching 0.02 (still within the uncertainty of CIMEL) in winter. Based on the above analysis we estimate that the uncertainty in the AOD from Brewer#086 is 0.035 for 320-360 nm and 0.055 for 310 nm.
Errors in the calculation of the AOD are not however completely independent of errors in the calculation of the DG ratio because errors in the calculation of the CTF affect both toward different directions (e.g., higher CTF means higher DG ratio and lower AOD). Then the propagating errors in the calculation of the SSA are again toward different directions (e.g., underestimation of the SSA due to higher DG ratio and overestimation of the SSA due to lower AOD). Despite this partial counter-balancing, the uncertainties due to errors in DG ratio and AOD are treated as independent in the present analysis, since uncertainties in AOD are higher and not solely due to errors in the CTF. Furthermore, it is very difficult to isolate the effect of CTF and perform an independent investigation.
In order to investigate the effect of these uncertainties, the SSA and the AAOD were re-calculated for 2015 using different AOD input (± 0.035 and ± 0.055 relative to the reference AOD, for 320-360 nm and 310 nm respectively). Analysis for wavelengths 320-360 nm yielded wavelength-independent results for the SSA. When AOD is above 0.3 the error in SSA is below 0.2 at 310 nm and 0.1 at 320-360 nm. For very high AOD, above~0.6, the error in the SSA, even for 0.055 error in AOD is less than 0.05 Remote Sens. 2019, 11, 2179 7 of 26 (<10%). The increased optical path as the SZA increases makes the effect of errors in AOD slightly more significant. When AOD is very low, below 0.1, relative errors are generally above 40% and in extreme cases exceed 100%. In absolute numbers, errors in SSA are below 0.1 for AOD above 0.3 (for an error of 0.035 in AOD). When AOD is between 0.1 and 0.3 and it is overestimated by 0.035 or 0.055, errors in SSA are up to 0.2 or 0.3 respectively. Errors are smaller when AOD is underestimated by the same amount. Errors in SSA become more important when the AOD is below 0.1, reaching~0.5.
The absolute errors in AAOD corresponding to the aforementioned errors in AOD increase on average with increasing SZA and not (at least significantly) with decreasing AOD. Errors of 0.035 and 0.055 in AOD correspond to errors in AAOD ranging from 0.01 to 0.04 and 0.02 to 0.07 respectively. The fractional errors range from 5 to 100% and 10 to 150% respectively, reaching their maxima at the maximum SZA. These errors are minimal for maximum AAOD and become more important as AAOD decreases.
As a second step for the evaluation of the quality of the SSA and AAOD from Brewer measurements, the two products at 340 nm were recalculated using the AOD from CIMEL as input to the RTM. The derived quantities were then compared to those calculated using the AOD from Brewer#086. This way we investigated whether the differences between the AOD from Brewer#086 and CIMEL could introduce biases in the time-series of the SSA and the AAOD. The results are presented in Figure 1. Brewer#086 measures systematically lower AOD by~0.01 at SZAs below 55 • . Differences are larger (0.02-0.03) for SZAs near 70 • . These differences in AOD are generally not important for the calculation of the SSA when AOD is higher than 0.3 or the SZA is smaller than 60 • , introducing slightly lower SSA values (~5%) when AOD from Brewer#086 is used. The corresponding differences in SSA when AOD is~0.1 and SZA is near 70 • are of the order of −20%. Differences in the calculation of AAOD are on average below 10% for SZAs smaller than 55 • and increase up to~30% for SZAs near 70 • . The absolute errors in AAOD corresponding to the aforementioned errors in AOD increase on average with increasing SZA and not (at least significantly) with decreasing AOD. Errors of 0.035 and 0.055 in AOD correspond to errors in AAOD ranging from 0.01 to 0.04 and 0.02 to 0.07 respectively. The fractional errors range from 5 to 100% and 10 to 150% respectively, reaching their maxima at the maximum SZA. These errors are minimal for maximum AAOD and become more important as AAOD decreases. As a second step for the evaluation of the quality of the SSA and AAOD from Brewer measurements, the two products at 340 nm were recalculated using the AOD from CIMEL as input to the RTM. The derived quantities were then compared to those calculated using the AOD from Brewer#086. This way we investigated whether the differences between the AOD from Brewer#086 Considering that CIMEL is a well-calibrated instrument, developed specifically for AOD measurements, and that the agreement between AOD from Brewer#086 and CIMEL is within the uncertainties of the later, the results presented in Figure 1 provide a very strong indication that the quality of the SSA and AAOD from Brewer#086 has been optimized after re-evaluating the direct and global UV datasets.
Long term drifts in the sensitivity of Brewer#086 could however introduce drifts in the measured AOD and corresponding biases in the time-series of the SSA and AOOD. Thus, as an additional quality criterion, the long-term changes of the annual mean ratio between the AOD at 340 nm from Brewer#086 and CIMEL for the period 2005-2017 (for which measurements from CIMEL are available), and the AOD at 320 nm from Brewer#086 and Brewer#005 for the period 1998-2017 (for which measurements from both, Brewer#005 and Brewer#086 are available) were evaluated. In both cases the detected trends in the annual mean ratio of close-in-time AOD measurements (within ± 15 mins) were insignificant and smaller than 0.01/decade.

