Spatial and temporal variability of climate extremes in Romania and associated large-scale mechanisms

Authors


ABSTRACT

The simultaneous variability of several climate extremes in Romania on the one hand and the understanding of the large-scale mechanisms responsible for this variability on the other are examined. Ten indices associated with temperature and precipitation extremes computed at high spatial resolution for the period 1961–2010 are analysed. Significant increasing trends for the temperature extremes are detected in all seasons, except for autumn, with the highest increasing rate in summer and the lowest in spring. Regarding precipitation extremes, significant increasing trends over large areas in the frequency of very wet days and maximum daily amount during autumn and in the maximum duration of dry spells during summer are the only ones detected.

The large-scale mechanisms responsible for these characteristics of variability, especially the simultaneous variability of several climate extremes, are identified through the canonical correlation analysis applied to a combination of various large-scale predictors and to combined climate extremes.

In winter, it was found that the thermodynamic factor (represented by air temperature anomalies at 850 hPa) mainly controls the trends of temperature extremes in Romania, whereas the dynamic one (represented by the sea level pressure anomalies) controls the pattern of trend magnitude. Regarding precipitation extremes, the role of the two factors is reversed. The Carpathians' influence is noted for this season. In summertime, the thermodynamic factor is dominant for both temperature and precipitation extremes analysed in this article. For temperature extremes, the T850 alone could explain their variability characteristics, whereas for precipitation extremes (frequency and duration) the SH700 has the dominant role, except for the maximum duration of dry intervals, which is controlled by a combination of T850 and SH700 anomalies. The connections found in this study are strong and explain a great part of the total observed variance, showing that these results can be used in a future study to build skilful statistical downscaling models, simultaneously for several seasonal climate extremes, giving the results more physical coherence.

1 Introduction

Changes in the frequency and intensity of extreme climate events (such as heavy precipitation and resultant flooding, droughts and high temperatures) have been documented in many scientific papers (e.g. Alexander et al., 2006; Moberg et al., 2006; Kenyon and Hegerl, 2008; Kenyon and Hegerl, 2010), and a global comprehensive outlook has been presented in the latest IPCC Report (IPCC, 2007). Thus, on a global scale, Alexander et al. (2006) report widespread significant changes in temperature extremes associated with the warming trend, especially those related to daily minimum temperatures from an analysis for the period 1951–2003. Precipitation extremes showed a widespread and significant increase, but the changes are much less spatially coherent compared with changes in temperature extremes. Similar findings would have been found at a smaller number of sites if the entire 20th century had been considered. However, important differences are recorded on regional scales, especially for precipitation extremes. These differences are mainly related to the fact that for global/large-scale studies the available observation data sets were not enough to capture the local/regional features of the changes in climate extremes as a response to the large-scale modes of climate variability. Moberg et al. (2006) found that the overall global findings by Alexander et al. (2006) are not representative for Europe if the entire 20th century is considered. They also found an overall warming but only a small difference between average trends in daily minimum and maximum temperatures averaged for 75 stations across Europe. Klein Tank and Können (2003) found similar results but for a shorter period (1946–1999). On the other hand, analyses on various subregions documented some differences against the overall picture on the European scale presented above. For example, Domonkos et al. (2003) analysed the frequency of winter and summer temperature extremes at 11 sites in central and southern Europe (1901–1998) and found large long-term fluctuations, with a slight warming trend, statistically significant only at a few stations. For the Mediterranean basin, based on the analysis of 15 different indices of temperature extremes, Efthymiadis et al. (2011) found that the underlying trends are generally consistent with global trends of the temperature and its extremes: cold extremes decrease and warm/hot extremes increase. This consistency is better manifested in the western part of the Mediterranean where changes are the most enhanced since the mid-1970s. However, they found an interbasin discrepancy that is clearer in winter, whereas in summer changes are more uniform. Other studies have focused on even smaller subregions, such as Italy (Baldi et al., 2006; Bartolini et al., 2008) showing that the magnitude of the extreme temperature changes varies seasonally and regionally. Moreover, the changes are not always steady or monotonic. Therefore, it is important to carry out climate extremes studies on regional scales using a high-density observational data set to understand better the mechanisms controlling their variability. The present work aims to carry out such a study. Regarding Romania, several studies documented on the analysis of the climate extremes variability (Tomozeiu et al., 2002; Baciu et al., 2004; Busuioc et al., 2007, 2010; Păltineanu et al., 2009; Birsan and Dumitrescu, 2014) but most of them used a limited number of stations/short intervals for analysis or did not enter into the details of the mechanisms controlling their variability. The present study attempts to fill this gap.

Most studies on climate extremes revealed various large-scale circulation patterns/indices as main drivers of the climate extreme variability (e.g. Kenyon and Hegerl, 2008; Jones and Lister, 2009; Kenyon and Hegerl, 2010). Even if some of them show a global influence, their effects are strongest for some regions such as: El Niño–Southern Oscillation around the Pacific Rim and throughout all of North America; Pacific interdecadal variability for the Northern Hemisphere, especially around the Pacific region and North America (Kenyon and Hegerl, 2008). For Europe, the strongest influence is given by the North Atlantic Oscillation (NAO), especially in wintertime (Hurrell and van Loon, 1997; Bojariu and Paliu, 2001; Tomozeiu et al., 2002; Haylock and Goodess, 2004; Cassou et al., 2005; Scaife et al., 2005; Birsan and Dumitrescu, 2014). Folland et al. (2009) defined a summer counterpart of NAO called Summer North Atlantic Oscillation (SNAO) that exerts a strong influence on the northern European climate through changes in the position of the North Atlantic storm tracks, having a very important role in generating summer climate extremes (flooding, drought and heat stress) in northwestern Europe. Both NAO and SNAO exert an interannual/decadal influence (e.g. Dima et al., 2001; Gámiz-Fortis et al., 2010). On a multidecadal timescale, the European extremes are influenced by the Atlantic Multidecadal Oscillation (AMO) (Sutton and Hodson, 2005; Della-Marta et al., 2007, Gámiz-Fortis et al., 2010; Ionita et al., 2011, 2013), which exhibits a quasi-periodicity of about 60 years (Enfield et al., 2001). A deterministic mechanism for AMO is presented by Dima and Lohmann (2007). In most cases, the relationship between circulation patterns and climate extremes have been detected through the simple correlation between the two data sets (e.g. Efthymiadis et al., 2011; Ionita et al., 2011), composite maps (e.g. Tomozeiu et al., 2002; Ionita et al., 2013) or through multivariate techniques such as the canonical correlation analysis (CCA) (Haylock and Goodess, 2004; Xoplaki et al., 2004; Della-Marta et al., 2007).

Besides the NAO/SNAO, other well-documented large-scale atmospheric circulation patterns in the Euro-Atlantic domain (Arctic Oscillation, Mediterranean Oscillation) together with more regional patterns (Scandinavian, Central European, East European and East Atlantic) have been used to find their connection with the climate extremes for European subregions (e.g. Tomozeiu et al., 2002; Efthymiadis et al., 2011). The most recent circulation types developed by Philipp et al. (2007), using a more stable statistical scheme, have also been used to justify the trends in extremes on European regions (Jones and Lister, 2009; Ionita et al., 2013).

Della-Marta et al. (2007) investigated a combined influence of a number of forcing factors on the heat wave events over western Europe [sea level pressure (SLP), Atlantic sea surface temperatures (SST) and European precipitation] using the canonical correlation analysis (CCA) technique, considering precipitation as a proxy for soil moisture strengthening land–atmosphere feedback process. Xoplaki et al., 2004 used the same technique to explain the wet season precipitation variability over 292 sites across the Mediterranean area, showing that 30% from this variability can be accounted for by a combination of four large-scale geopotential height fields and SLP. These kinds of techniques seem to be more adequate to reproduce the complexity of the mechanisms controlling the regional climate variability. A similar but improved technique will be used in the present work.

In all the studies presented above, the connection between the forcing factors and climate extremes has been analysed separately for each extreme index. In the present study, we try to analyse the spatial/temporal characteristics of simultaneous variability of various climate extremes in Romania, on the one hand and to understand the large-scale mechanisms controlling this variability through the combined influence of dynamic and thermodynamic forcing factors on the other. The CCA technique is used to achieve this objective and a high spatial resolution daily data set (for the 1961–2010 period) is used to calculate the selected climate extremes (Section 'Data and methods'). This analysis will try to supply an integrated view on the mechanisms controlling the variability of climate extremes in Romania at a high-resolution spatial scale, contributing to the deepening of knowledge in this field. Details on methodology are shown in Section 'Data and methods' and the main findings are presented in Section 'Results'. Conclusions and discussions are summarized in Section 'Conclusions and discussions'.

