Introduction
During the last few decades, the danger of overexposure to solar UV radiation
has been well analyzed and a causal link has been established to skin
diseases and cancer since the mutation of DNA can be triggered by UV-B doses
(; ; ;
; among others). On the other hand, the cutaneous
production of vitamin D, a “vitamin” that is proven to be essential for
human health, is also activated by spectral UV radiation. Hence, accurate
knowledge of “safe” UV doses for humans is paramount in order to balance
the harmful and beneficial effects of UV exposure (e.g., ;
; ). Of particular
relevance is the Commission Internationale de l'Éclairage (CIE) action
spectrum as a model for the susceptibility of skin to sunburn (erythema)
. As a result of advances in the fields of photobiology
and ground-based measurements of UV using different types of instrumentation,
a variety of methods now exist to obtain erythemal, vitamin D, and DNA-damage
dose rates (; ;
; ; ).
In parallel, space technology has been making huge steps forward to monitor
the Earth's surface and atmosphere at higher spatial and temporal resolution
and erythemal, vitamin D, and DNA-damage dose rates and doses can now be
retrieved globally from solar backscattered radiation observations from
different satellite sensors. Subsequently, long, reliable, and high-temporal-resolution ground-based estimates of surface photobiological effective dose
quantities are of high importance in order to validate and characterize the
satellite-derived UV products. Ozone layer depletion and recovery in times of
climate change reinforce the need for establishing global long-term and
quality-assured climate data records of the incident solar UV daily doses at
the surface.
In this study, photobiological
UV daily doses retrieved from ground-based measurements using empirical
models and satellite estimates are cross-validated to assess their accuracy
and potential utility.
TEMIS satellite-based UV data products
Operational services
The Tropospheric Emission Monitoring Internet Service (TEMIS) was established
in 2001 at the Royal Netherlands Meteorological Institute (KNMI) as part of a
project from the European Space Agency (ESA), and the service has been
maintained since. The TEMIS UV data product services started in 2003 and are
available through the web portal at http://www.temis.nl/uvradiation/.
The UV products, currently version 1.4, are produced in near-real time on a
latitude × longitude grid of 0.5∘×0.5∘ and
consist of datasets, maps, and time series. The products are calculated
using operational satellite data streams of the global ozone distribution
and, over Europe, the diurnal variation in cloud cover fraction.
The TEMIS UV data products essentially exploit the empirically based
parametrization by of the amount of UV radiation incident
at the surface in Wm-2, as a function of the total ozone column and
the solar zenith angle at a given local solar time, taking into account an
appropriate action spectrum, i.e., the wavelength-dependent response to UV
radiation of health effects or otherwise.
Since the initiation of the TEMIS UV services, maintenance and updates have
been implemented following, for example, changes in the operational assimilated
global ozone distribution , which were based on the
SCIAMACHY instrument aboard ENVISAT up to April 2012,
and later on GOME-2 aboard MetOp-A . Recently, the global
ozone Multi-Sensor Reanalysis version 2 (MSR-2) by has
been used to create a reanalysis of the global clear-sky UV index for a
longer historical period (from November 1978 to December 2012).
Left panel: erythemal UV dose over Europe on 22 June 2016. Thessaloniki,
indicated by a black square, had an almost cloud-free day with an erythemal
UV dose of 5.77 kJm-2 and an erythemal UV index of 10.1. Right panel: action spectra of erythema (red solid line), generalized DNA damage (blue
dashed line), and production of vitamin D (magenta dotted line: draft version as
used within TEMIS ; green dashed line: final version as
adopted by the CIE ).
Cloud attenuation over Europe is prescribed using the near-real time cloud
mask product provided by the EUMETSAT Nowcasting
Satellite Application Facility (NWC-SAF), which has been received, processed, and
archived at KNMI since July 2005. The operational cloud cover dataset has
been based on the different SEVIRI instruments that have been operational
aboard the Meteosat Second Generation (MSG) satellites from January 2004
onwards using the Meteosat 8, 9, and 10 platforms. The effect of
grid cell average surface elevation, though not the actual 3-D topography, on
surface UV is taken into account in the calculations. Changes in surface
albedo are prescribed using a monthly climatology of surface reflectivity
. The effects of aerosols are included implicitly in the
parameterization but do not vary over time .
Products and algorithms
TEMIS provides two types of surface UV products: (i) the clear-sky erythemal
UV index and (ii) the daily UV dose (daily integral) related to different
health effects. The erythemal UV index (UVI) is determined using the action
spectrum adopted by the International Commission on Illumination (CIE) for
erythema or reddening of the skin due to sunburn. In the current v1.4 TEMIS
service, the UVI is based on the CIE action spectrum described by
. describe an improved version of that
action spectrum adopted by CIE in 1998. The effect of this improvement on the
UVI values is small, well below 1 % except for high solar zenith angle
situations . The improved CIE erythemal action spectrum will
be included in the forthcoming upgrade (v2.0) of the TEMIS service.
Following international agreements, the UVI represents the amount of UV
radiation at local solar noon, i.e., when the Sun is highest in the sky, under
clear-sky conditions. The UVI is usually given as a dimensionless index,
where 1 unit equals 25 mWm-2. Using the operational
meteorological data streams (temperature, pressure, winds) which are included
in the ozone data assimilation , the UVI is available in
forecast mode and TEMIS provides forecasts of both the global ozone field and
UVI for today and the coming 8 days.
The daily UV dose (UVD) is the total amount of UV radiation, usually given in
kJ m-2, integrated between sunrise and sunset, accounting for the
variation in the solar zenith angle (SZA) and cloud cover fraction (in TEMIS
version 1.4); this is available over Europe only) during the day (see
Fig. , left). The UV dose is calculated for three action spectra
(see Fig. , right): the erythemal UV dose (UVD-CIE) based on the
CIE erythemal action spectrum , identical to the one used
for the UVI-CIE, the generalized DNA-damage UV dose (UVD-DNA) based on the
action spectrum determined by and normalized at 300 nm
based on , and the vitamin D UV dose (UVD-VitD) based on the
action spectrum for the production of previtamin D3 in the human skin
.
