Theor. Appl. Climatol. (2007) DOI 10.1007/s00704-006-0275-z Printed in The Netherlands 1 2 ARPA-SIM, Bologna, Italy National Meteorological Administration, Bucharest, Romania Climate change scenarios for surface temperature in Emilia-Romagna (Italy) obtained using statistical downscaling models R. Tomozeiu1 , C. Cacciamani1 , V. Pavan1 , A. Morgillo1 , and A. Busuioc2 With 12 Figures Received February 9, 2006; revised July 29, 2006; accepted August 25, 2006 Published online December 28, 2006 # Springer-Verlag 2007 Summary Possible changes of mean climate and the frequency of extreme temperature events in Emilia-Romagna, over the period 2070–2100 compared to 1960–1990, are assessed. A statistical downscaling technique, applied to HadAM3P experiments (control, A2 and B2 scenarios) performed at the Hadley Centre, is used to achieve this objective. The method applied consists of a multivariate regression based on Canonical Correlation Analysis (CCA), using as possible predictors mean sea level pressure (MSLP), geopotential height at 500 hPa (Z500) and temperature at 850 hPa (T850), and as predictands the seasonal mean values of minimum and maximum surface temperature (Tmin and Tmax), 90th percentile of maximum temperature (Tmax90), 10th percentile of minimum temperature (Tmin10), number of frost days (Tnfd) and heat wave duration (HWD) at the station level. First, the statistical model is optimised and calibrated using NCEP=NCAR reanalysis to evaluate the large-scale predictors. The observational data at 32 stations uniformly distributed over Emilia-Romagna are used to compute the local predictands. The results of the optimisation procedure reveal that T850 is the best predictor in most cases, and in combination with MSLP, is an optimum predictor for winter Tmax90 and autumn Tmin10. Finally, MSLP is the best predictor for spring Tmin while Z500 is the best predictor for spring Tmax90 and heat wave duration index, except during autumn. The ability of HadAM3P to simulate the present day spatial and temporal variability of the chosen predictors is tested using the control experiments. Finally, the downscaling model is applied to all model output experiments to obtain simulated present day and A2 and B2 scenario results at the local scale. Results show that significant increases can be expected to occur under scenario conditions in both maximum and minimum temperature, associated with a decrease in the number of frost days and with an increase in the heat wave duration index. The magnitude of the change is more significant for the A2 scenario than for the B2 scenario. 1. Introduction Climate changes have a large impact on ecosystems, the environment and human activities. These changes consist of shifts not only in mean values, but also in the frequency and intensity of extreme weather events. All these components must be estimated in order to produce a complete evaluation of the impacts of climate change over a specific region. During recent years, great attention has been paid to the study of extreme weather events and their impacts, especially after the occurrence of several extreme events which affected large regions, such as the summer floods of Central Europe in 2002, and the intense climate anomaly which affected almost the entire European continent in summer 2003, including both heat waves and intense drought. R. Tomozeiu et al. A correct description of these events requires information on the main surface fields at the local scale, which is also the scale at which input is needed for impact studies. This is due to the fact that the greatest impacts are often linked to the occurrence of local phenomena, characterised by very high spatial and temporal variability. The best tools available from the climate community to evaluate the pattern and intensity of climate change are general circulation models (GCMs). Given the current computational resources, the production of long (order 100 years) integrations of these models makes it necessary to run them at spatial resolution of the order of 100 km. Information at this scale can be used, only to evaluate climate changes connected with shifts in mean values over relatively large regions. In order to evaluate changes at the local scale and changes in the frequency of extremes, further efforts are needed to increase the spatial and temporal resolution of the final prediction product. Dynamical, statistical–dynamical, and statistical downscaling methods are used to achieve this objective. Dynamical downscaling uses limited-area, highresolution models (regional climate models – RCMs) driven by boundary conditions from a GCM to derive local-scale information. RCMs cover a selected region with limited spatial extent at higher resolutions (up to 25 km). In the last decade, RCMs have been used to reproduce the European climate, both at present day (Giorgi et al., 2004a) and in the future under different possible scenarios (Christensen and Christensen, 2003; Giorgi et al., 2004b). Statistical–dynamical downscaling uses observational data, to identify specific events representative of weather types characterising the climate of a region; uses RCMs, to describe these events and their possible impacts at the local scale, and uses GCMs, to evaluate possible future changes in the frequency of occurrence of these weather types (Fuentes and Heimann, 2000). Statistical downscaling (SD) is a complementary technique which develops statistical relationships that link large-scale atmospheric variables (predictors) and local=regional climate variables (predictands). This relationship is then applied to the large-scale predictors simulated by GCMs, under various emission scenarios, in order to obtain local climate-change information. Some studies have compared the SD and RCM approaches (Murphy, 2000; Huth et al., 2001; Schmidli et al., 2006). The conclusion derived from these studies is that SD and dynamical downscaling techniques are comparable for simulating current climate but the climate change projections in some cases are quite different, especially regarding the magnitude of the climate change signal (temperature) or even in terms of the sign of the climate signal (e.g. precipitation). Both approaches are currently widely used to produce regional climate change scenarios, and the differences between them are used to better asses the uncertainty associated with the use of these different techniques. Both approaches have advantages and disadvantages. The main advantage of the dynamical approach is its physical basis, even if there are some simplifications due to some missing feedbacks. The main drawback is its computational cost, which still substantially limits the possible RCM resolution and length of the experiments produced, reducing the capability of RCMs to capture the frequency and, to a greater extent, the changes in frequency of extreme weather events. In contrast, the main advantages of SD techniques are that they are computationally inexpensive, can be used to derive variables not available from RCMs, and allow the direct downscaling of indices of frequency of extreme weather events, even up to local scale. The SD disadvantages refer to the fact that they need long and homogeneous observational time series for the fitting and validation of the statistical relationship. There are different types of statistical downscaling that can be grouped in three categories: weather generators, weather classification, and regression models. Weather generators (Busuioc and von Storch, 2003; Katz et al., 2003; Buishand et al., 2004) and SDs based on weather classification (Zorita and von Storch, 1999; Palutikof et al., 2002) are focused on the daily time scale. Regression models represent nonlinear or linear relationships between predictands and large-scale predictors. The non-linear approaches include artificial neural networks, which allow the fitting of a more general class of statistical model (Trigo and Palutikof, 2001; Cavazos et al., 2002). Linear models are the most popular, and include multiple regression (Palutikof et al., 2002; Hellstr€om and Chen, 2003; Hanssen-Bauer et al., 2003), models based on canonical correlation Climate change scenarios for surface temperature in Emilia-Romagna analysis (CCA) (Von Storch et al., 1993; Busuioc et al., 2001; Hellstr€ om and Chen, 2003; Busuioc et al., 2005) or singular value decomposition analysis (SVD) (Huth, 2002; Widmann et al., 2003). A comparison between the three linear methods (CCA, SVD and Multiple linear regression – MLR) applied to eight winter daily temperature at 39 stations in central parts of western Europe, has been made by Huth (2002). One advantage of CCA and SVD is that they seek pairs of patterns that are optimally correlated making possible a physical interpretation of the connection between predictands and predictors. Because of this advantage this method is also used to validate GCM output (e.g. Busuioc et al., 2001). The study also reveals that CCA is a good method for regionalization while MLR provide good results in the reproduction of time structure. Similar studies of comparisons of results from different downscaling techniques have been performed in Scandinavia by Hanssen-Bauer et al. (2005). They compared the above linear techniques and concluded that overall no method is superior as long as the information content in the predictors is similar. The choice of predictors, predictor domains and other strategic choices are more critical for the results than the choice of linear technique. The development of statistical downscaling methods, as well as the identification of the more robust downscaling techniques, and their application in order to obtain scenarios of extremes and mean values for different European regions was one of the main objectives of the European project STARDEX (http:==www.cru.uea.ac.uk=cru= projects=stardex=). STARDEX, together with PRUDENCE and MICE, formed a cluster of three projects, all related to changes in mean and extremes of climate and their impacts in different European areas. In the present study, a statistical downscaling method (SD) based on Canonical Correlation Analysis (CCA) has been used in order to construct future scenarios of changes in seasonal mean and extremes of minimum and maximum temperature at 32 stations in Emilia-Romagna. In order to build the model, that is to quantitatively identify the relation between large-scale variability and local climate, the approach used in the present paper is a ‘‘perfect – prog’’ approach, that is, first, the statistical downscaling model is built using observational data only, then, the model is applied to the output of GCM experiments so as to reproduce local climate characteristics or to evaluate local future scenarios. The models proposed here use a selection of fields between mean sea level pressure (MSLP), 500 hPa geopotential height (Z500), and temperature at 850 hPa (T850) as predictors. The predictands are seasonally averaged minimum and maximum air temperature (Tmin and Tmax, respectively), the 10th percentile of minimum air temperature (Tmin10), the 90th percentile of maximum air temperature (Tmax90), the number of frost days (Tnfd), and the heat wave duration (HWD), all evaluated at the station level. The influence of these predictors on the Italian climate during the winter season has been described by Cacciamani et al. (1994), Pavan et al. (2005), and Tomozeiu et al. (2005), and for summer by Colacino and Conte (1995), while the present study represents the first contribution on the same issue with respect to the others seasons. The main objective of this work is to produce local climate change scenarios (period 2070– 2100) of mean and extreme air temperature in Emilia-Romagna, using HadAM3P model simulations and statistical downscaling models. This aim is achieved through the realization of the following steps: (1) detect the connection between large-scale circulation patterns derived from NCEP=NCAR reanalysis and observed minimum= maximum temperature in Emilia-Romagna; (2) develop an optimal statistical downscaling model based on the above relationship; (3) validate the HadAM3P output in terms of its ability to reproduce the selected predictors and (4) construct scenarios of extremes and mean temperature based on optimal statistical downscaling models, applied to HadAM3P output. This paper is organized as follows: Section 2 presents data and methods used. The selection of the predictors as well as the set-up of the SD models using NCEP=NCAR reanalysis is presented in Sect. 3. The ability of the HadAM3P model to simulate the predictors is presented in Sect. 4. The robustness of the downscaling models is presented in Sect. 5. In Sect. 6, scenarios of changes in mean and extreme temperature over Emilia-Romagna for the period 2070–2100 compared to 1960–1990, are presented. The A2 and B2 IPCC (Intergovernmental Panel on Climate R. Tomozeiu et al. Change) SRES (Special Report on Emissions Scenarios) scenarios are considered. Conclusions are presented in Sect. 7. 2. Data and method 2.1 Local scale data The observational data set used in this study consists of time series of daily minimum and maximum temperature (Tmin and Tmax) at 40 stations located in Emilia-Romagna, a region of Northern Italy. Figure 1 shows the map of the region including orography (right) together with its location within the Italian Peninsula (left). Data have been collected by the offices of the former Italian Hydrographic Service (recently incorporated by ARPA-SIM) at Bologna and Parma, and cover the period 1958–2000. The data is quality controlled (Pavan et al., 2003) and homogeneity tested (Tomozeiu et al., 2005). Taking into account only the stations more than 80% complete and those that pass the homogeneity control, the number of stations for which extremes are computed is reduced to 31 (Fig. 1, right). As can be observed, the remaining stations are uniformly distributed over the region. A set of six indices the predictands are calculated at the seasonal level using daily data: mean Tmax, 90th percentile of maximum temperature (Tmax90), mean Tmin, 10th percentile of minimum temperature (Tmin10), number of frost days (Tnfd) and heat wave duration index (HWD). The extreme indices are computed for each station and each season during the period 1958–2000. Seasons are defined in the standard format: Winter as December to February (DJF); Spring as March to May (MAM); Summer as June to August (JJA); Autumn as September to November (SON). The number of frost days (Tnfd) is defined as the number of days with Tmin <0 C while the heat wave duration (HWD) is defined as the maximum number of consecutive days with Tmax greater than Tmax90. Spatial and temporal variability of these observed indices have been studied in a previous paper (Tomozeiu et al., 2005). 2.2 Large scale data The large-scale predictors include mean sea level pressure (MSLP), geopotential height at 500 hPa (Z500) and temperature at 850 hPa (T850), which have been extracted from the NCEP=NCAR re-analysis (Kalnay et al., 1996). This data set refers to the monthly mean values of these variables at 2.5 2.5 horizontal resolution for the window 90 W–90 E; 0 –90 N, and for the period 1958–2000. The GCM simulations used are those produced by the Hadley Centre using the HadAM3P model. Fig. 1. Map of Italy indicating the position of Emilia-Romagna (left), the stations used in this study (marked by cicles) and the orography of the Emilia-Romagna (shaded area) (right) Climate change scenarios for surface temperature in Emilia-Romagna They include present day (1960–1990) and A2 and B2 scenario experiments (2070–2100). The HadAM3P is an atmosphere only general circulation model with 19 vertical levels and a horizontal resolution of 2.5 3.75 , comparable to a spectral resolution of T42. The sea surface temperatures used to drive this model are obtained from observations, for present day simulations, and from the HadCM3 ocean-atmosphere coupled model, for scenario simulations. A detailed description of the HadAM3P model and of the experiments used in this paper is presented in Pope et al. (2000), and Johns et al. (2003). The daily Z500, MSLP, T850 data for the control run and scenario experiments have been provided by the STARDEX project at the same resolution as NCEP=NCAR reanalysis (2.5 2.5 ). The model daily data were interpolated to 2.5 2.5 using the ‘‘bivar’’ method described by Akima (1984). 