Trends in seasonal indices of daily temperature extremes in the Iberian Peninsula, 1929–2005

Authors


Abstract

Seasonal changes of extreme daily temperature indices (frequency of frost days, tropical nights, and ‘summer’ days) of 25 stations in the Iberian Peninsula (IP) from 1929 to 2005 are investigated. Cluster analysis is used to identify homogeneous groups of stations with similar thermal regimes. The Mann-Kendall test is applied to look for statistically significant trends. The results show a significant decreasing trend (−0.6 days/decade), especially from 1965 onwards, in the frequency of frost days for coastal stations, and significant increasing trends, from 1980 onwards, especially important to the southeast of the IP (+3.8 days/decade) for summer tropical nights, and in the southern stations (+2.3 days/decade) for summer days. The behaviour of the winter frost days index is related to the evolution of winter NAO and EA patterns (with increasing trends in the last decades of the 20th century). Although multiple regression models establish a relationship between summer indices and the SCAN pattern, the behaviour of residuals indicates that summer indices must be explained by other causal mechanisms. Copyright © 2011 Royal Meteorological Society

1. Introduction

Global surface mean temperature has increased by 0.7 °C since the late 19th century (Trenberth et al., 2007). The observed rise in global and regional average temperature during the last century (particularly in the last 50 years) has fuelled widespread interest on studies on climate change. Climatic change is manifested not only by changes of mean values, but also by changes in the occurrence of extremes. Observed changes in the occurrences of extreme temperatures are generally consistent with a global warming (Trenberth et al., 2007), although different patterns and intensities have been found in different regions of the world (Manton et al., 2001; DeGaetano and Allen, 2002; Vincent et al., 2005; Klein Tank et al., 2006; Vincent and Mekis, 2006). In Europe, numerous works in the last decade (e.g. Morberg and Jones, 2005; Moberg et al., 2006) have reported significant changes in extreme temperatures during the 20th century, more so in the second half of the century and in the upper tail of minimum temperature (Tmin) in summer. In the Iberian Peninsula (IP), from the mid-70s, detected trends towards warming are quite larger than on the global scale. In Spain, average series of maximum temperature (Tmax) and Tmin followed increasing rates of 0.51 versus 0.47 °C/decade (Brunet et al., 2007a), while for Portugal, estimated trends are of 0.49 versus 0.54 °C (Ramos et al., 2011). The two above cited works highlight higher warming trends for spring and summer in this recent period, increasing the (intra)annual temperature range.

The IP is situated in an area of climatic transition between temperate and subtropical latitudes. The existence of semi-desert, Mediterranean, Atlantic, and mountainous environments gives rise to pronounced spatial variations in temperature and precipitation, which are also compounded by a large temporal variability. Spatial distribution of temperatures in the IP is strongly influenced by orography, latitude, and distance from the sea. Continental sites usually register larger variability, with high (low) temperatures in summer (winter). Although less frequently, under certain atmospheric conditions, also coastal sites can experience very high temperatures, e.g. in the presence of Föhen effect (Font Tullot, 2000), as well as very cold and frost events, e.g. in the Mediterranean coast in episodes of strong advection of cold polar air masses, such as the phenomenon known as helada negra (‘black frost’). Fluctuations in climate parameters are of particular interest for semiarid regions of high natural variability, like parts of the Mediterranean, e.g. the southeastern IP. Extremes can lead to serious consequences for multiple sectors, such as agriculture, natural environment, and human health. Therefore, a comprehensive knowledge of such variations is necessary. The occurrence of extreme temperatures above/below certain thresholds which have impacts on agriculture and ecosystems are evaluated in this paper.

There are various methods of defining and characterising extreme events, for instance, by analysing the statistical behaviour of the tail of a weather element's probability distribution (Sillmann and Roeckner, 2008), or by means of indices; a list of indices defined by the Expert Team on Climate Change Detection Indices (ETCCDMI), includes percentile, threshold, absolute and duration-based indices (Alexander et al., 2006). Percentile-based indices, such us the occurrence of warm nights, warm days, cold nights, and cold days (Alexander et al., 2006), have been studied in the IP by Brunet et al. (2007a) at 22 Spanish stations during 1850–2005 and by Ramos et al. (2011) at 23 Portuguese stations during 1941–2006, on seasonal and annual basis, estimating significant decreasing (increasing) trends in the annual number of cold days and nights (warm days and nights). Rodríguez-Puebla et al. (2010) reported significant trends of about 1.6%/decade in the annual frequencies of warm days and − 1%/decade in cold nights for the period 1950–2006; Prieto et al. (2004) studied the behaviour of minimum winter extreme temperatures in 45 observatories across Peninsular Spain for the period 1955–1998, finding a generalized decreasing trend in the frequency of extreme cold days (Tmin < 5th percentile) between November and March.