Extinction Coefficient Profiles
A default climatological aerosol profile (default vertical profiles of aerosol optical properties) [69] was used as input in the RTM calculations. In order to evaluate the effect of using a default profile, the SSA was re-calculated for 16 days (504 spectra) in the period 2013-2015 using extinction coefficient profiles measured by a Raman/elastic backscatter LIDAR operating at LAP [84][85][86][87][88][89]. For each of these days, one of the following aerosol species was dominant: dust, biomass burning, and continental. The differences in the calculated SSA when different profiles are used instead of the default are in all cases within ±0.04. The use of the default profile leads to a systematic overestimation of the SSA for highly absorbing mixtures (i.e., SSA below~0.6) and low AOD (below 0.3). In this case the overestimation may be up to~10%, which however is well below the overall uncertainties for low AOD. Based on this analysis, and assuming that the investigated sample is representative of the overall uncertainties related to the used aerosol extinction profile, we estimate a standard uncertainty of 0.02 in SSA (~2% for high AOD and~5% for AOD below 0.3). The corresponding uncertainty in AAOD is 0.01. For SZAs larger than the limit of 70 • assumed in the present study the effect of using the default profile is slightly larger.

SSA Profiles
The use of a default SSA atmospheric profile for the RTM simulations also introduces uncertainties. The SSA in the used profile slightly decreases with increasing altitude (by~4% in the first 10 km), which however is not always representative for what really happens, especially in the UV region. In order to investigate the uncertainties related to the use of a climatological profile, the simulations were repeated using SSA profiles derived from the LIDAR measurements. The SSA profiles from LIDAR were retrieved using a simple method, providing extinction and SSA profiles which, although highly uncertain, are closer to reality than the default aerosol profile used for the RTM calculations, especially for cases where more than one of the predominant aerosol types are present.
Twelve different aerosol profiles of the SSA and the extinction coefficient at 340 nm were chosen and used. The used profiles represent cases when dust (4 of 12) or smoke (5 of 12) aerosol layers are present over and/or under less absorbing aerosols, as well as typical conditions when only continental aerosols are prevailing (3 of 12). The classification of different aerosol types was performed based on the minimum Mahalanobis distance from an ensemble of reference clusters (e.g., Papagiannopoulos et al. [90]). A more analytical description of the used methodology is provided in Appendix A. Although the SSA profiles from the LIDAR are uncertain, and not suitable for climatological analysis, they provide strong evidence that when a layer of highly absorbing (e.g., dust or smoke) aerosols is present over or/and under less absorbing (e.g., continental aerosols) the SSA can change significantly with altitude. When the constructed SSA profiles (combined either with the default or the corresponding measured extinction coefficient profiles) are used in the simulations, the difference in the calculated SSA is within ± 0.02 as long as the reference SSA is above 0.6 and AOD above 0.2. When the reference SSA is below 0.6 and AOD near 0.2, the default SSA profile leads to underestimation of the (effective) SSA (up to 0.05). In all studied cases the AOD was above 0.2. Due to the small number of cases with AOD below 0.3 (only 13 cases) only the uncertainties for large AOD are estimated. Although the present study is limited to SZAs below 70 • , it is worth mentioning that the effect of the SSA profile becomes significantly more important for SZAs above 70 • , especially for low AOD and SSA, when differences of up to~0.1 were found.

Tropospheric Ozone Profiles
Default ozone profiles [67] were scaled to the measured total ozone in the RTM calculations. Part of solar UV radiation scattered by aerosols in the troposphere is then absorbed by ozone. Inaccurate description of this effect in the RTM simulations might lead to under-or over-estimation of the absorption by aerosols. Although the relative distribution of aerosol and ozone in the troposphere can affect the DG ratio, we could not find any study quantifying this effect. Thus, in order to investigate the sensitivity of the algorithm to changes in the profile of tropospheric ozone, a number of simulations were repeated for variable aerosol conditions in 2017 using tropospheric ozone profiles from a MAX-DOAS instrument operating at LAP [91,92]. These ozone profiles were retrieved based on external data sets for the stratospheric component [93]. As ozone profiles from the MAX-DOAS extend up to~4-6 km, the average ozone mixing ratio at the highest available level was scaled to the corresponding value in the default ozone profile. Then, the new profile was again scaled to the total ozone measured by the collocated Brewer#005. In all cases, the differences in the calculated SSA were of the order of 0.01 (even for 310 nm), well below the overall uncertainties. Thus, this source of uncertainty was not taken into account in the calculation of the overall uncertainty budget. The methodology for the quantification of the uncertainties in SSA and AAOD from Brewer#086, as well as the results of the analysis, are discussed below. The uncertainties in the final products are mainly due to uncertainties in the measurements of Brewer#086, as well as uncertainties in other atmospheric parameters used as inputs for the RTM.