2.3 The statistical downscaling (SD) method A group of statistical models based on the Canonical Correlation Analysis (CCA) have been constructed in order to obtain future scenarios of mean and extreme temperature at the local scale in Emilia-Romagna using different subsets of all possible predictors. The CCA method was first introduced to climate research by Barnett and Preisendorfer (1987). This method identifies predictor-predictand pairs of patterns, which maximise the correlation between two corresponding patterns (Von Storch and Navarra, 1995). In addition, the method offers a physical interpretation of the mechanism that controls the regional climate variability (Von Storch et al., 1993; Busuioc et al., 2001). In order to reduce the noise of the fields involved, before the CCA is applied, the data sets are projected on EOFs (empirical orthogonal functions) and only those explaining the majority of the total observed variance are retained (Von Storch and Navarra, 1995). A subset of CCA pairs is then used in a multivariate linear model in order to estimate the predictand anomalies from the predictor anomaly field (Barnett and Preisendorfer 1987; Von Storch et al., 1993; Busuioc et al., 2001). In the present paper, models are built for each season and index, choosing each time a different subset of predictors from the fields extracted from the NCEP=NCAR reanalysis. All data are de-trended before use. All models are calibrated on the period 1958–1978 and 1994–2000 and validated on the period 1979–1993, and only the best performing model is retained. The performance (skill) of the downscaling model is quantified at the station level in terms of: (i) Spearman rankcorrelation coefficient (CORR) which is just the correlation coefficient calculated on the ranks of the two time series, (ii) root-mean square-error (RMSE) between observed and simulated index with bias removed and (iii) BIAS, defined as follows: RMSE ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u u1 X u ½indicesmod el ðiÞ indicesobs ðiÞ BIAS2 t N i 2 verification period ðiiÞ BIAS ¼ hindicesmod el iverification hindicesobs iverification ðiiiÞ The work done in order to select optimum statistical downscaling models, using predictors from NCEP=NCAR reanalysis, shows that the skill of the downscaling models is dependent on: 1) the predictors (large-scale field, single or combined); 2) the domain (area) of predictors; 3) the number of EOFs retained for the CCA and the number of CCA components used in the regression model. The selection of the predictors is done so as: a) to have strong, robust and physically meaningful relationships with the predictands; b) to have stable and stationary relationships with the predictands; Table 1a. Areas of definition of the predictors used within the SD models Code Area Area Area Area Area Area Area (Long.=Lat.) A B C D E F 90 W–90 E=0 N–90 N 60 W–60 E=20 N–90 N 35 W–35 E=30 N–60 N 12.5 W–30 E=30 N–55 N 5 E–35 E=30 N–50 N 5 W–20 E=37.5 N–50 N R. Tomozeiu et al. Table 1b. Predictors used in the optimal SD model for each predictand and season Predicatand DJF MAM JJA Tmax Tmax90 Tmin Tmin10 Fd HWD T850=area C T850 þ MSLP=area B T850=area C T850=area C T850=area C Z500=area D T850=area C Z500=area F MSLP=area D T850=area C T850=area C Z500=area E T850=area T850=area T850=area T850=area – Z500=area c) to be well reproduced by the HadAM3P model; d) the SD model is able to explain the observed low-frequency variability and trends of the predictands. Points a) and b) above are achieved by applying the CCA using different time period definitions for the calibration period, in order to check the robustness of the CCA results. The analysis done shows that the application of the method using calibration periods of either 1958–1980 or 1981– 2000 results in the identification of similar predictor-predictand patterns and in comparable strength of predictor-predictand connections, suggesting that the predictor-predictand relationships described in the following are robust and stationary. Finally, in all cases, the ability of the HadAM3P model to reproduce the observed variability of the predictors is checked by comparing the model EOFs with re analysis EDFs for each predictor field. The predictor definitions can be altered either by changing the large-scale field used or the domain used as space-window in the PC filter (Benestad, 2001). In order to maximise the performance of the statistical model, several possible domain extensions have been examined in the predictors definition, for each season and predictand. Table 1a reports the code used to identify the different domains for the large-scale predictors, depending on the season and the predictand considered, while Table 1b reports the predictor definitions used within the best performing model for all predictands and seasons. This last table reveals that, in general, all surface fields are best predicted using T850 except for: winter Tmax90 and autumn Tmin10, best predicted by a combination of T850 and MSLP, spring Tmax90 and, winter, spring and summer HWD, best predicted by Z500. SON C E C D E T850=area C T850=area C T850=area C T850 þ MSLP=area C T850=area C T850=area E The sensitivity of the models to the number of EOFs retained for the CCA, and to the number of CCA components used in the regression has also been checked. A model is considered optimal if it satisfies the condition that its skill (CORR, BIAS, RMSE) does not change significantly after the addition of one component (CCA). It is known that, in general, the addition of more canonical modes enhances the accuracy first, but after adding a certain number of EOF=CCA patterns the accuracy declines slightly (Busuioc et al., 2001; Huth, 2002). Results with respect to this issue are shown for a limited number of cases for reasons of brevity. Finally, the SD models derived from the observations are applied to the HadAM3P output, from present day experiments (1960–1990) and experiments done using different future IPCC SRES scenarios (2070–2100). In order to do this, the model anomalies with respect to the control model climate are projected onto each CCA predictor pattern. The full time series for the A2 and B2 scenarios is computed by adding downscaled anomalies to the observed mean. Climate changes in the predictands presented above are calculated as the difference between the mean of the downscaled values for the period 2070–2100 (A2 and B2 scenarios) and for present day climate. The statistical significance of the changes is also evaluated using the Student’s t-test. 3. The statistical downscaling models 3.1 Two examples of links between large-scale patterns and local climate As previously mentioned, a by-product of the CCA method is that it automatically identifies the link between large-scale variability and local climate. Given the great number of models presented here (different depending on season and predictand), Climate change scenarios for surface temperature in Emilia-Romagna Fig. 2. First two CCA pairs of T850 (a and c) and Tmin10 (b and d) for winter for brevity, only a limited number of models are described in detail, highlighting the dynamical and physical interpretation of the predictor-predictand link. In the following, the predictand-predictors connection is presented for two extremes: winter Tmin10 and summer HWD. Figure 2 (a–d) shows the first two CCA pairs of T850-Tmin10 connection obtained by retaining the first four EOFs of winter T850 and four EOFs of winter Tmin10. The variance explained by the first four EOFs of T850 and Tmin10 is presented in Table 2a. As can be observed, together the patterns explain around 90% of the total variance for both predictor and predictand. Figure 2a and b