Absolute indices represent maximum or minimum values within a year. Previous works (Serra et al., 2001; Burgueño et al., 2002; del Río et al., 2007) have found increasing trends of these variables in different areas of the IP during the 20th century, up to 0.7 °C per decade (Furió and Meneu, 2011). Duration indices such as cold or warm spell duration indices (CSDI and WSDI, respectively) were studied in the IP by Brunet et al. (2007b), who found that the reduction of occurrences of CSDI was a more dominant feature. Ramos et al. (2011) studied these annual indices for Portugal, and, by contrast, the results highlight more widespread trends for the recent increase in WSDI (especially in the period 1976–2006, significant at 9 out of 23 stations with positive trends) than in the reduction of CSDI (significant at 9 out 17 stations with negative trends during 1941–2006).

Absolute threshold indices are defined as the number of days in which a maximum or minimum temperature falls above or below a fixed threshold chosen as a relevant value in terms of major impacts on society, ecosystems or agriculture. The occurrence of frost days (FD), summer days (SU) and tropical nights (TR) are among these indices. This type of index has received less attention in the literature on extremes in the IP, with the exception of the recent works by Cuxart and Guijarro (2010), and Ramos et al. (2011), focused, respectively, on the island of Majorca (Balearic archipelago), and Portugal. In Majorca, Cuxart and Guijarro (2010) found a significant trend to diminishing the number ofFD. In Portugal Ramos et al. (2011) found similar results for FD, and Miranda et al. (2002) reported an increase of TR and SU indices from the 1970s to the end of the 20th century. A recent work by Sanchez-Lorenzo et al. (2011) focuses only on TR index, using 17 stations for the period 1961–2007 (summer season) and the relation with atmospheric patterns by means of Canonical Correlation Analysis. These authors find a significant increase of TR in the last decades of the record, more notable since the late 1980s in the east and central parts of the IP, although with some regional differences.

The first objective of this work is to complete the analysis of extreme temperature indices in the IP, studying absolute threshold indices for a set of 25 stations during the common period 1929–2005. These indices are studied on a seasonal time scale, due to the strong seasonality of climate in the IP.

Frost days index is defined as the number of days in which minimum temperature is below 0 °C. The occurrence of frosts has important impacts on agriculture, depending on the season of the year and the phenological characteristics of plants, in particular, for the IP, on grapes, cereal, and orange trees. Tropical nights index is defined as the number of days in which minimum temperature is above 20 °C. This index provides information about possible heat stress for organisms, an increase of frequency of TR could thus be an indicator for more heat stress (Sillmann and Roeckner, 2008). Another interesting index, the SU index, is defined as number of days with maximum temperature above 25 °C, and provides information on the behaviour of maximum temperatures, with impacts on droughts, forest fires, energy demand, and human comfort and health, among others (García-Herrera et al., 2005; Trigo et al., 2006, 2009).

The objective of this paper is to analyse the spatio-temporal variability of FD, TR, and SU indices in the IP, and to look for possible relationships with the main teleconnection indices affecting the study area. The data base (daily maximum and minimum temperatures of 25 stations in the IP during the period 1929–2005), and methods used to analyse the spatio-temporal variability of the indices are described in Section 2. Results are discussed in Section 3, and, finally, conclusions and challenges for future research are discussed in Section 4.

