Real-time UV index retrieval in Europe using Earth observation-based techniques: system description and quality assessment

Abstract. This study introduces an Earth observation (EO)-based system which is capable of operationally estimating and continuously monitoring the
ultraviolet index (UVI) in Europe. UVIOS (i.e., UV-Index Operating System) exploits a synergy of radiative transfer models with high-performance computing and EO data from satellites (Meteosat Second Generation and Meteorological Operational Satellite-B) and retrieval processes (Tropospheric Emission Monitoring Internet Service, Copernicus
Atmosphere Monitoring Service and the Global Land Service). It provides a
near-real-time nowcasting and short-term forecasting service for UV radiation over Europe. The main atmospheric inputs for the UVI simulations
include ozone, clouds and aerosols, while the impacts of ground elevation and surface albedo are also taken into account. The UVIOS output is the UVI at
high spatial and temporal resolution (5 km and 15 min, respectively) for
Europe (i.e., 1.5 million pixels) in real time. The UVI is empirically related to biologically important UV dose rates, and the reliability of this EO-based solution was verified against ground-based measurements from 17
stations across Europe. Stations are equipped with spectral, broadband or
multi-filter instruments and cover a range of topographic and atmospheric
conditions. A period of over 1 year of forecasted 15 min retrievals under
all-sky conditions was compared with the ground-based measurements. UVIOS forecasts were within ±0.5 of the measured UVI for at least 70 % of the data compared at all stations. For clear-sky conditions the agreement was better than 0.5 UVI for 80 % of the data. A sensitivity analysis of EO
inputs and UVIOS outputs was performed in order to quantify the level of
uncertainty in the derived products and to identify the covariance between the accuracy of the output and the spatial and temporal resolution and the
quality of the inputs. Overall, UVIOS slightly overestimated the UVI due to observational uncertainties in inputs of cloud and aerosol. This service
will hopefully contribute to EO capabilities and will assist the provision
of operational early warning systems that will help raise awareness among
European Union citizens of the health implications of high UVI doses.


at a spatial resolution of 1 o x 1 o -1 day with assimilated ozone fields from GOME-2 (METOP-B). We have to mention also 201 here that the selection of the RTM inputs has been decided based on their real-time availability. 202

UVIOS real-time processing concept 203
The LUT approach, despite its large size (almost 2.5 million spectral RTM simulations for clear and all sky conditions) 204 (Kosmopoulos et al., 2018), still provides estimates at discrete input parameters values. To overcome this mathematical issue, 205 we performed a multi-parametric interpolation technique to correct the input-output parameter intervals. This solution is 206 computationally more costly than a continuous function-approximation model, i.e. a Neural Network (NN) model 207 (Kosmopoulos et al., 2018), but the accuracy improvement is significant. Indicatively, using a test set of 1 million RTM 208 simulations for UVI from the developed LUT, we applied the NN developed in Kosmopoulos et al. (2018) and found a mean 209 execution time of around 144 seconds followed by a mean absolute error (MAE) of 0.0321, while by using the proposed 210 UVIOS multi-parametric interpolation exploiting the HPC and distributed computing benefits we found for the same test set 211 an execution time of 295 seconds with a MAE of 0.0001. The inclusion of many parameters (in this study we incorporated 212 eight, i.e. AOD, SZA, TOC, COT, ELE, ALB, AE, SSA) with small step sizes dramatically increase the LUT size, followed 213 by high computing requirements for the multi-parametric interpolation/extrapolation procedures. 214 For the UVIOS simulations performed in this study, a 32-core UNIX server was used equipped with 256 Gb of RAM and 12 215 Tb of storage system working in a RAID10 architecture. The combination of the HPC with the analytical LUTs, which were 216 developed by using the libRadtran RTM, allow a high speed multi-parametric interpolation and polynomial reconstruction 217 (Gal, 1986) to increase accuracy between the LUT records following a mathematical equation relating the UVIOS outputs to 218 the EO inputs. 219 An example of the UVIOS input output data is presented in Figure 1 through a flowchart illustration of the modelling technique 220 scheme. The inputs, including the solar and surface elevation, albedo, aerosol, ozone forecasts and the cloud observations as 221 described in Table 1, are fed to the real-time solver that results in spectrally weighted output of UVI for the European region. 222 Figure 2 shows the memory usage and error statistics for a range of different LUT sizes. The LUT error decreases as the LUT 223 size increases, regardless of the function being approximated. The LUT sizes in Figure 2 fit into cache on our HPC 224 environment, thus performance in terms of processing speed and overall output accuracy vary only slightly between the table 225 sizes shown. In our case, UVIOS shows that LUT transformation can provide a significant performance increase without 226 incurring an unreasonable amount of error, provided there is sufficient memory available. We note that the cache size is a 227 critical factor for LUT performance, while under a HPC environment practically there is no limit. Such techniques can be 228 implemented in hardware with distributed computing that operates in parallel to provide optimum performance. 229 230 https://doi.org/10.5194/amt-2020-506 Preprint. Discussion started: 1 March 2021 c Author(s) 2021. CC BY 4.0 License.

