Recent spatial and temporal variability and trends of sunshine duration over the Iberian Peninsula from a homogenized data set

Authors


Abstract

[1] This paper describes the construction of a new data set of sunshine duration over the Iberian Peninsula (IP), based upon 72 monthly instrumental series covering the 1931–2004 period. Particular emphasis is placed upon the quality control, gap filling, and homogenization procedures. We also produced a gridded version (1° × 1° resolution, latitude, and longitude) of the homogenized data set covering the whole IP with 84 cells. We subsequently analyzed sunshine duration data in search of temporal variability, long-term trends, and spatial differences over the IP. The analyses were carried out on the mean series calculated for the whole IP and for four subregions, at annual, seasonal, and monthly resolution. We defined the subregions using a principal component analysis applied to the sunshine duration data. The temporal evolution of the mean annual series for the whole IP shows a decrease in sunshine duration from the 1950s to the early 1980s, followed by a positive trend up to the end of the 20th century, which matches the “global dimming” and the subsequent and recent “brightening” described for other regions in international scientific literature. The most interesting seasonal significant trend for the IP series was found for summer (−0.9% per decade, over the 1951–2004 period); this trend is influenced by the months of July and August in two subregions. In addition, important trends ranging from +2.8 to +3.8% per decade were observed for the March series for all Mediterranean climate subregions.

1. Introduction

[2] Analysis of the solar radiation reaching the Earth's surface is very important in the context of recent climate change, as, apart from changes in cloudiness, variation therein might also be indicating certain anthropogenic disturbances such as variations in aerosol concentration [Ramanathan et al., 2001]. A reduction of solar radiation between the 1950s and 1980s has been particularly well established and documented and corresponds to a widespread reduction of −0.51 ± 0.05 Wm−2/a, equivalent to a decrease of 2.7% per decade [Stanhill and Cohen, 2001]. The phenomenon is commonly known as “global dimming,” although this decline shows important spatial variations [Gilgen et al., 1998; Liepert, 2002]. An opposite trend is, on the other hand, detected starting from the late 1980s, and the recent decades were characterized by a progressive brightening in solar radiation measurements [Wild et al., 2005; Pinker et al., 2005].

[3] Although the cause of the decrease in solar radiation in the 1950s–1980s period is not fully known and certainly cannot as yet be quantitatively explained, the most accepted reason at present is a change in the transmissivity of the Earth's atmosphere [Stanhill, 2005], owing to an increase in aerosol concentrations as a consequence of anthropogenic emissions. With regard to the recent brightening phenomenon, certain hypotheses suggest reductions of anthropogenic aerosol emissions over the last few decades, resulting from air pollution regulatory actions in developed countries, and also due to the declining economies of most Eastern European countries in the late 1980s [Streets et al., 2006]. These two causes may account for the measured increase in atmospheric transmission under clear-sky conditions since the 1990s [Wild et al., 2005]. Both the global dimming and the recent brightening, however, still involve a remarkable degree of uncertainty in their description. Alpert et al. [2005], for instance, suggested that the influence of human activities may indeed be significant at local, but not global, scales, and consequently the negative trend of solar radiation found in previous studies should be considered as local or regional dimming.

[4] Thus there is a need for more accurate measurements (with higher spatial resolution and not only measurements recorded exclusively in large urban areas) and to extend the analysis using longer series of records, as evidence of global dimming is thus far based upon few measurements of solar radiation [Stanhill, 2005]. For this purpose, the analysis should be supported and extended with the use of other climatic variables (proxy measures) such as evaporation, visibility, cloudiness, or sunshine duration, which are available for a longer time period and reach further back into the past. The advantage of sunshine duration, in relation to other variables, is that it is less subjective than visibility or cloudiness observations. Different studies have shown good correlations between this variable and solar radiation records [e.g., Iqbal, 1983; Almorox and Hontoria, 2004], the latter one taking into account some Spanish meteorological stations. In addition, few types of sunshine duration recording instruments have been used since the first measurements were made.

[5] Several studies have researched the spatial and temporal behavior of sunshine duration in recent decades, as measurements of this variable were initiated in the late 19th century. For example, Angell [1990] and Stanhill and Cohen [2005] analyzed the behavior of sunshine durations in the USA; Kaiser and Qian [2002] and Liang and Xia [2005] in China; Niu et al. [2004] in the Tibetan Plateau; Inoue and Matsumoto [2003] in Japan; Liu et al. [2002] in Taiwan; and Jones and Henderson-Sellers [1992] in Australia. Several studies involving Europe have also been published: Brázdil et al. [1994] for different areas of central and eastern Europe; Power [2003] for Germany; Moisselin and Canellas [2005] for France; Pallé and Butler [2001] for Ireland; Huth and Pokorná [2005] for the Czech Republic; Auer et al. [2007] for the Alps region; and Reinhard et al. [2005] for Switzerland.

[6] A systematic analysis of temporal evolutions of sunshine duration, based upon the large amount of data available for the Iberian Peninsula (IP), is still lacking. Thus the objective of our study is to set up and analyze a long and high-quality data set of sunshine duration records over the IP, in order to contribute to the knowledge of global dimming-brightening phenomena, their interactions with climate change, and their effects upon, for example, the hydrological cycle and solar energy resources. In section 2, we describe the original instrumental series used in this study. Section 3 explains the quality control and homogenization procedures applied to the series and describes the method used to generate a gridded version of the database. Section 4 provides information on the methodology used for clustering the station-mode and grid-mode data sets into subregions, which was based upon a principal component analysis. The evolution of annual, seasonal, and monthly sunshine duration series is then presented in section 5, where trend analysis is also discussed. Finally, in section 6, the discussion of our results is related to previously published works and section 7 presents the conclusions of this paper and their relevance for future studies.

2. Data

[7] Most of the series were obtained from the Spanish Instituto Nacional de Meteorología (INM) and corresponded to a number of stations across peninsular Spain, these series containing at least 30 a of data. The database was then completed with three Portuguese sunshine duration series from the Instituto de Meteorologia (IM, Portugal), and with the series for Gibraltar (Wheeler [2001]; updated by MetOffice) and Perpignan (southeastern France), provided by European Climate Assessment & Data Set (ECA&D) [Klein Tank et al., 2002]. In some cases there is more than one station in close proximity but with different periods of data which do not overlap. In these circumstances we generated a composite series. In total, 22 series of the final data set were created by assembling two or three subseries.

