Validation Report of the CAMS UV processor Issue #22 December 2020/ January-February 2021 (DJF) CAMS-72: Solar radiation products

Full text

(1)

ECMWF COPERNICUS REPORT Copernicus Atmosphere Monitoring Service

Validation Report of the CAMS UV processor

Issue #22

December 2020/ January-February 2021 (DJF)

CAMS-72: Solar radiation products

Issued by: M.R.A. Pitkänen, W. Wandji, A. Arola, FMI Date: 06/06/2021

(2)

Copernicus Atmosphere Monitoring Service

This document has been produced in the context of the Copernicus Atmosphere Monitoring Service (CAMS).

The activities leading to these results have been contracted by the European Centre for Medium-Range Weather Forecasts, operator of CAMS on behalf of the European Union (Delegation Agreement signed on 11/11/2014). All information in this document is provided "as is" and no guarantee or warranty is given that the information is fit for any particular purpose.

(3)

Copernicus Atmosphere Monitoring Service

Contributors

FMI

Mikko R. A. Pitkänen William Wandji Antti Arola

(4)

Copernicus Atmosphere Monitoring Service

Contents

1. Source of data 7

1.1 Ground based measurements 7

1.2 CAMS UV data 10

2. Methodology 12

3. Results and discussions 13

3.1 Performance of the UV processor – UV Index 13

3.2 Performance of the UV processor – spectral UV 22

4. Conclusions 28

5. References 29

(5)

Copernicus Atmosphere Monitoring Service

Summary

The CAMS UV processor produces global UV index and spectral UV forecasts, whose performance is regularly monitored in a comparison against high quality ground-based reference measurements starting from 2016. The reference data is measured using broadband radiometers and spectroradiometers presently at a total of 41 locations with wide latitudinal coverage. Overall, the performance of both UV index and spectral irradiance forecasts is high in terms of high accuracy and low bias as well as high correlation in comparison to the reference data. At some measurement sites, improvements near the CAMS model cycle upgrade can be identified from the improved accuracy, while at many sites the improvement is masked behind the high natural variability of UV and behind seasonal changes in UV irradiance. Detailed global and regional changes in UV forecasts between model cycles can be identified by comparing the overlapping periods of two model runs (see reports for SON 2017, JJA 2018, JJA 2019, and SON 2020), and the changes can be associated with, for instance, aerosols, but the improvements in model cycle performance cannot currently be confirmed very effectively due to limited spatial coverage of reference data. Overall, UVI and spectral UV index forecast accuracy (median relative root-mean-square difference rRMSE 0.37) and correlation (median Pearson coefficient 0.92) have improved since 2016, and a portion of that can be accounted for improvements in the UV processor itself and in the CAMS model cycle upgrades. Median relative bias reduced from 0.03 in DJF 2020 to -0.01 in DJF 2021. Overall improvement since 2016 can be seen in the CAMS spectral UV irradiance for the DJF season. Among the three stations with spectral UV reference data CAMS UV showed good performance at each site, but with all statistics considered, no site indicated better overall results for UV forecasts over the other stations during DJF 2021.

(6)

Copernicus Atmosphere Monitoring Service

Introduction

As part of supporting the development of the ultraviolet (UV) radiation products of CAMS Services for solar radiation, an evaluation of the UV processor against ground-based spectral UV data is carried out on a periodic basis. For this purpose, the best possible quality of ground measurements is required, and data must be provided in a reasonably short temporal delay after the acquisition. The data is obtained from both public and restricted access online databases for UV data and through direct contacts with institutes performing the measurements. These ground-based measurements are UV Index (UVI) and spectral UV irradiances.

In the previous report, CAMS UV radiation output was validated against 41 ground-based stations located in Europe, Israel, Thailand, Australia, New Zealand and Antarctic. All available data from September to November 2020 (hereafter referred as SON 2020) were used in the previous report. Furthermore, all available data from January 2016 until November 2020 were used for long-term time series assessments.

This report is the 22nd issue in a series of periodic reports, and it focuses on December 2020 to February 2021 (DJF 2021). After requesting UV measurements from data providers, a total of 41 ground-based sites were available for DJF 2021. The results are compared with those obtained by a similar validation for the same calendar months from the previous four years, 2016, 2017, 2018, 2019, and 2020.

(7)

Copernicus Atmosphere Monitoring Service

1. Source of data

1.1 Ground based measurements

Ground-based measurements are obtained from:

- 11 European stations from the COST-713 UV Index Database hosted by the Finnish Meteorological Institute (FMI)

- Marambio (Antarctic) and Sodankylä (Finland) provided directly by FMI

- Reading (United Kingdom) spectral UV from EUVDB. UV index from DEFRA, https://uk-air.defra.gov.uk/data/uv-data (valid on 2021-06-02, acknowledgement: Crown 2020 copyright Defra via uk-air.defra.gov.uk, licenced under the Open Government Licence (OGL).)