Earth-Observations 231
The Cloud Optical Thickness (COT) data from Meteosat was used, whose retrieval algorithm is based on 0.6 and 1.6 micron 232 channel radiances of Meteosat SEVIRI. MSG products have been described in Derrien and Le Gléau (2005)  comparison principles of (point) station UVI measurements with a 5 km MSG COT matrix are possibly responsible for at least 236 part of the observed deviations (e.g. Kazadzis et al. (2009a)). For instance, when a MSG pixel is partly cloudy, the ground 237 measurements of UVI could fluctuate more than 100%, depending on whether the sun is visible or whether clouds attenuate 238 the direct component of the solar irradiance. The result is that in cases of partly covered MSG pixels and in the absence of 239 clouds between the ground measurement and the sun, the ground truth UVI would be much higher than the UVIOS one. Of 240 course, the presence of small clouds which have not been identified by MSG and cover (part of) the sun disk, is plausible as 241 well, consequently causing an overestimation of the modelled UVI (Koren et al., 2007). Furthermore, sensors onboard 242 geostationary satellites suffer from the parallax error, which contributes to the spatial errors of the images and the overall 243 uncertainty of the products (Bieliński, 2020; Henken et al., 2011). The error depends on the altitude of the cloud and the 244 viewing angle (parallax errors are more significant for high viewing angles). 245 UVIOS calculations at high solar zenith angles (>75 deg) are retrieved assuming cloudless skies since the MSG COT product 246 is not available in these conditions. This has an effect on the quality of the UVIOS overall performance at high solar zenith 247 angles. However, since such measurements are associated with very low UV Index (<1), this inconsistency does not affect 248 UVIOS UV Index results associated with dangerous effects on human health. There is more discussion in the next section on 249 how we use these data for the UVIOS validation. 250 For the total aerosol optical depth, we used 1-day forecast data from the Copernicus Atmospheric Monitoring Service (CAMS) 251 as the basic input parameter. These forecasts are based on the Monitoring Atmospheric Composition and Climate (MACC) 252 analysis and provide accurate data of aerosol optical depth (AOD) at 550 nm with a time step of 1 h and spatial resolution of 253 0.4°. 254 For aerosol single scattering albedo properties climatological values from MACv2 aerosol climatology (Kinne, 2019) was 255 utilized. Monthly means of single scattering albedo at 310nm were acquired from global gridded data at a 1 o x 1 o spatial 256 resolution. Also, in order to derive the Angstrom exponent, monthly means of AOD at 340nm and 550nm were used. The 257 calculated Ångström exponent was then applied to the 550 nm AOD (from CAMS) in order to get AOD in the UV. 258 The surface albedo data were obtained from the Copernicus global land service (CGLS: Geiger et al., 2008; Carrer et al., 2010). 259 As a global surface ALB product is not available in the UV region, for this study we have used the climatological product of 260 CGLS (in the visible range) as follows: based on the findings of Feister and Grewe, 1995; we used a UV albedo of 0.05 for 261 non-snow cases and a UV ALB equal with CGLS when CGLS exceeded 0.5 (snow cover). The total ozone column forecasts 262 were obtained from Tropospheric Emission Monitoring Internet Service (TEMIS) which is a near-real time service which uses 263 https://doi.org/10.5194/amt-2020-506 Preprint.  (Eskes et al., 2003). The elevation data was obtained from the 5-minute Gridded Global Relief Data 266 (ETOPO5) database, which provides land and seafloor elevation information at a 5-minute latitude/longitude grid, with a 1-267 meter precision in the region of Europe and is freely available from NOAA (NOAA, 1988). Figure 3 shows an example of the 268 input-output UVIOS parameters. An extensive validation of the MACC analysis and forecasting system products were 269 performed by Eskes et al. (2015). The aerosol optical properties were validated against 3-year (Apr. 2011 -Aug. 2014) near 270 real time level 1.5 Aeronet measurements and for AOD at 550 nm an overall overestimation was exhibited. Due to dedicated 271 validation activity of the MACC service a validation report that covers the time period of this study (Eskes et al., 2018) is also 272 available, presenting an overall positive bias during 2017. This overestimation of AOD at 550 nm may explain some of the 273 UVI underestimation under clear sky conditions (see section 3.2.2). 274