[8] It was generally not possible to obtain information on the instruments used in the different observatories; however, it is very likely that most of these were Campbell-Stokes heliographs. This instrument detects sunshine if the beam of solar radiation concentrated by a spherical lens is able to burn a special, dark paper card. Sunshine duration is consequently the amount of time, usually expressed to the nearest 0.1 hour, in which direct solar radiation is above a certain threshold.

[9] Taking into account all the data sources, the final data set consists of 72 series (Table 1) of sunshine duration records distributed over the whole IP (Figure 1). The longest series are San Fernando and Tortosa, starting in 1880 and 1910, respectively, but most series start after the 1930s. This network essentially covers the entire IP and the two main climate types: Oceanic (in the north-northwest), and Mediterranean, and the subtypes and variants, according to the Martín Vide and Olcina [2001] Spanish climate classification. Table 1 also shows the annual normal (1971–2000 mean) sunshine duration for each station. There are noteworthy differences between the stations in the north of the IP (sunshine duration less than 1800 h/a in several stations) and those in the southern area (some cases with sunshine duration greater than 2800 h/a).

Figure 1.

Location of the 72 sunshine duration stations (black dots) and of the generated grid points (boxes with indication of the subregion to which the different grid points belong to); a schematic regionalization based on PCA (bold lines) is indicated too.

Table 1. Details of the Stations and Sunshine Duration Series Considered
StationLong., degLat., degAlt., mLength% DataCompositeaAnnual Mean,b hSubregion
  • a

    Indication of composite series (Y) and the number of subseries used to obtain them.

  • b

    Mean total annual sunshine duration (in hours) of the homogenized values estimated over the 1971–2000 period.

A Coruña Airport−8.3843.30971971–200493.4N1904.0W
A Coruña City−8.4243.37581951–200491.4N2009.9W
Albacete−1.8638.957041951–200483.6N2747.7E
Alicante Airport−0.5638.29311967–200499.1N2752.4S
Alicante City−0.4938.37821939–200499.9N2799.0S
Almería−2.3936.84201945–200499.7Y (2)2931.2S
Badajoz−6.8338.881851951−2004100.0Y (2)2833.3W
Barcelona2.0841.3061951–200492.4N2485.7E
Bilbao−2.9143.30391952–200495.4N1708.4N
Burgos Airport−3.6342.368901951–200497.5N2193.4W
Burgos City−3.7042.348541951–198192.5N2263.0W
Cádiz−6.2636.5081955–200487.8N2992.0S
Calamocha−1.3040.938891951–200392.9Y (3)2384.4E
Castellón−0.0239.95351943–200498.3Y (2)2689.4E
Córdoba−4.8537.84911959–200481.5N2816.5S
Ciudad Real−3.9238.996271951–200499.2Y (2)2690.7S
Cuenca−2.1440.079561951–200481.3N2559.0E
Daroca−1.4141.117791967–200499.3N2518.0E
Faro−7.9737.0581970–200499.8N3193.5S
Getafe−3.7240.306171951–200494.4N2779.0E
Gibraltar−5.4036.20-1933–200499.8Y (2)2684.7S
Gijón−5.6443.5431938–200196.9Y (2)1741.6N
Girona2.7641.901271942–200497.6N2102.3E
Granada Airport−3.7837.195701972–200496.7N2933.2S
Granada Aerial base−3.6337.146851941–200496.6N2852.7S
Huelva−6.9137.28191951–200490.4Y (2)2972.1S
Huesca−0.3342.085411951–200488.3N2690.9E
Jerez de la Frontera−6.0636.75271973–200499.0N2901.6S
León−5.6542.599161951–200494.4N2631.4W
Lisboa−9.1338.781041970–200499.8Y (2)2783.8W
Lleida0.6041.631921951–200496.6Y (3)2630.6E
La Molina1.9442.3317041956–199891.3N2010.3E
Logroño−2.3342.453521951–200498.5N2248.2E
Madrid Airdrome−3.7940.386871951–200494.0N2778.9E
Madrid Airport−3.5440.455821951–200499.8N2852.9E
Madrid City−3.6840.416671935–200483.0N2646.3E
Málaga−4.4936.6771935–200496.9Y (2)2868.0S
Molina de Aragón−1.8740.8410631949–200497.0N2274.6E
Morón de la Frontera−5.6237.16871956–200474.5N2942.7S
Murcia-Alcantarilla−1.2337.96851961–200498.7N2900.3S
Murcia-San Javier−0.8037.7921951–200499.4N2884.0S
Navacerrada−4.0140.7818901963–200499.6N2236.0W
Ourense−7.8642.331431954–200483.3Y (2)2042.1W
Oviedo−5.8743.353361972–200497.2N1702.5N
Pamplona−1.6442.774521953–200494.4Y (2)2253.3N
Perpignan2.8742.74421931–200493.4N2284.5E
Pontevedra−8.6242.441071963–200490.5Y (2)2171.2W
Porto−8.6841.23701970–200498.6N2543.3W
Ranón−6.0343.561271968–200497.5N1688.9N
Reus1.1641.15731953–200497.0N2525.0E
Salamanca−5.5040.957901951–200499.8N2587.6W
Santander Airport−3.8243.4361972–200489.9N1630.0N
Santander City−3.8043.49521931–200499.2Y (2)1754.8N
Santiago de Compostela−8.4342.903641956–200498.1N1979.8W
Segovia−4.1340.9510051951–200489.8Y (3)2577.8W
Sevilla−5.9037.42261931–200498.0Y (2)2905.3S
San Fernando−6.2136.47301933–199498.4N2497.5S
Soria−2.4741.7710821951–2004100.0N2587.9E
San Sebastián Airport−1.7943.3681964–2004100.0N1724.9N
San Sebastián City−2.0443.312521939–200499.5N1761.0N
Tarifa−5.6036.10411953–199997.3N2856.8S
Teruel−1.1240.359001971–200482.1Y (2)2509.4E
Toledo−4.0539.885161951–200499.5Y (2)2857.8E
Tortosa0.4940.82481910–200498.2N2593.5E
Valencia Airport−0.4739.49571966–2004100.0N2799.0E
Valencia City−0.3839.48111938–200499.1N2665.5E
Valladolid Airport−4.8541.708461951–200496.8N2601.3W
Valladolid City−4.7741.657351941–200498.8Y (3)2615.7W
Vigo−8.6342.222551951–2004100.0Y (2)2259.1W
Vitoria−2.7242.885081949–200493.0Y (3)1890.8N
Zamora−5.7341.526561951–200494.8N2423.2W
Zaragoza−1.0141.662471951–200499.4N2653.0E