2. Data and methods

Data from the database Spanish Daily Adjusted Temperature Series (SDATS), corresponding to 23 Spanish stations, were used. This data base includes the longest and most reliable daily record of maximum and minimum temperature in Spain (Brunet et al., 2006, 2007a). Data of Lisbon and Perpignan from the dataset of the European Climate Assessment and Dataset (ECAD) have been added. All the data were provided by the Climate Research Group of the University Rovira i Virgili (Tarragona, Spain). Figure 1 shows the location of the 25 stations, and Table I the geographical data (altitude above sea level, latitude, longitude) corresponding to each one. All records have precision of 0.1 °C, as recommended by the WMO on required observation accuracy and reporting resolution for surface air temperature (WMO, 1996). This precision involves less bias in the calculation of indices above thresholds than data of lower precision (Zhang et al., 2009).Therefore, one can expect that problems related to low data precision do not significantly affect trend estimation and detection in the index series. This network covers the whole IP reasonably well, encompassing the main Iberian climate domains, with 13 stations covering the northern half of the peninsula (above 40°N), and the other 12 stations in the southern half. Spatial coverage is limited by the need to use long temperature records (with the aim of assessing more robust trends), and more long-record stations are desirable to cover the north and western areas of the IP. The study period is 1929–2005, the longest and common period for the set of stations.

Figure 1.

Map of the study area (station codes in Table I). This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Table I. Data series in the Iberian Peninsula for the period 1929–2005
Station (CODE)Altitude (m a.s.l.)LatitudeLongitude
Albacete (Alb)69938°56′N01°51′W
Alicante (Ali)8238°22′N00°29′W
Badajoz (Bad)19538°53′N06°48′W
Barcelona (Bar)9441°25′N02°07′E
Burgos (Bur)85442°22′N03°38′W
Cádiz (cad)3036°45′N06°20′W
Ciudad Real (CRe)62938°59′N03°55′W
Coruña (Cor)6743°22′N08°25′W
Granada (Gra)68537°08′N03°37′W
Huelva (Huel)2637°15′N06°56′W
Huesca (Hues)54242°05′N00°19′W
Lisbon (Lis)7738°72′N09°15′W
Madrid (Mad)66740°24′N03°40′W
Málaga (Mal)736°40′N04°29′W
Murcia (Mur)7537°57′N01°13′W
Pamplona (Pam)44242°49′N01°38′W
Perpignan (Per)43042°73′N02°87′E
Salamanca (Sal)78240°56′N05°29′W
S.Sebastián (SSeb)25943°18′N02°02′W
Sevilla (Sev)3137°25′N05°53′W
Soria (Sor)108041°46′N02°28′W
Tortosa (Tor)4940°49′N00°29′E
Valencia (Val)1139°28′N00°22′W
Valladolid (Vall)73541°46′N04°46′W
Zaragoza (Zar)23341°39′N01°00′W

All the details about quality control and homogenisation of the Spanish series are explained in the papers by Brunet et al. (2006, 2007a). The two series from the ECAD (Klok and Klein Tank, 2009) were similarly checked and homogenized. Data were analysed on a seasonal basis. Seasons were considered as follows: winter: December to February; spring: March to May; summer: June to August; and autumn: September to November. For each station it was considered that a season was complete if the number of missing data was less than 10 days. In general terms, although with slight differences from one station to another, the percentage of gaps was of the order of 5 and 6% in all the seasons. The main gaps in the Spanish series correspond to the period 1936–1939 (Spanish Civil War), and to 1944–1945 (World War) for Perpignan. Lisbon minimum temperature series are available until December 2003.

FD index is defined as the number of days with minimum temperature below 0 °C. TR index is usually defined as the number of days with minimum temperature above 20 °C, and SU index as the number of days with maximum temperature above 25 °C. In this work, two threshold values have been used for TR (20 and 25 °C) and SU (25 and 30 °C) indices, defining TR20, TR25, SU25, and SU30 indices in a similar way to other works (for instance, Choi et al., 2009). All the indices were calculated on a seasonal basis, for each individual station and year, even though some of them will be less frequent or rarely occur (i.e. TR inland during winter months or frost in the coast out of winter).

