SEARCH

SEARCH BY CITATION

Keywords:

  • artificial neural network;
  • column ozone;
  • erythemal UV radiation;
  • sunshine duration;
  • time series reconstruction;
  • trend analysis

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

[1] Solar erythemal UV radiation (UVER) is highly relevant for numerous biological processes that affect plants, animals, and human health. Nevertheless, long-term UVER records are scarce. As significant declines in the column ozone concentration were observed in the past and a recovery of the stratospheric ozone layer is anticipated by the middle of the 21st century, there is a strong interest in the temporal variation of UVERtime series. Therefore, we combined ground-based measurements of different meteorological variables with modeled ozone data sets to reconstruct time series of daily totals of UVER at the Meteorological Observatory, Potsdam, Germany. Artificial neural networks were trained with measured UVER, sunshine duration, the day of year, measured and modeled total column ozone, as well as the minimum solar zenith angle. This allows for the reconstruction of daily totals of UVERfor the period from 1901 to 1999. Additionally, analyses of the long-term variations from 1901 until 1999 of the reconstructed, new UVER data set are presented. The time series of monthly and annual totals of UVERprovide a long-term meteorological basis for epidemiological investigations in human health and occupational medicine for the region of Potsdam and Berlin. A strong benefit of our ANN-approach is the fact that it can be easily adapted to different geographical locations, as successfully tested in the framework of the COSTAction 726.

1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

[2] Solar UV radiation plays a major role in many biological and chemical processes, influencing not only the human skin and immune system, but also impacting on animals and plants. Overexposure to UV radiation from the sun or from artificial sources is a considerable public health concern. There is increasing evidence that it suppresses the immune system and triggers premature skin aging, as well as melanoma and non-melanoma skin cancer in humans. On the other hand, beneficial effects of UV radiation such as vitamin D3production in the skin are reduced with low values of UV irradiance, as they occur with low sun elevations at mid- and high latitudes during wintertime [e.g.,Feister et al., 2011].

[3] The impact of UV radiation on biological systems usually shows a strong spectral dependence and is defined by a spectrally integrated value Ebiol, according to equation (1), where the solar spectral irradiance EUV(λ) is convolved with a biological weighting function Sbiol(λ).

  • display math

Here we use erythemally weighted UV radiation (UVER), applying for Sbiol(λ) the erythemal action spectrum [McKinlay and Diffey, 1987]. Erythemal weighting is used, since it is essential for many biometeorological aspects including human health damage, and it is the quantity that has been widely measured. Nevertheless, most of the existing measured UV time series are relatively short, and thus it is not possible to analyze long-term changes in the UV radiation caused by variations of total column ozone, aerosol optical depth, or cloudiness. But long-term time series of reconstructed UV data are of particular interest not only in bio-climatology, but also for epidemiological investigations in skin cancer and occupational health. Epidemiology accounts additionally for temporally changing patterns of exposure to the sun, e.g., an increase in outdoor activities, and to artificial UV sources, e.g., with the rising popularity of sunbed tanning.

[4] In this study the methodology of reconstructing long time series of erythemal UV radiation as presented by Junk et al. [2007]is refined and extended by integrating the essential ozone information from two different numerical models. An improved input data set is used which contains modeled daily total column ozone data for Potsdam, Germany, for the period before ground-based measurements started in 1964. Combined with measured daily sunshine duration and other parameters, daily totals of UVER have been reconstructed for the period from 1901 until 1999. In comparison to the results presented by Junk et al. [2007], uncertainties can be considerably reduced for the time-span in the reconstructed time series before ground based ozone measurements became available.

[5] A similar reconstruction methodology has been successfully applied in projects like the COST-Action 726 (“Long-term changes and climatology of UV radiation over Europe”) and the SCOUT O3 project (“Stratospheric-Climate Links with Emphasis on the Upper Troposphere and Lower Stratosphere”). Finally trend analyses of 30-year time slices of the new UVER data sets are performed and discussed.