[10] Figure 2 shows the temporal evolution of data availability for the final database (only after 1930). There are evident improvements in data availability in 1951 and at the beginning of the 1970s. Thus although series start before 1931, the low density of stations and nonhomogeneous spatial distribution of the longest series led us to limit the analysis to the 1931–2004 period, but with greater confidence for the 1951–2004 subperiod, as in the latter over 50% of stations and 100% of grid points (see section 3.3) are available.

Figure 2.

Temporal evolution of data availability of the final sunshine duration database. Availability of both stations (solid line) and grid points (dashed line) is indicated.

3. Improvement of the Data Set and Grid Construction

3.1. Quality Control

[11] Although the sunshine duration series were subjected to preliminary quality checks by the meteorological offices that provided the data, we applied different quality control checks at a daily resolution, following the recommendations of Aguilar et al. [2003], in order to improve our sunshine duration database. In particular, we attempted to: (1) detect and correct/remove gross errors, such as aberrant (more hours registered than the maximum possible) or negative values; (2) check the consistency of calendar dates (days per annum or month); and (3) remove false zeros.

[12] Following these preliminary quality checks, the daily series were converted into monthly values of absolute sunshine duration, by summing daily sunshine duration values. When more than 6 d in a month were missing, we did not compute monthly sunshine duration, and the whole month was considered to be missing. When less than 6 d were missing, the total sunshine duration for the corresponding month was obtained by adequately correcting the sum of the available days using the number of missing values.

3.2. Homogenization

[13] In the last decades the scientific community has become aware of the fact that meteorological observations and climatic series are often affected by nonclimatic noises (no real climate signal) caused by different factors that can introduce abrupt or gradual breaks. These inhomogeneities can introduce critical errors if they are not considered prior to the data analysis. Thus significant improvements in the homogeneity procedures have been developed and proposed by climate research scientists over the last decades. For an extensive review, several works summarize the importance, methods, and future perspectives of the homogeneity question [e.g., Peterson et al., 1998; Aguilar et al., 2003].

[14] There are many examples of homogeneity procedures applied to different climatic variables, such as temperatures, precipitation, pressure, or radiosonde station data, but few studies as yet apply homogeneity procedures to sunshine duration series. The most recurrent justifications are that there have been no changes in instruments and that the different types of sunshine duration recorders used during the period analyzed have made no difference [Stanhill and Cohen, 2005]. Recent literature on this subject, however, demonstrates that sunshine duration series are also affected by a number of non negligible errors [Moisselin and Canellas, 2005; Auer et al., 2007]. Consequently, in these latter works, two relative homogenization approaches were applied in order to homogenize long-term sunshine duration series for France and the Alpine region, respectively.

[15] In relation to our paper, owing to the few metadata obtained for the sunshine duration series (only relocations of the meteorological stations), indirect methods were required to support the homogenization procedure. We chose to apply a relative homogeneity test, and discarded absolute tests which are more suitable for analyzing long-term series when no reference series are available in the vicinity.

[16] The homogenization of the data set was based on the procedure described by Brunetti et al. [2006a]. This procedure rejects the a priori existence of homogeneous reference series and consists of testing each series against other series, by means of a multiple application of the Craddock test [Craddock, 1979], in subgroups of preferentially 10 series. When a break is identified in one series (test series), the series used to estimate the adjustments are chosen among those series that prove to be homogeneous in a sufficiently large subperiod centered on the break and that correlate well with the test series. We decided to use several series to estimate the adjustments in order to ensure their stability and to prevent unidentified outliers in the reference series from producing bad corrections. The adjustments from each reference series were calculated on a monthly basis and were then fitted with a trigonometric function in order to smooth the noise and to extract only the physical signal (the adjustments often follow a yearly cycle). We then calculated the final set of monthly adjustments by averaging all the yearly cycles, excluding from the computation those stations whose set of adjustments presented an incoherent behavior pattern in relation to the others. When a clear yearly cycle was not evident, the adjustments used to correct the monthly data were chosen as constant through the year and were calculated as the weighted average among the monthly values, where the weights were the ratios between monthly mean sunshine duration and total annual sunshine duration.

[17] Only 18 of the 72 series (25% of the total) proved to be homogeneous according to the homogeneity test. Figure 3a shows the number of detected breaks per year. The increase in the number of corrections between the 1950s and the 1980s appears to follow the temporal evolution of the data availability. Indeed, if we take a look at Figure 3b, which shows the number of breaks per station, it will be seen that the number of detected breaks is less time-dependent, despite the greater variability in the first period (due to the lower data availability) and a decrease in the last decade. The total number of corrected breaks is 292, which, considering only the series that were homogenized, corresponds, on average, to 5.4 corrections per series.

Figure 3.

Number of detected breaks per year in the sunshine duration database (solid line, left axis) (a) in absolute values and (b) in relation to the number of available series. The number of available series is also plotted (dotted line, right axis).

[18] Figure 4 shows the mean adjustment curve averaged over all single series, the standard deviation, and the absolute range of the adjustments. The mean of all adjustments (the bold line) is characterized by values systematically lower than 1 before 1951 and systematically higher than 1 after this date, indicating that the inhomogeneities are not completely random. This behavior leads to a long-term trend of 2+0.64%/decade if evaluated over the 1931–2004 period. This positive trend is mainly due to few corrections (mostly negative) performed in some of the few series available starting from 1931; indeed, if we analyze the long-term behavior of the mean adjustments starting from 1951 (when the density of the stations is higher) the trend is −0.2%/decade. In any case, in the adjustments, a long-term signal persists, indicating that the use of original data in estimating long-term sunshine duration evolution can produce biased results. Moreover, the wide absolute range of adjustments (the dotted lines in Figure 4), along with the standard deviation range (thin lines in Figure 4), highlights the need to homogenize most series in order to obtain reasonable results. After the homogenization, all gaps between the first available year of each series and December 2004 were filled with estimates based upon the highest correlated reference series.

