Erythemal daily doses measured at the Polish Polar Station, Hornsund
(77

The importance of the solar UV radiation on human health and ecosystems is
widely discussed in the literature since the ozone hole discovery in the
early 1980s (e.g., WMO, 2014). The Montreal Protocol was signed by UN
countries in 1987 to protect the ozone layer, which acts as a shield against
the solar UV. Since 1980, especially large ozone depletion was observed every
year in the late winter and spring (the so-called ozone hole) over
Antarctica (e.g., WMO, 2014). However, severe ozone losses appeared
occasionally over the Arctic, e.g., in 2011 (Garcia, 2011; Bernhard et al.,
2013) and in 2016 (

The amount of column ozone and its vertical distribution have been measured
using a ground-based and satellite network. Nowadays, the ozone distribution
over the whole globe is available for scientific purposes. The surface UV
radiation in the UV-B range also depends on the Sun's elevation, cloud and
aerosol characteristics, and the surface albedo, which are widely variable
from site to site. There are a limited number of ground-based stations
measuring erythemaly effective doses continuously for longer than 20 years.
These include only five of the northernmost stations above 70

Maintaining homogeneity of long-term UV time series (20

The erythemal UV measurements at Hornsund were carried out from 1996 up to 2001 by an improved version (with temperature stabilization) of the classic Robertson–Berger (RB) UV meter. This was a prototype of the presently widely used broadband Solar Light Model 500 (denoted SL 500) radiometer produced by Solar Light Co. RB meter was designed in the early 1970s to measure erythemal solar irradiation as its spectral characteristic resembled that of the human skin (McKinlay and Diffey, 1987). The prototype was designed in the Institute of Geophysics (IG), Polish Academy of Sciences (PAS), Belsk, in the late 1980s and since then was used in the UV monitoring at the Central Geophysical Laboratory, IG PAS, Belsk, Poland. It was moved to the Hornsund observatory in 1995 and put into regular UV monitoring in 1996 (Krzyścin and Sobolewski, 2001) that lasted up to autumn 2001. Since spring 2004, a new UV broadband meter Kipp & Zonen UVS-AE-T (Fig. 1) has been installed at Hornsund and started continuous UV monitoring in April 2005. In spring 2006 and 2007, it was calibrated by the IG PAS substandard Kipp & Zonen UVS-AE-T (no. 616), which was frequently adjusted to the Belsk's Brewer spectrophotometer (Sobolewski and Krzyścin, 2006). There were logistical difficulties with the calibration of the Hornsund meter by a higher-level standard (e.g., the Brewer spectrophotometer) as the station could only by reached by snowmobiles (in spring), helicopters, or ships (in summer). Thus, we decided to apply radiative transfer (RT) model simulations for clear-sky conditions to calibrate output of the UV radiometer during cloudless days and perform a homogenization of the past UV data.

The following ancillary data routinely measured at Hornsund is used in the model simulations: snow depth, aerosol characteristics (aerosol optical depth, single scattering albedo from the Cimel Sun photometer observations since 2004), and the sunshine duration (by a Campbell–Stokes recorder).

The observing platform at the Polish Polar Station Hornsund
(77

Clear-sky conditions over the Hornsund observatory were identified by the
examination of the 1 min erythemal irradiation daily pattern. The
smoothness of the pattern and the steady increase (before local noon) and
decrease (after local noon) of the irradiances provided a criterion for
cloudless day. The Tropospheric Ultraviolet-Visible (TUV) RT model by
Madronich (1993) is implemented to calculate hypothetical clear-sky daily
doses for the selected cloudless days. The TUV input consists of the column
ozone amount (taken from the site overpasses by the Solar
Backscatter Ultraviolet (SBUV) instrument onboard the NOAA satellites and
aerosol characteristics from the AERONET database. The ground albedo in UV
range is approximated by a local formula:

For each year (2005–2016) ratios between the modeled and observed daily doses
were averaged to provide the annual correction factor, which is applied to
all measured daily doses. The annual correction factor (ACF) was calculated
separately for selected ranges of the noon solar zenith angle (SZA);
SZA

The calibration constants for the Kipp & Zonen UVS-AE-T instrument in the period 2005–2016 derived from a comparison of the modeled clear-sky doses with the observed ones in cloudless conditions for four solar zenith angle (SZA) ranges.