2 Data and methods

2.1 Data

Using daily measurements of minimum/maximum air temperature and precipitation from all available stations in Romania with full data records for the period 1961–2010, ten seasonal indices associated with temperature and precipitation extremes have been computed for all seasons, quantifying their intensity, frequency and duration, as follows: maximum temperature (Tmax); minimum temperature (Tmin); frequency of very warm days/nights (maximum/minimum temperature ≥ 90th percentile): Frtmax90, Frtmin90; longest period of very warm days/nights: Dtmax90, Dtmin90; longest dry period: Dmaxpp0; frequency of very wet days (daily precipitation amount above the 90th percentile): Frpp90; longest very wet period: Dmaxpp90; maximum daily precipitation (Ppmaxd). The daily data were a priori quality controlled using the MASH software (Szentimrey, 1997). The number of available stations for each index is different from one index to another. For the six temperature indices (Tmax, Tmin, Frtmax90, Frtmin90, Dtmax90 and Dtmin90), 85 stations were available, whereas for the precipitation ones (Frpp90, Dmaxpp90, Dmaxpp0 and Ppmaxd) full data records from 98 stations were used. The location of the 98 stations is presented in Figure 1. The 90th percentile threshold of daily temperature and precipitation has been calculated using the empirical distribution function, similar to the method presented in the European STARDEX project (Haylock and Goodess, 2004). In the case of precipitation, only daily amounts >0.1 mm day−1 have been considered. For the Dmaxpp0, the days without precipitation have been used. In the following, all these indices will be addressed as climate extremes or extreme indices. To overcome the shortness of the data set in identifying a plausible climate signal, some climate extremes computed over a longer period (1901–2010) at Bucuresti-Filaret station have been additionally analysed. The grid point monthly mean of four large-scale predictors from the NCEP/NCAR reanalysis (Kalnay et al., 1996) have been used to understand the large-scale mechanisms controlling the variability of climate extremes in Romania, as follows: SLP, temperature at 850 hPa (T850) and specific humidity at 850/700 hPa (SH850/SH700).

Figure 1.

Name and location of stations used in this study.

2.2 Methods

The main objective of this article is to analyse the spatial/temporal characteristics of simultaneous variability of several climate extremes in Romania on the one hand and to understand the large-scale mechanisms controlling this variability on the other. Compared to previous studies, this type of statistical analysis is applied for the first time in Romania for climate extremes and, as far as the authors are aware, it has been used rather rarely by the international scientific community in climate research studies. An attempt in this respect was carried out by Busuioc (2001) when the connection between the simultaneous variability of seasonal winter mean temperature and precipitation totals at 14 stations in Romania and the large-scale circulation was analysed. In this article, a high-resolution data set has been used (with the disadvantage of covering a shorter period) to understand the complexity of the mechanisms controlling a wide range of indices describing the climate extremes in Romania. Based on the results achieved in this study, it is expected to add value to the statistical analysis that is generally criticized for the lack of physical coherence between various climate variables. On the other hand, if the results are promising, it is expected that statistical downscaling models could be constructed simultaneously for more extreme indices, giving the results more physical coherence; this issue will be developed in the next step. Similar techniques have been used by Xoplaki et al. (2004) but for a single predictand (Mediterranean precipitation).

To solve the major objective presented above, more specific objectives will be carried out as follows:

  • analysis of long-term trends and shifts in the mean for all extreme indices presented above; to reach this objective, the nonparametric tests Mann-Kendall (Sneyers, 1975; Kulkarni and von Storch, 1995) and Pettitt (Pettitt, 1979) will be used; a comparison with the Kendall-Theil method (Sen, 1968; Helsel and Hirsch, 1992), less affected by outliers, and logistic regression (Frei and Schar, 2001), will be highlighted.
  • synthesis of spatial and temporal characteristics of the climate extremes variability using empirical orthogonal functions (EOF) analysis (Wilks, 1995; von Storch and Zwiers, 1999).
  • connection between regional variability of selected climate extremes in Romania and large-scale climate variability using the CCA (Barnett and Preisendorfer, 1987; Zorita et al., 1992; von Storch et al., 1993; Busuioc and von Storch, 1996) to understand the large-scale mechanisms controlling the spatial and temporal variability of the selected climate extremes in Romania.

The trend estimator used is the ordinary least squares method and the significance level is 5%. It is known that there are various robust nonparametric trend estimators (e.g. Helsel and Hirsch, 1992; Moberg and Jones, 2005; Moberg et al., 2006) but the trend magnitude is quite similar, whereas the corresponding statistical significance is less certain (Cohn and Lins, 2005). Special attention is paid to the statistical modelling of the temporal variation of weather events represented as count data (e.g. frequency of heavy precipitation) as presented by Frei and Schar (2001). They used the statistical model of logistic linear regression (LLR) for estimating and testing long-term trends in count records. The S-plus statistics package (Venables and Ripley, 2002) can be used to estimate the statistical significance of the estimated trend parameter that can be inferred from the p-value testing against the null hypothesis (slope of the linear regression = 0). In this article, the R language has been used to compute the LLR and to estimate the p value.

We are first interested in the explanation of large-scale mechanisms leading to some changes (trends, shifts) simultaneously to various types of climate extremes (e.g. temperature and precipitation-related) and less in the accuracy of the trend magnitude or significance level. The method presented by Sen (1968) has also been tested for some climate extremes used in this study and led to a similar signal as the ordinary least squares method, except for cases when many equal values are presented in data (e.g. Frpp90, Dmaxpp90 and Dmaxpp0). In this case the slope is computed between all i pairs of the variable x:

display math

where i = 1…N. For n values in the time series x this will result in N = n(n − 1)/2 values of β. The slope estimate b is the median of βi, i = 1…N. As the slope is less affected by outliers, the largest values are smoothed. For the time series of Frtmax90 and Frpp90, the LLR has been also tested using the algorithm presented by Frei and Schar (2001) and S-plus statistics package (Venables and Ripley, 2002). The statistical significance of the linear trend has been estimated with the Mann-Kendall test, two versions: MK1 (Sneyers, 1975) and MK2 (Helsel and Hirsch, 1992) with corrections for equal values (common for count data related to the frequency of extreme events). The statistical significance of the trend estimated by the LLR is inferred from the p-value testing against the null hypothesis (slope of the linear regression = 0). The comparison between the results derived from the three methods will be highlighted in this study.

The EOF analysis is applied to each of the mentioned climate extremes and then to various combinations between them, in order to find their main variability modes. As for the combined indices, their time series are first standardized by dividing the anomalies (deviation from the long-term mean) by the standard deviation. Possible significant changes (linear trends, shifts in the mean) in the evolution of the time series associated with the first EOF pattern (PC1) of climate extremes in Romania synthesize (over the entire country) the possible changes in the climate regime of the selected climate extremes. Detailed station-based analysis is also carried out to identify the regional characteristics of the climate signal. The PC1 contains the most part of the climate signal (e.g. significant trend), if this is present in the data set. Therefore, if PC1 has a strong trend, a large spatial coherent area or the entire country shows a significant trend, which is physically controlled by large-scale mechanisms analysed further.

Secondly, the large-scale mechanisms controlling the variability of the climate extremes in Romania will be carried out through the CCA (Barnett and Preisendorfer, 1987; Busuioc and von Storch, 1996; Busuioc et al., 2001). By CCA, the optimum linear combination for two multidimensional vectors (predictand – the spatial vector of climate extremes and predictor – the spatial vector of certain large-scale variables) and pairs of patterns are selected such that their associated time series be maximum correlated. By construction, the CCA method allows a physical interpretation of the mechanism controlling the variability of the regional climate parameters (Busuioc and von Storch, 1996; Busuioc et al., 2001, 2010). Dynamic (SLP) and thermodynamic (T850, SH850/SH700) variables will be used as large-scale predictors. In this article, the connection between the simultaneous variability of more large-scale predictors (up to three) and one or more extreme indices in Romania will be analysed through the CCA. The CCA will use as inputs the outputs from the multifield EOF analysis applied to combined standardized anomalies of predictor/predictands. Therefore, this method will be used to supply an integrated view on the mechanisms controlling the simultaneous variability of climate extremes in Romania at a high-resolution spatial scale, never used so far. The analysis of the time series associated with the first two CCA pairs in terms of their trends and shifts, will provide a physical explanation to the trends and shifts found in the regime of analysed extreme events in Romania. Before the CCA, the predictors and predictands are projected onto their EOF space. It is known that the CCA pairs are ranked in terms of magnitude of the canonical correlation coefficient between the time series of the CCA pairs, and not in terms of explained variance of the CCA pair patterns. Only when the connection predictor–predictand is given by the principal variability mode (EOF1) of the large-scale predictor, the associated explained variance of each pattern of the first CCA pair is highest as canonical correlation coefficient. Considering the main objective of this article, the optimum combinations of EOFs is considered in the CCA, such that the explained variance of the predictor/predictand in the first CCA pair be as large as possible and the predictand associated time series have, as much as possible, the same characteristics of variability (similar trend rate and shift in the mean) as PC1 of the selected predictand. Thus, it could be stated that the selected CCA pattern of the predictor is responsible for the main variability mode of the predictand. This mechanism has been presented by Busuioc and von Storch (1996), in the explanation of winter precipitation change at 14 stations in Romania for the 1901–1987 interval, only the SLP is considered as predictor. Similar plausible mechanisms have been later found by Busuioc et al. (2001) for the monthly precipitation in Sweden and by Xoplaki et al. (2004) for the wet season Mediterranean precipitation variability.