2. Data and Methods

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

2.1. Description of Measurements and Models

[6] The required daily data for the Potsdam site (52°22′N, 13°5′E, 107 m a.s.l.), South-West of Berlin, were taken from the central data archive of the German Meteorological Service.

[7] Sunshine duration (SD) is defined as an accumulated time of bright sunshine during a specified period, here one day. Bright sunshine is defined as direct solar irradiance ≥120 W m−2. SD was measured with Campbell Stokes sunshine recorders. The time resolution of recorded values is one hour and the values are read-off to tenths between 0 and 1. All the values of one hour are summed up to hourly and then to daily sunshine duration. Horizon obstructions can be excluded for the Potsdam site. In the 99-year time series (01/1901 until 12/1999) only three complete months were missing (08/1998, 08/1999 and 09/1998). These gaps were replaced by 30-year climatological means, centered on each missing value.

[8] Daily column ozone data were taken from three sources: (a) modeled ozone from the CCM SOCOL [Rozanov et al., 2008], called model 1 hereafter (data available for the period from 1901 until 1999); (b) ozone reconstructed by a statistical model, hereafter model 2 [Krzyścin, 2008] (data available for the period from 1950 until 2003); and (c) ground based Dobson ozone observations available for the period from 1964 until 2003 [Spänkuch et al., 1999].

[9] Model 1: CCM SOCOL, the Climate-Chemistry Modeling tool for SOlar Climate Ozone Links studies is a combination of the general circulation model MA-ECHAM4 [Manzini et al., 1997] and the chemistry-transport model MEZON [Rozanov et al., 1999; Egorova et al., 2003]. CCM SOCOL is a spectral model with T30 horizontal truncation resulting in a grid spacing of about 3.75°. In the vertical direction the model has 39 levels in a hybrid sigma-pressure coordinate system extending from the surface to 0.01 hPa pressure height. The chemical-transport part treats 54 chemical species of the oxygen, hydrogen, nitrogen, carbon, chlorine and bromine groups. Their mixing ratios are determined by gas-phase, photolysis and heterogeneous (in/on aqueous sulfuric acid aerosols, water ice and nitric acid trihydrate) reactions. CCM SOCOL in version 2.0 [Schraner et al., 2008] was applied to simulate ozone and climate change during the entire 20th century. It was driven by prescribed time evolving sea surface temperature, sea ice distribution, greenhouse gases and ozone destroying substances, stratospheric sulfate aerosols, solar spectral irradiance, anthropogenic and natural sources of carbon monoxide and nitrogen oxides, as well as land use changes [Fischer et al., 2008]. A temporally invariant aerosol distribution with the aerosol parameters surface area density, extinction coefficient, single scattering albedo and asymmetry factor has been used for the troposphere. Last three parameters were defined separately for all spectral intervals of the model radiation code. The use of a temporally invariant aerosol distribution in the model disregards the seasonal variations of anthropogenic aerosol and could lead to higher uncertainties in the modeled ozone data. The output data were stored at 12-h intervals for all model grid cells. A spatial subset of 3 × 3 grid cells around Potsdam was extracted from the ozone data fields. Spatial means of daily averages were calculated. This averaging procedure was also applied for the ozone data set from model 2.

[10] Model 2: In the framework of the COST-Action 726,Krzyścin [2008] presented a statistical model to reconstruct daily values of total ozone over Europe since 1950. The model was trained on the satellite derived total ozone over the period from 1979 to 2004 to select an optimal set of total ozone proxies from various indices of the atmospheric circulation, as well as from meteorological variables derived from the NCEP/NCAR reanalysis [Kalnay et al., 1996]. Model 2 parameterizes the long-term solar effects on total ozone using the so-called Penticton solar radio (10.7 cm) flux measured in Canada, since 1947. The 11-year solar signal in total ozone was found to be small in middle latitudes, ∼1.5% increase from the solar minimum to solar maximum [e.g.,Austin et al., 2008]. The COST-726 O3 database contains this ozone data set with a temporal resolution of one day (01/1950 until 12/2004) and a spatial resolution of 1° × 1° covering all of Europe from 25°W to 35°E and 31°N to 80°N (http://private.igf.edu.pl/∼jkrzys, last accessed 15.02.2012). The value of one grid cell represents an area of approximately 1500 km2 at the latitude of Potsdam.