Note that the 2005 (draft) version by used for UVD-VitD
within TEMIS differs from the CIE-adopted vitamin D action spectrum
(see Fig. , right). The difference, which
includes a wavelength shift of 3 nm (the applied action spectrum peaks at
295 nm and not at 298 nm as proposed by CIE), would increase the TEMIS data
by a factor of about 2.2 (2.1 in summer, 2.3 in winter) when using the CIE
vitamin D action spectrum – an important change that will be implemented in
a forthcoming update (v2.0) of the TEMIS UV operational data streams.
For each
action spectrum, a parametrization is applied following
for the UV solar irradiance as a function of SZA(t) and total ozone column
providing a first guess of the UV irradiance weighted with a specific action
spectrum (UVI′), at time t using the global assimilated ozone
field at local solar noon (t=12h).
The final UVI(t), which can be seen as the UV index at time t (i.e., with a
time-dependent SZA), is then calculated from UVI′(t) by applying a set of
correction factors:
UVI(t)=UVI′(t)⋅fD⋅fC⋅fH⋅fA(Wm-2),
where fD is the correction for the day-to-day variation in the
Sun–Earth distance, fC the correction for the attenuation due to
clouds (in the case of clear-sky conditions: fC=1), fH
the correction for the surface elevation, and fA the correction
for the ground albedo.
The UVI index at local solar noon, UVI (t=12h), follows
directly from Eq. (1), i.e., using SZA (t=12h), after division by
25 (mW m-2). The TEMIS products' uncertainty can currently only be
estimated from the errors reported in the ozone total amount; thus, it
reflects the lower boundary of the errors seen in the UV doses. Based on this
fact, TEMIS products include an uncertainty of 2–3 % in the daily doses.
The UVD products, in kJm-2, are determined from a 10 min step
integration of UVI(t), with a time-dependent SZA, over time t
between sunrise and sunset, which are assumed to lie symmetrically around local
solar noon.
For the calculation of fC the NWC-SAF cloud mask
is converted to a cloud fraction (Cf) by
counting the clear vs. cloudy instances per UV grid cell of
0.5∘×0.5∘ (latitude × longitude).
The cloud correction factor in Eq. () is then given by
fC=1.0,Cf<0.020.9651-0.2555⋅Cf,0.02≤Cf≤0.980.5,Cf>0.98,
a relationship that has been determined from the effect of clouds on surface
UV at the location of KNMI at De Bilt in the Netherlands
(; ). For the calculation of
fH a 5 % increase of the incident UV irradiance per km
surface elevation above sea level is assumed:
fH=1+0.05⋅H,
where the surface elevation H (in km) is determined from the GTOPO30
database (https://lta.cr.usgs.gov/GTOPO30/), resampled to the
0.5∘×0.5∘ UV grid. For the calculation of
fA the following function of ground albedo (Ag) is
applied, taking into account multiple reflections between the surface and the
overlying atmosphere:
fA=1-0.25⋅0.091-0.25⋅Ag.
The function derives from the series 1+xy+(xy)2+…=1/(1-xy), where
x=0.25 is the UV albedo of the overlying atmosphere for upward-reflected
UV radiation and y=Ag. Since the UV index
parameterization is empirically based on UV data collected at De Bilt and
Paramaribo, the Ag at these (urban) sites – with a 12-month
average value of 0.09 – is used as a normalization factor for the
calculation of fA. The data for Ag at each UV grid
cell are taken at 335 nm from the monthly TOMS/GOME climatology,
which uses the spectral dependency of the GOME database
but with a scaling to match the TOMS 340/380 nm database (; ).
Note that there is no explicit correction in Eq. () for the
variable presence of aerosols in the TEMIS UV data products. However, the
empirically based parametrization includes an implicit
aerosol correction due to the average aerosol load over these two urban
sites: an aerosol optical depth (AOD) at 368 nm of 0.3 and an aerosol single-scattering
albedo (SSA) of 0.9 . For situations where the real
aerosol load is lower (higher) than the assumed load, the UV data products
will underestimate (overestimate) the UV index and UV dose. With potential
future near-real time availability of aerosol optical parameters at a global
scale, the correction factors derived by could be applied
within future updates of the TEMIS UV services.
Ground-based data products
Instruments at Thessaloniki
The calculation of the photobiological doses over Thessaloniki
(40.63∘ E, 22.96∘ N) are based on measurements taken by
three different types of instruments in continuous operation at the
Laboratory of Atmospheric Physics of the Aristotle University of Thessaloniki
(LAP/AUTh: http://lap.physics.auth.gr).
Firstly, a Brewer MKIII spectrophotometer with serial number #086 (B086) is
equipped with a double monochromator and measures the UV solar irradiance
spectrum (286.5–363 nm) with a wavelength step of 0.5 nm.
Every scan lasts 7 min, while the use of a triangular-like slit results in a
bandwidth of 0.55 nm full width at half maximum (FWHM). The spectra
used in this study have recently been subjected to quality control and
re-evaluation , after which the remaining 1σ
uncertainty is estimated to be 5 % for wavelengths
longer than 305 nm and for SZAs smaller than
80∘. For lower wavelengths and larger SZA the uncertainty is larger
as a consequence of the photon noise that dominates due to the low recorded
signal . The simpler, single-monochromator Brewer
instrument with serial number #005 (B005) has been operational in Thessaloniki since
1982 and has been providing continuous, well-calibrated, and documented total
ozone column measurements (; ;
).