Figure 4.

Mean annual adjustment series obtained by calculating the yearly average ratios between the homogenized and the original sunshine duration series (bold line). Standard deviations (thin lines) and total correction ranges (dotted lines) are indicated too.

3.3. Gridding Sunshine Duration Data for the Iberian Peninsula

[19] In order to minimize possible errors resulting from eventually persistent inhomogeneities/outliers in some of the series and to prevent localized spots with high station density from dominating the climatic signal, we also generated a gridded version of the database. In addition, the gridded data allow for easier mathematical treatment in several types of data analysis, and improve future comparisons with other data sources or other variables (such as cloudiness).

[20] The grid has a 1° resolution both in latitude and longitude and was built with an improved version of the interpolation technique described by Brunetti et al. [2006a]. This improvement consists of the introduction of an angular weight. The total weight applied in the interpolation is the product of the radial and the angular terms. In short, the radial term takes the following form:

equation image

with c = −equation image where i runs along the stations, di (x, y) is the distance between station i and grid point (x, y). Parameter c was chosen to have weights of 0.5 for station distances equal to 2equation image from the grid point we wished to calculate. Here equation image is defined as the mean distance of one grid point from the next one obtained by increasing both longitude and latitude by one grid step.

[21] The angular term is that used by New et al. [2000]:

equation image

where θ(x,y) (i, l) is the angular separation of stations i and l with the vertex of the angle in grid point (x, y).

[22] Sunshine duration was calculated for each grid point that met one of the following conditions: (1) a minimum of two stations at a distance lower than equation image or (2) a minimum of one station at a distance lower than equation image/2. The grid value was then computed by considering all stations within a distance of 2equation image.

[23] The grid was calculated both in monthly, seasonal, and annual resolution, based upon monthly, seasonal, and annual single station anomaly series (obtained as ratio to the corresponding 1971–2000 normals). This prevents bias in the grid series as the number of available series tends to vary over time [Brunetti et al., 2006b] or owing to differences in the absolute mean values. The conversion of these anomalies into absolute values requires knowledge of the monthly normal at the grid point.

[24] The grid was constructed from 9°W to 3°E longitude and from 36° to 44°N latitude only over land. The final grid-mode data set consists of 84 grid points as shown in Figure 1. Figure 2 shows the time evolution of the grid-mode data set, together with that of the station-mode data set. Over 80% of the grid point series are available as from the 1940s, and 100% are available as from 1951.

4. Regionalization

[25] The regionalization was performed by means of a Principal Component Analysis (PCA), with the objective of clustering the records into subregions with similar sunshine duration variability. The analysis was applied both to the station-mode and to the grid-mode data sets, starting from the correlation matrix, and calculated considering all the 12 months of the year and using the series of monthly normalized anomalies (ratios to the 1971–2000 mean) [Preisendorfer, 1988]. We used all months of the year in order to obtain only one regionalization, and avoid discussing different subregions for different temporal resolutions. The analysis was performed on the 1971–2004 subperiod, for which the grid-mode data set is complete and the station-mode data set is almost complete.

[26] The results of the PCA applied to the stations series (Table 2) show that 7 Empirical Orthogonal Functions (EOF) account for more variance than the original variables (i.e., their eigenvalues are greater than 1) and explain more than 85% of the total variance of the data set. We selected the first four EOF, which have eigenvalues greater than 2 (more than 3% of explained variance). These selected EOF were rotated by means of a VARIMAX rotation to redistribute the variance into the components and to obtain stable and physically meaningful patterns [Von Storch, 1995]. In Figure 5 the geographical representation of the loadings obtained from the PCA is plotted (stations that have their maximum loading in the corresponding component are also indicated), enabling the following four subregions to be identified: (1) the central-east (E), a region with Mediterranean climate clearly influenced by the proximity of the Mediterranean Sea; (2) the north (N), that corresponds to the region of the IP with an Oceanic climate of midlatitudes; (3) the central-west (W), a region with a variety of climates (from Oceanic in the north to Mediterranean in the south) that are barely affected by the Mediterranean Sea; (4) the south (S), a region with a typical Mediterranean climate covering the southern sector of the IP. Each station was assigned to the region with maximum loading (Table 1).

Figure 5.

First four rotated EOF obtained from the PCA applied to the sunshine duration monthly series (normalized by the annual cycle). The stations with the maximum loading in the corresponding EOF are also plotted.

Table 2. Eigenvalues, Explained Variances, and Cumulative Explained Variances of Nonrotated and Rotated EOF Obtained From the PCA Applied to the Stations Series
EOFNonrotatedVarimax Rotated
EigenvalueVar %Tot Var %EigenvalueVar %Tot Var %
139.5554.9354.9319.6627.3027.30
211.0115.2970.229.2812.8940.19
35.317.3877.6013.6418.9559.14
42.393.3180.9115.6821.7780.91
51.612.2383.14   
61.331.8484.99   
71.101.5386.52   

[27] In order to check whether the PCA results were affected by the nonhomogeneous spatial distribution of the stations throughout the IP, we also applied the analysis to the gridded data set. The eigenvalues of the correlation matrix obtained revealed that in this case, only six EOF account for more variance than the original variables and that they explain more than 91% of the total variance. Once again in this case, however, only the first four components have eigenvalues greater than 2 and account for more than 3% of the total variance. These four VARIMAX rotated and normalized EOF were graphically represented (not shown here), providing evidence that the results of the PCA applied to the grid and to the stations are very similar and ensuring that the analysis was not influenced by the spatial distribution of the stations. Each point of the gridded data set was assigned to the component with maximum loading (as shown in Figure 1). Figure 1 also shows a schematic representation of the boundaries of subregions established by the PCA.