All the maps presented in this article were constructed with the Ordinary Kriging interpolation method as implemented in ArcGIS 10.2 environment. Based on minimizing the mean square error, an optimization procedure was used for fitting all the semivariogram models.

3 Results

3.1 Spatial characteristics for variability of climate extremes in Romania

The first two EOF patterns (hereafter abbreviated as EOF1 and EOF2) of the ten climate extremes analysed in this article have been calculated and Figure 2 shows, as an example, the EOF1 and EOF2 for the Frtmax90 (winter and summer); Figure S1, Supplementary Information, shows, as an example, the first two EOF patterns for Frpp90 (winter). The EOF1 presents similar spatial characteristics (same sign of variability over the entire country, higher magnitude over the southern and eastern areas in wintertime and generally over the southwestern-western areas in summertime) for all indices, showing that a large-scale mechanism is responsible for their variability over the entire country. If this mechanism is common to more indices or what the specific mechanisms responsible for the variability of each or more climate extremes are, will be analysed in Section 'Mechanisms controlling the variability of climate extremes in Romania' The EOF2 shows a dipole pattern with some differences between winter and summer seasons. In winter, for all climate extremes the EOF2 presents a dipole structure separating the intra-Carpathian regions from the extra-Carpathian ones, showing the influence of the Carpathian Mountain in this season, acting as a barrier in the air mass flow over Romania. A different pattern is presented by the maximum duration of dry intervals (Dmaxpp0) showing a north–south dipole structure with larger spatial extension of the northern part, which suggests a smaller influence of the mountain chain. In summertime, a dipole structure with a northeast-southwest gradient can be revealed, except for the Frpp90 that shows a north–south dipole structure.

Figure 2.

The first two EOF patterns of Frtmax90 (very warm days) for winter and summer, with explained variance included. Zero line is marked.

The variance cumulated by the first two EOFs ranges between 35% (Dmaxpp90, summer) and 86–87% (Frtmax90, summer and spring); the highest values are generally noted for temperature extremes and lowest for precipitation extremes. The number (N) of the EOFs explaining more than 1% of total observed variance and variance explained by the first two EOFs are presented in Table 1. The N values range between 5 (Frtmax90) and 28 (Dmaxpp90), showing very high spatial variability for Dmaxpp90 captured by many EOF patterns, whereas for the Frtmax90 only a few patterns are enough to explain the total observed variance, the EOF1 explaining about 80%. This result shows that, as will be seen in Sections 'Temporal characteristics of the climate extremes variability in Romania', for the Frtmax90 (and generally for all temperature extremes) a single large-scale mechanism can be found to be responsible for its (their) spatial and temporal variability, whereas for the precipitation extremes it is more difficult to find such a mechanism. This result is in agreement with the conclusions drawn from previous papers related to the mean state (seasonal mean temperature, precipitation totals: Busuioc and von Storch, 1996; Tomozeiu et al., 2005; Busuioc et al., 2010) or seasonal mean of the maximum temperature recorded to a fewer number of stations (Tomozeiu et al., 2002).

Table 1. Results of the EOF analysis and trends/shifts of the PC1 for various seasonal extreme indices analysed in this article: N number of EOF patterns explaining more than 1% from the total variance; v1, v2 explained variance of the first two EOFs; Mann-Kendall statistic Z for PC1; shift in PC1. For the multifield EOF analysis, the EOF1 sign of the component patterns are shown in the second row. The Mann-Kendall statistics and shift for the Frtmax90 calculated at Bucuresti-Filaret station for the period 1901–2010 are also shown (last row). Bolded values show the 5% significance level, while italics show 10% significance level. Z values around 0 are noted by ‘-’.
ParameterWinterSpringSummerAutumn
Nv1v2Z, shiftNv1v2Z, shiftNv1v2Z, shiftNv1v2Z, shift
Frtmin9067752.8, 1987106682.7, 199377663.9, 198510685-, -
Dtmin90125672.6, 1987154892.3, 1999116573.3, 1995154881.1, 1998
Frtmax9057783.0, 198758161.4, -48153.9, 19846786-, -
Dtmax90966112.9, 19851065101.7, 199386683.8, 19867718-, -
Frpp90183610−0.7, 1971213010−0.7, -243370.5, -1741131.6, 1994
Dmaxpp9026179-, -27198-, -28176-, -242681.9, 1993
Dmaxpp0154812-, -16469-, -2035101.8, 198416509−1.1, -
Ppmaxd212511 25209 261980.72125112.2, 1993
Frtmax90 + Frtmin906 71 + +93.1, 1987    6 71 + +113.9, 1985    
Frtmax90 + Dmaxpp013 38 + +232.9, 1987    16 50 + +63.5, 1986    
Frtmax90 + Frpp9014 38 + −193.0, 1987    19 41 + −143.2, 1984    
Frtmax90 + Frpp90 + Dmaxpp017 28 + − +202.3, 1986    23 35 + − +103.1, 1984    
Frtmax-Bucuresti-Filaret    4.2, 1.9 1947,1984   2.7, 1980   1.9, 1986   -, -

As mentioned in Section 'Data and methods', the main objective of this article is trying to find plausible mechanisms responsible for the simultaneous variability of more climate extremes. The multifield EOF analysis reveals, as principal mode, the same sign of simultaneous variability for all thermal extremes and opposite variability between thermal and precipitation extremes, except for Dmaxpp0 (Table 1), which is of opposite sign to the other precipitation extremes but partially in the same phase with temperature extremes. Examples for winter and summer are presented in Table 1. The mechanisms responsible for the simultaneous variability of various climate extremes are presented in Section 'Mechanisms controlling the variability of climate extremes in Romania' using the CCA analysis.

3.2 Temporal characteristics of the climate extremes variability in Romania

The long-term linear trends for the period 1962–2010 have been calculated for all extreme indices presented in Section 'Data and methods'. To compare the rate of change between them the trends of the PC1 were first computed, the results are shown in Table 1 where the Mann-Kendall statistics and shift in PC1 mean are presented for all seasons (winter, spring, summer and autumn). It is noted that the trend statistics for PC1 are similar to those calculated using the two versions of Mann-Kendall test: MK1 (Sneyers, 1975) or MK2 (Helsel and Hirsch, 1992. The PC1 evolution for winter and summer is shown in Figure 3. Significant increasing trends of PC1 for the six temperature extremes (Frtmax90, Frtmin90, Dtmax90, Dtmin90, Tmax and Tmin) have been detected in all seasons, except for autumn and Dtmax90-spring. The increase rate is more pronounced in summer and less pronounced in spring. The linear trend seems not very appropriate (it is not steady or monotonic) in summer when a significant upward shift around 1986 separates a subinterval with a slightly decreasing trend (1962–1986) by a subinterval (1986–2010) with a marked increasing linear trend. The same shift was identified in winter (Figure 3), but in this season the linear trend is more appropriate. This result is in agreement with those presented by Efthymiadis et al. (2011) for the Mediterranean temperature extremes. This finding also suggests that for summertime a multidecadal variability of the temperature extremes could be noted, which is in agreement with the conclusion found by Ionita et al. (2013) from the analysis of the summer temperature in Romania at 14 stations over a longer period, as well as by Della-Marta et al. (2007) for the summer heat waves over western Europe. We will return to this issue later.

Figure 3.

Time series associated with the first EOF patterns of the temperature extremes (top) and precipitation extremes (bottom) for winter (left) and summer (right).The values associated with Dmaxpp0 are multiplied by −1.