Cluster analysis, one of the most useful data mining procedures for discovering groups and identifying patterns (Halkidi et al., 2001), was applied to reduce the dimensionality of the problem and find for each index different regions within the IP. Cluster analysis is an effective statistical tool for grouping stations into climatologically homogeneous regions (Gong and Richmann, 1995; DeGaetano, 2001; Bartolini et al., 2008; Joliffe and Philipp, 2010).Ward's minimum variance method (Ward, 1963) was used to cluster the indices data. Ward's method chosen in this study is the most frequently used hierarchical clustering technique for climatic classification (Muñoz-Díaz and Rodrigo, 2004; Bednorz, 2008). The most basic stage before applying a clustering algorithm is to establish a numerical similarity or dissimilarity measurement to characterize the relationships among the data. Euclidean distance is the most commonly used measure, although many other distance measurements exist (Gong and Richmann, 1995). An important practical problem is the choice of the intermediate step at which clusters will be formed and the number of clusters to be retained, but there are no universally accepted objective techniques by which to accomplish this. A common subjective approach is to inspect a plot of the distances between merged clusters as function of the stage of the analysis. When similar clusters are being merged early in the process, these distances are small, and they increase relatively little from step to step. Late in the process there may be only a few clusters, separated by large distances. If a point can be discerned where the distances between merged clusters jumps markedly, the clustering process can be stopped just before these distances become large (Wilks, 1995). The inspection of the agglomeration distance plot and the tree diagram, or dendrogram, where the successive clustering steps are summarized, provides the number of clusters and the members belonging to each cluster.

Time series of the indices were subjected to the nonparametric Mann-Kendall test to detect any significant trend. The linear model was used as a first approach to analyse the changing rate in the cases where the Mann-Kendall test detected a significant trend. The slopes of the trends were calculated by least-square linear fitting. These procedures were applied to regional time series, considering a regional series as the mean value of the indices corresponding to all the stations included in each cluster. In the analysis of the time series of the indices, the sum of anomalies (index value of each season minus the long-term mean of the complete period 1929–2005) accumulated over time, starting from the first year until the last year of the series, has been used. An advantage of cumulative anomalies is that they point out periods of predominant negative or positive anomalies more clearly: in case of predominant negative anomalies the curve is decreasing while it is increasing in phases of predominant positive anomalies (Philipp et al., 2007). This method allows distinguish between different periods, which may be compared using usual statistical tools (t-test for difference between means, F-test for variance ratios, and Kolmogorov-Smirnov to compare distribution functions). All statistical tests were evaluated at the 95% confidence level.

The influence of large-scale variables on extreme indices was examined by means of multiple regression analyses. It was calculated as a stepwise multiple regression analysis between the regional series determined by the cluster analysis of the indices (predictands) and the seasonal North Atlantic Oscillation (NAO), East Atlantic (EA), East Atlantic/Western Russia (EA/WR) and Scandinavian (SCAN) patterns (predictors), from 1951 to 2005, when there are pattern values available. The analysis is focused on the teleconnection indices that impact on IP climate (Rodríguez-Fonseca and Rodríguez-Puebla, 2010). The teleconnection indices were taken from the Web page of the Climate Prediction Center (http://www.cpc.ncep.noa.gov/). Stepwise regression was applied in order to investigate the relative importance of the predictors. A significance level of 0.05 was used as condition for entry a new predictor into the models. The p-value of the model estimation indicates the level of statistical significance. The adjusted R-squared statistic indicates the percentage of variance explained by the model. Residuals of the linear models were inspected to determine the role of the teleconnection indices in the detected trends: Durbin-Watson statistic (DW) tests the residuals to determine if there is any significant autocorrelation. If the DW value is greater than 1.4, there is probably not any serious autocorrelation in the residuals.

3. Results

An exploratory analysis of the FD, TR20, TR25, SU25, and SU30 indices for each station on a seasonal basis was made. In some cases, it was not possible to estimate the index, simply because the event did not occur (for instance, occurrence of TR or SU in winter, FD in summer, FD in spring for some coastal stations). In this section, we examine the spatio-temporal variability of the indices that have shown a more general behaviour, that is, winter FD, summer TR20, and summer SU30. The behaviour of the indices in spring and autumn will be studied in a future research. In each case, the cluster analysis was used to classify the stations network in different clusters. Ward's method with Euclidean distance was used, and the number of clusters was established by means of the agglomeration distance plot and the dendrogram. Further analyses were made using the series obtained as the average of the stations belonging to each cluster.