Total Ozone
The effect of the total ozone uncertainties in the retrieval of the AOD, mainly at 310 nm, has been taken into account in the estimation of the overall AOD uncertainties [59]. We additionally tried to quantify the effect of the uncertainties in total ozone on the simulations for the retrieval of the SSA. Simulations for 2015 were repeated keeping all the parameters the same, except total ozone. Based on the studies of Carlund et al. [83] and Fragkos et al. [94] an overall uncertainty of 2.3% in the retrieval of total ozone was estimated (due to uncertainties in the used cross sections, stratospheric temperature variations, calibration and measurement uncertainties). The change of the calculated SSA due to a change of 2.3% in the total ozone column used in the simulations is in all cases (even at 310 nm) lower than 0.01 (thus not detectable).

Surface Albedo
The available measurements of the surface albedo at 440 nm from AERONET (version 3-level 2) for the period 2005-2017 were analyzed resulting to a median value of~0.08 which varies randomly by about ±0.03 throughout each year. A number of studies show that over urban environments surface albedo is gradually decreasing from VIS to the UV-B (e.g., Feister and Grewe [95]). In Corr et al. [33] it is shown that if a spectrally dependent surface albedo which decreases by~0.035 from 368 to 332 nm is used in the RTM, the average SSA at 332 nm is higher by~0.03 relative to the corresponding SSA for a spectrally flat surface albedo. The results of the later study refer, however, to measurements for average AOD of 0.2-0.3. By performing RTM simulations we confirmed that a difference of 0.03 in the albedo corresponds to a difference of~1% in the DG ratio, which is not important for AOD above 0.3 but may lead to differences of up to 5%-8% in SSA when AOD is~0.1. Since measurements of the surface albedo in the UV for Thessaloniki are not available, and the variability in the surface albedo for VIS is ±0.03, we decided to use a spectrally independent albedo of 0.04. Although surface albedo is larger when the city or the surroundings are covered by snow, these occasions are rare (zero to a few days in a year) in Thessaloniki, and usually coincide with overcast sky. Assuming a maximum error of 0.03 in the used value of the albedo, an uncertainty of 5% is estimated for AOD below 0.03. For higher AOD the uncertainty is though below 1%. In all cases the uncertainty in AAOD is 0.01.

Absorbing Gases
The two main absorbing gasses in the UV region, excluding ozone, are sulfur dioxide (SO 2 ) and nitrogen dioxide (NO 2 ). The former was not taken into account in the simulations for the calculation of the SSA. For the later, default monthly averages for the period 2012-2016 [91] were used for the simulations, as well as for the retrieval of the AOD. The effect of the particular gases in the retrieval of AOD [96] has been already taken into account in the estimation of the overall uncertainty budget for AOD. In order to estimate the effect of not (accurately) taking into account the absorption by the particular gases in the simulations for the calculation of the SSA, the direct to global ratio for different NO 2 and SO 2 optical depths was calculated assuming SZAs between 20 • and 70 • and AOD between 0 and 1. For total columns of NO 2 between 0 and 10 DU and SO 2 between 0 and 5 DU (total columns of SO 2 and NO 2 over Thessaloniki are generally smaller), the change in the ratio is always below 1%, and the corresponding uncertainties in the calculation of the SSA are negligible.

Aerosol Asymmetry Parameter
A constant asymmetry factor of 0.7 has been used for the simulations. This value was chosen after analyzing the asymmetry factor at 440 nm from AERONET for the period 2005-2017, and calculating an average value of 0.69 ± 0.03. Given that the asymmetry parameter is generally increasing with decreasing wavelength in the visible, a slightly higher value of 0.7 is used for the UV. Discussion regarding the uncertainties due to the use of a default value for the particular parameter can be found in Khatri et al. [46] and Kazadzis et al. [36], concluding in both studies that the relative uncertainties are negligible. Their results are confirmed if we accept that the uncertainties are solely due to the spread in real values of the asymmetry parameter at 440 nm (in this case the standard deviation is 0.03 introducing negligible uncertainty). However, there are two additional sources of uncertainty [97]: (1) There are no available measurements allowing to accurately quantify how the asymmetry parameter changes as a function of wavelength in the UV spectral region. Although for the calculation of the SSA in the UV we use a representative value for 440 nm, the corresponding asymmetry parameter for wavelengths 310-360 nm may differ significantly. (2) There are measurement uncertainties and biases (e.g., finite field of view of the CIMEL), as well as uncertainties due to the assumptions and the simplifications in the AERONET algorithm (e.g., assumption of spheroids, used refractive indexes and uncertainties in particle size distribution), which are introduced in the retrieval of the asymmetry parameter [98,99].
Accurate quantification of the above uncertainties is a very hard task and is out of the scope of the present study. The later Assuming that the standard uncertainty in asymmetry parameter can be as large as 0.1, the relative uncertainties for SSA and AAOD would be of the order of 0.04 and 0.03 respectively. The later quantities show that uncertainties due to the used asymmetry factor may be important. Due to the lack of more reliable estimates of the uncertainty related to the used asymmetry parameter, we use the later values for the calculation of the overall uncertainty budget.