The spatial distribution of the linear trends at the 85 stations have been analysed in details, Figure 4 showing an example for the Frtmax90 (winter and summer). The overall climate signal is similar to those obtained from the PC1 (Table 1) with some spatial details, showing patterns similar to EOF1 with the respect to signal magnitude. The number of stations with statistical significance trends varies between seasons and with the statistical method used for the trend test (MK1, MK1 and LLR): significant trends over the entire country in summer, followed by almost the whole country in winter, less extended area in spring and no significant trends in autumn. However, the three methods display a similar climate signal (with respect to the spatial distribution of significant changes) when this signal covers the entire country whatever the case: strong significant (summer) and not significant (autumn). For winter and spring, the LLR gives the lowest number of significant changes at the 5% level (41 in winter, 8 in spring), whereas the MK1 test gives the highest number (82 in winter, 38 in spring). The MK2 test gives an intermediate result (74 in winter and 13 in spring). However, the three methods display similar spatial distribution of the signal, except the spatial extension, showing that the climate signal is reliable. The increase is more enhanced for the frequency of very warm nights during spring and for the frequency of very warm days during summer. Similar characteristics have been identified for the other temperature extremes, except for the maximum duration of consecutive very warm days in spring, when the signal is not statistically significant. For all six temperature extremes, the trends are higher over southern and eastern region in winter and western, southern and southeastern regions in summer. The similarity between the spatial and temporal evolutions of the six temperature extremes over the entire country is obvious, confirming a common mechanism controlling their variability. For simplicity reasons, only the results related to the Frtmax90 will be retained in the following analysis.

Figure 4.

(a) Linear trends (number of days for the period 1962–2010) of the Frtmax90 (top) and Dmaxpp0 (bottom) for winter (left) and summer (right). In case of the Dmaxpp0, the cross pattern areas show trends at the 5% significance level. In case of the Frtmax90, the trends are significant at the 5% level over the entire country. (b) Rates of change (mm per degree temperature rise) in mean and extreme precipitation (maximum daily amount) in response to changes in near-surface temperature for summer (left-daily mean precipitation and right-maximum daily precipitation).

Considering the quite short interval (49 years) used in this analysis, to conclude if the found trends are reliable or they are only short-term (decadal/multidecadal) variations and to enforce these findings, the Frtmax90 computed for 110 years (1901–2010) at the Bucuresti-Filaret station has been analysed. The findings presented above were confirmed in all seasons (Table 1, last row) with some differences related to the climate signal intensity, as expected, considering different time intervals: significant increase in winter, spring and summer, with higher values for winter that is different compared to the results presented above for a shorter interval when summer trends are enhanced. Figure 5 (top) shows the temporal evolution of standardized Frtmax90 at Bucuresti-Filaret for winter and summer. Strong multidecadal variability (especially in summer) can be revealed in both seasons that can be also seen in PC1 (Figure 3). These findings suggest a connection with the AMO as many previous papers have for the European climate extremes (Sutton and Hodson, 2005, Della-Marta et al., 2007; Folland et al., 2009; Gámiz-Fortis et al., 2010; Ionita et al., 2011) as well as for the mean summer temperature at 14 Romanian stations (Ionita et al., 2013). To prove this conclusion, the standardized AMO index (Enfield et al., 2001) is also shown in Figure 5. The two curves are very coherent only on a multidecadal timescale, whereas on an interannual scale there are some differences in the anomaly amplitude, even in the anomaly sign. The physical reasons of the interannual variability will be explained in Section 'Mechanisms controlling the variability of climate extremes in Romania'

Figure 5.

The standardized values of the Frtmax90 (top) and Dmaxpp0 (bottom) calculated for Bucuresti-Filaret station (1901–2010), in comparison with the standardized values of AMO, for winter (left) and summer (right).

As regards precipitation extremes, the climate signal is not as clear as for temperature extremes. First, it should be noted that compared to the temperature extremes presented above (considered as moderate extremes) in the case of precipitation, the 90th percentile threshold gives very intense extremes due to the characteristics of the precipitation regime in Romania represented by many days without rainfall. This makes the application of trend significance tests to these time series more difficult. According to the comments presented in Section 'Methods', the LLR has been applied to Frpp90, MK2 to all the four precipitation extreme indices and MK1 to Dmaxpp0 and Ppmaxd. An example for the spatial distribution of precipitation extreme trends (Dmaxpp0) is presented in Figure 4. Significant increasing trends in the Frpp90 at more stations (11 stations according to MK2 and 6 stations according to LLR) during autumn, Dmaxpp0 during summer (mainly recorded over coherent southwestern areas according to MK2 and more extended to the central part according to MK1) have been detected; the readers can refer to Figure S2 (compared to Figure 4(a), bottom-right). For Ppmaxd, the trend is significant (5% level) at 11 stations in autumn and 8 stations in summer (MK2 test). In the remainder of cases, the linear trends are not significant (according to all the three methods). The overall climate signal is similar to those obtained from the PC1 (Table 1). This result is also in agreement with the characteristics of the temporal variability of the total seasonal precipitation presented by Busuioc et al. (2010) for the same station network and almost the same period (1961–2007). Figure 3 shows the PC1 evolution for the four precipitation extremes, e.g. for winter and summer. Frpp90, Ppmaxd and Dmaxpp90 show very similar variations, whereas Dmaxpp0 reveals an opposite phase but is about in same phase with the temperature extremes. This result shows that the Dmaxpp0, associated with persistent drought intervals, could be (at least partially) controlled by mechanisms that are also common to temperature extremes (more details will be presented in Section 'Mechanisms controlling the variability of climate extremes in Romania'). However, a strong decadal–multidecadal component could be revealed (especially in summer) that is in agreement with previous results presented by Ionita et al. (2011) for the summer drought over Europe. Dmaxpp0 longer time series (1901–2010) computed at Bucuresti-Filaret station revealed components of multidecadal variability that are more enhanced for summer. Figure 5 presents the standardized time series for the Dmaxpp0 in summer and winter, compared to standardized AMO index. It is seen that for summer the picture is similar to those for the Frtmax90, proving the AMO influence on summer persistent drought in Romania at a multidecadal timescale, as Ionita et al. (2011) has shown on the European scale. On an interannual scale there are discrepancies, their reasons are explained in Section 'Mechanisms controlling the variability of climate extremes in Romania'

Considering that the increases of maximum daily precipitation cover large areas over the country (even if only a restricted number of locations are statistically significant at the 5% level), namely, 49% in spring, 57% in winter, 62% in summer and 89% in autumn, we calculated the rates of change (mm per degree temperature rise) in mean and extreme precipitation in response to changes in near-surface temperature for summer (Figure 4(b), left-rate of change in daily mean precipitation and right-rate of change in maximum daily precipitation). It is seen that the signal is at an increasing rate over the majority of locations (more spatially coherent for mean precipitation, as expected) but it is close to 0 for the mean values and much larger for the maximum ones. This result suggests the relative contribution of thermodynamics [related to the Clausius-Clapeyron (CC) Equation] to changes in precipitation characteristics in summer as highlighted by previous theoretical and application studies (Pall et al., 2007; Allan and Soden, 2008; Scoccimarro et al., 2013). According to the CC Equation, an increase in the moisture-holding capacity of the atmosphere of approximately 7% per degree temperature rise is expected. There is some model evidence from global climate models (GCMs) that indeed precipitation extremes increase at the rate predicted by the CC Equation. However, a CC scaling is not obtained in every model and there exists no observational evidence for such a scaling (Lenderink and Meijgaard, 2008). Lenderink and Meijgaard (2008) analysed the 99-year record of 1-hour precipitation observations at one station in The Netherlands with respect to changes in the high percentiles (90th, 99th and 99.9th) in response to observed changes in near-surface temperature. They found that the 99th and 99.9th percentiles exhibit a temperature dependency close to the CC relation for daily mean temperatures roughly below 12 °C, while for higher temperatures this dependency increases to two times the CC relation. A similar study is now in preparation for Romania by analysing subdaily precipitation extremes (that are more connected to the CC scaling, see Lenderink and Meijgaard, 2008) and will further be sent for publication.

To understand the large-scale/regional-scale mechanisms responsible for the characteristics of climate extremes variability presented above (trends and shifts in the mean), the CCA is applied to find plausible connections between spatial patterns of climate extreme anomalies over Romania and spatial patterns of anomalies for various large-scale predictors. These results are presented in detail in the following section for winter and summer. For the reasons presented above, only the Frtmax90, Frpp90 and Dmaxpp0 will be considered as predictands.

3.3 Mechanisms controlling the variability of climate extremes in Romania

The main mechanisms controlling the variability of the three representative climate extremes (considered as predictands) presented above (Frtmax90, Frpp90 and Dmaxpp0) are analysed through the CCA, applied either between one predictor data set and one predictand data set or between combined predictands and/or combined predictors. The tested large-scale predictors have been mentioned in Section 'Data and methods', which were found to be skilful in previous papers, either for Romania mainly in statistical downscaling models (Busuioc et al., 1999, 2006, 2010) or for various studies related to other European regions (e.g. Zorita et al., 1992; von Storch et al., 1993; Xoplaki et al., 2004; Tomozeiu et al., 2007; Busuioc et al., 2008). These predictors refer to dynamic variables (SLP) and thermodynamic ones (T850, SH700 and SH850), their combination is considered as optimum to find plausible connections with the regional climates (Huth, 2003). However, an important aspect is related to the optimum area selected for these predictors to capture the most features of the spatial and temporal variability for the regional climate of interest. This aspect is analysed in Section 'The selection of the optimum predictors' area' and the obtained results are then used in Section 'CCA results', where the CCA is performed.