Secondly, a Norsk Institutt for Luftforskning
(NILU)-UV multi-filter radiometer has been fully operational in Thessaloniki
since 2005 and forms part of the UVNET network of NILU-UV radiometers
(http://www.uvnet.gr, ). The NILU-UV with serial number
04103 (NILU103) provides 1 min measurements in five UV channels with
nominal central wavelength at 302, 312, 320, 340, and 380 nm and a FWHM of 10 nm, while its sixth channel measures the photosynthetically active
radiation (PAR) and is used here to determine cloud-free cases based on the
cloud detection algorithm proposed by . Although the B086
measures the UV spectrum with high spectral resolution, the time frequency of
the scans usually varies from 20 to 40 min. Nevertheless, Brewer
spectrophotometers are a very powerful means for calibrating other UV
measuring instruments that provide higher temporal resolution measurements.
Specifically, for the calibration of NILU103 raw data, cloud-free response-weighted irradiances were derived from B086's measured spectra. Since B086
scans the UV solar spectrum within approximately 7 min, the time period
needed to scan the spectral range of each NILU103's channel spectral
response is approximately 3 min. The coincidences of NILU103's raw data to
B086's weighted spectra were performed based on the time that B086 measured
the wavelength at which each channel peaks. Subsequently, the time difference
that can be introduced between the two datasets is normally less than
±1 min. To account for this time window, the mean values of three
consecutive NILU103 measurements were analyzed, with the central one chosen
to be the closest to B086's time scan of the peak wavelength of each channel.
Then, NILU103's data were corrected for possible drifts in time via a time-dependent smoothing spline fit. Furthermore, the drifts of the channels were
monitored through monthly lamp measurements. Both methods resulted in the
same patterns for the drifted channels. After correcting for time drifts, a
time-independent absolute calibration factor was derived through scatter
plots based on linear regression through origin. To evaluate the validity of
the calibration procedures, the NILU103 calibrated data were compared once
again with B086 response-weighted irradiances and the time series were checked
for time drifts and SZA dependence. By calibrating the NILU103 measurements
with the B086 coincident response-weighted irradiances, we estimate that the
uncertainties of the NILU103 measurements used in this study are 5.6%
.
Thirdly, a Yankee Environmental Systems (YES) UVB-1 radiometer operating also
in Thessaloniki, provides 1 min erythemal dose measurements with
a spectral response very similar to the erythemal action spectrum
. Using model simulations with the libRadtran radiative
transfer proper weighting factors are calculated with
respect to SZA and the total ozone column (TOC). These factors are used to
transform the UVB-1 measurements into erythemal irradiance .
A similar transformation is applied for the vitamin D and DNA-damage weighted
irradiances (see Sect. ). In addition, the Brewer measurements
have been used to correct the UVB-1 observations for the degradation of its
absolute spectral response and for sudden changes in the behavior of the
instrument. Thus, the datasets from the UVB-1 and the NILU-UV radiometers are
not completely independent since the Brewer instrument was used for the
calibration of both instruments.
Model selection. Top: boxplots of the z scores of the input
variables and the erythemal UV dose (CIE) with mean values denoted by μ.
Bottom: the pairwise linear Pearson correlation coefficient for each
combination of the input variables and/or the output variable. The results
are unnoticeably different in the case of vitamin D and DNA-damage doses. To
save space, we have used abbreviated labeling of the sinusoidal terms so
that, for example,
sin(DOY) refers explicitly to sin(DOY×2π/T).
In addition, at Thessaloniki, a CE318-N sun–sky photometer, also known as a CIMEL, provides continuously atmospheric
observations through the NASA Aerosol Robotic Network (AERONET)
. The CIMEL provides aerosol optical depth at the UV
wavelength of 340 nm, amongst other aerosol properties, which is used
to investigate the effects of aerosol variability at Thessaloniki on
comparisons with the satellite-derived UV products.
Products and algorithms
Effective UV doses from the Brewer spectrophotometer
The B086 spectra were processed by the SHICrivm algorithm and extended to
400 nm . The extended spectra were validated with a
collocated EKO UV-A instrument and weighted with the
action spectra for (i) the erythemal dose , (ii) the
formation of vitamin D in the human skin , and (iii)
DNA damage . The corresponding effective doses have been
calculated by integrating the weighted spectra over the nominal wavelength
range, while the time of measured doses was matched to the time that B086
scanned the wavelength where the highest sensitivity of each action spectrum
is found. Since DNA-damage action spectrum peaks at the lower measured
wavelengths, the correspondent time was chosen to be the starting point of
the scan. It appears that in most cases the three doses have time differences
less than 1 min. The 1σ uncertainty of the derived effective doses
for the erythema and the vitamin D is estimated to be 5% since the
contribution of photons with wavelengths shorter than 305 nm (where
the signal may be very low) is small. However, the uncertainty in the
calculated effective dose for the DNA damage is larger at SZA greater than
60∘ because of the important contribution of shorter wavelengths
(very low signal levels) and may reach 20% for SZAs near 80∘
in overcast conditions. The B086 provides measurements with a time frequency
of 20 to 40 min, but atmospheric circumstances can change considerably
within this period. It is therefore better to base the evaluation of the
TEMIS UV dose rate (available at 10 min intervals) on the NILU103 data,
which have a better temporal resolution; thus, they suffer much less from
changes in atmospheric conditions (like clouds) during one measurement than
the Brewer measurements.
Left: the robustness analysis on the grid of 100 NN models using
the minimum validation mean squared
error (MSE) as the criterion for selection of the optimal NN
architecture (which was found to have 13 hidden neurons and a
training : validation data split of 90% : 10%). Right: the
progress of the NN training of the optimal architecture with back-propagation
iteration out to 100 iterations “epochs”, where MSE <1.0e-4.