The calibration constants for the prototype of the Solar Light
instrument in the period 1996–2001 derived from a comparison of the modeled
clear-sky doses, assuming fixed aerosol optical depth (0.16 at 340 nm) with
the observed ones in cloudless conditions in the March–June period. Full
circles and bars represent the mean value and

The same procedure was used for the first period (1996–2001) of the UV
monitoring at Hornsund but constant aerosols of aerosol optical depth (AOD)
at 340 nm equal to 0.16 was assumed. During that period there were no Cimel
Sun photometer observations. Thus, for the 1996–2001 calibration, we select
AOD value representing the mean AOD value found for the period 2004–2016.
Moreover, only one ACF value was calculated regardless of SZA. Figure 3 shows
yearly ACF values for the period 1996–2001. It is seen that almost linear
instrument deterioration of

Radiative model simulations of daily erythemal doses for clear-sky conditions in 1996 using observed total ozone, surface albedo, and extreme high and low monthly aerosol optical thickness derived from all CIMEL Sun photometer measurements at Hornsund in the period 2004–2016. Uncertainty is calculated as the difference between the extreme daily doses expressed in percentage of the mean daily dose.

The uncertainty of UV observations by the prototype instrument induced by unknown aerosol AOD in ACF calculations could be estimated using extreme AOD monthly values in RT simulations, i.e., 2.5 and 97.5 percentiles of AOD values taken from all Cimel measurements in a selected month for the period 2004–2016. Figure 4 shows the differences between clear-sky erythemal daily doses calculated in 1996 for extreme high (2.5 percentile) and extreme low (97.5 percentile) AOD monthly values. Actual snow cover and satellite total ozone were used in these simulations. Because the AOD variability range depends on the month, we found that the uncertainty level varies between 2 and 7 %. Further in calculations we select 7 % as a characteristic of the instrument's uncertainty induced by a lack of precise information on aerosol loading in ACF calculations for the period 1996–2001.

Past variations of the surface erythemal radiation in periods without UV
measurements could be retrieved from statistical and radiative transfer
modelling using various proxies to describe attenuation of UV radiation in
the atmosphere (e.g., Lindfors and Vuilleumier, 2005; Koepke et al., 2006;
Lindfors et al., 2007; Rieder et al., 2008). Global solar radiation, cloud
cover, solar zenith angle, and the sunshine duration were usually used as
proxies to construct empirical formulas to determine cloud attenuation of UV
radiation. Junk et al. (2007) applied an advanced statistical technique,
artificial neural network, to find out the most effective combination of
proxies for surface UV estimation. It appeared that global solar radiation,
solar zenith angles, and the diffusive part of global solar radiation were
essential proxies for UV reconstruction giving 1–2 % bias and

The observed versus modeled erythemal doses:

Here reconstruction of the erythemal daily doses is derived using hypothetical clear-sky erythemal daily doses from a radiative transfer model simulation using the following input parameters: total ozone (for satellite overpasses), albedo (retrieved from snow depth), and aerosol optical depth at 340 nm (from collocated CIMEL Sun photometer measurements). The attenuation due to clouds is derived from an empirical formula based on the daily sunshine duration measured by a Campbell–Stokes recorder. The model's regression parameters were determined using the 2005–2006–2007 daily erythemal doses (by Kipp & Zonen UV radiometer). The model is verified using the 1996–1997–1999 (output of SL 500 prototype) and 2009–2010–2011 (output of Kipp & Zonen UV radiometer) data.

A semiempirical model is built to reproduce the measured daily doses
(erythemal J m

the daily total ozone (TO

the snow albedo according to Eq. (1) with the snow depth from the station meteorological data

daily observed aerosol optical depth (AOD) at 340 nm by the collocated Cimel Sun photometer or AOD equal to 0.16, i.e., equal to long-term (2004–2016) monthly means of AOD at 340 nm, for days without CIMEL measurements.

Monthly mean of the daily erythemal doses, participation of the monthly sum of daily doses in the yearly (February–October) sum of the daily doses, monthly mean difference between the observed and modeled doses, and corresponding root mean square error (RMSE), monthly mean difference between the observed and modeled doses as a percent of the observed doses, and corresponding RMSE. The results are from modeled and observed daily doses at Hornsund for the period 2005–2016.