- Thessaloniki (Greece). Data provided by the Laboratory of Atmospheric Physics, Aristotle University of Thessaloniki, through personal contact

- 17 Australian stations from ARPANSA network

- 3 Israelian stations from Israel Meteorological Service (IMS) - 4 stations in Thailand from Silpakorn University

- 3 stations in New Zealand from National Institute of Water and Atmospheric Research (NIWA) CliFlo: NIWA’s National Climate Database on the Web, http://cliflo.niwa.co.nz (valid on 2021-06-02). Acknowledgement: CliFlo: NIWA's National Climate Database on the Web, http://cliflo.niwa.co.nz/, retrieved 17-May-2021.

Figure 1a and 1b display three maps of all stations in Europe, Israel, Thailand, Australia, New Zealand, and Antarctic. Table 1 lists all the 41 stations, their corresponding abbreviations, location coordinates, and altitudes. Thessaloniki provides spectral UV measurements, while Sodankylä and Reading provide both spectral and UV index data.

Figure 1a: Maps showing 41 stations used for the periodic validation in Europe, Israel, Thailand, and Australia except four stations located in the Antarctic.

(8)

Copernicus Atmosphere Monitoring Service

The data from ground-based measurements consist of UVI and spectral UV irradiance. UVI is obtained from two different types of instruments: multi-band filter radiometer and broadband detectors, while spectral UV is measured with spectroradiometers. The broadband detectors account for a weighting function that mimics that of the erythemal one.

The very best target for the uncertainty of broadband instruments falls within 10-15% (Seckmeyer et al., 2006). However, for correcting the cosine error, in addition to the solar zenith angle (SZA), the correction method also requires the total column ozone (Bodhaine et al., 1998). It is a laborious task to determine such a calibration. To our knowledge, this correction has not been fully considered in those broadband measurements submitted to the COST-713 UV Index Database. Therefore, standard uncertainties can be estimated as being slightly higher with up to 20%. Norwegian ground-based UV (GUV) instruments, on the other hand, are somewhat different instruments (https://uvnett.dsa.no/instrumenter_en.aspx valid on 2021-06-06) and the network is very well maintained. Therefore, we anticipate those standard uncertainties being closer to the 10-15% range.

Figure 1b: map showing the four Antarctic stations. The vertical line from top to bottom indicates

the 0° and 180° longitudes, respectively.

(9)

Copernicus Atmosphere Monitoring Service

Table 1: the 41 stations that provided spectral UV (bold) and UV index during DJF 2021.

Station Abbr. Latitude Longitude Altitude

Ny-Ålesund, Norway NYA 79.0 N 11.83 E 15 m

Alomar, Norway AND 69.2 N 16.0 E 380 m

Sodankylä, Finland SOD 67.22 N 26.39 E 179 m

Trondheim, Norway TRH 63.42 N 10.38 E 70 m

Kise, Norway KIS 60.78 N 10.82 E 140 m

Finse, Norway FNC 60.58 N 7.57 E 1200 m

Bergen, Norway BRG 60.38 N 5.33 E 40 m

Kjeller, Norway KJE 59.98 N 11.05 E 143 m

Østerås, Norway OST 59.92 N 10.75 E 150 m

Landvik, Norway LAN 58.33 N 8.52 E 10 m

Reading, United Kingdom READ 51.44 N 0.94 W 66 m

Florence, Italy FLO 43.82 N 11.20 E 50 m

Thessaloniki, Greece THE 40.63 N 22.95 E 60 m

Bet-Dagan, Israel BET 32.0 N 35.21 E 25 m

Jerusalem, Israel JER 31.78 N 34.96 E 700 m

Eilat, Israel EIL 29.55 N 34.81 E 10 m

Chiang Mai, Thailand CMA 18.78 N 98.98 E 312 m

Ubon Ratchathani, Thailand UBO 15.25 N 104.87 E 120 m

Nakhon Pathom, Thailand NAK 13.82 N 100.04 E 72 m

Songkhla, Thailand SNG 7.2 N 100.06 E 15 m

Darwin, Australia DAR 12.42 S 130.89 E 30 m

Townsville, Australia TOW 19.33 S 146.76 E 10 m

Emerald, Australia EME 23.53 S 148.16 E 190 m

Alice Springs, Australia ALS 23.80 S 133.90 E 550 m

Brisbane, Australia BRI 27.45 S 153.03 E 20 m

Gold Coast, Australia GLD 28.00 S 153.37 E 10 m

Perth, Australia PER 31.92 S 115.96 E 15 m

Newcastle, Australia NEW 32.90 S 151.72 E 20 m

Sydney, Australia SYD 34.04 S 151.10 E 20 m

Adelaide, Australia ADE 34.92 S 138.62 E 10 m

Canberra, Australia CAN 35.31 S 149.20 E 580 m

Leigh, New Zealand LEI 36.27 S 174.80 E 27 m

Melbourne, Australia MEL 37.73 S 145.10 E 60 m

Kingston, Australia KIN 42.99 S 147.29 E 50 m

Christchurch, New Zealand CHR 43.53 S 172.61 E 6 m Invercargill, New Zealand INV 46.42 S 168.33 E 0 m Macquarie Island, Australia MIS 54.50 S 158.94 E 10 m