3.1. Winter FD

Figure 2 shows the results related to winter FD. Two clusters have been obtained: cluster 1 (C1) grouping all the inland stations, and cluster 2 (C2), with all the coastal stations. The mean value of C1 is 30.4 days (standard deviation s = 10.1 days), whereas the mean value of C2 is 5.2 days (s = 3.2 days), indicating the important role of the distance from the sea in the occurrence (or not) of frosts.

Figure 2.

Cluster analysis (Ward's method, Euclidean distance) of winter FD. Black line indicates the spatial grouping of stations. Bottom left: tree diagram; bottom right: agglomeration distance plot. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 3(a) shows the time series corresponding to each cluster, alongside the 11-year moving average (the 11-year moving average was estimated simply to obtain a clearer overview of the temporal behaviour of index). The Mann-Kendall statistic for C1 was − 0.55 (p-value = 0.5821), and − 3.61 (p-value = 0.0003) for C2, that is, there is not a significant decreasing trend in the inland stations, meanwhile coastal stations show a significant decreasing trend (slope of the linear trend − 0.6 day/decade). However, this result corresponds to the analysis of the whole period from 1929 to 2005. Figure 3(b) shows the cumulative anomalies corresponding to each cluster. The behaviour of C1 is more random, showing increasing and decreasing fluctuations. But in general terms, two different periods may be distinguished, with 1965 (the maximum of the curve) as the change point. Table II summarizes the comparison between the two periods. Results show that the mean value of the winter FD has decreased significantly only for C2, with a shift to lower values. In the case of C1, where the frosts are more frequent, the slight decrease, from 31.7 to 29.1 days, is not significant. The F-test also shows significant changes in the variance in the case of C2 at the 90% confidence level (slight decrease). As a result, empirical distribution functions corresponding to both periods are significantly different (KS statistic), showing a significant change in this variable for coastal stations.

Figure 3.

(a) Time series of the winter FD for C1 and C2, along with the 11-year moving average. (b) Sum of cumulative anomalies for C1 and C2

Table II. Comparison of different periods for winter FD, summer TR20, and summer SU30 (s = standard deviation; t = t-test for difference between means; F = F-test for ratio of variances; KS = Kolmogorov-Smirnov statistic for comparison of distribution functions, p = p-value). Bold numbers: differences significant at the 95% confidence level. Italic numbers: differences significant at the 90% confidence level
FD
ClusterPeriodMean (days)s(days)tFKS
C11929–196531.79.61.0970.8080.1979
 1966–200529.110.6p = 0.2738p = 0.5199p = 0.4457
C21929–19656.43.43.7151.8481.6204
 1966–20053.92.5p = 0.0004p = 0.0623p = 0.0104
TR20
ClusterPeriodMean (days)s(days)TFKS
C11929–198025.77.07.21160.34330.7469
 1981–200544.311.9p = 0.0000p = 0.0013p = 0.0000
C21929–19809.23.95.98580.35840.6892
 1981–200517.66.5p = 0.0000p = 0.0021P = 0.0000
C31929–19800.70.73.9040.24470.4861
 1981–20051.91.4p = 0.0005p = 0.0000p = 0.0007
SU30
ClusterPeriodMean (days)s(days)TFKS
C11929–198014.34.25.37760.50450.6300
 1981–200520.65.3p = 0.0000p = 0.0407p = 0.0000
C21929–198047.69.75.29261.00250.5438
 1981–200558.58.3p = 0.0000p = 0.9719p = 0.0001

3.2. Summer TR20

Figure 4 shows the application of cluster analysis to summer TR20. In this case a clear division from northwest to southeast is observed. Cluster 1 (C1) corresponds to the southern Mediterranean area, cluster 2 (C2) to a central region, oriented from northeast to southwest, and cluster 3 (C3) corresponds to the northwestern IP. This spatial distribution seems to reflect the intense heat advection from northern Africa and across the central Iberian plains towards the northwestern coast, responsible for some of the more extreme summer events in Iberia (García-Herrera et al., 2005). Alb also belongs to C3, perhaps due to its height above sea level (699 m a.s.l., higher than nearby stations, such as Mur or CRe (Table I). In addition, cluster analysis shows less clear differences between clusters C2 and C3, see tree diagram). Mean values are 31.7, 11.9, and 1.1 days for, respectively, C1, C2, and C3 (with standard deviations of 12.4, 6.3, and 1.1 days, respectively). The Mann-Kendall statistics were + 6.2 (p-value = 0.00000), + 4.3 (p-value = 0.00002), and + 2.9 (p-value = 0.00339) for, respectively, C1, C2, and C3, indicating increasing trends of TR20 for the three clusters, especially important in C1, where the frequency of TR is greater. Mediterranean warm influence is notable in the occurrence of TR. Linear trends indicate an increase of 3.8 days/decade for C1, 1.4 days/decade for C2, and 0.2 days/decade for C3.