[11] Ground-based ozone and UV observations: The daily mean values of the re-evaluated Dobson ozone measurements recorded at Potsdam from 1964 until 2003 had been re-evaluated bySpänkuch et al. [1999]. These measurements are based on direct sun and zenith sky measurements with Dobson spectrophotometers #64 and #71. For the period from 1987 to 2003, the few days with missing Dobson ozone values were filled by Brewer ozone measurements.

[12] UV irradiance was measured at Potsdam by Brewer spectroradiometers (#030 MKII, and #118 MKIII) once or twice per hour and is available for the period from 01/1995 to 12/1999. Regular on-site calibrations of the instruments were performed by 1000 W FEL quartz halogen lamps that had been calibrated according to the SI by the German National Metrological Institute Physikalisch-Technische Bundeanstalt with uncertainties of ±3% in the UV region. All measurements are corrected for the cosine error accounting additionally for changing cloudiness within the scan interval [Feister et al., 1997]. To account for short time variations of UV radiation, ratios between UV radiation and one minute values of global solar irradiance were used to determine hourly and daily totals of erythemal UV radiation. For further details and information about the instrument calibration see, e.g., Feister et al. [2008]. Spectral UV irradiance at Lindenberg (52°12′31″N, 14°7′17″E, 127 m a.s.l.) has been measured by Brewer spectroradiometer #078 (MKIV). The method of cosine correction applied to Brewer instruments at Potsdam could not be used, because the required concurrent broadband UV radiation data were not available Feister et al. [2008]. Therefore, the cosine correction method described by Bais et al. [2005] that does not require additional measured input data was applied.

2.2. Neural Networks

[13] Neural networks have been applied successfully to various studies for estimating solar irradiation [Chevallier et al., 1998; Dorvlo et al., 2002; Feister et al., 2008; Junk et al., 2007; Mohandes et al., 1998; Reddy and Ranjan, 2003]. We used this nonlinear statistical approach to reconstruct daily values of UVER back to 1901.

[14] Artificial neural networks (ANN) consist of simple elements that operate in parallel. The most important common characteristic of biological and artificial neural networks is their capability to learn from examples. The neuronal network technology imitates the human brain's own problem solving ability to apply knowledge gained from past or previous experience to new problems by building a system of “neurons” that makes new decisions, classifications or forecasts. ANNs learn the relationship between the input and the output data by studying previously recorded data without knowing the physical relationships [López et al., 2001]. A function is set up by adjusting the values (weights) of the connections between the elements (nodes) during the training process. This function is iteratively adjusted until the network output matches the target with an adequate accuracy.

[15] This study estimates daily doses of UVER by a neural network approach using as standard predictors temporal information of the month and the day of the year (DOY), the minimal solar zenith angle of a day, which is for a given location a function of the DOY, and the sunshine duration. Depending on the network structure, total ozone is used as an additional predictor. In order to find the best ANN configuration several network architectures with up to three hidden layers and 32 nodes were tested. To evaluate the performance of the different networks the results were compared to independent measurements, which were not used to setup and train the model. Hence this setup allows us to reconstruct UVER despite the relatively short period of direct measurements of UV radiation, as ozone data is available through the modeling studies.