Finally, the relative sunshine duration (in percent of the polar day
duration), SUN_DUR(

To determine how model (2) uncertainty influences trend estimates for the
whole period 1983–2016, we propose a Monte Carlo methodology to derive the
trend value and its uncertainty based on a hypothetical bootstrap sample
(

Model (2) with CMF defined by Eq. (3) performs almost perfectly (see
Fig. 5a). The model–observation correlation coefficients exceed 0.9 and the
smoothed pattern of scattered data obtained by LOWESS (locally weighted
scatter plot smoothing, Cleveland, 1979) matches the 1–1 line (perfect
agreement line – diagonal of the square). Slope by an ordinary
least-squares fit is 0.99

The regression coefficient of model (2) was computed using the multilinear
least-squares fit to the observed 2005–2006–2007 daily doses. Comparisons
of the modeled data to the observed ones taken in different periods will
provide a kind of model verification and will support the accuracy of the
calibration constants applied to total UV data (Sect. 3). Figure 5b and c
show the comparisons for the period 2009–2010–2011 and 1996–1997–1998,
respectively. The model–observation correlation coefficients are high
(

The model–observation agreement appears to be even better for the monthly averages
of daily erythemal doses (Fig. 5d) for three periods together:
1996–1997–1998, 2005–2006–2007, and 2009–2010–2011. Here the
correlation coefficient is equal to

Time series of normalized deviations of monthly means of daily doses
and normalized yearly sum of daily doses: observed values (red circles) and
reconstructed values (blue circles). Vertical lines show

We propose a Monte Carlo procedure to estimate a linear trend that accounts for
various uncertainties of the daily doses throughout the examined period. The
monthly and yearly trend values and their significance are derived averaging
linear regression coefficients and their errors taken from a standard
least-squares linear regression applied to a large number (

For the periods 1983–1995 and 2001–2004, we use the reconstructed data,
based on model (2), adjusted for the model (2) uncertainty,
DOSE

For the period 1996–2001 and 2005–2016 we use the Monte Carlo set of the
potential representatives of the observed time series that is also adjusted
for the observation uncertainty, DOSE

Monthly means of daily erythemal doses show significant intrayear variability (Table 1), with a late spring–early summer maximum. Therefore, to compare trend values in selected months, the trend analyses are applied to percentage differences of the monthly mean of daily doses from the long-term monthly mean (2005–2016). The yearly sum of erythemal doses are calculated as a sum of the daily doses between 1 March and 30 September as earlier, and later doses are small or zero because of high solar zenith angles and polar night (between 29 October and 11 February). Similarly, to the monthly means of daily doses, the yearly sum is converted to the normalized departures relative to the mean yearly sum in the period 2005–2016.

Figure 6 illustrates the time series of the monthly (March-September) and yearly normalized departures from the long-term (2005–2016) monthly means and a yearly sum of daily doses. The regression lines show the long-term tendency in the 1983–2016 period and in the shorter period (1996–2001 and 2005–2016) when only the results of measurements were taken into account.

There are large year-to-year fluctuations in the monthly fractional
deviations in the range between

Table 2 presents statistical characteristics of the Monte Carlo trend
estimates. The mean linear trend coefficients and their mean standard errors
together with pertaining range of estimates (between minimum and maximum of
the slope and standard deviation) were derived by examining all slopes and
their errors obtained by an ordinary least-squares fit to each
Monte Carlo representative of the original time series. The trends are
calculated for the period 1983–2016 (both observed and reconstructed data),
for the period 1996–2016 (with observed data for the 1996–2001 and
2005–2016 periods, and the reconstructed one for the period 2002–2004), and
for the period with the UV measurements only (with the gap for the period
2002–2004). The statistically significant decline at 2

The monthly and yearly mean slope of the linear fit and the
corresponding standard deviation (in percentage per year) from 10 000
samples of the hypothetical time series for the periods 1983–2016 (modeled
and observed data) and 1996–2016 (observed and modeled data), and for the
periods 1996–2001 and 2005–2016 (observed data only). The minimum and
maximum values of the slope and its standard deviation selected from 10 000
simulations are in the parentheses. Numbers in bold font represent
statistically significant estimates at 2

To find sources of the long-term UV variability at Hornsund, we also analyze
time series of yearly (March–September) sums of hypothetical clear-sky daily
doses by RT simulations (using total ozone, aerosols, and snow albedo as
model input) and the pertaining yearly cloud modification factor (i.e.,
actual yearly sum divided by the corresponding clear-sky value). Figure 7a
shows the yearly sum of daily doses values, for clear-sky and all-sky
conditions (kJ m