Ground-based measurements 275
In order to validate the UVIOS results 17 ground based stations were selected, for which measurements of the UVI were 276 available during 2017. The stations are shown in Fig. 4. Comparisons were performed with a 15-minute step. The ground based 277 measurements were obtained from spectrophotometers (Brewer), spectroradiometers (Bentham), filter radiometers (GUV) and 278 broadband instruments (SL501 and YES) as Table 2 shows. I Note that UV data in table 2 has been calibrated, processed and 279 provided directly by the responsible scientists for each station. References wherein more information for the data quality of 280 particular instruments can be found are also provided. Brewer spectrophotometers measure the global spectral UV irradiance 281 with a step of 0.5 nm, and a resolution which is approximately 0.5 nm (usually between 0.4 and 0.6 nm). Depending on their 282 type the spectral range is usually 290-325 nm (MKII, MKIV) or 290-363 nm (MKIII,). Since Brewer spectrophotometers 283 measure the spectrum up to a wavelength which is shorter than 400 nm, extension of the spectrum up to 400 nm in order to 284 calculate the UV index is usually achieved using empirical methods (e.g. (Fioletov et al., 2003;Slaper et al., 1995)). The 285 additional uncertainty in the UVI due to the latter approximation is well below the overall uncertainty in the measurements. 286 Bentham spectroradiometers measure the whole UV spectrum (290 -400 nm) with step and resolution which can be 287 determined by the operator. The spectra from AOS and LIN (measured by Bentham spectroradiometers) used in this study 288 have been recorded with a step of either 0.25 or 0.5 nm and a resolution of ~ 0.5 nm. The Brewer Spectrophotometer measures 289 the total column of ozone using the differential absorption method, i.e. measuring the direct solar irradiance at four wavelengths 290 and then comparing the intensity at wavelengths that are weakly and strongly absorbed by ozone (Kerr et al., 1985). Brewer 291 TOC measurements are used in the present document to validate the TEMIS forecasts. The Ground-based Ultraviolet (GUV) 292 instrument is a multichannel radiometer that measures UV radiation in five spectral bands having central wavelengths as 305, 293 313, 320, 340 and 380 nm. However, in addition to UV irradiances, other data that can be obtained from GUV instruments are 294 total ozone and the cloud optical depth (Dahlback, 1996  usually very cold winters, and short, cool to mild summers. SOD has subarctic climate having very cold winters and mild 308 summers. 309

Evaluation methodology 310
The time series period covers the whole year 2017 at 15-min intervals, following the MSG available time steps. A 311 synchronization between the UVIOS simulations and the ground-based measurements was performed in order to match the 312 15-min intervals of UVIOS to the measured data. The UVIOS data availability is 93%, while for the ground stations it reaches 313 almost 79% enabling a direct UVI data comparison of 77% of the 2017 time steps. For the comparison we used the closest 314 instrument measurements to the 15-min intervals with a maximum deviation of 3 minutes in order to avoid solar elevation and 315 cloud presence mismatches. Additionally, the UVIOS comparisons included measurements up to 70 degrees SZA. The 316 rationale for this cutoff was that UVIOS retrievals at high SZA are retrieved as cloudless as COT is unavailable from MSG. 317 In addition, the comparison is also impacted by limitation of the horizon of ground-based (GB) sites (e.g. Davos, Innsbruck, Where εi = xfxo are the residuals (UVIOS errors), calculated as the difference between the simulated values (xf) and the 329 ground-based values (xo), and where N is the total number of values. MBE quantifies the overall bias and detects whether the 330 UVIOS overestimates (MBE>0) or underestimates (MBE<0). RMSE quantifies the spread of the error distribution. Finally, 331 the correlation coefficient (r), as well as the coefficient of determination (R 2 ) were used to represent the proportion of the 332 variability between modeled and measured values. of the cases are not represented by a normal distribution and also small UVI differences (especially in high cloudiness instants 344 and high solar angles -e.g. differences less than 1 UVI having UVI values less than 1) can differ by up to 500%. Thus, we 345 have focused on absolute differences in order to have a more representative assessment of the actual effect (UV Index) and its 346 results based on low (less than 0.5), moderate (0.5-1) and high (more than 2) UVI differences between UVIOS and the ground 347 based measurements. However, in the Appendix A we also provide such differences in percent. 348 In Table 3, U1.0 and U0.5 represent the percentage of cases with absolute differences between modelled and ground based 349 UVI measurements within 1 and 0.5, respectively, for all comparisons between the 15-minute model retrievals and the 350 corresponding ground-based measurements. As shown in Table 3 Figure 6 shows the effect of 376 this TOC bias on the calculated UVIOS. As seen in Table 4, there is a mix of small underestimation and overestimation cases 377 in the TOCs used within UVIOS, with average absolute differences of 4-5 DU. Worst TOC UVIOS inputs were found in AOS 378 and ROM (7.6 and -9.9 DU) leading to maximum (at 30 degrees SZA) differences in UVI of -0.22 and 0.3 for AOS and ROM, 379 respectively. In general, in most of the cases UVI mean differences are less than 0.1. It has to be noted that the TOC differences 380 have a larger impact when expressed in percent at higher SZAs, while in Figure 6 higher absolute differences for low SZA's 381 are associated with higher UVIs at these SZAs. Detailed comparisons for each station are shown in the Appendix A figures. 382