Effective UV doses from NILU-UV irradiances using a neural network model
A feed-forward function-approximating NN model was coded
using MATLAB's object-oriented scripting language in conjunction with its
Neural Network Toolbox . As inputs, the NN has time series
vectors of NILU103 irradiance measurements at 302, 312, 320, 340, and
380 nm together with temporal variables: the SZA, the day of the week
(DOW), the day of the year (DOY), and its sinusoidal components
sin(DOY×2π/T) and cos(DOY×2π/T), where
T is the number of days in the year. As outputs, the NN calculates time
series for the biological UV products resulting from B086 response-weighted
spectra: i.e., erythemal CIE, vitamin D, and DNA-damage effective doses. The
rationale behind including temporal variables in the inputs is that
geophysical variables very often exhibit periodicity associated with an
annual or diurnal cycle and are now commonly incorporated into atmospheric
chemistry models . From the NILU103 data, a matrix of
n=47 908 co-located input–output vectors was extracted to train and
validate the model. All output variables were found to correlate strongly and
positively on all five of the irradiances (0.922≤r≤0.995), strongly
anti-correlate with SZA (-0.891≤r≤-0.909), and weakly
anti-correlate with the temporal variables. Figure shows the
z scores of the input variables and the erythemal UV dose (“CIE”) together
with the pairwise linear Pearson correlation coefficient.
The input and output vectors used in our study were connected via two network
layers, the first containing hidden neurons with hyperbolic tangent (tanh) activation functions and the second containing linear activation
functions. The mathematical details of this input–output structure is
described in Appendix A. Key to the success of the modeling approach is
signal-to-noise separation. The NN model is constructed using denoised time
series of the NILU-UV irradiances and denoised time series of the
photobiological products. Once constructed, the original (noisy) data are
input to the model to calculate the photobiological outputs. In order to
achieve this, we applied singular spectrum analysis to separate the signal
(total trend plus periodicity) from the total noise component for each of the
irradiance and photobiological product time series (see , for
a review of the singular spectrum analysis methodology). In this work we
calculated the unbiased estimator for the lag-covariance matrix using the
method of . The window length was rounded to
[log(n)]1.5=36 following the recommendation of and the
minimum distance length criterion they introduce was applied. This was found
to give a consistent separation of the signal from noise for the NILU103
irradiance measurements at 302, 312, 320, 340, and 380 nm at
eigenvalue ranks 9, 7, 7, 5, and 5, respectively, and in the case of the
photobiological products at eigenvalue ranks 7, 8, and 8, respectively, for CIE,
vitamin D, and DNA. This denoised data structure enables the NN model to
determine the underlying relation between the input and output parameters
most efficiently.
The optimal NN architecture was then found by minimizing the mean squared
error (MSE) between the NN estimates and Brewer reference output data for
each NN in a grid of 100 NN architectures where the number of hidden neurons
was varied from 5 to 15 and the proportion of training data (t/n) was
varied from 50 to 95% in steps of 5%. The subset of t vectors
was chosen randomly with a random number generator applied to the vector of
indices [1:n] and the remainder being used as a validation set that
contained (n-t) vectors. During each of 100 iterations of the learning
process, the weights and biases of each NN are tuned with the
back-propagation optimization algorithm to minimize the
MSE cost function over the set of input–output vectors. We have used the
Bayesian regularization scheme based on a Laplace prior .
As a result of this initial robustness analysis, the optimal NN was found to
require 13 hidden neurons and a training to validation ratio of
90%:10% as seen in Fig. , which also shows the
result of applying the model selection procedure as well as the progression
of training of the optimal NN architecture towards convergence at the
horizontal asymptote for the “best” validation MSE after 100 epochs of
back-propagation learning using Bayesian regularization. Note that for the
rather long time series used here, there is almost no visual dependence on
the training fraction above 50 % with a gradient in the optimization
surface only being apparent in the direction of increasing number of neurons.
It is important to note that the optimal NN is valid for the range of
parameters determined by the training data shown in Table .
Temporal variables other than SZA are not listed and have the following
expected ranges: DOY=[0,366], sin(DOY×2π/T)=[-1,1],
cos(DOY×2π/T)=[-1,1], and DOW =[1,7].
Range of validity of the trained optimal NN as determined by its
input parameters (upper list) and output parameters (lower list).
Parameter
Min
Max
Mean
SD
Ir(302)
0
0.017
0.003
0.004
Ir(312)
0
0.229
0.064
0.055
Ir(320)
0
0.333
0.108
0.079
Ir(340)
0
0.678
0.252
0.159
Ir(380)
0
0.871
0.327
0.208
SZA
15.63
81.162
54.373
16.120
CIE
0
0.234
0.056
0.054
Vitamin D
0
0.460
0.103
0.107
DNA
0
0.011
0.002
0.002
For validation, this optimally trained NN was then fed with the remaining
“unseen”) input vectors from the 10% of the training data and its
estimates are compared against the target measurements of the output vector
to evaluate the network performance. The correlation of NILU103 NN estimates
with target outputs was high (r=0.988 to 0.990) and found to have a very
low bias (0.000 to 0.011 absolute units) as shown in Fig. .
NN validation. Upper panels: regression of the NILU103 NN estimates
on the coincident Brewer-derived erythemal UV dose (CIE) (left), vitamin D
(center), and DNA-damage dose (right). Lower panels: histograms of the
difference between NN estimates and the Brewer-derived quantities. The mean
(μ) and standard deviation (σ) are indicated.
Neural-network-based estimates of retrieval uncertainty is still embryonic
(see, for example, ) due to the difficulty associated
with propagating errors through a nonlinear function. In order to provide a
rough estimate, we calculated the median absolute percentage error (MAPE) for
the difference between the target values and the NN outputs and obtained the
following estimates of the NN uncertainty: Δ(CIE)=3.6,
Δ(vitaminD)=4.5, and Δ(DNA)=5.1%. The
uncertainties seen in the NILU NN products are well aligned with the
uncertainties introduced by the NILU and B086 irradiances, 5.6 and 5 %,
respectively. An estimation of the uncertainty in the NILU NN
products based on error propagation results in absolute errors of less than
7.5 % for all three products.