A procedure for the examination of the UV data homogeneity is proposed based on RT simulations for clear-sky conditions. It allows the introduction of the yearly calibration coefficient showing the instrument sensitivity loss (1996–2001) and stable behavior in the period of measurements by the Kipp & Zonen UVS-AE-T instrument (2005–2016). For all-sky conditions, the regression model is built using 3-year data (2005–2006–2007), and comparisons of the modeled data with earlier (1996–1997–1998) and later (2009–2010–2011) data show the same model performance as for the model building period that supports the data homogeneity and its usefulness for the long-term trend analysis. The regression model allows the reconstruction of UV doses since 1983, i.e., in the period when the daily total ozone (from the satellite observations) and the sunshine duration data, which represented a proxy for the cloud effects on surface UV, were both available. The reconstruction model is also used to fill the data gaps in the UV observing period (since 1996).

Previous studies showed that the sunshine duration was a worse proxy for a parameterization of the cloud attenuation when compared to the global solar radiation. Using global solar irradiance is more appropriate to parameterize cloud effects on UV (Koepke et al., 2006). However, this variable was measured at Hornsund in some disjointed periods since 1983 but the pyranometer data were not calibrated by a higher-ranking instrument. Sunshine duration measurements by a Campbell–Stokes instrument seem to be less influenced by deterioration of the instrument's sensitivity, and its calibration is very simple, as during cloudless conditions the sunburn track on a recorded cart should appear throughout the whole day.

Analyses of the yearly sums of daily erythemal doses at Hornsund reveal a
nonstatistically significant trend in the period 1983–2016. Two phases of
the long-term behavior of total yearly doses could be identified, i.e., a
positive tendency in total yearly doses in the period 1983–2000 and
afterward a leveling off. The linear trend calculation by a standard
least-squares fit applied to the measured (1996–2016 with the 2002–2004
gap) data shows a statistically significant declining tendency in monthly means
of daily doses (May and June) and in the yearly sum of the erythemal doses.
However, such a declining tendency is forced by 2–3 years of high positive
fractional deviations of the erythemal doses around 2000. Longer time series
(since 1983) do not show any sign of the declining tendency starting around
1996. Bernhard et al. (2011) analysis of the monthly trends at Barrow
(Alaska) for the period 1990–2010 revealed a statistically significant trend
only in October (decline of about

The stratospheric ozone changes appear as a less important driver of the UV
long-term variability in the whole analyzed period. Figure 8 shows the
long-term (1979–2016) pattern of the total ozone mean (using SBUV merged
data) for the period May–August at Hornsund, Barrow, and Resolute, i.e., in
the part of the year with naturally high UV radiation (

Smoothed time series (by LOWES smoother, Cleveland, 1979) of annual fractional deviations of mean total ozone in the May–August periods for the period 1979–2016 at Hornsund (Svalbard, Norway), Barrow (Alaska, United States), and Resolute (Cornwallis Island, Canada).

It seems that excessive UV radiation over the Arctic will be unlikely during the 21st century as a prolonged decrease in ozone will not be possible due to the declining tendency in the concentration of the ozone-depleting chemicals in the stratosphere and anticipated intensification of the Brewer–Dobson circulation loading a higher amount of ozone into the Arctic stratosphere (WMO, 2014). The downward UV tendency in the Arctic will also be induced by the increase in the cloudiness, and lowering of the ground albedo due to the snow and sea-ice melting (e.g., Bais et al., 2015).

A continuation of UV measurements at Hornsund seems to be necessary as it is located in a region vulnerable to climate changes with the local climate strongly dependent on the heat arriving with the Gulf Stream. A projection of the weakening of the Atlantic meridional overturning circulation (Boulton et al., 2014) will lead to the surface cooling at the location. The possibility can not be excluded that high-reflectivity areas (sea-ice and snow) will extend over west Svalbard and the present climatic contrast between the west (warm) and east (cold) part of Svalbard will disappear. Any projection for erythemal irradiance by the end of the 21st century is the most uncertain for this part of the Arctic.

The total ozone overpass data were acquired from the data
archive of SBUV merged ozone at

The authors declare that they have no conflict of interest.

Funding for this study was provided by the Ministry of Science and Higher Education, Republic of Poland, statutory activity no. 384 1/E-41/S/2017 and from the funds of the Leading National Research Centre (KNOW) received by the Centre for Polar Studies for the period 2014–2018. We are grateful to numerous dedicated individuals who have collected meteorological data over many decades.Edited by: Paul Young Reviewed by: three anonymous referees