Marambio, Antarctic MAR 64.24 S 56.63 W 198 m

Casey, Australia CAS 66.28 S 110.52 E 40 m

(10)

Copernicus Atmosphere Monitoring Service

1.2 CAMS UV data

The UV processor is part of the CAMS Integrated Forecasting System (IFS). It provides UV estimates from July 2012 onwards. A revised version of the UV processor has been implemented and used since 3rd September 2015 (Bozzo et al., 2015). The CAMS model output is provided with 5 nm spectral

resolution. Knowing the erythemal action spectrum given by McKinlay-Diffey’s definition (McKinlay and Diffey, 1987), spectral data are integrated to get the UVI. UV irradiances integrated into 5 nm spectral bands from 280 nm to 400 nm are also provided and result in 24 spectral bands. Both UVI and UV irradiances are given for clear and all-sky conditions separately. The outputs are instantaneous values.

Before 3rd September 2015, UVI estimates, and UV irradiances were produced every 3 hours. Since

then, UVI has been provided every hour (MARS keywords: type=fc, stream=oper, param=2.214/3.214, class=mc, expver=0001, levtype=sfc) whereas the UV irradiances are still provided in the 3–hourly output frequency (MARS keywords: type=fc, stream=oper, param=55.210, class=mc, expver=0001, levtype=ml). The modeled UVI and spectral irradiances evaluated in this report are the forecasted (00 UTC base time every day) UV values closest to observations.

In the past years the CAMS-IFS UV forecasts have been affected by several upgrades, the important ones have been listed in Table 2. Note that all model cycle upgrades, except 41R1_CAMS_hires, include also meteorological changes originating from the upgrades of the underlying IFS weather prediction model. Hence, the differences between model cycles are a combination of improvements in both meteorology and UV related data assimilation and modeling. Further information on changes in both IFS and CAMS-IFS models can be found at https://atmosphere.copernicus.eu/node/326 (valid on 2021-06-02). For information on the dissemination of the current global NRT products see

https://confluence.ecmwf.int/display/COPSRV/CAMS+Global (valid on 2021-06-02). UV index

forecasts can also be accessed in chart form at

https://atmosphere.copernicus.eu/charts/cams/uvindex-forecasts (valid on 2021-06-02)

The cycle upgrade to 47r1 took place on 2020-10-06 including modifications in aerosols, ozone and the UV processor parameterization, and so a notable change in CAMS UV performance took place. Further details on the changes can be found in the Validation Report of the CAMS UV processor Issue #21 for SON 2020 (ref. CAMS72_2018SC3_D72.2.1.1-2021Q1_UV_VAL_202103_v1) and at

(11)

Copernicus Atmosphere Monitoring Service

Table 2: Important UV related changes of CAMS IFS NRT forecasts

CAMS cycle Implemented on Implementations potentially affecting UV forecasts

47r1 2020-10-06 Updated anthropogenic emissions inventories and volcanic outgassing. Updated biomass-burning emissions and introduction of Hybrid Linear Ozone scheme. Updated dust sources and new sea-salt emission scheme. Revised coefficients in UV processor, based on ATLAS3 spectrum

46r1 2019-07-09 Updated emissions for anthropogenic, biogenic, dust and sea salt aerosols. New aerosol species: nitrate and ammonium. Coupling of sulfur species between chemistry and aerosol schemes. 45r1 2018-06-26 Assimilation of NO2 from GOME-2 satellite instrument. Changes

in emissions and modeling of sea salt, volcanic and biomass burning and secondary organic aerosols

43r3 2017-09-26 The extraterrestrial UV spectrum Kurudz upgraded to newer ATLAS3. Updates optical properties of organic and sea salt aerosols. Improvements to aerosol modeling. Improved use of ozone information in the UV processor.

43r1 2017-01-24 Improved use of cloud overlap information in UV processor. Changes to dust, sulfate and SOA aerosol modeling. Assimilation of satellite based vertical profiles of ozone (OMPS).

41r1_hires 2016-06-21 Increase in horizontal resolution from T255 to T511

41r1 2015-09-03 New UV processor. UV coupled with prognostic aerosols instead of climatological aerosols. UV index forecasts produced now hourly in addition to the 3-hourly forecasts. Assimilation of MODIS deep blue AOD and GOME-2 SO2, affecting UV

attenuation by aerosols. Modifications in emissions of organic matter and black carbon.

(12)

Copernicus Atmosphere Monitoring Service

2. Methodology

Before evaluating the performance of the CAMS UV processor, the best quality of the ground-based measurements should be selected. In order to achieve this goal, two main constraints have been applied on ground-based measurements. The first one is that only UV measurements with solar zenith angle (SZA) lower than 88° are used. This criterion has been also applied on ECMWF outputs. Realistically, UV measurements cannot be equal to zero. Therefore, the second constraint is that measurements should be greater than an infinitesimal threshold set empirically to 0.005 UVI. Additional constraints have been applied based on the thresholds for range of 2.5° SZA range based on the 99% and 1% percentiles of long-time series of measurements accounting for the fact that the presence of broken cloudiness in the sky may increase UV fluxes.