[28] Following regionalization, we computed the annual, seasonal, and monthly mean series for the four subregions and for the whole IP, starting from the grid-mode data set, by averaging all available data in each subregion (and in the whole IP). Use of mean series provides a more synthetic description of the climatic signal than the single station or the single grid point series and permits a higher signal-to-noise ratio, enabling better identification of long-term trends. The computation of regional series from the gridded data set, rather than stations, enables us to obtain series that are more representative of the different regions: regional series from the station-mode, despite being highly correlated with those from the grid-mode regional series, may be influenced by those areas with the highest station density. Thus further analysis was restricted to the subregional series obtained from the gridded data set.

5. Time Evolution and Trends of Sunshine Duration

5.1. Annual and Seasonal Results

[29] There is a high correlation (Table 3) between the subregional series and the IP one, indicating that the IP series captures a high portion of the sunshine duration variability. The correlation coefficients are higher than 0.92, on an annual basis, for all subregions but the N (0.78). On a seasonal basis the correlations are also very high, with common variance always greater than 50% (i.e., correlation coefficient greater than 0.71), the only exception being the N in winter. With regard to the correlations among the subregional series, the analysis highlights once more the peculiarity of the N series in relation to the other subregions; in fact, the common variance among the annual subregional series is in general greater than 50% with the exception of the N.

Table 3. Correlation Coefficients Among Subregional Sunshine Duration Series and Between Subregional and IP Series (1951–2004)a
 NWSIP
  • a

    Values significant at a level greater than 99% in bold; significant at a level greater than 95% in italics.

AnnualE0.590.780.860.92
N 0.690.570.78
W  0.800.93
S   0.92
WinterE0.540.800.910.94
N 0.660.380.69
W  0.750.94
S   0.89
SpringE0.790.850.860.96
N 0.790.620.86
W  0.770.94
S   0.89
SummerE0.440.660.740.86
N 0.670.280.74
W  0.660.91
S   0.79
AutumnE0.570.740.660.89
N 0.620.170.71
W  0.650.94
S   0.75

[30] Figure 6 shows the annual and seasonal IP series, together with their 11-a window 3-a σ Gaussian low-pass filter for a better visualization of long-term and decadal variability. In addition, the amplitude of the interannual variability is presented in Table 4; this interannual variability changes from one season to another, with the highest and lowest variability in winter and summer, respectively. Finally, the overall trends of the series, calculated over the most confident period (1951–2004) by means of least-square linear fitting, are shown in Table 5 together with their significance estimated by means of the Mann-Kendall nonparametric test [Sneyers, 1992].

Figure 6.

Average IP sunshine duration series (thin line), plotted together with the 11-a window 3-a σ Gaussian low-pass filter (thick line), showing (a) annual; (b) winter; (c) spring; (d) summer; (e) autumn. The series are expressed as relative deviations from the 1971–2000 mean. Dashed lines are used prior to 1951 owing to the lower number of records for this initial period.

Table 4. Interannual Variability of the IP and Subregional Series (1951–2004) Expressed by Their Standard Deviation (Y, Annual; Wi, Winter; Sp, Spring; Su, Summer; Au, Autumn)
 YWiSpSuAu
IP0.0400.1260.0860.0480.072
E0.0430.1150.0890.0490.075
N0.0550.1490.1220.0910.105
W0.0460.1710.0990.0550.099
S0.0410.1300.0790.0470.067
Table 5. Sunshine Duration Trends Over the 1951–2004 Perioda
 IPENWS
  • a

    Values are expressed in percentage per decade. Bold values indicate trends with significance level higher than 99%; italic values indicate trends with significance level higher than 95%; other values with significance level greater than 90% are also indicated; for lower levels of significance only the sign of the trend is indicated.

Year-+-1.0 ± 0.4−0.6 ± 0.4
Winter+++--
Spring++--+
Summer0.9 ± 0.4--−1.2 ± 0.5−1.1 ± 0.4
Autumn---1.9 ± 0.8-
January+++++
February+++-+
March+3.0 ± 1.4+2.8 ± 1.3++3.8 ± 1.8+3.0 ± 1.3
April-+--+
May2.2 ± 1.1−1.7 ± 1.1-2.6 ± 1.32.4 ± 1.0
June++-++
July−1.3 ± 0.5--−1.9 ± 0.6−1.4 ± 0.5
August−1.6 ± 0.5−1.3 ± 0.5-−1.6 ± 0.5−2.1 ± .0.5
September-++-1.3 ± 0.7
October---3.5 ± 1.6-
November-+--+
December--+--

[31] The IP annual sunshine duration series (Figure 6a) starts with an initial decade without relevant variations, followed by an increasing tendency in the 1940s and reaching the absolute maximum at the beginning of the 1950s. Subsequently, annual IP sunshine duration shows a negative trend lasting some decades up to the absolute minimum of the series at the beginning of the 1980s. The final decades are characterized by a new increase in sunshine duration reaching a relative maximum at the beginning of 2000s, followed by lower values in the final years, which are insufficient with regard to establishing whether a new decreasing shape is starting. The trend, estimated over the 1951–2004 period, is slightly negative, but not significant.

[32] In the winter seasonal series (Figure 6b), the long-lasting decrease from the 1950s to the 1980s is interrupted by a relative maximum at the beginning of the 1970s; moreover the 1980s minimum is anticipated at the end of the 1970s, due to three winters with very low sunshine duration from 1977 to 1979. Spring and summer seasonal series (Figures 6c and 6d) are more similar to the annual one, the most relevant differences being the larger maximum at the beginning of the 1950s which lasted up to the end of the 1960s. Moreover, for the spring series, the recent maximum is more important and anticipated to the end of 1990s, followed by a more evident decrease in the final years of the series. For the summer series, the two lowest values reached in 1982 and 1983 are quite noteworthy. On the other hand, in summer the recent increase lasts until the end of the series and does not compensate the previous decrease. Finally, the autumn seasonal series (Figure 6e) shows a decadal tendency which is very similar to the winter one, as in this case too a relative maximum at the beginning of the 1970s interrupts the 1950s–1980s decrease, although in autumn the 1980s minimum is shifted forward at the beginning of 1990s. Regarding long-term trends of seasonal series, the only significant value is the summer negative trend (−0.9% per decade; see Table 5).