present the first canonical correlation anomaly pattern (CCA1) of T850Tmin10, resultant from the above combination, characterised by a correlation coefficient between fields of 0.73 and by a fraction of explained Table 2a. Fraction of total variance (%) explained by the first four EOFs of winter T850 and Tmin10 EOF1 EOF2 EOF3 EOF4 T850 (%) Tmin10 (%) 42 33 10 5 79 6 2 2 Table 2b. Mean skill over Emilia-Romagna of winter Tmin10 for different SD set-up No. EOFs predictor= predictand; No. CCA in SD BIAS RMSE Spearman correlation 4EOFs T850=4EOFs Tmin10; 2CCA 6EOFs T850=7EOFs Tmin10; 3CCA 0.74 0.38 0.6 0.71 0.39 0.4 R. Tomozeiu et al. variance around 15% for T850. The CCA1 of Tmin10 explains a great part of the total variance, 55%, and represents the main pattern of covariability for the T850-Tmin10 couple. This is associated with a negative (positive) anomaly of T850 centred over the Eastern Mediterranean and Turkey, extending to Northern Italy, with anomalies of Tmin10 with the same sign over Emilia-Romagna (Fig. 2b), with maximum amplitude located over the plains. The T850 anomaly is typically associated with a large-scale circulation pattern characterised by a strong positive Z500 anomaly which extends from Eastern Europe to Russia, advecting cold continental air over the Eastern part of the Mediterranean and South-Eastern Europe. This pattern leads to a northward shift of the European end of the Atlantic mid-latitude jet, and over Northern Italy it is often associated with clear sky nights and with temperature inversions. During these episodes, anomalously low values of Tmin10 can be observed over the plain, while minimum temperature can reach very high values over the mountains, with anomalies up to þ10 C, well over the 50th percentile. This justifies the fact that the maximum amplitude of this pattern is located over the plain. Under the opposite phase, this CCA is linked with a negative Z500 anomaly over Eastern Europe and Russia, with an increase in T850 over the Eastern Mediterranean, positioned well under the southern flank of the jet, and, in Emilia-Romagna, with milder Tmin10 values, especially over the plains, protected by clouds. The second CCA (Fig. 2c and d) explains 40% of the T850 total variance and 18% of the Tmin10 total variance. The time series associated with the predictor=predictand patterns are significantly correlated at 0.55. The T850 pattern is characterised by a negative (positive) anomaly centred over south-central Mediterranean and Northern Africa. This leads to negative (positive) anomalies of Tmin over Emilia-Romagna with maximum amplitude over the mountains, and large-scale circulation characterised by the presence of an anomaly in the intensity (weakening in the phase represented in Fig. 2c) of the westerly jet, with its main axis located over the British Isles and Northern Europe. Under these conditions, the temperature is normally stratified, the region often experiences transient synoptic systems of Atlantic origin, and greater anomalies of minimum temperature are generally observed over the mountains, more exposed to upper air cold (warm) advection from North-West (South, South-East). The CCA1 and CCA2 of T850 are the most important patterns controlling the variability of winter T850. The third and fourth CCA patterns are characterised by lower values of the predictorpredictand correlations than the first two CCA pairs and present a dipolar structure (maps not shown). The second row of Table 2b shows the skill in terms of BIAS, RMSE and Spearman correlation for the SD model obtained using only the first two CCAs. In order to examine the influence of the number of EOFs retained in the construction of CCA, the analysis is repeated using the first six EOFs of T850 and seven EOFs of Tmin10 explaining 97.5% from the total variance of predictors= predictands. Then, the first three CCA patterns with the associated time series (that are significantly correlated) are used in the construction of the SD model. Analysing the skill from Table 2b that presents both SD versions (row 2 and row 3) it can be observed that, while no significant changes are present in the BIAS and RMSE, an improvement is observed in the Spearman correlation when a combination of four EOFs of both predictor and predictand is used in the CCA analysis and only the first two CCA pairs are selected in the construction of the SD model. This type of sensitivity test has been done for each predictand=predictor couple in order to find the optimum SD model. The relationship between HWD and Z500 during summer is presented in Fig. 3. In this case, the first three EOFs of Z500, explaining together 80% of the total variance and the first three EOFs of HWD, explaining together 75% of the total variance, are used to filter the data prior to CCA. Figure 3a and b show the CCA1 of the Z500HWD pair. The time series associated with these patterns (not shown) are significantly correlated (0.7) while the fraction of total variance explained by each pattern is 32% for Z500 and 49% for HWD. The CCA1 of Z500 (Fig. 3a) is characterised by a dipole with maxima located over Central Europe and minima over Northern Africa. This is the typical anomaly circulation pattern associated with long summer heat waves over Central and Southern Europe, similar to the Climate change scenarios for surface temperature in Emilia-Romagna CCA HWD (JJA) CCA1 Z500 (JJA) 50N 45N 40N 35N 30N 5W 0 5E 10E 15E 20E 25E 30E 35E a b CCA1 Z500 (JJA) CCA HWD (JJA) 50N 45N 40N 35N 30N 5W 0 5E 10E 15E 20E 25E 30E 35E c d Fig. 3. The first two CCA pairs of Z500 (a and c) and HWD (b and d) for summer large-scale anomaly which characterised summer 2003 over the Euro-Atlantic sector (Cassou et al., 2005; Colacino and Conte, 1995). In EmiliaRomagna it leads to a substantial increase in the length of the HWD index, with maxima of about two days located over the hills in the central-eastern part of the region, the areas which are most frequently affected by the convective break up of dry spells. The second CCA pair (Fig. 3c and d) is characterised by a Z500 pattern explaining 28% of the total variance, which presents a dipolar structure with centres located over North-Eastern Europe and Northern Africa-Western Mediterranean, leading to a prevailing south-easterly wind anomaly over Emilia-Romagna. This flow is typically associated with moist warm air from the Adriatic Sea, which leads to an increase in surface air temperature everywhere but in a few locations where it favours the onset of strong convective events. These events may explain the local shortening of the HWD index over the eastern plain. This HWD pattern explains 16% of the total variance. The time series associated with the second CCA pair is significantly correlated at 0.55. CCA1 and CCA2 are used in the construction of the SD model in order to estimate summer HWD for the 1979–1993 period. 