Detection of dimming/brightening in Italy from homogenized all-sky and clear-sky surface solar radiation records and underlying causes (1959–2013)

A dataset of 54 daily Italian downward surface solar radiation (SSR) records has been set up collecting data for the 1959–2013 period. Special emphasis is given to the quality control and the homogenization of the records in order to ensure the reliability of the resulting trends. This step has been shown as necessary due to the large differences obtained between the raw and homogenized dataset, especially during the first decades of the study period. In addition, SSR series under clear-sky conditions were obtained considering only the cloudless days from corresponding ground-based cloudiness observations. Subsequently, records were interpolated onto a regular grid and clustered into two regions, northern and southern Italy, which were averaged in order to get all-sky and clear-sky regional SSR records. Their temporal evolution is presented, and possible reasons for differences between all-sky and clear-sky conditions and between the two regions are discussed in order to determine to what extent SSR variability depends on aerosols or clouds. Specifically, the all-sky SSR records show a decrease until the mid1980s (dimming period), and a following increase until the end of the series (brightening period) even though strength and persistence of tendencies are not the same in all seasons. Clear-sky records present stronger tendencies than all-sky records during the dimming period in all seasons and during the brightening period in winter and autumn. This suggests that, under all-sky conditions, the variations caused by the increase/decrease in the aerosol content have been partially masked by cloud cover variations, especially during the dimming period. Under clear sky the observed dimming is stronger in the south than in the north. This peculiarity could be a consequence of a significant contribution of mineral dust variations to the SSR variability.

transforming the original records into homogeneous ones. In particular, the homogenization of the series was performed 123 using a procedure based on a relative homogeneity test (Brunetti et al., 2006b), comparing by means of the Craddock 124 test (Craddock, 1979) each test series against 10 other reference series, that well correlate with the test one. The 125 common variance between two stations depends on their distance and for the Italian territory, as previously found for 126 SD (Manara et al., 2015), it falls to 50% at a distance of about 150 km. Homogenization was performed at monthly time 127 scale. However, a daily version of the adjustments was also generated in order to homogenize the daily series.

128
All analyzed series showed at least one inhomogeneous period highlighting the importance of homogenization,  (Table 1), an instrument with higher quality, as recommended by WMO (Italian Air Force, 2012).
133 Figure 3a shows the Italian average annual SSR anomaly series (relative anomalies with respect to the period 1976-134 2005) before and after homogenization with corresponding Gaussian low-pass filters (11-year window; 3-year standard 135 deviation) that allow a better visualization of the decadal variability and long-term trend, while Fig. 3b shows the curve 136 obtained by averaging the mean annual adjustments over all single records, together with their absolute range. The 137 details on how we obtained regional SSR anomaly series will be explained in the following sections.

138
The average annual records before and after the homogenization show a different decadal variability during the 1959-

145
We filled the gaps in each monthly record using a procedure similar to that described in Manara et al. (2015). In 146 particular, the median of a set of five estimated values, corresponding to the five highest correlated reference records, 147 was selected in order to avoid outliers resulting from peculiar climatic conditions of the reference station. When less 148 than five reference records fulfilling the requested conditions (distance within 500 km from the record under analysis 149 and at least six monthly values in common with it in the month of the gap) were available, the median was calculated 150 with the available reference series. After the gap filling procedure, all series had at least 90% of available data during and February fall and, for the first year, the winter and the annual means are calculated using also the monthly mean of et al. (2015), is based on an Inverse Distance Weighting approach with the addition of an angular term to take into 163 Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-206, 2016 Manuscript under review for journal Atmos. Chem. Phys. Published: 24 May 2016 c Author(s) 2016. CC-BY 3.0 License. account the anisotropy in the spatial distribution of stations. The resulted grid spans from 7° to 18° E and from 37° to 164 47° N, with 58 grid points over the Italian territory (Fig. 1).

165
Then, the monthly gridded anomaly records were subjected to a Principal Component Analysis (PCA) in order to 166 identify areas with similar SSR temporal variability. With this technique, it is possible to identify a small number of 167 variables, which are linear functions of the original data and which maximize their explained variance (Preisendorfer, 168 1988;Wilks, 1995). The analysis focused on the 1976-2005 period, the same reference period used to calculate the 169 anomaly series. The analysis shows that the first five eigenvectors have an eigenvalue greater than 1 and they explain 170 more than 91% of the total variance of the data set. Then, we selected to rotate the first two empirical orthogonal 171 functions (EOFs), which are those that account for 59% and 23% respectively of the original variance of the data set, in 172 order to obtain a more physically meaningful pattern (Von Storch, 1995). We decided to select these two EOFs because 173 they account for 82% of original variance while the other three account only for 9%. This procedure allowed to divide 174 the Italian territory in two regions: Northern Italy (29 grid points) and Southern Italy (29 grid points) ( Fig. 1). Finally,

175
we calculated the monthly, seasonal and annual mean anomaly series for the two regions by averaging all corresponding 176 grid point anomaly records. From here, we refer to these series as the SSR anomaly series obtained under all-sky 177 conditions to distinguish them from the series presented in the next section (Sect. 2.4) obtained selecting only the clear-  days. In particular, we considered only the days with a daily TCC mean of 0 okta. For the SSR series without a 184 corresponding TCC series, we considered the data from nearby stations. The main limitation of the previous procedure 185 is that the condition adopted to select clear-sky days (0 okta of TCC) allows often to select a very low number of days.