The validation in this report covers the full years 2016 to 2020 of CAMS estimated UV data as well as all the available ground-based observational data for the respective stations. It also covers 2021 until end of February and includes the most recent three-month DJF period being the focus of this validation report.

Modeled UV values are compared with measured ones. For spectral comparisons, measured spectral irradiances are integrated over each spectral range of 5 nm to match with the spectral resolution from the UV processor. According to the spectral range of UV outputs, comparisons are then carried out over the spectral range between 290 nm and 400 nm. That results in maximum 22 spectral bands for European stations, each of 5 nm width as spectral resolution.

For each station and for both UVI or UV irradiances, deviations between modeled (estimated) and measured values are computed. Pearson correlation coefficient (CC), bias (Bias), root mean square error (RMSE), relative bias (rBias), and relative RMSE (rRMSE) normalized by the mean of measured values are derived:

𝐵𝑖𝑎𝑠 =' & ∑' 𝑌+,-./0-12 (4)− 𝑌710,8912 (4) 4:& (1) 𝑅𝑀𝑆𝐸 =? &'∑'4:&@𝑌+,-./0-12(4)− 𝑌710,8912 (4)AB (2) 𝑟𝐵𝑖𝑎𝑠 = D & EFGHIJFK LLLLLLLLLLLLLLL 𝐵𝑖𝑎𝑠 (3) 𝑟𝑅𝑀𝑆𝐸 = D & EFGHIJFK LLLLLLLLLLLLLLL𝑅𝑀𝑆𝐸 (4)

where j indicates each value, and n the total number of values. The value Y can be UVI or UV irradiances for each spectral band and Ȳ represents the average value of Y. These statistics are computed monthly and seasonally.

(13)

Copernicus Atmosphere Monitoring Service

3. Results and discussions

3.1 Performance of the UV processor – UV Index

There are 40 stations that provided UV index measurements for this report, Thessaloniki being the 41st station and providing spectral UV. Figure 2 shows two examples of the scatterplot between ground-based measurements (horizontal axis) and the CAMS estimates (vertical axis) for the period of DJF 2021 by cyan dots and similarly the same periods of 2016, 2017, 2018, 2019, and 2020 by orange, yellow, light green, dark green, and purple dots, respectively, in Jerusalem, Israel (left side) and in Alice Springs, Australia (right side). Similar plots for each station are shown in the Appendix. Jerusalem (Fig. 2, left side) represents hot-summer Mediterranean climate conditions (Köppen-Geiger climate class Csa, according to Beck et al., 2018), where CAMS UV index had improved forecast performance compared to DJF 2020. During DJF 2021 CAMS UVI shows correlation with measurements of 0.95, rBias of -0.07 and rRMSE of 0.28, which indicates more accurate forecasts compared to DJF 2020 (correlation coefficient 0.90, rBias -0.04, and rRMSE 0.41). This change can be seen in Fig. 2 as reduced scatter around the bulk of data points near the 1-1 line, especially at measured UVI values below 2. In Jerusalem UVI index reached its maximum value of 6.7 in this season, which is approximately half of the highest UVI values typically measured at Jerusalem during a full year.

In DJF 2021 UV comparison Alice Springs is an example of a subtropical hot desert station (Köppen-Geiger climate class BWh, according to Beck et al., 2018), where CAMS UV indicates improved statistics for DJF season from DJF 2020 to 2021. In Figure 2 (right side) a persistent bias during earlier years appears to have greatly reduced for DJF 2021 (rBias 0.05) compared to DJF 2020 (rBias 0.23), and similarly rRMSE was notably reduced from 0.38 to 0.27 from DJF 2020 to DJF 2021, respectively. Correlation coefficient changed little from DJF 2020 (0.96) to 2021 (0.95).

(14)

Copernicus Atmosphere Monitoring Service

Figure 3 shows the frequency distributions of the measurements in dark blue, CAMS estimates in orange, and the corresponding differences in light blue for the period of DJF 2021. The highest frequencies in Jerusalem (Fig. 3 left side) correspond to typically slightly negative differences, confirmed by the 25-, 50 and 75-percentiles of UVI difference (-0.28, -0.07 and 0.02, respectively). In Alice Springs (Figure 3 right side), however, the 25-, 50 and 75-percentiles of UVI difference -0.11, 0.14 and 0.98, respectively, reveal a positively biased distribution. This positive bias appears mostly for UVI above 7. Contrasting with Jerusalem, Alice Springs has considerably higher UV error variability, which could be partially explained by the fact that UVI in stations with smaller UV levels tend to indicate smaller absolute biases compared to UVI at stations with higher UV levels.