Effective UV doses from the UVB-1 radiometer
As described in Sect. , the measured doses by the YES
UVB-1 radiometer are converted to erythemal doses by applying proper
correction factors which depend on the values of SZA and TOC for each
measurement. The parametrization suggested by is then
applied to convert the erythemal dose to vitamin D effective dose. Based on
the measurements of B086 we found that when the UV index is lower than 2, the
vitamin D is overestimated significantly and should be divided by the
following correction factor (cf) obtained empirically by a least squares fit to the
data:
cf=-0.086⋅UVI3+0.379⋅UVI2-0.575⋅UVI+1.317.
In a similar way, the DNA-damage effective doses were estimated from a more
complex empirical relationship that was developed using data from B086 for
the period 1993–2010 and evaluated using data for the period 2011–2014. The
relationship for the DNA-damage effective doses consists of TOC, CIE, the
cosine of the SZA (cosθ) and the ratio between the CIE and the
climatological value of CIE on each day and SZA (CIEclim):
DNA=g(TOC,CIE,cosθ,CIEclim)=f(CIE,TOC)/(cf1(cosθ)⋅cf2(r)),
where
r=CIE/CIEclim,f(CIE,TOC)=a1+a2⋅CIE+a3⋅TOC+a4⋅CIE2+a5⋅CIE⋅TOC+a6⋅CIE3+a7⋅TOC⋅CIE2a8⋅CIE2+a9⋅CIE+a10,cf1(cosθ)=b1⋅eb2⋅cosθ+b3⋅eb4⋅cosθcf2(r)=1,r>2c1⋅r2+c2⋅r+c3,r≤2.
The values of the constant terms in Eqs. (–)
are a1=-2.703×10-5, a2=0.01245, a3=1.428×10-8, a4=0.1151, a5=-1.736×10-5, a6=-0.1505,
a7=-9.527×10-5, a8=-3.523, a9=0.9388, a10=0.9611, b1=1.022, b2=-3.994, b3=0.7306, b4=0.2755, c1=-0.3026, c2=0.8971, and c3=0.401. The empirical rule given by
Eq. () was found to be valid for UVIs greater than 0.5. The
daily mean TOC from the single-monochromator B005 was used in the empirical
equations and in cases of missing data, daily climatological means derived
from the 30-year record of B005 were used. Using the effective doses from the
double monochromator B086, we estimated that the 1σ uncertainty in the
determination of vitamin D is smaller than 3% for UVI values greater
than 2 and exceeds 10 % for UVIs lower than 1. The 1σ uncertainty
in the calculation of the effective dose for the DNA damage is smaller than
7% for the range of used UVIs (i.e., greater than 0.5). The mean ratio
between semi-simultaneous measurements of the clear sky erythemal irradiance
from the B086 and the pyranometer (±1 min differences between the mean
time of the spectral scan and the UVB-1 measurements) for SZAs below
80∘ for the period 2004–2014 is 1.00±0.04, indicating that the
uncertainty in the erythemal irradiance from the pyranometer is similar to
that of the Brewer B086.
Comparison of the NILU-UV and UVB-1 data products
Following the appropriate methodologies already discussed in
Sect. and , erythemal, vitamin D, and
DNA-damage daily doses can be obtained from the NILU103 and an erythemal-like
measuring instrument, in this case a UVB-1 radiometer. Even though the UVB-1
data were corrected for the degradation of its absolute response with B086
data, the validity of its measurements as absolute values can be used to
evaluate the performance of the NN used to derive all of the biological dose
products based on NILU-UV measurements.
In order to have comparative results for the satellite data evaluations,
daily doses of all three quantities under investigation were calculated and
their agreement was evaluated. For these evaluations both the UVB-1 and the
NILU103 1 min data were matched in order to avoid discrepancies due to
random time gaps in the original time series. Then, the daily integrals were
calculated for both NILU103 and UVB-1 datasets, without any other constrains
on the data. The UVB-1 erythemal daily doses are underestimated on average by
∼2% when compared to NILU103 retrievals, with a standard deviation
of 5.4%. When limiting the data to those during which more than
70 % of the original measurements were classified as cloud-free, the
average agreement is close to perfect (average difference of 0.5%)
with a corresponding standard deviation of 4.2%. This cloud
classification criterion, according to which days with more than 70 %
abundance of cloud-free measurements are characterized as cloud-free, is used
throughout the study, unless stated otherwise. As seen in the lower panel of
Fig. a, during the winter months UVB-1 tends to underestimate the
erythemal daily doses, while during the summer months the opposite behavior
is observed.
Daily integrals relative percentage differences of erythemal
(a), vitamin D (b), and DNA-damage (c) doses
estimates from the UVB-1 and NILU103 radiometers (upper panel) and the same
datasets averaged on a monthly basis along with the 1σ error bars
(lower panel).
The daily integrated data for vitamin D retrievals show that there is a good
agreement between the UVB-1 and NILU103 sets. In both subsets, i.e., for all
and clear skies, respectively, the standard deviations of the differences
between the two datasets are 7.4 and 5%, respectively, while the
differences between the datasets is of the order of 4% for all skies
and approaching zero (0.2%) for the cloud-free days only. But, as
observed in Fig. b, the number of cloud-free days is limited to
only 25% of the originally available amount of days. Again, there is a
seasonal pattern for vitamin D which is similar to the seasonal pattern
observed for the daily erythemal doses.
Time series of the relative percentage differences between the
SCIAMACHY/GOME2A and NILU-UV effective daily doses under all skies (upper
panel) and the seasonality of the differences based on the average month
along with the 1σ error bars (lower panel).
Concerning the DNA-damage daily doses
(Fig. c), the comparisons show that in general UVB-1
underestimates the daily dose on average by ∼ 5%, with a
standard deviation of about 18%. For the cloud-free days, UVB-1 show
an underestimation of ∼ 2% with a standard deviation of about
∼ 16%. The seasonal pattern observed at the lower level of
Fig. c is similar to the one depicted for the aforementioned
daily doses but enhanced to ±20%, especially for the winter months
where the UVB-1 significantly underestimates the doses derived from NILU103,
probably due to the fact that the DNA action spectrum peaks at shorter
wavelengths.