Overall Uncertainties in the Retrieval of the SSA and the AAOD
A summary of the uncertainties in the calculation of the SSA and the AAOD due to the effect of all factors discussed above is provided in Table 1. The uncertainties for 310 nm are discussed independently from those for 320-360 nm, and are reported only for SZAs smaller than 55 • . Separate quantification has also been performed for two classes of AOD values, above and below 0.3. Standard uncertainties due to errors in the DG ratios and AOD have been assumed equal to the error in the calculation of the SSA introduced by the ± 1σ uncertainties in the DG ratios and the AOD. Since the relative uncertainties may be extremely large (and in some cases misleading regarding their actual impact), only the absolute uncertainties are reported in Table 1.

Wavelength Dependence
The limitations applied on the data used in this section (see Section 2) lead to decreased uncertainties relative to those reported in Table 1, especially for the AAOD. The standard uncertainty in AAOD for the data presented in Figure 2 is~0.03 for wavelengths 320-360 nm and~0.05 for 310 nm. Results for four selected Brewer wavelengths (310, 320, 340, and 360 nm) are shown in Figure 2, while results for all available VIS and NIR wavelengths (440, 675, 870, and 1020 nm) are presented for CIMEL. From Figure 2 it is clear that the absorption efficiency of aerosols over Thessaloniki increases fast in the UV, resulting to much higher AAOD relative to that for VIS wavelengths. This increase of the AAOD with decreasing wavelength is typical for aerosol mixtures dominated by accumulation mode fine dust particles [100][101][102]. At 310 nm the median AAOD (~0.11) is an order of magnitude higher than the corresponding value for 675 nm (~0.01). In all cases, the differences between the AAOD in the UV and VIS are larger than the standard uncertainty in the retrieval of the AAOD.
The higher absorption efficiency in the UV results in lower SSA values. The median SSA is 0.92-0.93 for VIS wavelengths, and 0.91-0.92 for NIR. The corresponding values in the UV-A region are ~0.86-0.87, while at 310 nm the median is ~0.84. SSA decreases from 440 to 360 nm (by 0.05-0.06) and with a lower rate from 360 to 320 nm. Important decrease is again noticed from 320 to 310 nm (~0.02). However, in all cases the differences between UV and VIS are of similar magnitude with the uncertainties in the retrieval of the SSA. The median values shown here are not representative for the whole year, as most winter time measurements are not included due to the SZA limitation. Additionally, the number of available measurements for summer is larger than for other seasons due to much lower cloudiness. The results presented in Figure 2 are in good agreement with the results reported in other studies [33,36,38].

Annual Variability
In Figure 3, the monthly climatological values for the AOD, the SSA, and the AAOD are presented for 2005-2017, as well as the corresponding measurements at 440 nm from the CIMEL for the same period. The monthly averages presented have been derived from all available measurements of CIMEL and Brewer#086 (i.e., no temporal matching has been applied). The optical From Figure 2 it is clear that the absorption efficiency of aerosols over Thessaloniki increases fast in the UV, resulting to much higher AAOD relative to that for VIS wavelengths. This increase of the AAOD with decreasing wavelength is typical for aerosol mixtures dominated by accumulation mode fine dust particles [100][101][102]. At 310 nm the median AAOD (~0.11) is an order of magnitude higher than the corresponding value for 675 nm (~0.01). In all cases, the differences between the AAOD in the UV and VIS are larger than the standard uncertainty in the retrieval of the AAOD.
The higher absorption efficiency in the UV results in lower SSA values. The median SSA is 0.92-0.93 for VIS wavelengths, and 0.91-0.92 for NIR. The corresponding values in the UV-A region are~0.86-0.87, while at 310 nm the median is~0.84. SSA decreases from 440 to 360 nm (by 0.05-0.06) and with a lower rate from 360 to 320 nm. Important decrease is again noticed from 320 to 310 nm (~0.02). However, in all cases the differences between UV and VIS are of similar magnitude with the uncertainties in the retrieval of the SSA. The median values shown here are not representative for the whole year, as most winter time measurements are not included due to the SZA limitation. Additionally, the number of available measurements for summer is larger than for other seasons due to much lower cloudiness. The results presented in Figure 2 are in good agreement with the results reported in other studies [33,36,38].