Figure 4.

As in Figure 2, but for summer TR20. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 5(a) shows the time series of the summer TR20 indices for C1 and C2 (and Figure 5(b) for C3, with very low values compared to the other clusters). Increasing trends are clear from around 1980 for all the clusters. These results are in congruence with the regionalisation of TR—by PCA—found by Sanchez-Lorenzo et al. (2011) for the period 1961–2004 and extended summer (JJAS), also detecting a change point by around the 1980s, and larger trends for the Mediterranean region, although the identified regions are slightly different (i.e. due to the use of different stations, periods and methods of regionalisation). Figure 5(c) shows the sum of cumulative anomalies for C1 and C2 (Figure 5(d) for C3). There are slight differences in the change point of each cluster, with the change of slope of C1 curve before the change corresponding to C2 and C3. In general terms, we can establish two clearly defined periods, 1929–1980 and 1981–2005. Table II shows the comparison of the statistics corresponding to these periods. In the three clusters, the differences between both periods are significant, with an important increase in the order of about 20 days for C1, 9 days for C2, and 1 day for C3. Also, the distribution functions corresponding to both periods are significantly different (KS statistic), with significant increases in mean value (t-test) and variances (F-test). This sharp increase in minimum temperatures since the last decades of the 20th century has been reported in previous works (e.g. Brunet et al., 2007a; Ramos et al., 2011).

Figure 5.

(a) Time series of the summer TR20 for C1 and C2 along with the 11-year moving average. (b) As in (a) but for C3. (c) Sum of cumulative anomalies for C1 and C2. (d) As in (c) but for C3

3.3. Summer SU30

Figure 6 shows the application of cluster analysis to summer SU30. In this case, two clusters can be distinguished: C1 corresponds to northern area of the IP, and C2 corresponds to southern area. Proximity to the sea seems to moderate the behaviour of maximum temperatures, including coastal stations (Lis, Cad, Mal, and Val) in C1. The mean value for C1 is 16.3 days (s = 5.2 days) and the mean value for C2 is 51.3 days (s = 9.7 days). For both clusters, Mann-Kendall statistic indicates significant positive trends, with values of 4.9 (p-value = 0.000001) and 4.7 (p-value = 0.000002) for, respectively, C1, and C2. Linear trends indicate an increase of 1.2 days/decade for C1, and 2.3 days/decade for C2.

Figure 6.

As in Figure 2, but for summer SU30. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 7(a) shows the time series of the summer SU30 series for the two clusters. An increasing trend of the index in the last decades of the 20th century is observed. Figure 7(b) shows the sum of cumulative anomalies for each cluster. The beginning of the increasing trend is around 1980 for C2, and it is slightly delayed for C1. Again, similar to summer TR20, the positive trend seems to begin earlier in southern areas than in northern areas of the IP. Two periods have been compared: 1929–1980 and 1981–2005. Results are shown in Table II. Differences in mean value are of the order of 6 days for C1 and 11 days for C2. The F-test shows a significant difference in the variances for C1 (with a slight increase of s since 1981), but there are not significant differences for cluster 2. The shift of mean value to higher values implies a significant change in the distribution functions (KS statistic). This increase is lower than that of the TR20 index, indicating that Tmax is not increasing as quickly as Tmin, which is a result found over Europe and many places worldwide (Alexander et al., 2006). However, according to other works (for instance, Brunet et al., 2007b), the magnitude of the changes in the IP of minimum and maximum temperatures has been of the same order since the last decades of the 20th century. This slight discrepancy may be caused by the use of different extreme indices (percentile-based and absolute threshold indices) and different periods.

Figure 7.