Aerosol effect 383
Aerosol optical depth measurements used for the UVIOS aerosol input evaluation have been collected from the AERONET- providing for a real time UV Index for Europe is using the CAMS (for AOD) and the TEMIS (for TOC) data. 398 In order to evaluate the effect of AOD on UVI, UVI differences between the UVIOS using both AOD datasets (CAMS and 399 AERONET) as UVIOS inputs were analyzed. Figure 7 shows the mean bias error of the CAMS -AERONET AOD impact on 400 UVI for all stations with available ground based AOD data as a function of SZA together with the uncertainty range (± 1 σ). 401 It can be seen that UVIOS with CAMS AOD input underestimates UVI compared to the UVIOS with AERONET data, except 402 for the UCC station. This is consistent with CAMS overestimations of AOD compared to the AERONET measurements, 403 except for the station UCC as shown in Table 5. Higher aerosol levels in the atmosphere tend to lower the UVI. Highest 404 difference in UVI is observed for the stations HEL, SOD, VIE. Since, the aerosol level at the stations HEL and SOD is very 405 low, the percent difference between the AOD from CAMS and AERONET is larger for these stations (although the absolute 406 difference is similar) relative to stations with higher AOD, leading to higher differences in the UVI. Aerosol content for VIE 407 is higher than HEL and SOD but still within 0.2 which might be the reason for the higher UVI difference. In terms of SZA, it 408 is observed that the mean bias decreases with an increase in the SZA as the values of UVI also decrease with SZA and the 409 most deviation is for station VIE which is consistent with the poor correlation between the CAMS forecasted input and the 410 measurements for this station as seen from Table 5.

Albedo effect 421
Surface albedo at UV wavelengths is small (2 -5%) for most types of surfaces (Feister and Grewe, 1995;Madronich, 1993)  around the station that does not necessarily coincide with the "effective" albedo area affecting UV measurements; and finally 435 that the monthly albedo product represents a monthly average while a real time CGLS product represents the last 12 days 436 (dynamically changing albedo). In order to investigate this last point we have compared the UV effects from the use of the two 437 albedo datasets for DAV station. In Fig. 10, the effect of surface albedo correction is shown for the Davos station, for a period 438 with snow cover and low percentage cloudiness. The climatological and the dynamically changing albedo are presented in 439 terms of percentage differences between modelled and ground measurements as a function of SZA. In the case of climatological 440 albedo, most of the percentage difference between forecasted and the measured UVI value is found to vary from -30% to 10% 441 for SZA between 20° to 70°, showing more underestimation than overestimation from the UVIOS simulations. Similarly, in 442 the case of dynamically changing albedo, most of the percentage difference between forecasted and the measured UVI value 443 is found to vary from -20% to 10% for SZA between 20° to 70°. The mean percentage difference between the results using the 444 two different albedo inputs is -2.76% in terms of accuracy improvement. However, beyond 70 degree SZA, there is a huge 445 variation in the percentage difference with mostly underestimations from the UVIOS simulations (not shown in Fig. 10). 446