3. Results

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

3.1. Comparison of Modeled Versus Measured Ozone Data

[16] Preceding the reconstruction of UVER, the differences between modeled and measured daily means of total column ozone were analyzed. Figure 1 shows the ratios between monthly means of daily ozone values from model 1 and measured ozone for the period 1964 to 1999.

image

Figure 1. Ratios between monthly ozone values predicted and ozone values measured at Potsdam, Germany in relation to the annual cycle for the period from 1964 to 1999 (N = 432) by model 1 (CCM SOCOL) in gray and model 2 (statistical model) in red. The median values are indicated as small black lines. The standard deviation ranges between 0.09 (January) and 0.03 (August) for model 1 and between 0.04 (January) and 0.02 (August) for model 2.

Download figure to PowerPoint

[17] The ratios for model 1 show a clear seasonal cycle revealing that model 1 predicted total ozone at Potsdam approximately 5% too low in the long term March/April maximum and 10% too high in the November minimum. The seasonal cycle has been parameterized by a third-order polynomial (blue line inFigure 1), which was applied to correct the daily values of model by subtracting the seasonally changing systematic differences (bias correction).

[18] The ratios between ozone from model 2 and from measurements do not show a seasonal cycle (red dots in the upper graph in Figure 1), and hence no correction was applied. In Figures 2a and 2b scatterplots of measured ozone for Potsdam and ozone predicted by model 1 and model 2 (corrected) are shown (1964 until 1999). In Figure 2cthe measured values of Potsdam were additionally compared to ozone measurements at Lindenberg, which is 73 km east-northeast of Potsdam in order to give an impression of the close spatial correlation of column ozone.

image

Figure 2. Scatterplots and relative differences of measured ozone at Potsdam, Germany compared to (a, b) ozone from model 1 (CCM SOCOL), (c, d) ozone from model 2 (statistical model), and (e, f) ozone measured at the DWD Meteorological Observatory Lindenberg, Germany, which is 73 km away from Potsdam. Please note the difference in time periods in Figures 2e and 2f.

Download figure to PowerPoint

[19] It can be seen that model 1 has a much higher scatter compared to the measurements at Potsdam than model 2. One reason for the increased scattering might be attributed to the lower spatial resolution of the model (3.75° in model 1 versus 1° spatial resolution in model 2); but the different model approaches - numerical versus statistical model – are also an important source for the different results.Table 1 shows descriptive statistics of the differences between the measurements at Potsdam and the two models as well as between the measurements in Lindenberg and Potsdam for a shorter time span. In addition, the correlation coefficients between Potsdam and Lindenberg (0.96), Potsdam and model 1 (0.78) and between Potsdam and model 2 (0.94) were calculated.

Table 1. Descriptive Statistics of the Differences Between the Ozone Measurements at Potsdam, Model 1 (CCM SOCOL), Model 2 (statistical model), and the Measurements at Lindenberg in DUa
 NMeanStDP25P50P75P95
  • a

    StD = standard deviation and P = quantile.

Potsdam - Model 1 (CCM SOCOL)110725.660.1−24.21.226.175.0
Potsdam - Model 2 (statistical model)110721.517.1−7.02.310.129.4
Potsdam - Lindenberg3106−2.012.2−7.1−2.24.216.3

[20] In general, the two modeled ozone records and the measured monthly ozone show similar long-term patterns as illustrated in the time series of 12-month running averages inFigure 3.

image

Figure 3. Twelve-month running averages (RA) of ozone from models 1 (CCM SOCOL, corrected) and 2 (statistical model), as well as from measurements at Potsdam, Germany (left-hand scale and lower graph). Twelve-month running mean monthly ozone values from model 1 and from ozone measurements at Arosa, Switzerland are shown in the upper graph (right-hand scale).