(a) Time series of the summer SU30 for C1 and C2, along with the 11-year moving average. (b) Sum of cumulative anomalies for C1 and C2

3.4. Relationships with teleconnection indices

Owing to the effects of atmospheric circulation on regional climate, the variations of the indices winter FD, summer TR20 and summer SU30 can be studied considering their association with large-scale atmospheric patterns, such as the teleconnection indices. The teleconnection indices studied are the NAO, EA, EA/WR, and SCAN. Monthly values of these indices since 1950 are available in the Web page of the Climatic Prediction Center. These monthly values were averaged to obtain seasonal values. Afterwards, a stepwise multiple regression analysis was made. The resulting models are summarized in Table III.

Table III. Regression models of the winter FD, summer TR20 and summer SU30 indices derived from large-scale variables (DW = Durbin-Watson statistic)
Model equationp-valueR2adjDW
FD(C1) = 30 + 5NAO − 10EA0.000051%2.22
FD(C2) = 4.7 − 2.7EA0.000039%2.46
TR20(C1) = 34 + 6NAO + 5EA − 8SCAN0.000426%1.14
TR20(C2) = 12.2 − 5.0SCAN0.000420%0.92
TR20(C3) = 1.1 − 0.7SCAN0.002514%1.59
SU30(C1) = 16.7 + 2NAO − 5SCAN0.000127%1.22
SU30(C2) = 52 + 4NAO − 10SCAN0.000044%1.34

The adjusted R-squared statistic indicates the percentage of variance explained by the model. In this sense, winter FD models (and summer SU30 model for the cluster C2) are the best with percentages higher than 39%. This result is consistent with the fact that large-scale circulation better explains temperature variability in winter than in summer (baroclinity), as obtained in the application of regression models for estimating observed temperature variability based on circulation-pattern classifications (Philipp et al., 2007). The DW tests the residuals to determine if there is any significant autocorrelation. If the DW value is greater than 1.4, there is probably not any serious autocorrelation in the residuals. Only in the cases of FD(C1), FD(C2), and TR20(C3) we can reject the existence of autocorrelation, i.e. because DW > 1.4. Therefore, in the other cases it may be necessary to look for other causal mechanisms to explain the trends detected.

The model for FD(C1) indicates that positive (negative) trends in the NAO (EA) index imply an increase of the number of FD for the inland stations of the IP. Anticyclonic conditions associated with the NAO positive phase imply clear skies, absence of clouds, and therefore loss of energy through nighttime outgoing long-wave radiation (Trigo et al., 2002). This situation may provoke the appearance of frosts in the country, especially in winter. The EA pattern is structurally similar to NAO, but with a strong subtropical link, leading to above average surface temperature in Europe during its positive phase. Therefore, there is a negative correlation between EA pattern and frosts days. In this case, the winter EA pattern contributes to the total variance with around 40%, while the contribution of the winter NAO is around 11%. The analysis of Mann-Kendall statistics in winter NAO and EA indices indicates significant positive trends in both indices. The absence of autocorrelation in the residuals confirms the important role of these teleconnection indices in the variability of FD for the cluster C1. A similar result was found by Rodríguez-Puebla et al. (2010), who establish a link between percentile-based indices (frequency of cold nights) and the EA pattern. The model for FD(C2) show similar results, except for the non-significant role of the NAO for coastal stations.

In the case of summer models, the leading role of the SCAN pattern is notable. The SCAN pattern is characterized by a primary center over Scandinavia, with weaker centers of opposite sign over Western Europe and eastern Russia/western Mongolia. The positive phase of the SCAN pattern is associated with below average temperatures across Western Europe. Therefore, a negative relationship between summer temperature indices and SCAN pattern must be found. This result has been found in all the models related to summer indices. NAO pattern also contributes to the total variance with around 6% for TR20 (southern stations), and SU30 indices. In summer, anticyclonic conditions lead to increased maximum and minimum temperatures. EA pattern is important for TR20 index and cluster 1 (southeastern stations), showing the same positive correlation, but contributing only on 6% of the total variance. The percentage of variance explained by models is lower than 30%, except for SU30 in southern stations. Meanwhile, summer EA pattern shows a significant positive trend (Mann-Kendall statistic = + 2.87, p-value = 0.0041), summer SCAN pattern do not show any significant trend (Mann-Kendall statistic = − 1.81, p-value = 0.0707). In addition, the DW statistic indicates serious autocorrelations in the residuals of the summer models (except in the case of TR20 for northwestern stations), with the SCAN pattern as main significant predictor. Therefore, the positive trends detected in summer indices must be explained by other causal mechanisms.