Cloud effect 447
For the evaluation we used measurements at SZA lower than 70 degrees, based on the lack of cloud input from MSG for higher 448 SZAs. The lack of MSG data results in an overestimation of UVIOS in high SZAs and the UVI is systematically overestimated 449 for long periods during winter at high latitude regions when SZA does not get below 70 degrees during the day. However, this 450 overestimation is low in terms of absolute UVI and does not usually exceed 0.2 UVI because maximum UVIs at such SZAs 451 rarely exceed UVI=1. As shown in Table 6 there are 45.4% of cases with underestimations and 54.6% cases with overestimations for cloudless 466 conditions (COT=0). For all the other cases, overestimations (62.5%) are more predominant than underestimations (37.5%). 467 The difference in the modelled and the measured values goes beyond ±1 UVI for only 5.1% cases for cloudless condition and 468 14.7% for all other cases. In general, under cloudy conditions, UVIOS shows an overestimation for UVI in contrast to the 469 ground measurements. One explanation for the overestimations could be the erroneous determination of COT from MSG above 470 the ground-based stations, giving cloud input that can be overestimated or underestimated. The results show that there is a 471 general tendency for a small underestimation of MSG COT that leads to a systematic but small UVIOS UVI overestimation 472 under cloudy conditions. Another possible explanation is the spatial representativeness of MSG COT. The MSG COT 473 determination is available at 5 by 5 km pixels that may differ from the actual situation of the cloud prevailing above the station, 474 especially in broken cloud conditions and a case when it blocks the direct radiation from the sun. Moreover, for lower solar 475 elevations, the direct sun irradiance can be blocked by cloud in neighbouring pixels. The first effect has been explored in the 476 relative frequency distribution of Figure 11 that shows a higher number (~ 63%) of data on the right of the zero UVI difference 477 vertical line for cloudy skies. When comparing data outside the 0.5 and 1 difference limits we also see that 1 -4 times more 478 data show a UVIOS overestimation as compared to the clear sky case. This shows that in general there is a small (in UVI 479 terms) but significant UVIOS overestimation for non-zero COT conditions. Moreover, for clear skies, as determined from the 480 MSG, we observe a less pronounced UVIOS overestimation that corresponds to the fact that even if MSG defines the situation 481 as completely cloudless, in reality there may be some cases where clouds near the GB station affect the measured UVI. This 482 effect is easier to understand when showing these differences as a function of solar zenith angle which is explored through 483 It is observed that there is an obvious pattern of scattered data for UVI differences higher than 1.5 compared with the ones for 490 differences less than -1.5. These data represent UVIOS overestimation for UVI retrievals due to the underestimation of the 491 cloudiness just above the stations. These data illustrate the well-known spatial representativeness issues whereby a COT value 492 for a satellite grid is not fully representative of a point measurement station. In addition, absolute and percentage relative 493 differences are shown in Fig. 14 (c) and (d) respectively for SZA up to 65 degrees. The differences between the UVIOS and 494 the GB UVI decreases in absolute level but increases in percent with an increase in SZA. This is due to the decrease of UVI 495 with increasing SZA. Modelled and the measured UVI difference is close to zero both for mean and median values. For SZA 496 below 30 degrees differences are 0 to -0.2, while 20 to 80 percentiles range from -0.6 to -0.2. Percentage difference increases 497 with SZA as absolute UVI decreases with the 20 to 80 percentiles showing differences between -10% and 10% 498 Fig. 15 presents the effect of surface elevation on UVI as a function of the percentage difference for various total ozone 500 columns. The UVI percentage difference is found to increase almost linearly with the increase in elevation for a particular total 501 ozone column. The percentage difference is similar for all ozone columns up to 1km, after which the differences with ozone 502 column become more apparent. That is, at a particular elevation, the percentage difference is higher for less total ozone column. 503 A 1% fluctuation (decline or increase) in column ozone can lead to about a 1.2% fluctuation (increase or decline) in the UV 504

Surface elevation correction 499
Index (Fioletov et al., 2003;Probst et al., 2012). Indicatively, the surface elevation correction in UVI for the DAV station (due 505 to UVIOS input deviation from to actual elevation) was of the order of 15%, while for INN and AOS it was 6% and for the 506 VAL station close to 8%. 507