3.3.1 The selection of the optimum predictors' area

To select the optimum area of the four predictors, the Spearman correlation coefficient between the PC1 of each climate extremes (presented in Section 'Data and methods') and PC1 of each predictor, considering various areas, has been computed and the results are presented in Table 2 as example for winter and summer. The rank (Spearman) correlation coefficient has been chosen instead of the ordinary (Pearson) correlation, as the rank correlation is more robust (e.g. is not sensible to the linear trend). For the simplicity of the presentation, details related to more areas are shown only for the winter season, whereas for the summer season only the optimum areas are included. As the spatial and temporal characteristics of the extremes related to minimum temperature are very similar to those related to maximum temperature, only indices related to maximum temperature are included in Table 2.

Table 2. Spearman correlation coefficient between PC1 of the climate extremes in Romania and various large-scale predictors. For winter, various areas have been tested: A25 (20°–30°E, 40°–50°N), A42 (17.5°–30°E, 37.5°–52.5°N), A49 (17.5°–32.5°E, 37.5°–52.5°N), A187 (5°–45°E, 30°–55°N), A231 (5°W–45°E, 30°–55°N), A275 (15°W–45°E, 30°–55°N). For summer, only the results for the optimum areas are shown. Bolded values show the 5% significance level. Not statistically significant correlations are not included.
PredictorWinterSummer
A25A42A49A187A231A275A25A187
 Frtmax90
SLP0.23       
SH700      0.3 
SH850      0.26 
T8500.840.830.830.38  0.790.58
 Frpp90
SLP   0.740.680.54  
SH7000.60.57    0.57 
SH8500.44     0.41 
T8500.30.430.40.78    
 Dmaxpp0
SLP   0.680.4   
SH7000.550.53    0.4 
SH8500.49       
T8500.20.30.270.67  0.550.37

For winter, T850 is the best predictor for the temperature extremes (correlation between 0.74 and 0.84, stronger connection for Frtmax90) and a smaller area covering Romania ranging between 20°–30°E and 40°–50°N is the optimum one. The SLP (5°–45°E, 30°–55°N) is also statistically significant correlated to temperature extremes but the correlation is much lower than for T850 (0.23–0.28). Regarding precipitation extremes, the SLP and T850 present about similar strong correlation (negative for Frpp90 and Dmaxpp90 and positive for Dmaxpp0), for both predictors the optimum area ranging 5°–45°E, 30°–55°N. The optimum SLP area found here is in agreement with those presented in previous papers when optimum statistical downscaling for winter precipitation total at 14/33 stations have been developed using SLP as predictor (Busuioc et al., 1999, Busuioc et al., 2006). It can be noted that for precipitation extremes, the optimum T850 area is larger than for temperature extremes, being similar to the SLP area. The SH700 is also a good predictor for the precipitation extremes even if the correlation is lower (between 0.35 and 0.60) but it is significant at the 5% level. The optimum area is a smaller one covering the Romanian territory (20°–30°E, 40°–50°N). The specific humidity at 850 hPa has also been tested but the correlation is a little bit lower. Among the three extreme precipitation indices, the frequency of very wet days (Frpp90) seems to be stronger connected with all selected predictors, whereas the maximum duration of very wet intervals (Dmaxpp90) shows the weakest connection. This result seems plausible considering the EOF results (characteristics of spatial and temporal variability) presented in Section 'Spatial characteristics for variability of climate extremes in Romania' (Table 1): Dmaxpp90 shows the highest variability (26 EOFs compared to Frpp90 – 18 EOFs) and the variance explained by EOF1 is twice lower (17% against 36%). This is an additional reason to consider the Frpp90 as representative for precipitation extremes for the next analysis presented in Section 'CCA results'

For summer, SLP is not a good predictor for any extreme indices. T850 is the best predictor for temperature extremes (with the same optimum area as for winter) and a good predictor only for Dmaxpp0 (same optimum area with the temperature extremes), which is in agreement with the results presented in Sections 'Spatial characteristics for variability of climate extremes in Romania' (a part of Dmaxpp0 variability is in-phase with the temperature extremes variability). For the other two precipitation extremes, T850 is not a good predictor that is different for the winter case. SH700 is the best predictor for the precipitation extremes. Concluding for the summer case, we could say that the thermodynamic predictors are optimum for both temperature and precipitation extremes analysed in this article. For temperature extremes, T850 alone could explain their variability characteristics, whereas for precipitation extremes the SH700 has the dominant role, except for maximum duration of dry intervals (Dmaxpp0) that is controlled by a combination of T850 and SH700 anomalies. More plausible explanations will be given in the following section.

3.3.2 CCA results

Before the CCA, we noted the main characteristics of the spatial and temporal variability of the predictors selected (Section 'The selection of the optimum predictors' area') for the CCA in this section. These refer to the PC1 trends for these predictors that have been analysed by Busuioc et al. (2010) and the conclusions can be summarized as follows: PC1 shows a significant increase of SLP in winter, a decrease of SH700 in all seasons (except for summer) and an increase of T850 for all seasons except for autumn. It should be mentioned that the EOF1 pattern shows the same (positive) sign over the selected areas.

3.3.2.1 Single predictand

First, the connection between each climate extreme index (predictand) and each predictor has been established using the CCA, and pairs of the spatial patterns of predictors optimally correlated with spatial patterns of the considered climate extremes have been selected. The time series associated with the most important CCA pairs (mainly the first two pairs showing the highest correlations) are analysed in terms of their trends and shifts in the mean. Secondly, a combination of more predictors was considered for each predictand. For simplicity of presentation only winter and summer are analysed in this article. From the reasons presented before, as representative predictands for the climate extremes mentioned above, the Frtmax90, Frpp90 and Dmaxpp0 have been selected. The summary of the CCA results (canonical correlation coefficient r, explained variance of predictors/predictands – v1/v2, Mann-Kendall statistic Z and the shift in the time series) are presented in Table 3.

Table 3. Results obtained through the CCA applied to the combination of the various predictand (climate extremes) and large-scale predictors, presented for the first two CCA pairs: r-canonical correlation coefficient, explained variance of predictors (v1) and predictands (v2). The last four columns refer to the Mann-Kendall statistic Z and shift in the mean for the time series associated with the first CCA pair. T850* corresponds to A25 area, T850** corresponds to A187 area, using the notations presented in Table 2. Bolded values show the 5% significance level. The number of EOFs of the predictands and predictors used in CCA is presented in brackets. Z values around 0 are noted by ‘-’.
PredictandPredictorCCA pairsCCA1 time series
CCA1CCA2PredictorPredictand
rv1v2rv1v2ZShiftZShift
   Winter       
Frtmax90 (2)T850*(2)0.8380750.231510319852.71987
Frtmax90 (2)T850** (2)0.8133740.4342112.919872.61987
Frtmax90 (3)SLP + T850** (3)0.8929690.7325103.519873.21985
Frpp90 (7)SLP (5)0.8356300.701382.319711.91971
Frpp90 (3)SLP + T850** (5)0.8144300.537121.819710.41971
Frtmax90 + Frpp90 (4)SLP + T850** + SH700 (3)0.8742270.8423253.219872.51982
Frtmax90 + Frpp90 + Dmaxpp0 (6)SLP + T850** + SH700 (4)0.9042240.8425182.719873.01982
   Summer       
Frpp90 (3)SH700 (7)0.7530200.582615
Frpp90 (2)SH700 (2)0.5559430.45158
Frtmax90 (3)T850* (3)0.8876650.7611224.119853.51985
Dmaxpp0 (3)SH700 (5)0.6233340.432461.219702.01984
Frtmax90 + Dmaxpp0 (3)T850* + SH700 (3)0.9036490.512964.719853.41986

For winter, after the application of the CCA between the Frtmax90 anomalies and T850 anomalies (area 20°–30°E, 40°–50°N), it has been found that CCA1 pair (not shown) associates positive Frtmax90 anomalies over the entire country (explained variance 75%) with positive T850 anomalies over the entire area (explained variance 80%) and the time series corresponding to these patterns are strongly correlated (a correlation coefficient of 0.83) with a trend and shift in the mean similar to those found in the PC1 of these parameters (Table 1). The CCA patterns of the two parameters are similar to the corresponding EOF1: T850 anomalies are almost uniform over the analysed area, with higher values over the northern part; Frtmax90 anomalies are higher over the southern-southeastern regions (Figure 2). This result shows that the main mode of winter Frtmax90 variability is mainly controlled by the main mode of the T850 variability over the selected area (anomaly sign) but the spatial pattern of the anomaly magnitude cannot be explained by T850 alone. When a larger T850 area (5°–45°E, 30°–55°N) is considered, the first CCA pair (not shown) associates T850 EOF2 (same sign over the entire area) with Frtmax90 EOF1 but the mechanism is the same as has been presented above, proving that the selected optimum area for this parameter is appropriate. Therefore, the increase in frequency of very warm days in Romania could be explained by the increase in temperature at 850 hPa over the entire area covering Romania. Similar results have been obtained for the mean temperature (Busuioc et al., 2010). The second CCA pair (not shown) associates dipole structures of the two parameters but the time series associated with these patterns display lower correlation (0.23) and do not reveal any trend. Two EOFs have been retained for the CCA in this case. It can be noted that when more EOFs are retained for the Frtmax90, a higher canonical correlation coefficient is obtained but with less explained variance for the canonical Frtmax90 patterns (not shown).