4. Conclusions

The main results of this study are summarized as follows:

  • - Winter FD index: there is not a significant decreasing trend in the inland stations, whereas coastal stations show a decreasing trend (−0.6 days/decade), especially from 1965 onwards.

  • - Summer TR20 index: there is a significant increasing trend, from 1980 onwards, especially important to the southeast of the IP (+3.8 days/decade in the long-term period 1929–2005).

  • - Summer SU30 index: there is a significant increasing trend, from 1980 onwards, particularly in the southern stations (+2.3 days/decade in the period 1929–2005).

  • - The behaviour of winter FD index is related to the evolution of NAO and EA patterns (with increasing trends in the last decades of the 20th century), which explain 51% of the variance of inland stations.

  • - Although multiple regression models establish a relationship between summer indices and the SCAN pattern, the behaviour of residuals indicates that summer indices must be explained by other causal mechanisms.

The trends detected in the winter and summer indices are consistent with previous studies analysing maximum and minimum temperatures, as well as other extreme indices based on percentiles, in the IP (Serra et al., 2001; Burgueño et al., 2002; Miranda et al., 2002; Prieto et al., 2004; García-Herrera et al., 2005; del Río et al., 2005, 2007; Brunet et al., 2007a, 2007b; Rodríguez-Puebla et al., 2010; Martínez et al., 2010; Furió and Meneu, 2011; Ramos et al., 2011) and in other regions of the world, such as Tuscany in Italy (Bartolini et al., 2008), Athens in Greece (Founda et al., 2004), western Europe (Heino et al., 1999; Klein Tank and Können, 2003), Middle East (Zhang et al., 2005), the Asia-Pacific region (Choi et al., 2009) or the northeast United States (Easterling, 2002; Griffiths and Bradley, 2007). In addition, change points found here (1965 for FD, 1980 for TR and SU indices) are similar to those previously reported in the literature (Alexander et al., 2006; Ivanov and Evtomiv, 2010; Sánchez-Lorenzo et al., (in press)).

The changes of winter FD index are linked to the EA index and, in minor degree to the NAO index. This conclusion is consistent with the result by Rodríguez-Puebla et al. (2010), indicating a decrease of the frequency of cold nights linked to more positive phases of the EA. These authors also find that the increases in the frequency of warm days (at an annual basis) are linked to a decrease in the SCAN pattern. Although this relationship has been found in our models (Table III), it is not conclusive for summer indices, because of the low percentage of explained variance and the behaviour of residuals, showing strong autocorrelation. Therefore, an additional causal mechanism must be studied to explain the behaviour of summer indices. A possible candidate is the influence of the sea breeze on daily extreme temperatures. We have seen the smoothing effect of the vicinity to the sea on the regionalisation of winter FD and summer SU30 indices: coastal stations have very low frequency of FD in comparison to inland stations, and some coastal stations are included in the cluster corresponding to northern stations for SU30 index. In fact, the smoothing effect on temperature trends due to the vicinity to the Mediterranean Sea has been especially observed for maximum spring and summer temperatures in Catalonia, to the northeast IP (Martínez et al., 2010). According to Xoplaki et al. (2003) the Mediterranean sea surface temperature (SST) plays an important role as thermal predictor in the Mediterranean Basin: warmer summer conditions are connected with positive SST anomalies in the Mediterranean Sea. The relationship between SST anomalies and extreme temperature indices, as well as a deeper analysis of spring and autumn indices, will be the subject of future works.

Acknowledgements

This work was supported by the Spanish Ministry of Education and Science (Project CGL2007-65546-C03-01) and the Junta de Andalucía (Project GLObal CHange in ARID Zones, GLOCHARID). The authors are indebted to Dr E. Aguilar from the Climate Research Group of the University Rovira I Virgili (Tarragona, Spain) for providing the original temperature data series (SDATS).

Ancillary