[33] Figure 7 shows the annual and seasonal sunshine duration series for the four subregions. As in the whole IP series, the variability (quantified in Table 4) is maximum in winter and minimum in summer. Besides the differences from season to season, some differences in the interannual variability of the different subregions are evident too, with the highest and lowest variability observed in the N and S subregions, respectively. As for decadal variability, the four subregions present a pattern similar to that of the IP series, with some peculiarities that differentiate one subregion from another. The most relevant differences on an annual basis concerns the N subregion, where sunshine duration in the 1930s is very low, with a strong increase in the 1940s, before reaching its maximum at the beginning of 1950s. This shape is also evident, although less pronounced, in the W and S subregions, while in the E, decadal variability is very similar to that of the IP. It should be noted, however, that in the 1930s and 1940s the data availability is lower and that decadal variability might be more sensitive to single station behavior. On the other hand, the E subregion differs from the IP for its highest maximum in the 2000s, whereas the other subregions show the highest maximum in the 1950s as in the IP series.

Figure 7.

As in Figure 6 but for the four subregions (Y, annual; Wi, winter; Sp, spring; Su, summer; Au, autumn).

[34] The seasonal analysis of the subregional series reveals that the N is again the subregion with the most relevant differences in relation to IP decadal variability. Besides the very low values in the 1930s and the sharp increase in the 1940s evident in all seasons, there is a remarkable difference in the 1980s, where the winter N series is characterized by a sharp increase in sunshine duration which culminated around 1990, when the absolute maximum was reached. Another remarkable difference with regard to the IP series concerns the high 1930s and 1940s winter values for the S series. On the other hand, while the E, W, and S subregional summer series show the same characteristic shape for the IP, the N is characterized by a more or less continuous decrease from the 1950s to the end of the series, with a very brief interruption in the second half of the 1980s.

[35] Finally, the trend analysis applied to the subregional series (Table 5), over the 1951–2004 period, highlights a dominant decrease in sunshine duration, with only few exceptions, when positive but not significant trends were detected. The most significant trends concern the W and S subregions, with decreases of −1.0% and −0.6% per decade, respectively, on an annual basis, mainly due to the summer season (−1.2% and −1.1% per decade, respectively) and, only for the W, due to autumn (−1.9% per decade).

5.2. Running Trend Analysis

[36] Significance and slope of the trends strictly depend on the selected period (1951–2004 in the previous section). Thus in order to provide results that are more comparable with other studies in the future and to extract as much information as possible from the data, a trend analysis was also applied on running windows of variable width [Brunetti et al., 2006b] for the annual and seasonal series and for the IP and subregional series. The slopes of the trends were estimated within temporal windows of widths ranging from 20 a up to the entire series length. The trends obtained were plotted for better visualization in graphs where the y axis represents the window width, and the x axis the central year of the window the trend refers to. The value of the trend is represented by the color of the corresponding pixel. Only trends with a significance level (calculated by Mann-Kendall nonparametric test) greater than 90% are plotted. These figures capture the whole possible spectrum of significant trends present in the series, thus providing the best possible detailed and quantitative description of the peculiarities observed in Figures 6 and 7, and giving evidence of which features are most important in terms of trends at decadal and longer timescales.

[37] Figure 8 shows the results of the running trend analysis applied to the IP series. On an annual basis, the most interesting feature highlighted by this analysis is the negative trend that characterizes almost all the timescales greater than 40 a, and the positive-negative-positive sequence of significant trends at shorter timescales. Positive trends for time windows of 20–30 a are centered on the 1940s and the second half of the 1980s. To the contrary, negative trends are centered on the 1960–1980 period and reach significant values even for time windows as long as 65 a, as discussed above. The absence of significant trends around 1950 and 1980 is due to the presence of maximum and minimum values in the series, respectively. On a seasonal basis, spring and summer trend patterns are similar to the annual ones (with the same positive-negative-positive sequence), with the main differences that summer also shows significant negative trends for timescales greater than 65 a (up to the whole series length 1931–2004); on the contrary, spring shows significant trends only for timescales of less than 50 a, and the positive trends centered on the 1980s are stronger than in the annual and summer series. Winter and autumn trends differ more from the annual ones; these seasons are characterized by few significant trends of opposite signs.

Figure 8.

Running trend analysis for the IP annual (top) and seasonal (lower graphics, from left to right and top to bottom: winter, spring, summer, autumn) series. The y axis represents window width, and the x axis represents the central year of the window over which the trend is calculated. Only trends with significance level greater than 90% are plotted.

[38] On a subregional basis the most interesting differences from the IP results concern the N and S subregions (Figures 9 and 10), while the E and W (not shown here) have similar trend patterns to the IP one. For the annual series, the N shows stronger positive trend phases, centered on the 1940s and the late 1980s, for timescales of up to about 40 a, and negative trends only over time periods of less than 60 a centered on the 1960s and 1970s. On the contrary, the S shows positive trends only for time windows of less than 30–35 a but negative trends over all time windows centered on the 1960s and 1970s, the negative trend over the whole 1931–2004 period also being significant. This is mainly due to the increases before the 1950s maximum and after the 1980s minimum, which are steeper in the N than in the S (see Figure 7).

Figure 9.

As in Figure 8 but for subregion N.

Figure 10.

As in Figure 8 but for subregion S.

[39] However, the biggest differences from the IP results concern the seasonal analysis, and winter in particular. In this season, the S shows the same shape as the IP, but with a strongly emphasized negative trend phase that involves all subseries starting in the 1950s maximum and with trends up to values of less than −12% per decade, while the N shows a completely different behavior pattern, with almost exclusively positive trends and only for subseries starting in the 1930s or ending in the 2000s. Other important differences from the IP results concern only the N subregion: the summer season in particular is characterized by the absence of positive trends at the beginning and at the end of the series, while the spring season shows trends similar to the IP ones, but stronger in magnitude, with increases greater then +12% per decade and negative trends up to −10% per decade.