When the combination of SLP and T850 is considered as predictor, the first two CCA pairs both show physically coherent mechanisms (Figure 6), which mainly associate positive T850 anomalies over the Romanian territory with an above normal frequency of very warm days, the circulation type given by the SLP anomalies modulating the magnitude of the Frtmax90 anomalies over various spatial areas as will be shown further. The first CCA pair (correlation of 0.89) associates a dipole structure of the T850 anomalies (with a large area of positive values and a nucleus of the highest magnitude covering Romania, except for a small area with negative anomalies in the southeastern area) and zonal surface circulation (these simultaneous patterns explaining 29% from the total observed variance) with above normal Frtmax90 values in Romania (explained variance of 69%), the highest anomalies are recorded in extra-Carpathian regions (except for the small southeastern area). This behaviour of the predictand spatial distribution is owed to the zonal surface circulation that brings Atlantic warmer air masses to Romania, first affecting the western regions and then, due to the Carpathians' obstacle, the extra-Carpathian regions. The time series associated with these patterns reveal significant increasing trends with similar shifts as presented above. The Frtmax90 pattern is similar to the EOF1 (but with less explained variance, 69% compared to EOF1, 77%, see Figure 2) and trend pattern (Figure 4(a)). This result shows that a higher fraction (69%) of the Frtmax90 increase is explained by this mechanism given by the first CCA pair. The second CCA pair (correlation of 0.73) associates a dipole structure of the T850 anomalies (with larger area of negative values covering Romania) and northerly-northwesterly surface circulation, induced by a positive SLP anomalies covering Romania (common explained variance of 25%), with negative Frtmax90 anomalies over almost the entire country (10% explained variance), stronger negative values are noted in the intra-Carpathian region; considering only the T850 influence (as was presented above), the strongest Frtmax90 negative anomalies should be recorded in the southern area (where the strongest T850 negative anomalies are shown) but due to the northerly surface circulation, colder air masses are transported to Romania, the intra-Carpathian areas are mainly affected. The times series associated with this pair do not show significant trends. The cumulated Frtmax90 explained variance from the first two CCA pairs (69% + 10%) is about the same with the variance explained by the EOF1 (77%) and CCA1 (74%) when only T80 is considered as predictor (Table 3). In conclusion, even if T850 alone could explain the observed temporal variability (increasing trend) of winter Frtmax90 in Romania, the combination of the SLP and T850 in the two mechanisms revealed by the first two CCA pairs can better explain the spatial pattern of trend (including magnitude) for the frequency of very warm days in Romania during the winter season, showing plausible physical mechanisms. This result shows the combination of dynamic (surface circulation) and thermodynamic (T850 anomalies) contribution in controlling the frequency of extreme temperature in Romania, with the former prevailing.

Figure 6.

Patterns of the first two CCA pairs of the combined predictors (left, SLP – contour, T850 – shaded) and Frtmax90 (right). The explained variance is listed on each pattern and the canonical correlation coefficient is placed in the left column. These values are also shown in Table 3.

For precipitation extremes, as presented in Sections 'Spatial characteristics for variability of climate extremes in Romania' and 'Temporal characteristics of the climate extremes variability in Romania', the two indices quantifying the frequency and maximum duration of very wet days (Frpp90, Dmaxpp90) show similar characteristics of spatial and temporal variability and for the simplicity of presentation, only the CCA results for Frpp90 are presented. The maximum duration of the dry intervals (Dmaxpp0) is in opposite phase with the other two precipitation indices and this parameter will be analysed separately and/or in combinations with the other indices. The CCA applied between the Frpp90 anomalies and SLP anomalies reveals two CCA pairs (not shown) explaining plausible physical connections between the two parameters, which are similar to those presented before for the connection between winter precipitation totals in Romania and large-scale circulation represented by the SLP anomalies (Busuioc and von Storch, 1996; Busuioc et al., 1999, 2006), which can be summarized as follows (Table 3): the CCA1 pair associates negative Frpp90 anomalies over the entire country (explained variance 30%) with positive SLP anomalies over the entire area (representing anticyclonic structures, explained variance of 56%). The time series associated with these patterns show very coherent temporal evolutions (correlation of 0.83) and both present a significant increasing linear trend (higher rate for SLP) and the same upward shift in the mean around 1971. The two CCA patterns are similar to the EOF1 patterns for these parameters but with a little less explained variance, especially for Frpp90 (30% against 36% in EOF1, see Table 1). This result could be interpreted as follows: the decrease in frequency of more than normal (equivalent to the increase of less than normal), very wet days is associated with an increase in frequency of anticyclonic structures (equivalent to a decrease in frequency of cyclonic structures) covering Romania. Considering the results presented in Section 'Temporal characteristics of the climate extremes variability in Romania', PC1 of the Frpp90 shows a slightly decreasing trend but not statistically significant. Therefore, the main mode of winter SLP variability alone does not completely explain the temporal characteristics (e.g. trend intensity) of the main variability mode for very wet days' frequency in Romania. An additional mechanism that could explain this might be given by another CCA pair or by including an additional predictor (T850, SH700) that would diminish the significant decrease of Frpp90 revealed by the connection between Frpp90 and SLP alone. Both aspects are analysed. The time series associated with the second CCA pair does not show any significant trend.

In the second step, the combination of T850 and SLP has been considered as predictor for the Frpp90 predictand. The patterns of the first two CCA pairs are presented in Figure 7. The first CCA pair (correlation of 0.81) associates a simultaneous anticyclonic structure over the entire area (with the nucleus covering Romania) and a dipole structure of the T850 anomalies (the whole of Romania is covered by positive anomalies), accounting for 44% explained variance, with negative Frpp90 anomalies over the entire country (30% explained variance, the same as when the SLP alone is the predictor). The time series associated with these patterns reveal the same increasing trend, statistically significant only for the predictor time series. This mechanism seems to reproduce better the observed behaviour of winter Frpp90 temporal variability (namely no significant trend, as well as the spatial variability (the CCA1 pattern for this parameter is similar to the EOF1 pattern).This result shows that the dynamic factor (represented by the SLP anomalies) gives the main mechanism controlling the variability of very wet days' frequency anomalies (i.e. the sign of these anomalies) but their intensity is modulated by the thermodynamic one represented by the T850 anomalies. The CCA1 time series (Figure 7(b)) show a very coherently temporal evolution and also exhibit a decadal variability that could be associated with the NAO. To prove this, the Spearman correlation between CCA1 time series of the Frpp90 and NAO index (also displayed in Figure 7(b)) was computed and a value of 0.46 (statistically significant at the 5% level) was obtained. This result shows that the mechanism given by the CCA1 pattern presented above controls very well the interannual variability of the very wet days' frequency in Romania but, on a longer (decadal) timescale, it seems to be modulated by the NAO. However, as seen in Figure 7(b), this connection is not stable with time. The second CCA pair also shows a plausible physical mechanism (Figure 7): a dipole SLP structure (representing a northwesterly circulation over Romania) simultaneously with a dipole structure of the T850 anomalies (with Romania placed in a larger area of positive anomalies), associates a dipole structure of the Frpp90 anomalies: positive in intra-Carpathian region and negative in the extra-Carpathian regions. This mechanism seems to be plausible from the physical point of view: the northwesterly circulations transport Atlantic moisture air masses to Romania, affecting mainly the intra-Carpathian areas and are blocked by the Carpathian chain with respect to the extra-Carpathian regions. The time series associated with these patterns reveal significant trends (5% level) only for the predictor time series.

Figure 7.

(a) Patterns of the first two CCA pairs of the combined predictors (left, SLP – contour, T850 – shaded) and Frpp90 (right). The explained variance is listed on each pattern and the canonical correlation coefficient is placed in the left column. These values are also shown in Table 3. (b) The time series associated with the first CCA pair (left) and second CCA pair (right) presented in Figure 7(a). The NAO index (multiplied by −1) is included at the CCA1.