186
We tried therefore to apply a less restrictive condition and we extracted also the clear-sky days using as threshold 1 187 okta.

188
Then, the monthly mean was calculated when at least two clear-sky daily values were available in the considered 189 month. After the gap-filling, the monthly/seasonal/annual relative anomaly series were obtained with respect to the filling. This reduced the series to 44 in the first case (Table 1). As final step, we interpolated the anomaly series as 193 described in Sect. 2.3 and we obtained the regional anomaly series, for the two regions and the two thresholds, by 194 averaging all corresponding grid point anomaly series.

202
Specifically, the length of the period considered for the analysis and the starting year of the window that the trend refers 203 to, are represented on y and x axes, respectively. Slopes are showed by means of the colors of the corresponding pixels 204 Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-206, 2016 Manuscript under review for journal Atmos. Chem. Phys. Published: 24 May 2016 c Author(s) 2016. CC-BY 3.0 License. with large squares for trends with a significance level p<=0.1 and with small squares for trends with a significance level 205 p>0.1 (from here, we refer to these trends as non statistically significant) evaluated by the Mann-Kendall non 206 parametric test (Sneyers, 1992).

207
At annual scale both regions show a decreasing tendency until the mid-1980s and a following increase until the end of 208 the series, with a period of stabilization during the second half of the 1990s. The trends of the two periods are quite 209 comparable, especially in the Southern region. In fact, the trend for the whole period under analysis (e.g., 1959-2013 210 period) is not statistically significant, as highlighted by the running trend analysis (Fig. 5). However, the intensity of the 211 dimming is slightly higher for the South than the North, especially for some windows starting in the mid-1960s (e.g.,

215
For the winter season, the records show a well-defined behavior with a dimming and a following brightening only in the

216
Northern region where the record shows statistically significant negative and positive tendencies for some periods 217 starting in the 1960s and 1980s, respectively. On the contrary, the behavior is not well defined in the Southern region, 218 with a minimum in the mid-1980s and two secondary maxima during the 1970s and 1990s (Fig. 4). As a consequence, 219 the series shows only some windows of less than 30 years and starting in the 1960s with negative trends as well as one 220 21 year long sub-period starting in 1990 with a significant decreasing tendency (-4% per decade) and very few positive 221 trend windows starting in late 1970s/early 1980s (Fig. 5).

222
The spring season has a pattern similar to the year with a clear negative-positive sequence and a period of stabilization    The autumn records (Fig. 4) start with a period without any trend for both regions and then they show a decrease 237 between the beginning of 1970s and the beginning of 1990s and a following increase which is stronger in the Northern 238 than in the Southern region. This picture is evident also in the running trends (  Atmos. Chem. Phys. Discuss., doi:10.5194/acp-2016-206, 2016 Manuscript under review for journal Atmos. Chem. Phys. Published: 24 May 2016 c Author(s) 2016. CC-BY 3.0 License.
In order to give a more accurate information on the variations highlighted by Fig. 4 and Fig. 5, we estimated the trends 245 in Wm -2 per decade for some key periods (Table 2) then we calculated the seasonal and annual regional records, by averaging all corresponding grid point absolute series.

251
The seasonal/annual normals were calculated averaging the corresponding monthly normals obtained using the same 252 data and the procedure explained by Spinoni et al. (2012). In particular, the monthly normals were obtained starting 253 from a database of SD normals for the Italian territory representative of flat and non-shaded sites. These normals were 254 at first transformed into SSR normals by means of the Ǻngström law and then interpolated onto the USGS GTOPO 30 255 Digital Elevation Model grid by means of an Inverse Distance Gaussian Weighting spatialization model.

256
As previously shown, the trend over the whole period under analysis (1959-2013) is significant only in summer in the 257 North (+3.1Wm -2 per decade) and in autumn for both regions (-1.5Wm -2 and -2.2Wm -2 for the North and the South 258 respectively), as a consequence of a weak dimming period in the first case and a weak brightening period in the second 259 one. Considering the 1959-1985 period, as reference for the dimming period, the trend is significant in all the seasons 260 both for the North and the South with values ranging between -7.2Wm -2 per decade in spring and -3Wm -2 per decade in 261 winter for the North and between -8.5Wm -2 per decade in spring and -4.2Wm -2 per decade in autumn for the South. As  increases if the correlation between the filters is considered (it is always higher than 0.96). The decadal variability 272 shown by the trends using the two thresholds is very similar with the exception of few periods where one of the two 273 records shows a higher interannual variability causing slight differences in the resulting trends. This is evident for 274 example before the 1980s during the spring season in the North and during the winter season in the South; in both cases 275 the 0 okta series shows higher variability than the 1 okta series. The difference in the resulting variability over the 276 whole period under analysis  is particularly evident in the North where it is always higher in the 0 okta than 277 in the 1 okta series (the standard deviation of the residuals from the low-pass filter is comprised between 0.02 and 0.04 278 for the 0 okta series and between 0.01 and 0.03 for the 1 okta series), while it is less evident in the South where the two 279 series show the same variability (the standard deviation of the residuals from the low-pass filter is comprised between 280 0.02 and 0.05 for both the thresholds). The advantage of 0 okta as threshold is that it allows to select only the real clear-281 sky conditions but the limitation of this choice is that it allows to select only a low number of days (in the North it is 282 particularly due to the higher frequency of cloudy days), thereby increasing the variability of the obtained series. On the contrary using 1 okta as threshold allows to obtain a more stable series selecting a higher number of days, which are 284 however not completely clear.
In order to better understand the magnitude and the length of the tendencies shown in Fig. 6, we subjected the clear-sky 286 records to a running trend analysis (Brunetti et al., 2006a), as previously illustrated for the all-sky series. The running 287 trend obtained for the two different thresholds is very similar, so we show and discuss only the one obtained with 0 okta 288 as threshold (Fig. 7).