3.2 Evaluation of the performance of the SD models-mean and extreme indices In this section a description of the performance of all statistical downscaling models built for different seasons and predictands in terms of Spearman correlation, BIAS, and RMSE is presented. These skill coefficients are computed for the period 1979–1993 while the construction of the model has been made for the periods 1958– 1978 and 1994–2000. The mean value of each skill indicator over Emilia-Romagna is computed for each season and index and is shown in Fig. 4. The Spearman correlation values for Tmax and Tmin (Fig. 4a) R. Tomozeiu et al. Fig. 4. Mean skill over all stations of the statistical downscaling models for each season and predictand: Spearman correlation (a), BIAS (b and c), RMSE (d and e) emphasises that these fields are well downscaled during winter, spring and summer (correlation significant above 95%) and less so during autumn (correlation significant at 80%). The BIAS (Fig. 4b) of Tmax varies between 0.1 C (summer) and 0.4 C (winter) while that of Tmin reaches 0.6 C (winter). The RMSE (Fig. 4d) reveals similar values for both parameters, with lower values in summer, around 0.2 C, and close to 0.3 C for the other seasons. The skill of the SD models for extreme events of minimum temperature – Tmin10 and Tnfd (Fig. 4a) – reveals good performances in all seasons for Tmin10, with significant correlation, BIAS and RMSE values (Fig. 4b and d) around 0.5 C (in winter and autumn), while Tnfd presents good performances during winter and autumn. The skill of the SD models for the 90th percentile of maximum temperature in Fig. 4a, b and d indicates good performance during winter, spring and summer, while small skill is obtained in autumn (correlation significant at 80%). Concerning the HWD index, spring is the season with the best performances, followed by winter, while during summer and autumn the performance is poorer (Fig. 4a, c, e). An overview of the above results reveals that the mean fields are better downscaled than the extremes, although it should be noted that a correlation value of 0.4 is significant at the 80% level, and that all fields, except summer and autumn HWD, are reasonably downscaled (Fig. 4a). The BIAS that occurs in the modelled indices could be due to the different periods used in the Climate change scenarios for surface temperature in Emilia-Romagna construction and validation of the model (different climate). This BIAS has been removed when the root-mean square error has been computed. The work presented above helps to select the optimum downscaling model for each index= season and to establish which indices can be downscaled. 4. The HadAM3P validation The ability of HadAM3P to reproduce predictors under present-day climate conditions is very important when assessing the reliability of climate change projections under scenario conditions. The simulations of the T850, MSLP and Z500 provided by HadAM3P for the period 1960–1990 are compared with the same fields provided by NCEP=NCAR reanalysis for the same period. The comparison is done by performing an empirical orthogonal function (EOFs) analysis and comparing the corresponding observed and simulated patterns. The analysis is done for all seasons and using data averaged over each season. The main characteristics of the EOF patterns taken into consideration within this study are: the variance explained by each EOF pattern; the spatial correlation between model and reanalysis EOF patterns. The verification is done at the seasonal level and for the area identified to be the ‘‘optimum’’ area in the construction of the SD model (Table 1b). The significance of the spatial correlation has been tested using the Fisher’s Z transformation (at 95% confidence level). In the following, the verification of the seasonal predictors for the area 35 W–35 E and 30 N–60 N is presented. Similar results to those presented below, have been obtained using different validation techniques that range from the simple analysis of long-term seasonal mean fields and daily standard deviation to more complex analyses such as composite patterns of circulation types or Principal Component analysis using singular value decomposition of standardised anomalies. Diagrams with different methods of validation of these predictors and for the whole European area are assembled on the STARDEX web site (http:==www.cru.uea.ac.uk=cru=projects=stardex= deliverables=D13=). 4.1 Z500 variability Z500 is identified as the best predictor of the heat wave duration index during winter, spring, and summer, and of spring Tmax90. Comparison of the HadAM3P data with the reanalysis reveals quite an accurate model representation of the main patterns of large-scale variability of Z500 in all seasons. The analysis has focused on the phase space representation by standard EOFs. Figure 5 presents the first pattern (EOF1) of Z500 for each season, derived from reanalysis (a, c, e, g) and model (b, d, f, h) data. The winter EOF1 pattern derived from NCEP (Fig. 5a) is reproduced by the model (Fig. 5b) but is shifted eastward, and the spatial correlation between the two patterns is statistically significant at 0.74. In particular, the main centre of the positive anomaly is not centred over the Atlantic just west of the European coasts, implying a North-Westerly large-scale flow over Central Europe, but exactly centred over Central Europe. The amplitude of the pattern is also more intense in the observations than in the model data. A good resemblance is found for the other three winter patterns (figures not shown), with significant correlation values between patterns (around 0.9). The fraction of total variance explained by the first four EOFs (Table 3) of the model is very close to that of the reanalysis. During spring, significant skill in terms of spatial correlation is found for all four EOFs, closer for NCEP EOF1 (correlation 0.9) and EOF4 (correlation 0.73) while EOF2 and EOF3 are inverted and the correlation is around 0.6. Figure 5c and d present, as an example, the spring EOF1 of Z500 derived from NCEP and controlrun data. The dipolar structure of NCEP EOF1 is well represented by the control run. The fraction of variance explained by the first four EOFs is higher for the model than for the reanalysis, the model overestimates the variance explained by the first two EOFs (Table 3). In summer, the patterns are reasonably well reproduced by the model, with a correlation between EOFs lower than in winter and spring, and statistically significant for the first three EOFs (between 0.6 and 0.7). Although the EOF1 control run pattern shares similar characteristics with the EOF1 NCEP pattern, some differences can be identified. In particular, the main positive anomaly centred over the British Isles in the re- R. Tomozeiu et al. Fig. 5. EOF1 of winter (a and b), spring (c and d), summer (e and f) and autumn (g and h) Z500 for the period 1960–1990 for NCEP=NCAR reanalysis (left column) and HadAM3P model (right column) analyses is located westward in the model, well over the North Atlantic. Furthermore, the model circulation is somewhat more intense than the observed and has a main northerly direction over Central Europe and Northern Italy, while in the observations it is orientated in an easterly-north Climate change scenarios for surface temperature in Emilia-Romagna Table 3. Explained variance (%) of the first four EOFs of the seasonal Z500 (NCEP and control run) Winter EOF1 EOF2 EOF3 EOF4 Spring Summer Autumn NCEP ctr. run NCEP ctr. run NCEP ctr. run NCEP ctr. run 37.2 34.6 12.4 8.3 35 31 16 9.3 26 19.3 17.6 13.5 32 30 15 7.7 44 14 12.6 11 45 20 15 6 40 24 12 10 33 31 14 8 easterly direction. Finally, the model tends to overestimate the fraction of total variance explained by the first four EOFs (Table 3). During autumn, significant skill in terms of spatial correlation has been found for all four EOFs, higher for EOF2 (0.9) and EOF1 (0.75) while EOF3 and EOF4 are inverted and the correlation is around 0.7 (statistically significant at the 95% level) for both patterns. Figure 5g and h present the patterns of the autumn EOF1 as derived from NCEP and control run data, respectively. Both the model and the reanalysis EOF1 pattern present a main positive anomaly centred in the vicinity of the British Isles (West of Ireland in the reanalyses and over Scotland in the model) and the intensity of this anomaly is captured reasonably well. The main difference between the two patterns is that the location of the most intense geopotential gradient is east of the positive anomaly in the NCEP data, implying a northerly flow over Central Europe and Northern Italy, while in the model it is south-west of the main positive anomaly, leading to a upper-air northeasterly circulation anomaly over Northern Italy and Central Europe and an intense return flow (easterly) anomaly over the North Atlantic. Finally, the model tends to underestimate the variance explained by EOF1 (Table 3) while the fraction of variance explained by the first four EOFs of control run data is equal to that explained by the NCEP patterns. The features of the simulated patterns of Z500 presented above from both model and reanalysis are in general