Figure 4 shows the full time series of UV measurements and CAMS estimates and their corresponding absolute bias, relative bias and RMSE in a moving time window from January 2016 to February 2021 in Alice Springs. The statistics were calculated for measurements close to local noon when the Sun is at its highest elevation in the sky. While the seasonal variability of UV index tends to mask some of the developments of CAMS UVI performance, Fig. 4 shows that rBias averages are considerably reduced from earlier seasons to DJF 2021 (rBias 0.05) compared to DJF 2020 (rBias 0.23), for instance. Also, rRMSE reduced from DJF 2020 (rRMSE 0.38) to 2021 (rRMSE 0.27), which indicates notable improvements from persistent positive bias to only slightly biased conditions. Figures like Fig. 2, Fig. 3 and Fig. 4 for each station are shown in the Appendix.

Figure 2: Scatterplots between measurements and CAMS estimates of UV Index for DJF in 2021, 2020, 2019, 2018, 2017 and 2016: Jerusalem, Israel (left side) and Alice Springs, Australia (right side).

(15)

Copernicus Atmosphere Monitoring Service

Figure 3: Frequency distributions of the measurements, CAMS estimations and their corresponding deviations of UV Index for DJF 2021: Jerusalem (left side) and Alice Springs (right side).

Figure 4: Time series of measurements and CAMS estimates of UV Index and their corresponding absolute bias, rBias and rRMSE in terms of moving average from January 2016 to February 2021 in Alice Springs. The

(16)

Copernicus Atmosphere Monitoring Service

DJF seasonal CAMS UVI performance statistics on correlation coefficient, rBias and rRMSE are shown in Figure 5 for different years. 2016, 2017, 2018, 2019, 2020, and 2021 are given in orange, yellow, light green, dark green, purple, and cyan, respectively. Ny-Ålesund (NYA) and Kise (KIS) have only 9 and 5 collocated data points, respectively, and therefore those results do not provide a reliable estimate on CAMS UVI performance. Also, due to the high latitude low sun conditions, Ny-Ålesund (NYA), Alomar (AND), Sodankylä (SOD), Trondheim (TRH), Kise (KIS), Bergen (BRG), Kjeller (KJE), Østerås (OST), and Landvik (LAN) observed only very low UVI values below 1, which adds uncertainty to the related data analysis.

Compared to DJF 2020, and excluding the stations with maximum measured UVI less than 1, the largest decrease in rRMSE occurred at Canberra (CAN, -0.27), along with +0.04 increase in correlation coefficient. Likewise, excluding stations with less than 1 maximum UVI, the largest increases from DJF 2020 to DJF 2021 in rRMSE was seen at Darwin (DAR, +0.15), where correlation reduced by -0.09.

Based on this evaluation data set, the CAMS model UVI in DJF 2021 has a near zero negative bias, as the 25-, 50- and 75-precentiles of rBias were -0.09, -0.01 and 0.04, respectively. The smallest rBiases were observed at Darwin (DAR, -0.26), and Mawson (rBias -0.21) while the highest rBias was observed at Macquarie Island (MIS, 0.21), when excluding Landvik (LAN), Østerås (OST) and Ny-Ålesund (NYA) due to their small UV index values. rBias at the 34 remaining sites fell within range ±0.20, of which 14 stations showed rBias within ±0.05. From DJF 2020 to DJF 2021 the median rBias was decreased from 0.05 to 0.03, respectively, for the overlapping subset of measurement stations. The scatterplots and Figure 5: Statistical indicators of comparisons between CAMS estimates and measurements of UV Index at

(17)

Copernicus Atmosphere Monitoring Service

time series in the Appendix give a more detailed view of changes in UV bias and scatter between years.

For the overlapping measurement sites, the median correlation coefficient of 0.92 is high, meaning that CAMS estimates are again well correlated with measurements in DJF 2021, as was the case in 2020 (0.91). Median rRMSE was 0.39 and 0.41 for DJF 2021 and DJF 2020, respectively, which indicates slightly improved forecast accuracy at the overlapping measurement sites.

When comparing the total seasonal statistic measures for all sites in different years, not all years may include the same measurement stations, which slightly complicates the interpretation of the results. However, when expressing seasonal total values using medians instead of averages, the total statistics are more robust against inclusion and exclusion of individual sites, which yields a statistic measure better suited for comparing the overall CAMS UV performance between years.

When including all stations for DJF 2021, and not just the ones common with DJF 2020, median rBias was -0.01, correlation 0.92, and rRMSE 0.37.

The heatmap in Figure 6 aims for a qualitative overview of the seasonal correlation between CAMS UVI and UVI measurements starting from MAM 2016. As the grey value corresponds to the total average correlation coefficient 0.91 over all sites and seasons, then red and blue colors indicate anomalies from the mean value. Firstly, correlation increases over time, because the dominance of blue colors shifts to the dominance of red colors. Secondly, an annual cycle can be seen at the northern high latitude sites with typically negative correlation anomalies in the summer and positive anomalies in the winter seasons, although this feature has weakened since 2019. Since 2017 CAMS upgrade 43r1 a cyclic pattern in correlation can also be seen at stations Florence (FLO) through Townsville (TOW). The stations in Thailand (CMA, UBO, NAK and SNG) experience annually cyclic weather due to monsoon, which appears to cause a similar pattern in UV correlation coefficient, however the cycle is in the opposite phase in Thailand compared to its latitudinal neighbors.