5.3. Overview of Trends for the Monthly Series

[40] The monthly series trends over the 1951–2004 period were also studied, in order to better understand the spatial and temporal behavior of sunshine duration on the IP. This detailed analysis has shown important differences with respect to the seasonal and annual analysis (see Table 5). For the whole IP, the January, February, March, and June trends are positive, while negative values were obtained for the other months. However, there are only 4 months with significant trends. The most interesting month is March, with an important positive trend leading to an increase in sunshine duration of +3.0% per decade (i.e., more than +15% for the period analyzed). This trend is mainly due to an increase in sunshine duration from the 1960s to the end of the 20th century (figure not shown), an evolution which is quite different from that found for the spring series (see Figure 6c). To the contrary, the other three significant trends are negative: −2.2%, −1.3%, and −1.6% per decade, for May, July, and August, respectively. These four significant trends are mainly due to the E, S, and W subregions. July and August negative trends are greater in the W and S subregions than in subregion E. The former subregions also show negative trends in some autumn months (September in subregion S and October in subregion W). There are no significant trends at monthly resolution in subregion N, confirming once again its different behavior.

6. Discussion

[41] From the analyses of the temporal evolution of sunshine duration over the IP, we can confirm that the so-called global dimming (from the 1950s to the 1980s) observed worldwide [Stanhill and Cohen, 2001] and the successive brightening [Wild et al., 2005; Pinker et al., 2005] also affected the IP. Indeed, throughout the whole IP, and also in the four subregions derived from the PCA, sunshine duration declines from the 1950s to the 1980s and then partially recovers at least until the end of the 20th century. When seasonal series are analyzed, spring and summer show the best correspondence with the annual series, which is a feature already found for other parts of Europe [Brázdil et al., 1994]. It should be mentioned, however, that few previous studies analyze sunshine duration (or solar radiation) at seasonal or monthly resolution.

[42] The only significant trend found for the whole IP for the overall 1951–2004 period is the one for summer (−0.9% per decade). This decreasing trend, converted to radiation units (mean daily irradiance) with the use of the relationship suggested by Almorox and Hontoria [2004] based upon an adjustment of the classical Angström-Prescott formula with Spanish data, corresponds to −1.4 Wm−2 per decade, which corresponds to −7.5 Wm−2 for the whole 54 a period. More interestingly, from the running trend analysis the decline and subsequent increase in mean daily irradiance, associated with the dimming and brightening periods, can be estimated. Specifically, for the whole IP, annual basis, and in the 1950–1980 period, the sunshine duration linear trend is about −3% per decade, which means −9.2 Wm−2 for the 30 a, whereas in the 1980–2000 period the trend of about +4% per decade accounts for a recovery of +8.1 Wm−2.

[43] Recent emission inventories of anthropogenic pollutants suggest that both scattering sulfur and absorbing black carbon aerosol showed large changes in accordance with surface solar radiation, with decreasing tendencies since the 1980s, after decades of increase, owing to effective air pollution regulation [Streets et al., 2006]. The increasing aerosols can enhance the backscatter and absorption of incoming solar radiation and weaken the direct solar radiation needed to activate the Campbell-Stokes sunshine recorder and therefore reduce the sunshine duration on any given day [Kaiser and Qian, 2002]. Unpublished results of a preliminary analysis performed by means of a radiative transfer model (J.A. González, personal communication, 2007) indicate that a change of 10% in aerosol optical thickness (AOT) can change daily sunshine duration by 5 min, owing to the effect of aerosols on the solar beam just after (before) sunrise (sunset). This involves, for a mean daily sunshine duration of 6.8 h on the IP, a relative change of 1.2%. The AOT changes needed to explain the trends in sunshine duration on the IP should be quite notable, but unfortunately few studies analyze decadal changes in AOT: for the global oceanic areas a decrease of 20% has been reported between 1991 and 2005 [Mishchenko et al., 2007], while higher decreases appear to arise from other regionalized studies for the same period [Geogdzhayev et al., 2005]. Thus we cannot rule out the possibility that even more important changes had been taking place over some land areas, such as the IP. Nevertheless, we must not forget the role of clouds, which may have an even more significant impact on solar radiation and sunshine duration. Cloudiness trends on the IP are studied elsewhere [Calbó and Sanchez-Lorenzo, 2006], but the conclusions are not definitive, since results differ depending on the data source used.

[44] Another remarkable feature of the temporal evolution of sunshine duration is the absolute minimum reached in 1983, with additional low values in 1982 and 1984, both in the annual series and, in particular, in the summer series. The cause of these negative anomalies of sunshine duration on the IP might be the El Chichón (Mexico) volcanic eruption (April 1982). It is well known that large explosive eruptions can inject aerosols into the lower stratosphere, thus reducing incoming solar radiation, and producing stratospheric warming and cooling at the surface for 2 to 3 a following the volcanic eruption [Robock, 2000]. This effect has also been detected previously using sunshine duration records [Lamb, 1977]. In spite of this global impact, there are seasonal and regional differences in the response of the climate system. For example, for the western part of Europe, circulation after tropical volcanic eruptions shows a strengthening of the Azores high that might be associated in winter with a more persistent positive phase of the North Atlantic Oscillation (NAO) pattern; and in summer the circulation pattern suggests a more frequent southwesterly flow over the British Isles which guarantees the penetration of more frequent cyclones with frontal systems [Prohom et al., 2003]. In addition, Olmo and Alados-Arboledas [1995] found, with the use of radiometric records, that a few weeks after the Pinatubo eruption (June 1991), diffuse solar radiation increased by up to 43% while beam and global radiation decreased by 10% and 4% respectively, at a site in the southeast of the IP. It should be noted that in 1992, sunshine duration on the IP also reached relative minimum values, both at annual resolution and in summer. Another possible explanation for the early 1980s minimum of sunshine duration could be found in some influence associated with the strong El Niño event of 1982–1983. Indeed, Mariotti et al. [2002] showed some relationships between the El Niño phenomenon and precipitation in Europe. Moreover, other El Niño events in recent decades (1972, 1977, 1993, 1997, 2002) show a good degree of correspondence with relative sunshine duration minima for the IP series, at annual resolution.