289
At annual scale the clear-sky records show a comparable decreasing and following increasing tendency both for the

296
During the winter season the trends show a dimming and a following brightening for both regions. The dimming is 297 slightly stronger than the brightening and, as a consequence, the trend over almost the whole period under analysis is 298 negative (e.g., -2% per decade during the 1959-2009 for both regions). The records show a period of stabilization 299 during the mid-1970s and during the 2000s (Fig. 6) and, as a consequence, some sub-periods especially in the South do 300 not show a significant trend after the beginning of 1980s.

301
The spring season has a pattern similar to the year with a clear negative-positive sequence after a starting period without

341
The differences between all-sky and clear-sky trends are less evident in spring and summer. Nevertheless, some

344
The differences between the all-sky and clear-sky anomaly records are highlighted also considering the ratios between 345 the latter and the former records. Before 1980, the low-pass filters applied to these ratios (figures not shown) give  Table 2 for the all-sky 357 series, were calculated and the results are reported in the same table.

358
These values confirm what already highlighted by Fig. 8. Assuming no changes in cloudiness, at the yearly scale the 359 trend during the dimming period  decreases to -6.3Wm -2 per decade in the North and to -8.4Wm -2 per 360 decade in the South (the corresponding all-sky trends are -4.4 and -6.4Wm -2 per decade, respectively), confirming that influenced in a significant way both periods and regions. In particular, during the brightening period the trend changes 367 in the North from +4.8Wm -2 per decade (all-sky) to +6.5Wm -2 per decade (estimated constant-cloudiness) and in the 368 South from a not significant value (all-sky conditions) to +5.4Wm -2 per decade (estimated constant-cloudiness).

369
Without the contribution of the clouds the correlation between the North and the South becomes higher: the correlation

386
Starting from the homogenized daily records, besides SSR series under all-sky conditions, SSR series under clear-sky 387 conditions were obtained selecting clear days from corresponding ground-based TCC observations. Then, these series 388 were projected onto a regular grid (1° x 1°) covering the entire Italian territory and clustered in two regions (Northern 389 and Southern Italy) by means of a Principal Component Analysis. The records of these areas were averaged in order to 390 get the corresponding regional all-sky and clear-sky SSR records for the 1959-2013 period.

391
The clearest feature of the Italian all-sky SSR is a significant dimming from the beginning of the series to the mid-

419
The stronger spring, summer and autumn dimming in the South than in the North under clear-sky may challenge the 420 above hypotheses as the North is more affected by air pollution due to higher emissions. Nevertheless, it is worth noting 421 that Southern Italy is more affected by coarse aerosols (Bonasoni et al., 2004), causing a significant contribution of 422 natural aerosols to SSR variability as for example mineral dust intrusions from the Sahara and Sahel (Prospero, 1996).

423
In particular, we highlight that a comparison between the Northern and Southern Italian clear-sky SSR variations and Europe shows a pronounced seasonal cycle with a maximum in summer and a minimum in winter (Pey et al., 2013;431 Varga et al., 2014) and a distinct gradient with the highest values near the northern coast of Africa (Gkikas et al., 2013;432 Prospero, 1996).  Italy where the brightening especially in the Northern region appeared to be stronger as compared to Europe.

457
The all-sky SSR trends presented for the Italian territory show some discrepancies with respect to the trends of SD

466
The fact that the dimming in SD is weaker than in SSR could indicate that the long-term increase in aerosols affects the 467 two variables in a different way inducing a more significant reduction in the intensity of SSR than in SD. The 468 discrepancies between SSR and SD trends could also be a consequence of a different sensitivity to changes in the 469 diurnal cycle and decadal variability of cloud cover, temperature and humidity that could modify the measurements of 470 SD differently than SSR, but the reasons for these differences need further research. However, Sanchez-Romero et al.

471
(2014) in a review reported some studies that found similar discrepancies between SD and SSR trends in different areas 472 of the world as for example Germany (Power, 2003) We sincerely thank all the institutions that allowed access to the data for research purposes and contributed to set up the   (7), 2425-2432, doi:10.5194/angeo-23-2425-2005, 2005.