(18)

Copernicus Atmosphere Monitoring Service

Figure 7 illustrates the same as Figure 6, but for rRMSE and color anomalies centered on total mean rRMSE of 0.40. Similar features can be seen as in Figure 6 for the correlation: 1) increased accuracy (reduction of rRMSE) from 2016 towards 2021, especially after 43r1 and 2) increased accuracy during winter at high latitude sites.

Figure 6: Heatmap of Pearson correlation coefficient for each site and for each quarterly period with DJF 2021 as the rightmost column. Grey color is centered at the total mean value 0.91 and white color indicates seasons with less than 50 data points. The implementations of new CAMS

(19)

Copernicus Atmosphere Monitoring Service

A similar heatmap for rBias is shown in Figure 8, but here anomalies in blue and red color are with regard to rBias of 0.00. The dominance of red and greyish colors indicates an overall shift towards near zero and positive biases near DJF 2017. Again, an annual cycle is apparent in the northern high latitude sites with negative rBiases in the summer and positive in the winter, except in DJF and MAM 2020, when a slight negative bias is typical.

In Fig. 8 the high rBias values in Ny-Ålesund (NYA) in MAM 2016 as well as in Landvik (LAN) during DJF 2018, 2019, and 2020 stand out from the heatmap, which is likely contributed to by the pronounced effect of data uncertainties at low wintertime UVI levels. However, the scatter is also higher at low UVI at these sites during DJF than during JJA (see scatterplots in the Appendices of CAMS72_2018SC1_D72.2.1.1-2019Q2_UV_VAL_201906_v3), which suggests there are also forecast biases present. SON 2020 and DJF 2021 stand out from in Fig. 8 at southern latitudes, as rBias is grey blue in color at several stations, where red grey dominated the during previous seasons. In the north such systematic shift cannot be seen. As shown in the previous Validation Report of the CAMS UV processor Issue #21 for SON 2020 (ref. CAMS72_2018SC3_D72.2.1.1-2021Q1_UV_VAL_202103_v1) this coincides with the CAMS cycle upgrade to 47r1 indicating, that 47r1 contributed to overall reduction in rBias.

Figure 7: Similar heatmap as in Figure 6, but for CAMS UVI rRMSE. Grey color is centered at the

(20)

Copernicus Atmosphere Monitoring Service

Figure 9 shows the relative bias as a function of the solar zenith angle (SZA) for DJF 2021. 14 (REA, FLO, JER, EIL, DAR, BRI, GLD, NEW, LEI, CHR, INV, MAR, CAS, DAV) sites showed overall increasing rBias with increasing SZA and 5 of the sites (NAK, ALS, ADE, MEL, MIS) show a decreasing rBias with SZA. At the 12 remaining sites CAMS UVI shows no consistent rBias-SZA dependency (NYA, AND, SOD, TRH, KIS, BRG, KJE, OST, and LAN are not included in these numbers due to very small maximum UVI values). All in all, the relative bias is within approximately -0.3 to +0.3 for most of the stations and for the most of SZA range except at OST, DAR, LEI, and MAW at selected SZA intervals.

Figure 8: Similar heatmap as in Figure 6, but for CAMS UVI rBias. The grey color is centered at

(21)

Copernicus Atmosphere Monitoring Service

The previous Validation Report of the CAMS UV processor Issue #21 for SON 2020 (ref. CAMS72_2018SC3_D72.2.1.1-2021Q1_UV_VAL_202103_v1) inspected the Antarctic spring season 2021, when a large and intense ozone hole was observed causing exceptionally high UV intensities for that region. The high UV season continued into the summer season DJF 2021, as can be seen in the Appendix for Marambio, Casey, Mawson and Davis. There the maximum UVI index values during DJF 2021 in the collocated data set ranged between 11.4 UVI and 13.6, which are very high values for such high latitude. During DJF 2020 the range of maximum UVI was 7.4 to 8.9 in the collocated data set. Then again, compared to the summer season for similar northern latitudes and altitudes at Ny-Ålesund, Alomar, Sodankylä and Trondheim (see Appendix), where the typical noon time UV indices ranged between approximately 2 UVI to 6 UVI, the Antarctic conditions were indeed extreme.