In Table an analytical overview of the NILU103 and UVB-1
comparison statistics is presented. All three quantities present high R2
values (0.99 to 1.00), while the Pearson coefficients (R) reveal a strong
linear correlation between the two ground-based datasets with values equal to
almost unity. The DNA data are subjected to higher sensitivity in lower
wavelengths and exhibit the largest differences between NILU103 and UVB-1.
Statistical analysis of the daily integral comparisons between NILU103 and UVB-1 retrievals.
Daily integrals
Erythemal (%)
Vitamin D (%)
DNA damage (%)
All skies
NILU clear
All skies
NILU clear
All skies
NILU clear
N counts
3013
731
3013
731
3013
731
R
0.998
0.996
0.998
0.996
0.997
0.997
R2
1.00
0.99
1.00
0.99
0.99
0.99
Mean (%)
-1.9
0.9
-3.6
0.2
-4.8
-2.3
SD (%)
5.4
4.2
7.4
5.0
18.3
16.4
Generally, the agreement between the two instruments is quite remarkable
given the different nature of the original measurements using different
spectral resolution and different angular responses, which could be major
parameters affecting the comparisons, especially for the seasonal and SZA
dependence, while the different retrieval methodologies could lead to further
discrepancies.
Evaluation of TEMIS satellite-based UV products with NILU-UV data products
The satellite-based TEMIS UV products are evaluated for the grid cell
containing Thessaloniki (grid cell center: longitude = 22.75∘,
latitude = 40.75∘). This evaluation uses a specifically
reprocessed dataset (version 1.4) to provide TEMIS UV dose rate values,
calculated at the 10 min steps of the time integration of the daily dose UV
products which are standard provided to the TEMIS data users. Time series
analysis and correlation statistics are performed on the daily UV dose for
erythema, vitamin D, and DNA damage over a 6-year period (2009–2014).
As seen in Fig. , for all skies the TEMIS UV doses agree
within 13 % on average and achieve rather high correlations of 0.92,
0.93, and 0.93 for erythema, vitamin D, and DNA damage, respectively
(Fig. ). The standard deviation of the differences for the
three datasets under all skies is 47.3, 45.7 and 47.1 % for erythema,
vitamin D, and DNA damage, respectively. The large variations between the
satellite-based and ground-based UV daily dose data records can be attributed
to different factors. For the full uncertainty, budget contributions relate, for example, to the uncertainty in the B086 originally used spectra, the uncertainty
caused by the application of the NILU-UV NN retrieval algorithm, the aerosol
climatology assumed in the satellite-based algorithm and total ozone column
retrieval errors. However, as will be demonstrated below, the greatest part
of the observed spread in the ground-based and satellite-based differences in
UV dose is related to the representation of clouds in the satellite algorithm
and selection of cloud-free days for the ground-based datasets.
The NILU103 and TEMIS datasets have high coefficients of determination and
low biases (small y intercepts) as seen in Fig. , while
the slopes are close to unity. Although most points seem to cluster evenly
around the y=x line especially for the higher values, some overestimation
of the satellite products at the lower values results in slopes that are
slightly less than unity.
Scatter plot of daily UV dose values provided by the joint
SCIA/GOME2A UV products (y axis) and NILU103 (x axis) in
kJm-2 under all-sky conditions.
One important aspect for the evaluation is the determination of cloud-free
days. The optical geometry of the two monitoring systems is different and the
point measurements of the NILU at Thessaloniki compared to the
0.5∘×0.5∘ spatial analysis of the satellite-based
product may be an important source of discrepancies. Since the
satellite-based estimates are based on only one total ozone column value
throughout the day, we expect that this could further increase the
uncertainty in the satellite-derived daily doses estimates.
Obviously, rapidly changing
cloudiness conditions can also lead to large discrepancies between the ground
and satellite retrievals. Currently the TEMIS satellite doses over Europe are
obtained using the cloud cover fraction per 0.5∘×0.5∘
grid cell as derived from SEVIRI/Meteosat cloud information (see
Sect. ). This information is incorporated in the TEMIS
retrieval algorithm on a half-hourly basis, but the frequency of this
information might need to be even higher when dealing with high-frequency
changing cloudiness conditions as shown in Fig. for two
specific cloudy days at Thessaloniki.
The evolution of the 10 min erythemal dose over the day as provided
by satellite (blue circles) and at the ground (red triangles) for two
days in 2009 showing a large temporal variability in cloudiness. The
satellite-derived UV daily dose is lower than the NILU103-derived UV dose by
23 % for the case on 30 May 2009 (left panel), while it is larger by
120 % for the case on 18 June 2009 (right panel).
The time evolution illustrated for the two days in Fig. show
that satellite cloud information cannot capture the rapid changes of
cloudiness on these days: the satellite retrievals may either overestimate or
underestimate the impact of clouds. Therefore, in order to evaluate the
performance of the satellite-based products, the cloudiness effects should be
further analyzed. For this, four different cases are examined in more detail:
all-sky cases (whose statistical analysis is given in
Fig. ), days with more than 10 % of the measurements
characterized as cloud-free (excluding overcast days), days with more than
70 % of the measurements characterized as cloud-free (relatively
cloudless days), and days with more than 90 % of the measurements
characterized as cloud-free (cloudless days). At this point it should be
mentioned that for the characterization of the cloud-free 1 min data,
the cloud screening detector proposed by was applied on
the NILU103 PAR measurements.
An overview of the impact in limiting the percentage of cloud-free cases per
day (Ncl) is provided in Fig. for the erythemal UV doses. The
relative percentage differences clearly improve considerably when excluding
the overcast days (Ncl > 10%). The original 12.5% average
overestimation of the satellite erythemal daily doses is reversed to
1.8% underestimation, while the standard deviation is less than
15%. When posing the 70% limitation, as applied on the
(UVB-1)-NILU comparisons in Sect. , the underestimation of the
satellite erythemal doses seems to be even less while the standard deviation
is similar. However, this limitation is affecting significantly the available
number of days fulfilling this restriction through a reduction of number of
days by 75%. However, when studying the cloudless days
(Ncl > 90%), the satellite product is overestimated on average by only
∼0.6% with a corresponding standard deviation of 11.5%. For
these cloud-free cases, the interpretation of aerosol effects into the
satellite algorithm could be an additional parameter affecting these
comparisons (see below).