Download figure to PowerPoint

[21] However, there are differences between the ozone as simulated by model 1 and model 2 from the mid-1950s to the beginning of the 1960s. The quality of model 2 was assessed through a comparison of the reconstructed total ozone with independent ground-based data for Dobson stations operated in Europe in the early 1950s and 1960s byKrzyścin [2008]. The differences are mostly within the range of model uncertainty. The ozone minimum after the El Chichon volcanic eruption in Mexico in 1982 is nicely reproduced synchronously by the two models and is in coincidence with the measurements. After the Mt Pinatubo volcanic eruption at the Philippines in 1991, there are more pronounced deviations between measurements and model results. For example, model 1 shows a noticeable temporal delay in the ozone minimum. The low ozone values in 1992 and 1993 resulted in UV radiation enhancements by 5% to 10% [Feister and Grewe, 1995], which were further enhanced by 5% due to a reduced cloudiness [Feister et al., 2002]. For comparison, the upper graph of Figure 3shows 12-months running mean monthly ozone values from model 1 for the station Arosa (Switzerland) at 46.7789°N and 9.6735°E, which is about 667 km Southwest from Potsdam. The modeled long-term ozone variations are similar at both sites with values at Arosa being by 8 DU (2%) smaller than at Potsdam. 12-month running averages of monthly means derived from ozone measurements at Arosa from August 1926 onwards show a similar long-term pattern and short-term variations as the model 1 output for that site. The high ozone peak values measured during the winter months of the period 1940 to 1942 at Arosa have been explained by anomalies in the stratospheric circulation over Central Europe in the winter months 1940, 1941, and 1942 [Brönnimann et al., 2004]. High ozone peaks during these years have been also detected by Griffin [2006] for Mount Wilson (U.S.). They do not appear in the model 1 ozone values. A first record of column ozone measurements using Dobson spectrophotometer No 9 was obtained at the Potsdam Meteorological Observatory between September 1, 1941 and March 12, 1945 [Hinzpeter, 1952]. Unfortunately, no ozone measurements are available at Potsdam for the winter months of 1940 and 1941. The ozone values measured in January and February 1942 do show ozone values higher by 30% to 40% than the overall mean between 1942 and 1945, but they are not higher than in the subsequent winter months 1943 to 1945. Therefore, it cannot be decided, whether the extreme ozone enhancement as reported for Arosa for the winter months from 1940 to 1942 did occur to the same extent at Potsdam. However, the winter months are of lesser importance for the reconstruction of UV radiation. The ozone measurements at Arosa that go further back in time than the measurements at Potsdam give us some more confidence on the reliability of the ozone input to the UV reconstruction.

[22] Based on the results of the comparison between modeled and measured ozone, the following data were used for the ANN calculations: model 1 data was used for the period 1901 until 1949, model 2 for the period 1950 until 1963, and measured ozone for the period 1964 until 1999. Ozone data from model 1 were also used for those few days, when ozone measurements are missing.

3.2. Preparation of the Test, Training and Validation Data Sets

[23] For the setup of the ANN, the UV measurements of the Brewer spectroradiometers (#030 MKII, and #118 MKIII) of the period 1995 until 1999 were used. This data set was combined with the other input data (ozone, sunshine duration, zenith angle, DOY, month). Only those days were extracted, where values of all parameters were available. A total of 1349 days remained. This data set was split into four data sets, a training data set (N = 792, 58.7% of the data), a test and a validation data set (N = 169, each 12.5%). Additionally, the first five days of each month were removed from the original data set to get a second independent test data set (N = 220, 16.3%). This data set is used for a second independent model evaluation, too.

[24] In order to see whether there is a significant gain in information using the modeled ozone data set, three neural networks with different predictors were used (Table 2). They apply the four standard predictors, but differ in ozone as an additional predictor. The first network (A) uses the measured ozone data to predict daily values of erythemal UV radiation. The capability to predict these values is used as the reference for the next two models. In the second network (B), modeled ozone values (model 1) were used instead of the measured ones. To estimate the influence of ozone as a predictor in the last network (C), ozone is omitted.