Figure 9: Relative bias of UV Index in centered 10° SZA bins for all stations in DJF 2021, divided in four smaller figures for clarity. Only SZA bins with ten or more data points are shown. Marker shapes and bordering correspond to the country of the measurement site, except Antarctic sites that are all shown

(22)

Copernicus Atmosphere Monitoring Service

CAMS UVI performance during DJF 2021 in the Antarctic, as shown in the Appendix, reveals typically underestimated highest UV index values above approximately 5 UVI. rBias was -0.15, 0.07, -0.21, and -0.18 for Marambio, Casey, Mawson, and Davis, respectively, and these large differences indicate difficult conditions for UV forecasting in terms of biases (median rBias for DJF 2021 was -0.01). However, correlation coefficient ranged between 0.92 to 0.94, while median was 0.92 and rRMSE ranged between 0.34, and 0.45, while median was 0.37. So, in terms of correlation, CAMS UV performance was as not weaker than elsewhere and rRMSE was approximately between 25- and 75-percentiles of rRMSE in the data set for DJF 2021.

3.2 Performance of the UV processor – spectral UV

There are three stations that provided spectral UV measurements for this report, Reading (UK), Sodankylä (Finland), and Thessaloniki (Greece), with 465, 170, and 412 observations collocated with CAMS UV, respectively. The study of the spectral UV comparisons is first done for the spectral wavelength band 310-315 nm and then for the full spectrum.

The left side of Figure 10 shows the scatterplot of measured vs. CAMS UV irradiance in the band 310-315 nm in Reading. The overall UV accuracy in DJF 2021 is high and visually similar to DJF 2020 and rRMSE reveals an improved accuracy in 2021 DJF (rRMSE 0.43) than in 2020 DJF (rRMSE 0.48). The histogram of differences in surface irradiance on the right side of Figure 10 shows that the forecast error in Reading had a slightly negatively biased distribution around 0.0 W/m2, with rBias of -0.06.

Likewise, Thessaloniki (see Appendix) indicates a small negative bias for CAMS UV irradiance at 310 nm to 315 nm compared to the measurements and rBias of -0.02, while Sodankylä shows a larger negative rBias of -0.20 and considerably smaller maximum irradiance values than at the two other stations.

(23)

Copernicus Atmosphere Monitoring Service

The results of spectral comparisons of UV irradiances are summarized in Figure 11 for DJF 2021. Two spectral bands below 300 nm are not displayed, because of very high relative measurement errors due to the very low absolute values of the irradiance and in Sodankylä this limit is set to 305 nm. Also, spectra from Thessaloniki and Sodankylä at 325 nm due to limitations of the instrument and data processing.

Firstly, for each site the correlation coefficient is high in all spectral bands and decreases slightly as a function of wavelength, except in Sodankylä, where correlation coefficient varies only little (0.92 to 0.93). Reading indicates the highest correlation in the overlapping part of spectrum at the shortest wavelengths, and the total range for correlation coefficient is 0.87 to 0.96. rBias is between -0.57 to 0.12 which is larger compared to the range of UV index rBias -0.26 to 0.21 (excluding stations with very low maximum UV values) in Figure 5 during DJF 2021. Thirdly, rRMSE varies as a function of the spectral bands in all stations and shows a similar tendency for each site. The overall range for rRMSE of spectral irradiance in DJF 2021 is 0.31-1.25, which is higher than the rRMSE range for UVI 0.18 to 0.57. This indicates higher accuracy for UV index forecasts than for spectral UV. Overall, none of the three locations indicate an overall best station for CAMS spectral UV forecast performance, but the best statistics are found for different sites depending on the wavelength.

Figure 10: Scatterplot (left side) between measurements and CAMS estimates of UV irradiance [W/m2] over 310-315 nm for DJF in 2021, 2020, 2019, 2018, 2017, and 2016 in Reading. Frequency distributions

(24)

Copernicus Atmosphere Monitoring Service

Figure 12, Figure 13, and Figure 14 show the comparison statistics for DJF 2021 and the same for the 3-month period from the previous four years Reading, Sodankylä, and Thessaloniki, respectively. CAMS Irradiance in Reading (Fig. 12) during DJF 2021 compared to DJF 2020 shows systematically improved performance with increased correlation up to by 0.04 to 0.05 during 2021. Also, rBias and rRMSE have improved in certain parts of the spectrum, namely from 310 nm to 325 nm and above 335 nm.

In Sodankylä and when measuring the change from DJF 2020 to DJF 2021, Figure 13 shows reduced of 0.02 to 0.05 in correlation coefficient, a weakened negative rBias by +0.09 to +0.14 and small changes in rRMSE between -0.01 and +0.03.

Finally, Fig. 14 indicates overall slightly reduced CAMS spectral UV performance in Thessaloniki, with DJF 2021 showing slightly weakened correlation by -0.03 to -0.04 and a slightly increased rRMSE typically by +0.04 to +0.10. rBias weakened similarly in Thessaloniki as in Reading at the same particular wavelengths.

Figure 11: Statistical indicators of UV irradiance in each spectral band for DJF 2021 for Reading, Sodankylä and Thessaloniki. Each marker is centered in its 5 nm band.

(25)

Copernicus Atmosphere Monitoring Service

Figure 12: Statistical indicators of UV irradiance for all spectral bands for Reading for DJF 2021 and the same for years 2016, 2017, 2018, 2019, 2020, and 2021. Each marker is centered in its 5 nm band.