Annual Variability
In Figure 3, the monthly climatological values for the AOD, the SSA, and the AAOD are presented for 2005-2017, as well as the corresponding measurements at 440 nm from the CIMEL for the same

Diurnal Variability
The diurnal cycle of absorbing aerosols has been studied with respect to local time (LT) instead of universal time (UTC), in order to correlate the results with the local activities. The used range of SZAs (less than 70°) allows the study of the variability between 8:00 and 19:00 LT in summer and 9:00-17:00 in winter. For each day, the difference (in %) of each hourly value (average for each hour of the day, e.g., 8:00-8:59, 9:00-9:59, …) from the noon value (average of data within ± 1 hour around the local noon) was calculated and were subsequently used to derive monthly hourly differences. The monthly AOD in UV wavelengths over Thessaloniki is minimum in winter and maximum in the summer, ranging from~0.35 to 0.5. Comparison between 360 and 440 nm shows that the spectral dependence of the AOD is stronger in the summer resulting to less pronounced annual cycle for longer wavelengths. This is also evident from the comparison between different UV wavelengths, although differences are small, well below the standard deviation of the monthly averages. This behavior could be related to increased fine aerosol during summer. These results are generally in agreement with the results of previous studies [62,84,86,87].
The monthly average SSA is greater in summer than in winter. In winter, the difference between the SSA in the UV and the VIS is much larger than the standard uncertainty in the retrieval of the former, proving that the SSA in the UV is much lower during this particular season. Again, the annual cycle of SSA in the UV is more pronounced than the annual cycle for 440 nm. The SSA at 440 nm ranges from~0.9 (in winter) to~0.93 (in summer). The SSA at 360 nm ranges from~0.7 in winter tõ 0.87 in summer, while for 320 nm, the SSA is different only in summer (~0.86).
In the case of AAOD the spectral dependence is stronger than for SSA and AOD, though within the uncertainty of the AAOD retrieval. The pattern of the annual variability of the AAOD in the UV exhibits two maxima, one in December-January (AAOD~0.07-0.08), and one in August (AAOD 0.06-0.07). The maximum in August corresponds to high SSA and to the maximum monthly average AOD in the year, which means that it is mainly due to the very high AOD although aerosols are not as absorbing as in winter. The very high AOD in summer results mainly from the meteorological conditions prevailing over the city during this season as has been analytically discussed in previous studies [62,103,104]. The winter maximum of AAOD corresponds to the lowest AOD and SSA values in the year, indicating an increased amount of highly absorbing particles in the aerosol mixture during this particular season. There is no obvious annual pattern for 440 nm, for which the average AAOD varies randomly between 0.015 and 0.023. The high AAOD levels in winter are possibly due to increased emissions of highly absorbing aerosols from domestic heating [105].

Diurnal Variability
The diurnal cycle of absorbing aerosols has been studied with respect to local time (LT) instead of universal time (UTC), in order to correlate the results with the local activities. The used range of SZAs (less than 70 • ) allows the study of the variability between 8:00 and 19:00 LT in summer and 9:00-17:00 in winter. For each day, the difference (in %) of each hourly value (average for each hour of the day, e.g., 8:00-8:59, 9:00-9:59, . . . ) from the noon value (average of data within ±1 h around the local noon) was calculated and were subsequently used to derive monthly hourly differences. Finally, the monthly values were averaged over the whole 15-year period to derive the climatological diurnal cycle for each month. The analysis was performed for data of Brewer#086 at 320-360 nm for the period 2003-2017. Similar analysis using data from CIMEL is not feasible due to the limited range of SZAs for which measurements are available. The pattern of diurnal variability does not change significantly for wavelengths between 320 and 360 nm. Thus, only the results for 340 nm are presented in Figures 4  and 5. Values on the x-axis correspond to the exact hour of the day (e.g., 8 corresponds to 8:00).
The diurnal variability of the SSA (Figure 4) follows a standard pattern throughout the year: Average values are low in the morning (8:00-11:00 LT), they increase by~5%-8% until~15:00 LT and slightly decrease in the evening by 1%-5%. The diurnal variability of the AAOD ( Figure 5) is almost identical in spring and summer: A gradual decrease of~50%-60% from 8:00 to 15:00 LT, and then an increase of similar magnitude in the following four hours. In winter and autumn, the average % change in AAOD is smaller than in spring and summer for the same hours. This behavior is possibly related to seasonal differences in the relative contribution of the main sources of absorbing aerosols.
Remote Sens. 2019, 11, x FOR PEER REVIEW 15 of 28 diurnal cycle for each month. The analysis was performed for data of Brewer#086 at 320-360 nm for the period 2003-2017. Similar analysis using data from CIMEL is not feasible due to the limited range of SZAs for which measurements are available. The pattern of diurnal variability does not change significantly for wavelengths between 320 and 360 nm. Thus, only the results for 340 nm are presented in Figures 4 and 5. Values on the x-axis correspond to the exact hour of the day (e.g., 8 corresponds to 8:00).  The diurnal variability of the SSA (Figure 4) follows a standard pattern throughout the year: Average values are low in the morning (8:00-11:00 LT), they increase by ~5%-8% until ~15:00 LT and slightly decrease in the evening by 1%-5%. The diurnal variability of the AAOD ( Figure 5) is almost identical in spring and summer: A gradual decrease of ~50%-60% from 8:00 to 15:00 LT, and then an increase of similar magnitude in the following four hours. In winter and autumn, the average % change in AAOD is smaller than in spring and summer for the same hours. This behavior is possibly related to seasonal differences in the relative contribution of the main sources of absorbing aerosols. Measurements at the center of Thessaloniki show that on an annual basis the main contributing source of aerosol pollution is vehicular traffic [105]. Comparison of the results shown in Figure 5 with traffic patterns in Thessaloniki [106] or other European cities (e.g., Järvi et al. [107]), clearly shows that the diurnal variability of the AAOD in spring and summer resembles the traffic pattern at the center of Thessaloniki. In winter however, highly absorbing organic aerosols emitted by oil combustion or biomass burning for domestic heating are an important fraction of the aerosol mixture [105,108] and possibly have a strong effect on the diurnal variability of AAOD. The very large standard deviation in the results of Figure 5 indicates, however, that during individual days the diurnal course may differ significantly from the average. Furthermore, aerosols which strongly absorb UV radiation (such as biomass burning aerosols or Saharan dust) are frequently transferred from remote areas [62,86,103,104,109], and possibly disturb the diurnal cycle of SSA and AAOD, which as discussed above, seems to be mainly attributed to aerosols emitted from local sources. Finally, it should be mentioned that the same patterns for the SSA and the AAOD are found also for smaller time periods (e.g., 2009-2013) and/or only for AOD larger than 0.3.