Table 2. Setup of the Three Different ANN Models
NetworkMonthDOYZenith AngleSunshine DurationMeasured OzoneModeled Ozone
Ayesyesyesyesyesno
Byesyesyesyesnoyes
Cyesyesyesyesnono

3.3. Performance of the Artificial Neural Networks

[25] Figures 4a–4c show the correlations between calculated and measured daily doses of erythemal UV radiation for the period from January 1995 to December 1999 at Potsdam. As expected, a clear relationship between the calculated and measured UVER values can be observed, with an increasing scatter when ozone is not used as a predictor (Figure 4c).

image

Figure 4. Scatterplots and absolute differences of daily values of calculated versus measured data of erythemal UVER [J m−2]ER, Potsdam, Germany, 1995 until 1999. (a, b) Network A: measured ozone data, (c, d) network B: modeled ozone data, and (e, f) network C: no ozone data.

Download figure to PowerPoint

[26] The ANN number A using the measured ozone (Figure 4a) shows the best results with a correlation coefficient of 0.987 and a root mean square error (RMSE) of 10.8% relative to the measured daily totals. The RMSE is reduced to 3.5% for the monthly totals.

[27] The correlation coefficient decreases to 0.96 using network B with the modeled ozone data (Figure 4b). The performance of network B is nearly the same in all seasons with slightly lower correlations during summer (seasons not shown here). In general this network underestimates the yearly totals of UVER radiation by 3%. More than 71% of the calculated monthly mean values are found within ±10% of the measured ones and 93% within ± 20%. This is comparable with the results of other studies [e.g., Lindfors et al., 2003] for Sodankylä in Finland.

[28] The highest RMSE values, 22.9% for daily data and 19.0% for monthly values, shows network C without any ozone data used for the calculations. Also the correlation coefficient decreases to 0.947 (Figure 4c). This result shows the benefit of the modeled column ozone for the reconstructed daily doses of UVER.

[29] Snow cover and snow depth can strongly modify UVER and have been successfully used in other studies to predict UVER for high altitude Alpine and for high latitude sites, respectively [Blumthaler and Ambach, 1988; Kylling et al., 2000; Lindfors and Vuilleumier, 2005]. Although a long time series of snow cover and snow depth is available for Potsdam, we decided not to use it as an additional predictor. In the ANN training period, only a few days per year occurred with a complete snow cover in the winter season with its low UVER, contributing little to the annual totals. Including these days would have considerably reduced the size of the training set. A test that includes a variable containing 0 for no snow cover and 1 for snow cover did not lead to better results.

3.4. UVER Time Series Reconstruction

[30] The ANNs are capable to predict daily totals of UVER, but their random error is considerably large, as input data are restricted to SD, observed or modeled ozone. Therefore, our statistical analysis concentrates on monthly values and annual totals derived from reconstructed daily values. For monthly and annual totals accumulated from the reconstructed UVERdaily doses, long input time series without missing values are essential. Therefore, a proper handling of missing values is important. Gaps of only one day in the time series were closed by a linear interpolation of neighboring values. The three missing months in the sunshine duration data set (August 1998 and 1999, as well as September 1998) were filled with long-term monthly means (30-years) centered on each missing value.

[31] Figure 5 shows as one of our main results the estimated UVER annual totals as well as the anomalies with respect to the reference period 1961 until 1999 (mean value 491 KJ m−2). The annual totals vary considerably since 1901. Long-term variations are better captured by the 5-year running mean (thick black line inFigure 5). A strong increase in the annual totals of UVERfrom the beginning of the reconstructed time series up to the mid-1950s is followed by a less pronounced decrease between 1955 and 1990.

image

Figure 5. Reconstructed annual UVER time series and anomalies (mean values 1961 until 1990: 4.91E + 05 [J m−2]ER), Potsdam, Germany, 1901 until 1999.