For the summer season, according to Table 2 it has been found that the dynamic factor (represented by surface circulation given by the SLP anomalies) is less important for the principal mode of the climate extremes variability in Romania. Therefore, the CCA has been applied separately between each predictor anomalies (excluding SLP) and each representative climate extreme index presented above (Frtmax90, Frpp90 and Dmaxpp0), according to the significant correlations found between PC1 of each of them and the results are presented in Table 2. The results show that the frequency of very warm days is mainly controlled by the T850 field (i.e. the increase in the Frtmax90 is induced by the increase in the T850, similar to winter), whereas the frequency of very wet days (Frpp90) is mainly controlled by the SH700 field; the no significant trend in summer SH700 covering Romania could explain the no significant trend in Frpp90. This result shows the predominance of the thermodynamic contribution on the temporal variability of the very wet days' frequency over Romania in the summer season that is plausible from the physical point of view, considering the mainly convective character of precipitation in this area. The results presented in Section 'Temporal characteristics of the climate extremes variability in Romania' regarding the rate of the extreme precipitation change against the rate of mean precipitation change in summer confirm the results presented above. However, the significant increasing trend in the Dmaxpp0 over large areas in southwestern, western and central regions cannot be explained by the SH700 variability alone, otherwise it would show no significant trend. It seems that the increase in temperature could be responsible for this. This result is in agreement with the results presented by Ionita et al. (2011) on the variability of the European drought. In this article, we try to justify this characteristic in the following through CCA applied between the combination of Frtmax90 and Dmaxpp0 on the one hand, and combined large-scale predictors on the other.

3.3.2.2 Combined predictands

For winter, a combination of the Frtmax90 (frequency of very warm days) and Frpp90 (frequency of very wet days) has been considered as predictand to understand the mechanisms producing simultaneously a significant increase of Frtmax90 over almost the entire country and a decrease (not statistically significant) of Frpp90. The combination of the SLP, T850 and SH700 has been considered as predictor. The patterns of the first two CCA pairs are presented in Figures 8, 9 and show plausible physical mechanisms. The CCA1 pair (Figure 8) associates simultaneously positive SLP anomalies, dipole T850 structure (with larger positive anomalies covering Romania) and negative SH700 anomalies with a pattern of simultaneous negative Frpp90 and positive Frtmax90 over Romania, with the highest magnitude of the anomalies for both parameters in southern and eastern areas (as in EOF1 for both parameters, see Figure 2 for Frtmax90). The explained variance for both combined predictors is 42% (almost the same as for the EOF1 pattern for combined predictors) and predictands is 27% (Table 1). The mechanisms are similar to those presented above when the two parameters have been analysed separately, considering the same combined predictors and they are not reiterated again. The time series associated with these patterns are strongly correlated (0.87) and show a significant increasing trend that could lead to the following conclusion: the significant increase in frequency of simultaneous occurrence of surface anticyclonic structures, above normal temperature at 850 hPa and below normal specific humidity at 700 hPa over Romania, are associated with a significant increase of Frtmax90 and a decrease of Frpp90 over the entire country. Considering the results presented above (cumulated with those presented in Sections 'Spatial characteristics for variability of climate extremes in Romania' and 'Temporal characteristics of the climate extremes variability in Romania'), the found Frpp90 trend is too strong but the predictors' trends are in agreement with those obtained separately for each of them. The second CCA pair (Figure 9) associates simultaneously structures of negative SLP anomalies (with the nucleus placed in the northwestern part), positive T850 anomalies and positive SH700 anomalies (23% explained variance) over the entire area, with simultaneous patterns of positive Frtmax90 and positive Frpp90 anomalies over Romania (25% explained variance). The time series associated with this CCA pair are also strongly correlated (0.84) and show a significant increasing trend for both predictors and predictands, which means a simultaneous increase in the frequency of Frtmax90 and Frpp90. This result, cumulated to those obtained from the first CCA pair (significant increase of Frtmax90 and decrease of Frpp90), leads to a diminishing of the Frpp90 decreasing trend found for the CCA1 pair (while the Frtmax90 increase is amplified), which is in agreement with the result obtained from observations. Therefore, the physical justification of the simultaneously significant increasing trend of very warm winter days' occurrence and no significant trend of very wet days' occurrence could be achieved by a combination of the mechanisms given by the two CCA pairs presented above and shows the combination of the dynamic and thermodynamic contribution to the simultaneous variability of the frequency of extreme temperatures and extreme precipitation in Romania.

Figure 8.

Patterns of the first winter CCA pair of the combined predictors (top, SLP (contour) + T850 (shaded) – (a), SH700 – (b)) and combined Frtmax90 (contour) and Frpp90 (shaded) (bottom, (c)). The explained variances and canonical correlation coefficients are displayed (Table 3).

Figure 9.

As in Figure 8 but for the second CCA pair.

When Dmaxpp0 is added to the combined predictand data set (i.e. Frtmax90 + Frpp90 + Dmaxpp0), similar plausible mechanisms are obtained, given by the first two CCA pairs resembling those presented above (Figures 10, 11, compared to Figures 8, 9), to which Dmaxpp0 is added in the same phase with Frtmax90 in the first CCA pair (Figure 10, bottom-right) and opposite one in the second CCA pair (Figure 11, bottom-right), but always in an opposite phase with Frpp90. The sign of Dmaxpp0 is given by the sign of the SLP and SH700, which are in opposite phase, showing plausible mechanisms: more frequent longer dry intervals are associated with more frequent below normal specific humidity at 700 hPa and surface anticyclonic structures (positive SLP anomalies). The time series associated with the first two CCA pairs, explain in a similar way as presented above when only Frtmax90 and Frpp90 are combined in the predictand data set, the observed trends of the three analysed climate extremes during winter: significant increase in Frtmax90 (more enhanced in southern and eastern regions resulting from overlapping of positive anomalies from CCA1 and CCA2 pairs) and no statistically significant trend in Frpp90 and Dmaxpp0 (resulting from the overlapping of opposite signs of anomalies from the CCA1 and CCA2 pairs). This result could explain the trend pattern of the Dmaxpp0 in winter (Figure 4).

Figure 10.

Patterns of the first winter CCA pair of the combined predictors (left-top: SLP – contour, T850 – shaded; left-bottom: SH700) and combined predictands (right-top: Frtmax90 – contour, Frpp90 – shaded; right-bottom: Dmaxpp0). The explained variance is listed on each pattern and the canonical correlation coefficient is placed in the left column. These values are also shown in Table 3.

Figure 11.

As in Figure 10 but for the CCA2 pair.

For summer, the combination of Frtmax90 and Dmaxpp0 has been considered as predictand and the combination T850-SH700 as predictor, this option is justified by the results presented in Table 2. The first two CCA pairs are presented in Figure 12, showing plausible physical mechanisms. The first CCA pair associates simultaneous patterns of strong positive T850 anomalies (with a nucleus taking the highest values placed in southwestern Romania) and negative SH700 anomalies (except for a small part in southwestern area) with strong positive Frtmax90 anomalies over the entire country and positive Dmaxpp0 anomalies, highest values are recorded in southeastern area (due to largest magnitude of the negative SH700 anomalies) and larger southwestern-central area (probably due to the higher positive temperature anomalies). The smallest Dmaxpp0 anomalies are recorded over the northern and northeastern areas. These patterns explain 49% of the observed combined predictands variance and 36% of the combined predictors variance, being similar to the explained variance of their EOF1 patterns (Table 1). The time series associated with this CCA pair show a strong correlation (0.90) and significant increasing trends (higher for predictors) leading to the following conclusion: the increase in frequency of strong positive T850 anomalies over the entire Romania simultaneously with an increasing frequency of negative SH700 anomalies (equivalent with the SH700 decrease) in summertime, is the main reason explaining the increase of very warm days' frequency (Frtmax90) over the entire country and of the maximum duration of dry intervals (higher anomalies in the southeastern, central and southwestern areas). The predictand patterns are similar to their trend patterns presented in Figure 4. The time series associated with this CCA pair are shown in Figure 13 and exhibit a very coherent evolution but their long-term trends are not always monotonic, suggesting a multidecadal variation similar to AMO (Figure 5). This conclusion is proved by the addition of the standardized AMO index in Figure 13. It is seen that the time series associated with the combined vector of Frtmax90 and Dmaxpp0 exhibit similar behaviour as their PC1 (Figure 3). This result shows that the interannual variability of the summer temperature extremes and prolonged droughts in Romania are mainly controlled by the regional mechanisms given by the CCA1 presented above, overlapped by the Atlantic multidecadal variability. The second CCA pair also shows a plausible mechanism: simultaneous slightly positive T850 anomalies and positive SH700 anomalies (higher magnitude in the southwestern area) are associated with negative Dmaxpp0 anomalies (higher magnitude in the southwestern area) and slightly positive Frtmax90 anomalies. The time series associated with this CCA pair do not reveal significant trends.