Summary and conclusions 508
In this study, a fast RTM model of UVI, the so-called UVIOS, using inputs of the SZA, aerosol optical depth, total ozone 509 column, cloud optical depth, elevation and surface albedo that implicitly includes temporal effects and the effect of cloud and 510 aerosol physics, allows for the generation of high-resolution maps of UVI. Ground based measurements of UV are the most 511 accurate way to determine this important health related parameter. However, such stations are sparse and hence, satellite 512 observations can be used in order to have a nowcasted UV service. To date polar orbiting satellites like TOMS, OMI and 513 recently TROPOMI provided a global UV dataset with a major disadvantage being the temporal resolution (one measurement 514 per day). This, combined with the large temporal variability of clouds can lead to huge deviations from reality when a single 515 https://doi.org/10.5194/amt-2020-506 Preprint. Discussion started: 1 March 2021 c Author(s) 2021. CC BY 4.0 License. daily measurement is included. Geostationary satellite, MSG, have been used in order to try to improve on such limitations 516 using cloud information every 15 minutes. 517 Comparison of the forecasted and the ground-based measurements indicated that at least 70% and 80% of comparisons were 518 within 0.5 UVI difference for all sky condition and clear sky, respectively. The mean differences TEMIS TOC and the ground 519 measured TOC from the WOUDC for one year of comparison data showed that TEMIS tends to slightly overestimate the TOC 520 for some stations along with underestimating it for other stations. While, in general, in most of the cases UVI mean differences 521 are less than 0.1, the TOC differences have a larger impact in percent UVI differences at higher SZAs. Such small differences 522 can also be the result of daily TOC variation not captured in TEMIS. 523 CAMS AOD seems to be slightly overestimated as compared with AERONET data that leads to a UVIOS underestimation. 524 CAMS data are found to overestimate the AOD from AERONET measurements with a mean difference of 0.07 at 500 nm. 525 All the stations have a mean positive bias up to 0.071 except one station that had a mean negative bias of 0.007. The analysis 526 of the impact of the mean bias error of the CAMS -AERONET AOD impact on UVI for all stations showed that the mean 527 bias decreases with an increase in the SZA as the values of UVI also decreases with SZA. The greatest deviation is for station 528 VIE which is consistent with the poor correlation between the CAMS forecasted input and the measurements for this station. 529 The real time data provision approach of UVIOS requires using a maximum of one-day ozone and aerosol forecast using the 530 TEMIS and CAMS service respectively. 531 Cloudy conditions show high percentage differences but low UVI differences, and have a general tendency to lead to a UVIOS 532 overestimation. It was found that 45.4% of cases have underestimations while 54.6% cases have overestimations for the 533 cloudless conditions, while overestimations (62.5%) were more predominant than underestimations (37.5%) for all the other 534 cases. In general, UVIOS showed an overestimation for UVI in contrast to the ground measurements under cloudy conditions 535 with the difference in the modeled and the measured values going beyond ±1 for 5.1% cases for cloudless conditions and 536 14.7% for all other cases. At individual stations the results for cloudless sky conditions, which are the most important for 537 health related issues, showed good agreement. In general, ~85% of all and 95% of cloudless cases are within 1 UVI difference. 538 The relative percentage biases can be large for low UVI cases due to clouds or at high SZAs, above 75°, due to the absence of 539 accurate information for clouds. The results show that there is a general tendency of small underestimation of MSG COT that 540 leads to a systematic but small UVIOS overestimation under cloudy conditions. Another possible explanation is the spatial 541 representativeness issues between a satellite and a single point on the ground. 542 SSA validation is difficult to perform as there are no systematic SSA measurements in the UV region. Using climatological 543 surface albedo has little impact at low albedo sites but mainly leads to underestimations in UVIOS simulations for high albedo 544 situations (snow cover). Most of the percentage difference between forecasted and the measured UVI values varied from -30% 545 to 10% for SZA between 20° to 70° (climate albedo), while it was found to vary from -20% to 10% for dynamically changing 546 albedo. Since high surface albedo conditions correspond to winter months (i.e. high SZAs and relatively low UVI) for the 547 stations used in the study, the corresponding absolute differences in the UVI are generally smaller than 2 UVI. However, there 548 was a huge variation in the percentage difference beyond 70 degree SZA with mostly underestimations from the UVIOS 549 https://doi.org/10.5194/amt-2020-506 Preprint. simulations. Finally, for uncertainties in elevation inputs, the UVI percentage difference is found to increase almost linearly 550 with the increase in elevation for a particular total ozone column and beyond that, it is seen that the rate of increase in the 551 percentage difference decreases with increase in the total ozone column.

Acknowledgements 570
We acknowledge the Eumetsat SAFNWC, the Copernicus and TEMIS services as well as the Aerocom and GOME teams for 571 providing all the necessary data used in this study. We would like to thank the 17 site instrument operators and technical staff 572 that made the ground based measurements feasible. 573