[45] We should also highlight the possible relationship between solar radiation at the Earth's surface (or sunshine duration, which is a surrogate for solar radiation) and temperature in the lower atmosphere. For example, it is easy to establish a parallel between the evolution of temperatures during the 20th century at global scale (with its well-known pattern involving a rise in the first decades, a light decrease in the 1940s–1970s period, and a sharp increase from then on) and the above mentioned global dimming and brightening phenomena. In this sense, a recent paper [Wild et al., 2007] suggests that solar dimming was effective in masking greenhouse warming, but only up to the 1980s, when dimming gradually turned into brightening, thus allowing the uncovered greenhouse effect to reveal its full dimension. Specifically for the IP, Brunet et al. [2007] also found two periods of rising (1901–1949 and 1973–2005) and one of falling (1950–1972) temperatures during the 20th century, the most significant being the unprecedented and sharp rise in temperatures observed in recent decades, mainly associated with spring and summer seasons. These results match quite well our findings in the present paper, except for the date of the change in the tendency (1973 for temperatures and around 1980 for sunshine duration). Note that rising temperatures in spring and summer also match the recent clear brightening found here. Therefore we can speculate that a relation between the trends of both climate variables exists and that global dimming may be partially responsible for the period of cooling registered on the IP during the 1950s, 1960s, and 1970s. The plausibility of this relationship is reinforced by the fact that the temperature trends found by Brunet et al. [2007] are more marked for daily maximum temperatures, which logically are more influenced by solar radiation [Wild et al., 2007].

[46] Moreover, since sunshine duration should be quite well related to cloudiness, and the latter with precipitation, it is interesting to note that the significant positive trend found in March for sunshine duration clearly tallies with the decrease in precipitation detected by Paredes et al. [2006] from 1960 to 1997 in the central and western areas of the IP. To the contrary, the clear negative trends of sunshine duration in May, July, and August appear to constitute a more original result of the present paper, as the corresponding increase in precipitation was not described as significant in the same study [Paredes et al., 2006].

[47] Finally, we should mention the big differences found between subregion N and the other areas of the IP. Most characteristics analyzed (normal sunshine duration values, interannual variability, overall trends for the 1951–2004 period, and running trends) show different behavior patterns for subregion N. As we have already indicated, the reason for these differences is quite simple: this region corresponds to the area of the IP that is not under the influence of a Mediterranean climate. This is confirmed by the fact that, in general, the most important differences are found with subregion S, which is the one with the most Mediterranean climate. Note that similar differences arise when analyzing other climatic variables, such as temperature [Brunet et al., 2007] and precipitation [Rodriguez-Puebla et al., 1998].

7. Conclusions and Future Work

[48] In this paper we have presented the construction of a new data set of sunshine duration over the Iberian Peninsula, based on 72 monthly instrumental series covering the 1931–2004 period. Special emphasis has been placed on the quality control and homogenization procedures, which are both necessary to avoid biased results when estimating long-term sunshine duration evolution from the original data. These aspects are particularly interesting, since few previous works deal with sunshine duration data. Then, a gridded version of the homogenized data set was produced. The grid has a 1° resolution, both in latitude and longitude, and covers the entire IP with 84 cells. The database produced has been named Sunshine Duration Over the Iberian Peninsula (SUNDUIB).

[49] Subsequently, we analyzed the sunshine duration data, in search of temporal variability, long-term trends, and spatial differences over the IP. The analyses were carried out on the mean series calculated over the whole IP (i.e., by using data from all grid points) and also for four subregions identified by PCA. These four subregions obtained are geographically consistent and easily interpretable in terms of different climate types (Mediterranean, more or less influenced by the Mediterranean Sea, or Oceanic). The mean series were computed and analyzed at annual, seasonal, and monthly resolution.

[50] The temporal evolution of the mean annual series for the whole IP showed a decrease in sunshine duration from the 1950s until the early 1980s, followed by an increase until the end of the 20th century. This evolution, also confirmed by a running trend analysis, is quite similar to what has been described in other parts of the World as a “global dimming” and the subsequent and recent “brightening.” The mean series for spring and summer showed similar behavior, while in autumn and winter, sunshine duration was quite constant during the period analyzed. The clearest signal for the IP series concerns the summer season decrease (−0.9% per decade), which is mostly due to two subregions (W and S, i.e., the Mediterranean climate regions with a weaker Mediterranean Sea influence). Moreover, the most significant contributions to these trends originate in remarkable and significant trends in July and August in the two subregions (ranging from −1.4 to −2.1% per decade). On the other hand, important trends obtained for the March series (for all subregions except the Oceanic climate subregion, N) are also noteworthy: they range from +2.8 to +3.8% per decade and produce a trend for the whole IP of +3.0% per decade. These latter trends match the declining precipitation found by other authors for the IP.

[51] All the above analyses establish a methodology that can be helpful when studying sunshine duration data in other parts of the globe. In addition, our results are useful to better describe recent climatic changes over the IP. It should be noted that with temperature and precipitation, solar radiation (or sunshine duration) has a major impact on socioeconomic activities of principal importance in Spain and Portugal: agriculture and especially tourism. In addition, our results (along with an appropriate formula for extrapolating these to radiation) may be valuable for studies dealing with the solar energy resource, the hydrological cycle (e.g., evapotranspiration), etc.

[52] In order to overcome some uncertainties derived from the use of sunshine duration records, our future work will deal with (1) the relationship between sunshine duration and cloudiness data, using both conventional cloud observations and satellite measurements; (2) further research into the quantitative relationship between sunshine duration and solar radiation in our study area; (3) assessment of the influence of low-frequency circulation patterns such as the NAO or El Niño/Southern Oscillation (ENSO) on sunshine duration variability; and (4) extending the series as far as possible into the past (i.e., the beginning of the 20th century).

Acknowledgments

[53] This research was supported by the Spanish Ministry of Education and Science (MEC) project NUCLIER (CGL2004-02325). Sanchez-Lorenzo was granted an FPU predoctoral scholarship by the MEC and developed part of this work while performing research at the Institute of Atmospheric Sciences and Climate, Italian National Research Council (ISAC-CNR, Bologna). We would very much like to thank Teresa Nanni (ISAC) for facilitating this research stay. The sunshine duration series were provided by the Instituto Nacional de Meteorología (Spain), Instituto de Meteorologia (Portugal), the European Climate Assessment & Data Set Project, and the MetOffice (United Kingdom). We also thank Dennis Wheeler and Germán Solé for providing part of Gibraltar and Tortosa series, respectively. Josep-Abel González (Universitat de Girona) kindly helped us in discussing the effects of aerosols on sunshine duration. Finally, we would like to thank the useful comments of the three anonymous reviewers.

Ancillary