agreement, although some differences are seen. 4.2 MSLP variability Mean sea level pressure (MSLP) is the best predictor for a limited number of extreme temperature winter, spring and autumn indices (see Table 1b). In the following, the representation by the model of the large-scale MSLP variability is validated for the area 35 W–35 E, 30 N–60 N, and for the seasons of interest, in the same way as for T850 and Z500. This analysis emphasises that the general features of the simulated large-scale variability are reasonably well reproduced in the model. During winter, the spatial correlation of the corresponding patterns varies between 0.7 (EOF1) and 0.9 (EOF4). The winter EOF1 (not shown) resembles a similar configuration as winter Z500, shifted to the east (as in case of Z500). The variance explained by this pattern is 41% in the model and 45% in the reanalysis, while the fraction of total variance explained by the first four EOFs is 93% in the model and 94% in the reanalysis. During spring and autumn the model has good skill in representing MSLP variability. In spring, the first four EOFs of the model explain 91% of the total variance while the corresponding group of NCEP EOFs explains 86%. The spatial correlation is high, between 0.8 and 0.9, while the RMSE varies between 0.01 and 0.4. In particular, the EOF1 for this season, shown in Fig. 6a and b, have a spatial correlation of 0.7. The model pattern presents a high pressure anomaly centred slightly east of the observed and is associated with a more intense zonal gradient, with the low pressure in the north-east of the domain being overestimated. As a result, the surface pressure gradient over Europe is higher in the model than in reality, presenting a less optimistic picture than the upper air Z500 analysis for the same season. However, it can be said that the model captures the general features of the surface flow over the region of interest, is similar to the Eastern Atlantic pattern, and is quite well represented. The first four EOFs (not shown) of autumn MSLP explain a fraction of the total variance equal to that explained by the corresponding reanalysis patterns R. Tomozeiu et al. Fig. 6. EOF1 of spring MSLP (a and b), and winter T850 (c and d), period 1960–1990, for NCEP=NCAR reanalyses (left column) and HadAM3P model (right column) (91%). The spatial correlation of the patterns again reveals a sufficiently good representation of the four EOFs of MSLP (0.7–0.9). 4.3 T850 variability The patterns of the first four EOFs of winter T850 derived from HadAM3P data explain together 85.4% of the total variance, very close to the value derived from the reanalysis (85%). The spatial correlation between model and reanalysis reveals that all four winter patterns are very well simulated, the correlation varying between 0.92 (EOF1) and 0.80 (EOF4). The EOF1 of winter T850 derived from NCEP (49% variance) and HadAM3P (42% variance) are represented in Fig. 6c and d. They show the presence of a greater extension of the Northern European temperature anomaly into Central Europe including, to a certain extent, Northern Italy. This is consistent with the differences observed in the upper air circulation anomalies associated with the Z500 EOF1 pattern. The HadAM3P model simulates reasonably well all four EOFs of spring T850 (maps not shown). The correlation between the patterns is statistically significant and varies between 0.6 (EOF1) and 0.9 (EOF4). The fraction of variance explained by the first four EOFs of the model is 76%, while those explained by the first four EOFs of reanalysis is 72%. The analysis of the first four EOFs of T850 during summer, reveals an overestimation of the variance in the model (80% explained by the four EOFs) with respect to reanalysis data (66%). The best simulated patterns in summer are EOF2 (correlation 0.72) and EOF3 (correlation 0.6) of the model which coincide with EOF2 and EOF4 of NCEP. For EOF1, the spatial correlation is lower than for EOF2 and EOF3, but is still statistically significant at 0.4. The general problems in the representation of T850 pattern of variability are a reflection of the differences noted earlier between the main observed and modelled upper air circulation anomalies as described by the Z500 field. During autumn, the fraction of total variance explained by the first four EOFs of the model is 76% while that explained by reanalysis is 72%. The spatial correlation emphasises that the first Climate change scenarios for surface temperature in Emilia-Romagna three EOFs are well reproduced by the model, with the correlation between modelled and observed varying between 0.7 and 0.8. 5. Downscaling of predictands for the present climate (1960–1990) In order to test the ability of the HadAM3P model to simulate the variability of local climate for the period 1960–1990 and the robustness of the SD models, these are applied to the HadAM3P control-run data. Mean value, BIAS, trend and standard deviation are evaluated using the SD models for each predictand=station and compared with those computed from the observed data during the period 1960–1990. The results show that for all indices the mean seasonal value is, in general, well reproduced. An example of the comparison between observed and downscaled data is presented in Fig. 7a and b, for seasonal mean values of Tmin and Tmax at two stations, one in the plain (Bologna – 51 m), and the other in the mountains (Sestola – 1020 m). The downscaled values reproduce the seasonal cycle very well, in both minimum and maximum temperature. Similar results are obtained for other stations, and the mean over all stations of the difference between downscaled and observed data (BIAS) of Tmin=Tmax is 0.8 C during winter and lower for the other seasons. The mean values of the extreme indices (Tmin10, Tmax90) are also well reproduced by the SDs in all seasons, better for Tmax 90, when the mean BIAS over all stations and seasons does not exceed 0.5 C (spring), than for Tmin10, when the mean BIAS over all stations does not exceed 1 C (plots not shown). Concerning the number of frost days and heat wave duration indices, the mean values of downscaled data are close to those of the observed except for winter Tnfd, with differences around five days observed at some stations situated in the plain area. Trend analysis of downscaled and observed data reveals that in general the signal is captured by the SDs, sometimes with lower intensity in the downscaled than in the observational data. Figure 7c presents the trend coefficient for minimum temperature at the stations of Bologna and Sestola ( C=decade). As can be observed, except for summer minimum temperature at the Sestola Fig. 7. Observed and downscaled values of seasonal mean minimum (a) and maximum (b) temperatures, Tmin trend (c) and standard deviation (d) for all seasons, at Bologna and Sestola stations R. Tomozeiu et al. station, the sign of the trend is captured although, sometimes, with lower intensity. Similar results are obtained for maximum temperature, as well as for extreme indices. Figure 7d presents the values of the standard deviation of seasonal minimum temperature at Bologna and Sestola. The inter-annual variability of the downscaled data is underestimated in comparison with the observed, as in most other statistical downscaling models. 6. Climate change scenarios of mean and extreme temperature in Emilia-Romagna The statistical downscaling models built using observational data are finally applied to the following HadAM3P scenario experiments: A2, a medium-high emissions scenario, with atmospheric CO2, concentrations reaching 715 ppm, and B2, a medium-low scenario, with CO2 reaching 562 ppm. The predictor anomalies used as input in the statistical downscaling models have been computed as differences between the scenario full fields and the control run climate. Regarding the predictand anomalies, downscaled from A2 and B2 experiments, these have been added to the observed climate in order to obtain the full field time series of the indices for the period 2070–2100. Changes are computed as the difference between mean values of each index in the period 2070–2100 and those in the control run (1960–1990). The significance of changes has been statistically tested using the Student’s t-test. The analysis begins with a discussion of changes in seasonal minimum and maximum temperature, derived from the downscaling models for each station for both A2 and B2 scenarios, followed by the scenarios of extreme temperature indices. 