Download figure to PowerPoint

3.5. Linear Trend Analyses

[32] Annual mean values of measured sunshine duration, ozone concentrations as well as the reconstructed time series of UVER values were shown in Figure 6. The positive trend of the UVER values at the beginning of the last century up to 1920 coincides with increasing values of SD and decreasing ozone concentrations. The general shapes of the UVER and the SD curves were very similar for the whole period. A clear influence of the O3 concentrations on the UVER values could not be derived from the annual data. The overall shape of the annual SD and UVER curves are similar with the results presented by Wild [2009] for the annual man surface solar radiation. An increase up to 1950 (“early brightening”) is followed by a slight decrease (“dimming”) and again positive trends 1990 onwards (“brightening”) [Wild, 2009; Wild et al., 2007].

image

Figure 6. Reconstructed annual total of UVERtime series (blue line, left-hand scale, lower graph), mean annual O3values in DU (gray line, right-hand scale) and mean annual sunshine duration (SD) in minutes per day (red line, left-hand scale, upper graph) for Potsdam, Germany, 1901 until 1999 (thin lines, annual values, thick lines 7-years running average).

Download figure to PowerPoint

[33] Trends within the reconstructed UVERare analyzed for 30-year time periods based on annual totals accumulated from the reconstructed daily values. The Statistical Program Library of the Potsdam Institute for Climate Impact Research PIK [Österle et al., 1999] and the TimeStats program [Udelhoven, 2005] are used for the trend analysis. The significance of the trends is assessed by the non-parametric Mann-Kendall test.

[34] Figure 7 shows the results for the annual totals of UVER. A period dominated by highly significant positive trends between the 1901/1930 and 1924/1953 long-term mean intervals is followed by an episode without statistically significant trends. This positive UVER trend at the beginning of the last century is in line with an observed trend in the cloud cover (data not shown here). Observations of total cloud cover were available three times per day since 1893 at Potsdam. The frequency of low cloud cover (0 up to 2 Octa) is reduced by 22% between 1901/1931 and 1961/1990. Between the 1931/1962 and 1959/1988 interval smaller negative trend values can be observed. From the 1965/1994 intervals onwards again positive trends occur (not significant). An additional trend analysis of the seasonal totals indicates different trend patterns throughout the meteorological seasons (Figure 8).

image

Figure 7. Linear trend analysis of the reconstructed annual totals of UVERfor time spans of 30-years each, starting in 1901, different significance levels (P) of the trend values are indicated by colors.

Download figure to PowerPoint

image

Figure 8. Linear trend analysis of the reconstructed seasonal totals of UVER for time spans of 30 years each, starting in 1901, different significance levels (P) of the trend values are indicated in color.

Download figure to PowerPoint

[35] Strong UVER increases in the summer season, month June to August, substantially contribute to the long period with positive trend values in the first half of the 20th century, while the period with negative trend values between the 1931/1962 and 1959/1988 intervals is mostly influenced by the trend values of September, October and November. Highly significant positive trend values for time spans starting at 1964/1993 until the end of the reconstructed time series can only be found during the winter months. The overall patterns of the trend values, as well as their level of significance show plausible distributions without extreme values in all seasons.

4. Discussion and Conclusion

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

[36] The knowledge of biologically effective UV radiation doses is important, since they play a significant role in many processes in the biosphere, including the influence on the human skin and immune system. In this overall context we investigated UVER, because it is relevant for human health damage and is the quantity that has been widely measured [Koepke et al., 2008].

[37] Long-term changes in the surface level of UVERradiation from ground-based measurements can hardly be discussed on a global scale, since reliable and long UVER time series (longer than 50 years) are not available at a sufficient number of sites. Therefore, models to reconstruct are necessary to analyze past UVER variability and trends based on time series of ozone, clouds, and aerosols serving as input to such models. The action spectrum for erythemal UVERincludes a relatively marked UVA portion compared with other action spectra, e.g., for DNA damage, Previtamin-D3-synthesis, or for general plant disease. Thus, a UV trend due to total column ozone can be more pronounced in the UV with other biological weightings. A UV trend only due to changes in cloudiness would have a similar magnitude as the UVER trend, because the spectral dependence on cloud optical depth is lower than on ozone.