Figure 13: Statistical indicators of UV irradiance for all spectral bands for Sodankylä for DJF 2021 and the same for years 2016, 2017, 2018, 2019, 2020, and 2021. Each marker is centered in its 5 nm band.

(26)

Copernicus Atmosphere Monitoring Service

Figure 15 displays the spectral rBias in 10-degree SZA bins for Reading and Sodankylä for DJF 2021 and Figure 16 shows the same for Thessaloniki. Each site shows a decreasing CAMS UVI bias value below 320 nm with decreasing wavelength. Also, in Reading rBias decreases with increasing wavelength. Figure 15 shows, that in DJF 2021 the overall spectral variation of rBias seen in Fig. 11 to Fig. 14 is mostly similar in all SZA bins above 310 nm, except in Sodankylä this cannot be concluded since the collocated measurements only cover the SZA range 70°-80°.

Figure 14: Statistical indicators of UV irradiance for all spectral bands for Thessaloniki for DJF 2021 and the same for years 2016, 2017, 2018, 2019, 2020, and 2021. Each marker is centered in its 5 nm band.

(27)

Copernicus Atmosphere Monitoring Service

Figure 15: Relative bias of UV irradiance as a function of wavelength, for different solar zenith angle bins, and for DJF 2021. Reading site is on the left side and Sodankylä on the right side. Each marker is

centered in its 5 nm band.

(28)

Copernicus Atmosphere Monitoring Service

4. Conclusions

For the DJF 2021 period, in general, UV processor estimates of UV index are well correlated to surface UV measurements. UVI correlation coefficients were higher than 0.87 at all but one station (excluding high latitude stations with very low maximum UV index) and the total median of 0.92 shows little difference to DJF 2021 (0.91). rRMSE median for all sites was 0.37, which also shows little difference to DJF 2020 (0.38). CAMS UVI rBias reduced from 0.03 (DJF 2020) to -0.01 (DJF 2021). A seasonal cycle was observed especially in correlation coefficient in the tropics and parts of midlatitudes with UV forecasts improving during certain seasonal conditions. Such a cycle has been also observed also at high latitudes, but those results suffer slightly from increased uncertainties in very low UV levels during winter.

As for UVI, the spectral UV irradiances from the UV processor also show a high forecast performance when compared against ground-based spectral measurements. The correlation coefficient was high, varying spectrally between 0.87 and 0.96 at the measurement sites in Reading, Sodankylä, and Thessaloniki during DJF 2021. Spectral rBias varied between -0.57 to 0.12, while rRMSE was in range 0.31 to 1.25.

(29)

Copernicus Atmosphere Monitoring Service

5. References

Bodhaine Barry A., E. G. Dutton, R. L. Mckenzie, and P. V. Johnston, “Calibrating broadband UV instruments: Ozone and solar zenith angle dependence”, Journal of Atmospheric and Oceanic Technology, 15(4):916-926, 1998.

Bozzo A., Arola A., Cesnulyte V., Pitkänen M. Report on implementation of spectral UV irradiance, MACC-III deliverable D57.2, work package 122, Feb. 2015.

Beck, H.E., N.E. Zimmermann, T.R. McVicar, N. Vergopolan, A. Berg and E.F. Wood, “Present and future Köppen-Geiger climate classification maps at 1-km resolution”, Nature Scientific Data, 2018. McKinlay, A. F. and Diffey, B. L., “A reference action spectrum for ultraviolet induced erythema in human skin” in: Commission International de l’Éclairage (CIE), Research Note, 6 (1), 17–22.

Seckmeyer G., A. Bais, G. Bernhard, M. Blumthaler, C.R. Booth, K. Lantz, R.L. McKenzie, P. Disterhoft, and A. Webb (2006), Instruments to measure solar ultraviolet radiation. Part 2: Broadband instruments measuring erythemally weighted solar irradiance. Available at: https://library.wmo.int/index.php?lvl=notice_display&id=12616, WMO/GAW No. 164 World Meteorological Organisation, Geneva.

(30)

Copernicus Atmosphere Monitoring Service

6. Appendix

This section shows the results for all stations ordered by decreasing latitude in Europe, Thailand, Australia as well as the Antarctic sites. For each station, three graphs are provided. The first graph on the top left is the scatterplot between measurements and CAMS estimates for DJF period and each year. The second graph on the top right is the frequency distribution of measurements (dark blue line), CAMS estimates (orange line) and deviations (light blue line). The third graph on the bottom is the full times series of measurements and CAMS estimates close to local noon, their corresponding bias on top and the rBias and rRMSE on the bottom in terms of moving average.

(31)
(32)
(33)
(34)
(35)
(36)
(37)
(38)
(39)
(40)
(41)
(42)
(43)
(44)
(45)
(46)
(47)
(48)
(49)
(50)
(51)
(52)
(53)

Copernicus Atmosphere Monitoring Service

ECMWF - Shinfield Park, Reading RG2 9AX, UK Contact: info@copernicus-atmosphere.eu

Figure

Updating...

References

Updating...