6.1 Changes in seasonal minimum temperature (Tmin) Climate change projections for the A2 and B2 scenarios indicate the possibility of an increase in minimum temperature in all seasons. Figure 8a presents changes in Tmin, averaged over all stations, for each scenario (A2 and B2) and each season. As can be observed, the increase is always greater in A2 than in B2, except in spring, and is always weaker in spring than in the other seasons. Fig. 8. Projected changes in seasonal averaged minimum (a) and maximum (b) temperatures for A2 and B2 scenarios. Mean over all stations At the station level, changes are greater for the stations situated in the plain while small spatial differences in changes (except summer) are noted over hills and mountains. A2 scenario changes in Tmin reach a maximum of 3.5–4 C in winter, summer and autumn at stations situated in the plain and 2.5 C at the stations over hill=mountains. These results are statistically significant at the 95% level for all stations. During spring, the changes in Tmin are less intense than in the other seasons, but are still statistically significant, with a maximum around 0.7 C for the stations situated in the plain. At some stations over the Apennines, minimum temperature is projected to decrease, although the result is not statistically significant. It is possible that the choice of MSLP as the only predictor for this SD model is responsible for the smaller change of this index under scenario conditions than in all other seasons. In fact MSLP, which mostly reflects changes in large-scale circulation and not radiation-induced temperature changes, may be less affected by changes in atmospheric composition than upper air temperature (Schubert, 1998; Huth, 2004). Given that this was identified as the predictor of choice, leading to the best Climate change scenarios for surface temperature in Emilia-Romagna performing SD model, it was decided to retain this variable. In the B2 scenario the projected changes reach a maximum value of 2.5 C in winter and autumn and 3.5 C during summer. 6.2 Changes in seasonal maximum temperature (Tmax) The seasonal mean values of Tmax, as in the case of Tmin, are projected to increase in all seasons. Figure 8b presents the changes in maximum temperature averaged over all stations for both scenarios (A2 and B2). Comparing these increases with those in minimum temperature, the projected changes of Tmin are slightly greater than Tmax during winter and autumn, suggesting, for these seasons, a general decrease in the daily temperature range, which is consistent with the expected changes under scenario conditions. The same is not true in spring and summer when the changes are greater for Tmax than for Tmin. At the station level, Tmax changes are characterised by small spatial differences over the whole region during spring, while in winter and autumn the predicted increase of Tmax is higher in the plain than over the hills and the mountains. In summer, the changes tend to be more intense over the mountains, while similar values are detected in the plain and over the hills. A comparison between A2 and B2 reveals, as in the case of mean minimum temperature, that changes are more intense in the A2 than in the B2 scenario. The values of the changes reach a maximum of 3 C and 2.5 C in winter and of 4.5 C and 4 C, respectively, in spring. In both seasons, the changes are statistically significant at nearly all stations, with the few exceptions located in the western part of the Apennines. During summer, significant changes are projected to occur at all stations in both experiments, with a mean change around 5 C (Fig. 8b) and local values up to 8 C. In autumn, under both scenarios, positive and significant changes in Tmax are projected to occur at almost all stations, with values up to 5 C and 3 C for the A2 and B2 scenario, respectively. An analysis of the probability density function of seasonal minimum and maximum temperature for present day (1960–1990) and future (2070– 2100) reveals a shift to the right of the distribution in both minimum and maximum temperature. Regarding the shapes of the minimum and maximum temperature distributions, these tend to become wider in autumn and sharper in winter and spring under scenario conditions. As for summer, the minimum temperature distribution widens, while the maximum temperature remains more or less similar. The results of changes in Tmin and Tmax are consistent with those presented for the mean surface air temperature over the Italian peninsula by Giorgi et al. (2004b), based on nesting RegCM with the HadAM3H model, for both A2 and B2 scenarios. The authors compared the scenarios for both models – RegCM and HadAM3H – and found that the most pronounced warming in Italy may occur in summer, and that the mean over the region is around 5 C in RegCM and 6 C in HadAM3H. For the other seasons, the projected increase is around 4 C. In all seasons, the increase is found to be greater in the A2 than in the B2 scenario. Räisänen et al. (2004) analysed the changes in the highest yearly maximum temperature (Tmax) and in the lowest yearly minimum temperature (Tmin), derived from the regional climate change simulations done at the Swedish Rossby Centre, and concluded that the changes are expected to be larger for the A2 than the B2 scenario and in most cases in the ECHAM=OPYC3-driven (RE) than in the HadAM3H-driven (RH) simulations. In all the scenario runs, the warming in Central and Southern Europe peaks in summer when it locally reaches 10 C in the RE and around 6–7 C in the RH. 6.3 Changes in seasonal Tmin10 Projected Tmin10 changes are significant in all seasons and most intense during winter, followed by spring and summer. Seasonal changes in Tmin10 are presented in Fig. 9, for the A2 scenario. Shading indicates the value of the change in Tmin10 (2070–2100 minus 1960–1990) while black dots indicate stations characterised by statistically significant results at the 95% level. As can be noted, significant increases during winter (Fig. 9a) are expected to occur everywhere, with maximum values (up to 7 C) in the plain area. The general pattern of the change closely resem- R. Tomozeiu et al. Fig. 9. Pattern of projected changes in Tmin10 (A2 scenario): winter (a), spring (b), summer (c) and autumn (d). Shading represents the value of changes, while stations characterised by significant results (at 0.05 level) are marked with black cycles bles that of the Tmin10 CCA1 pattern, suggesting that the observed changes in this index are mostly due to an increase in T850 over the eastern part of the Mediterranean, associated with a negative Z500 anomaly over Eastern Europe and Russia (see Sect. 3.1). The mean change averaged over all stations is projected to be around 4.5 C. Similar results are found for the B2 scenario, although in this case changes are less intense than for A2. Spring Tmin10 (Fig. 9b) is projected to increase significantly, with maximum changes around 3 C and 2.4 C in the A2 and B2 scenarios, respectively. During summer, Tmin10 (Fig. 9c) shows significant changes at all stations, with positive values for a great part of the region (maximum around 5 C) and negative values (up to 1 C) at some stations in the central plain and the Adriatic coast (Codigoro, Alfonsine, Modena, ReggioEmilia, and Parma). The summer change averaged over all stations is projected to be about 2 C less than for Tmin. During autumn (Fig. 9d), the changes are also significant over the whole region but less intense than in the other seasons (maximum around 1 C). Some negative changes, not statistically significant, are projected at some stations in the Apennine region. Comparing the projected changes in winter Tmin10 (Fig. 9a) with those of winter Tmin (Fig. 10b), the observed changes in Tmin10 are more intense than in Tmin suggesting a sharpening of the Tmin probability distribution in this season. Similar results have been obtained for spring, while during summer the magnitude of changes are slightly smaller in Tmin10 than for mean Tmin (maps not shown). 