Time series of the relative differences between the satellite-based
and ground-based retrieval of the UV erythemal doses; a classification
of the cloudless measurements per day is also shown along with the corresponding
statistics (upper panel). The seasonality of the data is also presented as
monthly mean values (lower panel).
A comprehensive statistical analysis of all three UV daily doses under
investigation for all cloudiness conditions is provided in
Table . All UV doses, erythemal, vitamin D, and DNA damage,
respectively, present high R2 values (≥0.9) for all the cloudiness
restrictions, revealing a high interconnection between the two datasets,
while the correlation coefficients denote that under all circumstances, the
UV effective doses present a high linear relationship. Although the
satellite-based retrievals overestimate for all-sky cases on average by
12.5, 13.0 and 12.4% for erythemal, vitamin D, and DNA damage
respectively (Fig. ), the percentages are much smaller when
considering only cloud-free days (in general less than 1.2%). Under
mixed cloudiness conditions (Ncl > 70 and > 10%)
satellite-based retrievals tend to underestimate the daily doses on average.
As seen in Table , the imposed cloudiness limitations do not
alter the standard deviations much.
Statistical analysis of the relative percentage differences
[(satellite - ground)/ground %] between the satellite and ground estimates
based on the cloudless instances within a day; the all-sky values are given
in Figs. and .
Erythemal doses
Vitamin D doses
DNA-damage doses
Cloudless instances per day (%)
> 90%
> 70%
> 10%
> 90%
> 70%
> 10%
> 90%
> 70%
> 10%
N counts
203
390
991
203
390
991
203
390
991
R
0.96
0.95
0.95
0.96
0.95
0.95
0.96
0.95
0.95
R2
0.92
0.9
0.9
0.92
0.91
0.91
0.92
0.91
0.91
Mean (%)
0.6
-1.0
-1.7
1.2
-0.4
-1.4
1.2
-0.3
-1.5
SD (%)
11.5
13.2
14.2
12.9
14.5
15.2
12.2
13.9
15.8
Table shows that even under cloud-free days there is a
scatter of almost ±13% between the two datasets for all three UV
doses. The seasonality seen in Fig. is also present when
limiting the datasets to cloud-free days, as seen in the lower panel of
Fig. , implying that apart from the cloud effects, there are
other factors affecting the agreement between the ground- and satellite-based
UV data products. One of the causes could be variability in aerosol load over
Thessaloniki which is neglected in the satellite-based retrievals.
At Thessaloniki, AOD values at 340 nm are provided by a CIMEL sun
photometer for the period 2011–2014.
In order to investigate the influence of aerosols on the satellite
retrievals, estimations of all three UV effective doses every 10 min were
obtained both from the satellite and NILU103 retrieval algorithms. These
datasets were limited to periods where the ground-based cloud screening
algorithm resulted in cloud-free cases. As seen in the upper level of
Fig. there is a strong dependence between the 10 min
doses for aerosol optical depth up to 0.4, while the differences show a slow
ascending slope for aerosol loads of more than 0.4. To further attest to
this aspect, linear fits were conducted for two datasets, one that comprised
data with AOD ⩽ 0.4 and the second with data with corresponding
AOD > 0.4. It was found that for all three UV effective doses, the slopes for
the first imposed limitation on AOD were higher than those calculated for the
second dataset. Specifically, the slopes for the two AOD limitations were
found to be 44.5 and 11.7 % for the CIE, 50.6 and 8.5 % for the DNA
damage, and 46.1 and 8.3 % for the vitamin D doses, respectively. This general
pattern is in compliance with the implicit climatological AOD and SSA values
applied in the satellite-based retrievals, where the AOD at 368 nm is
assumed to be 0.3 and SSA is set to 0.9 (Sect. ).
used a monthly aerosol climatology for the AOD and SSA at
315 nm in order to correct the OMI UV irradiances for absorbing aerosols.
SSA measurements in UV are not available for the under investigation period
in Thessaloniki; thus, a similar study, taking into account the parameter τabs=AOD×(1-SSA), cannot be performed.
To further investigate the AOD impact on the comparisons, the monthly means
were calculated for both AOD and relative differences. The pattern seen in
the monthly means of the AOD values is in general agreement with the
seasonality seen in the average monthly values of the relative percentage
differences between the satellite- and ground-based 10 min cloudless doses
(Fig. , lower panel), implying that there is a link between the
two observed seasonalities.
Relative differences of satellite-based and ground-based UV 10 min
doses as a function of AOD at 340 nm for cloudless cases at
Thessaloniki in the period 2011–2014. The statistics are provided in the
form of mean and standard deviation of the differences within 11 bins of AOD
values. The least squares linear fits for the three doses are also provided
(upper panel). Monthly mean values of AOD at 340 nm along with the mean
monthly values of the relative differences presented in the upper panel under
cloud-free cases are also provided (lower panel).
Model estimations performed with the model uvspec of the libRadtran library
(v. 1.7) reveal that, for typical aerosol optical properties for the site of
Thessaloniki, differences of 0.2 between the AOD values used in the
ground-based retrieval algorithm and the measured AOD may be responsible for
differences of the order of 10 % between the measured and retrieved
erythemal dose rates. Furthermore, other aerosol properties, like the single-scattering albedo, may vary significantly over urban sites such as
Thessaloniki , which can introduce extra uncertainties in the
effect of aerosols on the estimated UV irradiances which are of the same
order of magnitude as the uncertainty due to the variability in the AOD
(e.g.,
; ).