Long-Term Variability
Monthly and annual mean anomalies of the aerosol optical properties for the period 1998-2017 are presented in Figure 6. Again, only results for 320 nm are shown since the results for all wavelengths in the range 320-360 nm are very similar to each other. As already mentioned in Section 2, only data for SZAs between 50 • and 70 • were used in order to have a homogeneous dataset for the full period 1998-2017. Daily climatological values were calculated from the daily averages, as discussed in Section 5.1, and subtracted from the measurements to derive daily anomalies. Daily anomalies were then averaged for each month and monthly anomalies were calculated, only when at least ten daily values were available. Then annual mean anomalies were derived only when monthly means were available for at least six months. The results for AOD, AAOD, and SSA are presented in panels (a)-(c) of Figure 6 . This is in agreement with the results of several recent studies reporting increased emissions of biomass burning and wood smoke aerosols [106,108,110], as well as high concentrations of other pollutants [111] in the particular period relative to the period before 2009. The increases in aerosol reported by all these studies are attributed to the increasing use of wood and pellets instead of oil and natural gas for domestic heating, after the beginning of the economic crisis in Greece [112]. However, increased AAOD levels by~10%-20% relative to the long-term average have been calculated even for the warm months (April-September) of the same period. Part of this increase may be related to the increased road dust emissions due to the metro-railway construction works in Thessaloniki after 2008 [105,113]. The return of the AAOD to the 1998-2002 average levels (still higher by~0.01 relative to the 2003-2006 average) after 2012 could be due to reduced traffic activity (and consequently reduced emissions), related again to economic factors [114]. However, further investigation, which is out of the scope of the present study, is necessary in order to identify the causes of the above results more accurately. panels (a)-(c) of Figure 6, respectively. AOD decreases by ~0.1 from 1998 until the mid-2000. In the same period the annual mean SSA does not change significantly. However, when data only for the period March-November are used, SSA increases by ~0.03, in agreement with the results of Kazadzis et al. [26]. In the same period, the AAOD decreases by ~0.02 (when data for all months of the year are considered). From mid-2000 until 2017, AOD decreases again by ~0.1, while the SSA decreases by ~0.07. Using the calculated monthly averages of AOD and SSA as inputs in the RTM, the spectral UV irradiance at 320 nm was calculated for the entire period of study. Simulations of irradiance at 320 nm were also performed using a default value of 0.85 for the SSA and the monthly averages for AOD. For each of the two simulated datasets, the difference between the irradiance for each month and the corresponding month of 1998 was calculated. Then, the monthly differences were averaged for each year. The time series of annual means for the two simulated datasets are presented in Figure 6d and compared with clear-sky measurements of Brewer#086 at the same wavelength for 60 • SZA (59 • -61 • average). The annual anomalies for the measured irradiance were calculated using the methodology of Fountoulakis et al. [24] and were also normalized to 1998 values. Despite the differences in the methodology to derive the annual differences for the modeled and the measured irradiances, the short-and long-term variability of the measured irradiance follows quite accurately the modeled irradiance when the calculated SSA is used as input. In both cases an increase of~5% in UV is evident for the twenty-year period of the study. Though, when the default SSA is used in the simulations, a long-term increase in irradiance of~12% is derived. Figure 6d clearly shows that modeling and reconstruction of solar UV radiation from ancillary data may be misleading regarding the changes of UV radiation over polluted environments, if changes in the SSA are not taken into account.