Figure 12.

Patterns of the first two CCA pairs of the combined predictors (left, SH700 – contour, T850 – shaded) and combined predictands (right, Frtmax90 – contour, Dmaxpp0 – shaded) for summer. The explained variance is listed on each pattern and the canonical correlation coefficient is placed in the left column. These values are also shown in Table 3.

Figure 13.

Time series associated to the first CCA pair for summer (Figure 12): dotted line - combination between T850 and SH700; grey line - combination between Frtmax90 and Dmaxpp0; black line - standardised AMO index.

In conclusion, the results presented above show the importance of the combined predictors/predictands in the CCA, which provides, even through a statistical technique, more physically plausible explanations on large-scale/regional-scale mechanisms controlling the spatial and temporal variability of climate extremes in Romania. In winter, these mechanisms show that the thermodynamic factor (represented here by air temperature anomalies at 850 hPa) mainly controls the variability of temperature extremes in Romania (i.e. the anomaly sign), whereas the dynamic one (represented here by the surface circulation, i.e. SLP anomalies) roughly controls the magnitude pattern of the temperature extremes. Regarding precipitation extremes, the roles of the two factors are reversed. The specific humidity completes the picture of the mechanisms controlling the extreme precipitation variability. In summertime, the thermodynamic factor is dominant for both temperature and precipitation extremes analyzed in this paper. For temperature extremes, T850 alone could explain their variability characteristics, whereas for precipitation extremes the SH700 has the dominant role, except for the maximum duration of dry intervals (Dmaxpp0), which is controlled by a combination of T850 and SH700 anomalies. This result is in agreement with previous studies showing the convective nature of summer precipitation in Romania (NMA, 2008, Busuioc et al., 2010).

4 Conclusions and discussions

Ten indices associated with six temperature extremes and four precipitation extremes in Romania (quantifying their intensity, frequency and duration), computed at a high spatial resolution (85–98 stations) for the period 1961–2010 are analysed to identify the main spatial/temporal characteristics of their simultaneous variability on the one hand and to understand the large-scale mechanisms responsible for this variability on the other. Through the CCA, the connection between these indices (either used separately or in various combinations) and three large-scale predictors (represented by SLP, T850 and SH700), used both separately or in various combination, is examined. The main conclusions can be summarized as follows:

  1. Significant increasing trends for the six temperature extremes (Frtmax90, Frtmin90, Dtmax90, Dtmin90, Tmax and Tmin) have been detected in all seasons, except for autumn and Dtmax90-spring. The increase rate is more enhanced in summer and less enhanced in spring. Regarding the spatial pattern of the trend magnitude, some differences can be revealed between winter and summer. In wintertime, the increasing trend is higher over the extra-Carpathian regions (more pronounced in southern and southeastern regions) and is not significant over some small intra-Carpathian regions. In summer, the trend is significant over the entire country, showing an almost homogeneous magnitude, the results are in agreement with those presented by Efthymiadis et al. (2011). A multidecadal variability can be revealed, especially in summertime, when longer time series are examined (e.g. Frtmax90 at Bucuresti-Filaret station, 1901–2010) that could be associated with AMO, as many previous papers have shown for the European climate extremes (Sutton and Hodson, 2005, Della-Marta et al., 2007, Folland et al., 2009, Ionita et al., 2011) as well as for the mean summer temperature at 14 Romanian stations (Ionita et al., 2013).
  2. Regarding precipitation extremes, the climate signal is not as clear as for temperature extremes. Significant increasing trends over some areas in the Frpp90 during autumn, Ppmaxd during summer and autumn, Dmaxpp0 during summer (mainly recorded in a coherent southwestern, central and western areas) were detected. In the remainder of cases, the linear trends are not significant (or are significant only for small areas). However, a strong decadal/multidecadal component could be revealed, more evident for longer time series (1901–2010) that is in agreement with the findings presented on the European scale by Briffa et al. (1994, 2009), Ionita et al. (2011, 2013) and could be associated with NAO in wintertime and AMO in summertime. Even if the increase in maximum daily precipitation is significant at the 5% level only on a limited number of stations in summer, the rate of change in maximum daily precipitation in Romania in response to changes in near-surface temperature is increasing at most of the locations and is much larger than that in the daily mean precipitation (which is around 0). This result suggests the relative contribution of thermodynamics (related to CC Equation) to changes in precipitation characteristics in summer as highlighted by previous theoretical and application studies. However, this aspect will be thoroughly analysed in a future paper, which is in preparation.
  3. All indices show similar principal modes of variability (i.e. the same sign over the entire country, with a higher magnitude over the southern and eastern areas in wintertime and generally over the southwestern-western areas in summertime) but their fraction of explained variance is higher for temperature extremes (especially Frtmax90); this result shows that a single large-scale mechanism is responsible for the temperature extremes variability, whereas for the precipitation extremes it is more difficult to find such a mechanism. The second mode exhibits a dipole structure with differences between winter and summer: in winter the Carpathian Chain influence is noted, with an opposite variability between the intra-Carpathian and extra-Carpathian regions, whereas in summer a northeast-southwest gradient can be revealed.
  4. The multifield EOF analysis applied to the combination of several climate extremes reveals, as principal mode, the same sign of simultaneous variability for all thermal extremes and the opposite variability between thermal and precipitation extremes, except for Dmaxpp0, which shows opposite sign against the other precipitation extremes but generally is in-phase with temperature extremes.
  5. The CCA applied between combined large-scale predictors and combined selected climate extremes in Romania (Frtmax90, Frpp90 and Dmaxpp0) was noted to be a skilful tool to find plausible explanations from the physical point of view of the large-scale mechanisms responsible for the characteristics of the climate extremes variability, especially the simultaneous variability of several climate extremes, as presented below. To our knowledge, this technique has been applied for the first time, at least in the climate extreme analysis. Della-Marta et al. (2007) and Xoplaki et al. (2004) used a similar technique but only for combined predictors and a single predictand.
  6. For winter, it was found that the thermodynamic factor (represented here by air temperature anomalies at 850 hPa) mainly controls the temperature extremes in Romania, whereas the dynamic one (represented here by the surface circulation, e.g. SLP anomalies) roughly controls the magnitude pattern of the temperature extremes. Regarding the precipitation extremes, the role of the two factors is reversed. The specific humidity completes the picture of this mechanism. Therefore, the significant increase in the temperature extremes and not statistically significant trend in precipitation extremes in wintertime could be explained by the combination of two plausible mechanisms given by the first two CCA pairs, with very strongly correlated associated time series (0.90, 0.84), as follows: more frequent positive T850 anomalies over Romania simultaneous with more frequent anticyclonic structures and negative SH700 anomalies are associated with a significant increase in the frequency of very warm days and the length of dry intervals, as well as decrease in frequency of very wet days (CCA1); the predictand CCA1 patterns are similar to their trend patterns (Figure 4); more frequent cyclonic structures over Romania simultaneous with positive T850 anomalies and positive SH700 anomalies are associated with a significant increase in the frequency of very warm and wet days and shorter dry intervals (CCA2).
  7. In summertime, the thermodynamic factor is dominant for both temperature and precipitation extremes analysed in this article. For temperature extremes, the T850 alone could explain their variability characteristics, whereas for precipitation extremes (frequency and duration) the SH700 has the dominant role, except for the maximum duration of dry intervals (Dmaxpp0) that is controlled by a combination of T850 and SH700 anomalies. This result is in agreement with the previous studies related to the convective nature of summer precipitation in Romania (NMA, 2008, Busuioc et al., 2010). Therefore, the simultaneous significant increase in the summer frequency of very warm days and in the maximum length of dry intervals (higher anomalies in the southeastern, central and southwestern areas) can be mainly explained by the increase in frequency of strong positive T850 anomalies over the entire Romania simultaneously with an increasing frequency of negative SH700 anomalies (equivalent with the SH700 decrease). The predictand patterns are similar to their trend patterns presented in Figure 4.

Finally, we concluded that the analysis of the simultaneous variability of several climate extremes in Romania, in connection with the simultaneous variability of physically connected large-scale climate variables, provides a useful tool to find more plausible physical mechanisms explaining the observed changes in the regime of climate extremes in Romania. The connections found in this study are strong and explain a great part of the total observed variance, showing that these results can be used in a future study to build skilful statistical downscaling models, simultaneously for several seasonal climate extremes, giving the results more physical coherence, which could add value to these models.

Acknowledgements

This study was funded by the Executive Agency for Higher Education, Research, Development and Innovation Funding (UEFISCDI) through the research project CLIMHYDEX ‘Changes in climate extremes and associated impact in hydrological events in Romania’, cod PNII-ID-2011-2-0073 (http://climhydex.meteoromania.ro). Two anonymous reviewers are acknowledged for their useful comments.

Ancillary

Advertisement