Discussion and conclusions
In this work a cross-validation between ground-based measurements and
evaluation of TEMIS satellite-based estimates has been performed for three
important photobiological UV daily dose products: erythemal UV, vitamin D, and
DNA damage. The datasets to compare have been produced and compiled such to
allow a thorough discussion of their respective accuracies and limitations at
the mid-latitude UV and ozone monitoring station in the Laboratory of
Atmospheric Physics of the Aristotle University of Thessaloniki, Greece. A
neural network (NN) algorithm has been trained on NILU-UV multi-filter
radiometer irradiances at five different wavelengths together with weighted
action spectra from a Brewer MKIII spectrophotometer to produce 1 min time
series of erythemal UV, vitamin D, and DNA-damage dose rates. Further, the NN
estimated erythemal UV dose rates were compared with UVB-1 calibrated UV
measurements, and we show how appropriate methodologies can be applied to the
original UVB-1 dataset to also produce vitamin D and DNA-damage dose rates
at the same temporal resolution as the NILU-UV instrument. In this way we
could perform a ground-based verification and evaluation of the developed NN
algorithm for the NILU103 measurements. The cross-validation between the
NILU103 and the UVB-1 dataset revealed a very good agreement. In particular,
the following is found:
The temporally aligned NILU-UV NN and UVB-1 ground-based datasets (30 503
coincident “all-sky” dose rate data records) did not show differences of
more than 2 % in their daily integrals and these also had a moderately
low standard deviation of 5.4 %.
For vitamin D, the agreement was within 3.6 % for all-sky data with a standard deviation of about 7.4 %, largely associated with a SZA dependence at large zenith angles.
For cloud-free days this effect is reduced to about 5.0 %.
The DNA dose rates, the most demanding of the three doses discussed in this study because of their sensitivity to short wavelengths in the UV spectral region agree to within
about 5 %, dropping to 2.3 % for the cloud-free cases.
For the evaluation of the satellite-based TEMIS UV products with the NILU-UV-derived ground-based products, the following, in particular, is found:
The TEMIS UV daily dose products are, on average, 12.5 % higher than the NILU103 daily doses under all skies. Despite the presence of a visually apparent seasonal pattern,
the correlation was found to be robustly high (R2=0.92 and R=0.95).
For the vitamin D (DNA damage) UV daily doses the differences under all-sky cases between the satellite- and ground-based estimates are similar with differences of on average
13 % (12.5 %), again with the satellite overestimating the dose and
again with very good correlation of R2=0.93 and R=0.95 (R2=0.93
and R=0.95).
It is well possible that the implicit aerosol climatology used in the
satellite retrieval algorithm is at least partly contributing to higher UV
doses at a moderately polluted site as Thessaloniki. Further, in the shorter
wavelength part of the UV-B spectral region errors in measuring the total
ozone column can have a relatively higher impact for an accurately retrieval
of the DNA-damage UV dose and vitamin D UV dose compared to the erythemal UV
dose. However, the ratios and the standard deviations for the differences in
the three UV doses are similar, suggesting that the contribution of errors
related to the total ozone column retrieval may not be very important.
Uncertainties in the B086 spectra and the methodologies used for the
calculation of the vitamin D and DNA-damage effective doses might also be
partly responsible for the observed variability, but these factors can only
explain a small fraction of the total variability in the differences (in
general less than 7 % for all-sky conditions).
Through data selections for
different cloud cover conditions it was shown that the greatest part of the
variability is due to the differences between the cloud cover fraction
assumed in the satellite algorithm and the definition of cloud-free cases in
the ground-based retrievals because the different field of view between the
ground- and satellite-based instruments might lead to discrepancies regarding
the cloud influences on the UV daily doses. Three clusters of cloudiness
types were investigated in order to evaluate the cloud contribution on the
differences between the satellite- and ground-based UV daily doses. The
introduced clusters were identified based on the percentage of cloud-free
periods over a day: excluding overcast days (days with more than 10 %
cloudless measurements), moderate cloud-free days (days with more than
70 % cloudless measurements), and cloud-free days (days more than
90 % cloudless measurements).
The number of cloud-free days limits the dataset to one fourth of the original, while the mean relative differences are reduced for all daily UV doses. Remaining discrepancies are on
average less than 1.3 % for the vitamin D and DNA-damage doses, while the
agreement for erythemal UV is on average even smaller (0.6 %), revealing
the notable improvement of the comparisons when excluding the cloudiness
effects.
Differences of less than 2 % with moderate standard deviations (∼ 15 %) are found when excluding the overcast days, implying that the major source of the high differences observed
under all-sky cases can be attributed to the availability and treatment of the cloud information; for example, the satellite algorithm cannot distinguish between thin and thick clouds under overcast conditions
Finally, the influence of aerosol variability was investigated using the UV
doses from the cloud-free days only. Coincident AOD values at 340 nm from a
collocated CIMEL sun photometer were used in order to examine the dependence
of the observed differences to the aerosol load at the urban site in
Thessaloniki. The results showed that for AOD values up to 0.4 the
contribution of aerosols to the differences in UV dose is quite significant,
while for even larger AOD this contribution results to slowly ascending
slops. Furthermore, model estimations demonstrated that discrepancies between
the measured and assumed SSA values can also lead to high differences on the
retrieved irradiances which are equivalent to those attributed to the
variability in AOD. Thus, the discrepancies seen in the two datasets under
cloud-free conditions can be at least partly attributed to the implicit
aerosol information used in the satellite retrievals at the site of
Thessaloniki, which experiences significant variations in aerosol properties.
In conclusion, this comprehensive study has revealed the merits, limitations
and accuracy of both ground-based and satellite-based estimates of erythemal
UV, vitamin D, and DNA-damage daily doses and underlying dose rates. Although
calibration procedures, a priori information and constraints of the methods
applied in the original datasets can still limit the accuracy of the
calculated photobiological products, these types of data comparisons will
remain highly important for the validation of satellite-derived UV doses and
to further increase awareness of the harmful effects of overexposure to
UV radiation.