Conclusions
In the present study the SSA and the AAOD over Thessaloniki, Greece, for the period 1998-2017 have been calculated. The methodology described in Bais et al. [32] has been followed, which is based on the calculation of the ratio between nearly simultaneous direct and global spectral UV irradiance measurements from a double monochromator Brewer spectrophotometer, and comparison with the corresponding ratios calculated using a RTM. The uncertainties in the radiative transfer inputs, as well as those of the direct to global irradiance ratios have been assessed in order to quantify the corresponding uncertainties in the derived SSA and AAOD. For 320-360 nm the standard uncertainty in the SSA is 0.11 for AOD at 340 nm above 0.3 and exceeds 0.30 when AOD is below 0.1. The estimated uncertainty in the calculation of the AAOD is 0.05, independent of AOD. Uncertainties are greater for 310 nm. We conclude that the main contributor to the overall uncertainty budget is the uncertainty in the determination of AOD.
The uncertainties in the calculation of the SSA and AAOD from Brewer#086 are much larger relative to the uncertainties of the corresponding level 2 products of AERONET (0.03 for SSA and 30% for AAOD) [115], mainly due to the higher uncertainty in the AOD measured by Brewer#086 in the UV (0.035 at 340 nm) relative to the uncertainty in the AOD measured by CIMEL in the VIS (0.02 at 440 nm) and the more stringent filtering applied to AERONET level 2 products. SSA and AAOD in the UV are not available by AERONET, and even in the VIS region the availability of the particular products is low due to applied filtering.
The uncertainties in the SSA and AAOD from Brewer#086 are comparable with the uncertainties reported for wavelengths below 360 nm in other studies [36,38]. Thus, the results of the analysis show that well-maintained and calibrated Brewer spectrophotometers are able to provide measurements of the SSA and the AAOD in polluted environments (such as Thessaloniki), which are of similar quality to those of other instruments explicitly designed for aerosol measurements. It is however also obvious that there is a need to improve the accuracy of measurements of AOD in the UV, in order to achieve more reliable SSA and AAOD, since their uncertainties are currently very large. The recent progress in the QA/QC procedures of Brewer measurements [116], and in particular the improvement of the AOD retrieval [117], may result to further reduction of uncertainties in the future.
Despite the large uncertainties, no significant bias or dependence (e.g., change with time or SZA) when the AOD from Brewer#086 is compared to other instruments. This provides strong evidence for the reliability of the results presented in Sections 4 and 5, considering that AOD is the most important uncertainty factor for the calculation of the SSA and the AAOD. When the AOD from CIMEL is used for the calculation of the SSA and the AAOD instead of the AOD from the Brewer, differences depend slightly on SZA (within the uncertainty of the CIMEL) and are nearly normally distributed around their averages. The average differences are below 0.02 for the SSA and 0.01 for the AAOD. Additional comparison of the AOD and the spectral UV irradiance from Brewer#086 with measurements of Brewer#005 for the entire period of study ensures that there is no important change in the behavior of the former two quantities which could affect studies of the long-term changes of the SSA and the AAOD. It is also worth mentioning that the year to year variability of the annually averaged difference between the AOD at 340 nm from CIMEL and Brewer#086 for the period 2006-2017 never exceeds 0.01, which also provides strong evidence that the increasing annual averages of the AAOD observed in the period 2008-2013 are real.
Analysis of SSA and AAOD from Brewer#086 and CIMEL confirmed the findings of other studies reporting that the absorption efficiency of aerosols over urban environments increases significantly in UV relative to VIS wavelengths [33,36,38]. More specifically, it is shown that the SSA decreases from 0.92 at 440 nm to~0.87 at 360 nm. From 360 to 320 nm the SSA still decreases, though the decrease is slow and well below the uncertainty in the retrieval of the SSA. From 320 to 310 nm the SSA continues decreasing fast (from~0.86 to 0.84). Thus, linear extrapolation from VIS to UV wavelengths would lead to overestimation of the SSA. For AOD (at 340 nm) greater than 0.3 and SZA within 45-55 • the median AAOD at 310 nm for the period 2005-2017 is an order of magnitude larger than the corresponding median AAOD at 675 nm, showing that the absorption efficiency of urban aerosols in the UV-B region is much larger than in VIS. During the warm months of the year, the daily variability of AAOD follows the pattern of traffic in the city center, with maximum values in the morning and evening, and 30%-50% lower values between 14:00 and 16:00 LT. In the colder autumn and winter months, the diurnal variability is less pronounced, possibly because of the important contribution of domestic heating to the overall aerosol emissions, in addition to the emissions related to vehicular traffic. The monthly average SSA in the UV changes smoothly in the year, from~0.7 in January to~0.86-0.87 in August and then back to~0.7 in December. This change is larger than the uncertainty showing a clear annual cycle of the SSA in the UV. In the annual cycle of the AAOD two peaks were found, one in December-January when AOD is minimum but the contribution of absorbing aerosols in the mixture is important, and one in August when the AOD is maximum. The relative contributions of different types of aerosols in different seasons, as well as the role of the particular meteorological conditions prevailing during each month are currently under investigation. The results will be discussed in another study focusing on these particular issues.
The reduction of~20% of the AOD from mid-1990s to mid-2000s is accompanied by a reduction in AAOD of similar magnitude. Although the increase in AOD has ceased after the mid-2000s, the annual average SSA after 2008 is~0.07 below the 1998-2006 average. Very high values of AAOD, of~50% above the long-term average, have been recorded in the years 2009-2012. This increase is probably related to increased emissions of biomass and wood burning aerosols after the beginning of the economic crisis in Greece. Since 2014, AAOD returns to lower levels, possibly again as a result of the economic crisis, since vehicular traffic in Thessaloniki has been decreasing after 2012 due to increasing price of fuel [114].
Currently, most modeling studies of past and future changes of solar UV radiation assume constant SSA throughout the years, and in few cases changes in the AAOD have been considered [21,41]. The present study clearly shows the importance of taking into account long-and short-term changes in the absorption efficiency of the aerosol mixture. The study also provides strong evidence that, at least over Thessaloniki, aerosols absorb stronger in UV than in VIS wavelengths, and that this difference cannot be taken into account accurately by simple wavelength extrapolation of the aerosol properties. Finally, the fact that the detected variability of SSA and AAOD in the UV is of the same order of magnitude with the retrieval uncertainties highlights the need for further work in order to improve the accuracy of the retrievals of these products.