[38] The ANNs used in this study applies only SD and ozone as observed input to simulate the day-to-day variations of UVER. SD is reported as the time span for which the direct irradiation exceeds the threshold of 120 W m−2. Thus, information on cloud optical properties contained in the SD variable is much less than for example in global irradiation and thus increases model uncertainty [Koepke et al., 2008]. In order to quantify this uncertainty we calculated the correlation coefficient between the sunshine duration and the global radiation for the time span where both parameters were available (1937–1999). The correlation coefficient is 0.89 for the whole time span and varies in the different meteorological seasons (winter: 0.82, autumn: 0.88, spring: 0.91, and summer: 0.93). Hence, more complex models with global irradiation and additional important predictors such as aerosol optical depth and horizontal visibility show smaller uncertainties [Feister et al., 2008; Feister and Junk, 2006; Fioletov et al., 2001]. However, for a study extending over almost a century, the necessary input parameters for such more complex models are in general not available.

[39] In contrast to other methods [Lindfors and Vuilleumier, 2005], one advantage of the ANN approach is that there is no need to set up different models or regressions for the different seasons of the year. The seasonal variations of all parameters are directly taken into account in the model setup by the additional “day of year” (DOY) time parameter.

[40] The main objective of this study was to include as much as possible relevant and available input data in the reconstruction of a 99-year UV radiation time series. As shown byFeister et al. [2008]the consideration of ozone as an additional predictor leads to better results. Therefore, we included long-term time series of modeled ozone data in our reconstruction in order to derive trends and variations in time series of erythemal UV radiation for the period from 1901 until 1999.

[41] Linear trends should be interpreted with caution for predictions, because they are mostly not capable to describe true changes within a time series. However, they are an effective statistical method to describe and objectively compare past changes. Comparisons with similar studies are difficult as either different time spans or predictors are used. Our results for Potsdam agree nevertheless well with those presented by Lindfors and Vuilleumier [2005] for Davos (Switzerland), who showed low values of UVER in the late 1930s, followed by an increase up to the middle of the 1940s and a steady decrease from the early 1960s up to the late 1970s. Feister et al. [2002] also detected a slight decline for Potsdam in the late 1970s using a different reconstruction method. Chubarova and Nezval [2000], and Chubarova [2008] identified a decline in UVER radiation in Moscow for the summer months of the period from 1969 to 1997. Zerefos et al. [2011] showed comparable results for UVB radiation for selected sites in Canada, Europe and Japan with decreasing trends after the Pinatubo eruption followed by a pronounced UVB increase due to the brightening effect.

[42] A strong benefit of our ANN-approach is the fact that it can be easily adapted to different geographical locations, as successfully tested in the framework of the COST–Action 726.

[43] Two practical applications for the modeling of UVER should be mentioned. First, the erroneous measurements can easily be detected and second, gaps in time series of measurements can be filled with more accurate data than by an interpolation alone. The time series of monthly and annual totals of UVERprovide a long-term meteorological basis for epidemiological investigations in human health and occupational medicine for the region of Potsdam and Berlin. Finally we think it is important to state that despite the accuracy of the UVER radiation modeling, this does not render direct measurements of UV radiation.

Acknowledgments

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

[44] The authors gratefully acknowledge the financial support of the Ministère de l'Enseignement supérieur et de la Recherche (MESR) of the Grand Duchy of Luxembourg in the framework of the Air Quality Project (AirQ). We also greatly acknowledge the support by the SCOUT-O3 project for making parts of this study possible and the German Meteorological Service for providing local observations as input data to our model calculations. We also thank Rene Stübi (MeteoSwiss) for kindly providing the ozone data for AROSA, Switzerland.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information

Supporting Information

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Data and Methods
  5. 3. Results
  6. 4. Discussion and Conclusion
  7. Acknowledgments
  8. References
  9. Supporting Information
FilenameFormatSizeDescription
jgrd17969-sup-0001-t01.txtplain text document0KTab-delimited Table 1.
jgrd17969-sup-0002-t02.txtplain text document0KTab-delimited Table 2.

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.