Appendix A
The methodology which has been applied in order to obtain constructed SSA profiles at 340 nm is described below.
For the construction of the SSA profiles, aerosol extinction (at 355 and 532 nm) and backscatter (355, 532, and 1064 nm) coefficient profiles have been used. Measurements from the LIDAR [84][85][86][87][88][89]-which is located at a distance of few meters from the Brewer#086-in the period 2012-2018 are used. Measurements at the 1064 nm channel are not available before 2012, making impossible the construction of SSA profiles using the particular methodology. The extinction profiles, which are essential for this method, are only available during night-time when there is no skylight at the particular wavelengths (e.g., Siomos et al. [88]). Thus, only nighttime profiles are used. Since the profiles should ideally correspond to the conditions under which Brewer#086 performs measurements, i.e., only during daytime, profiles with a temporal difference of more than 6 hours with the last available spectrophotometer measurement are excluded, under the assumption that the aerosol vertical distribution is not changing much within this time interval (12 cases in total). The aerosol classification scheme based on the minimum Mahalanobis distance from an ensemble of reference clusters is applied per case (e.g., Papagiannopoulos et al. [90]) and per vertical level with a resolution of 100m.
Each level is classified in one of the following reference clusters: Continental clean (CC), desert dust (D), mixed marine (MM), polluted continual including smoke (PC+S). Layers that are not classified in any category (untyped) correspond mainly to parts of the profiles where either the aerosols are too mixed or where the aerosol load is not significant. Each of the other four categories is modeled by predefined mixtures based on the OPAC database [101,102]. CC mixtures are assigned to the "continental clean" category of OPAC (water soluble and insoluble particles), D mixtures are assigned to the "desert" category (mineral and water soluble), MM mixtures to the "maritime polluted" category (sea salt, water soluble, and soot), and finally PC+S mixtures to the "urban" category (water soluble, insoluble, and soot). Each category is consisted by components in different mixing ratios, resulting to different optical properties per wavelength (see Hess et al. [101]). Some species are assumed to be hygroscopic (sea salt, water soluble) and their optical properties are adjusted according to the relative humidity per layer. Relative humidity profiles from the European Center for Medium-range Weather Forecast (ECMWF) Reanalysis Interim (ERA-Interim) were applied.
An SSA at 340 nm from OPAC is assigned to each classified layer. Un-typed layers are interpolated if there are adjacent to classified layers. The resulting SSA profiles are scaled so that the columnar equivalent SSA at 340 nm matches the value produced from Brewer#086 (see Section 2 of the paper). This way the uncertainties due to the non-simultaneous measurements are reduced. Furthermore, the actual mixing ratio of the aerosol components per classified layer can be different than the predefined climatological mixtures of OPAC. The lidar extinction coefficient profiles at 355 nm are also scaled to the corresponding AOD measured by Brewer#086.