6.4 Changes in the seasonal number of frost days (Tnfd) Under both scenarios, the number of frost days (Tnfd) is projected to undergo a significant decrease during all seasons (no frost day is either observed or projected in summer). Figure 10a shows the changes in winter Tnfd (A2 scenario). As can be observed, significant decreases are projected to occur at all stations, with values locally more intense in the plain (except at two stations) and hill regions than over the mountains. This pattern is similar to that of winter Tmin Climate change scenarios for surface temperature in Emilia-Romagna 6.5 Changes in seasonal Tmax90 Fig. 10. Pattern of projected changes in winter Tnfd (a) and winter Tnav (b) for A2 scenario (c) Maximum and minimum values attained by Tnfd over Emilia-Romagna, in both A2 and B2 scenarios (Fig. 10b) under the A2 scenario, suggesting that the main reason for the decrease in this index is due to the general rise in the daily minimum temperature. Regarding the other seasons, Fig. 10c presents the minimum and maximum values of the seasonal changes projected to occur for Tnfd, under the A2 and B2 scenarios. As can be observed, in winter the number of frost days is projected to decrease up to 40 and 30 days in spring up to 25 and 20 under the A2 and B2 scenarios, respectively, while in autumn is the decrease is lower at around 10 days. This is due to the fact that the seasonal mean Tmin is greater in autumn over the whole region than in winter and spring (see Fig. 7a). All results are statistically significant at the 95% level over the whole region. Under both A2 and B2 scenarios, Tmax90 is projected to increase everywhere, except at a few stations located over the Apennines (Bedonia, Sestola, and Pavullo nel Frignano) where a significant decrease is found in all seasons. Figure 11 presents the changes in Tmax90 for winter (a), spring (b), summer (c) and autumn (d) for the A2 scenario. As can be observed, the season with greatest increase is summer, followed by autumn and spring. In winter, the increase of Tmax90 reaches 1 C (Fig. 11a), the pattern of changes are similar to those of winter Tmax, but less intense. For spring, the changes in Tmax90 reach values around 2.5 C (Fig. 11b) with a similar but less intense pattern to that of spring Tmax. During summer, Tmax90 is projected to decrease at two stations located over the western part of the Apennines (signal of decrease in this area is present in all seasons) and to increase in all other stations, with maximum values of increase in the south-eastern part of the region (up to 8 C). Changes in summer Tmax90 are similar to those in summer Tmax (map not shown) in terms of magnitude and pattern. Concerning autumn, a significant increase is projected in Tmax90 with values up to 5 C (Fig. 11d), very close to those of changes in autumn Tmax. The comparison between the maps of seasonal Tmax90 and Tmax changes suggests that, as in the case of Tmin, it is possible to predict a sharpening of the Tmax probability distribution under scenario conditions in both winter and spring. A comparison between A2 and B2 reveals similar changes, but with greater intensity in the A2 than in the B2 scenario. For example, the changes averaged over the region for the B2 experiment are equal to 1.7 C in autumn, 1 C in spring and around 0.5 C in winter. 6.6 Changes in seasonal HWD In general, HWD is projected to increase in all seasons for both the A2 and B2 scenarios. The signal is greater in summer when the maximum change reaches a value of 23 days. Taking into account that the performance of the statistical model (Fig. 4) is not statistically significant during summer and autumn, the following R. Tomozeiu et al. Fig. 11. As in Fig. 9, but for Tmax90 (A2 scenario) Fig. 12. Pattern of projected changes in spring heat wave duration (a) and spring Txav (b) for A2 scenario discussion is focused only on results for winter and spring. During these seasons, positive changes in HWD are projected to occur in both scenarios, more intense during spring than in winter when the increase is around two days. Figure 12a presents the pattern of changes in HWD (A2 scenario) during spring. As can be observed, the increase is statistically significant at all stations and reaches a maximum of about 10 days in the northern part of the region. The pattern of change is very similar to that for spring Tmax (Fig. 12b) with maximum change in the central part of the region. 7. Conclusions In the present study, statistical downscaling models (SDs) are built and used in order to provide future scenarios of temperature, mean and extremes, in Emilia-Romagna. The SD models are based on canonical correlation analysis, a technique that also helps to understand the large-scale mechanism that controls the local climate variability. The projection of changes in temperature under scenario conditions are constructed using as large-scale data input the HadAM3P model simulations, from the control, the A2 and the B2 IPCC SRES scenarios. Climate change scenarios for surface temperature in Emilia-Romagna The work done in order to select the optimum mode with respect to RMSE, BIAS and Spearman rank correlation values for each predictand reveals that: 1) T850 is the best large-scale predictor for reproducing local variability of seasonally averaged minimum and maximum temperature in Emilia-Romagna, with the exception of spring Tmin which is best simulated using MSLP only. T850 is also the best predictor for modelling the temporal and spatial variability of almost all local extreme temperature indices. All the same, winter Tmax90 and autumn Tmin10 are best estimated using a combination of T850 and MSLP, while spring Tmax90 and nearly all seasons of HWD (except autumn) are best predicted using only Z500 (see Table 1b); 2) the skill of the models is dependent on the combinations of EOFs=CCAs used to build the SD model and on the domain (area) of predictors; 3) the skill of the models varies between seasons, and indices, their performance being generally better for mean values than for extremes. The index with lower skill is the heat wave duration for summer and autumn; 4) HadAM3P is able to reasonably simulate the variability of the predictors – T850, MSLP and Z500 – in all seasons, although some differences have been noted between the model generated and the observed patterns and between the fractions of variance explained by the corresponding patterns. The main conclusions concerning the changes projected to occur in Emilia-Romagna under A2 scenario conditions can be summarized as follows: significant increases are projected in minimum temperature in all seasons. The signal has a greater magnitude in winter, summer and autumn than in spring. The average over all stations of the projected change is about 2 =2.5 C for all seasons, while the minimum=maximum values of increases vary from season to season, and are less intense in spring than in the other seasons; significant increases are projected in maximum temperature in all seasons, more intense in spring and summer when the mean over all stations of projected changes reaches 3 C and 5 C, respectively. During winter and autumn the magnitude of changes is around 2 C (mean over all stations); significant increases are projected in Tmin10 in all seasons, more intense in winter, when the mean change over all stations is around 4.5 C. For spring and summer the magnitude of increase is around 2 , while in autumn it is around 1 C; the number of frost days will decrease, more in winter (up to 40 days), and spring (up to 25 days), and less in autumn (up to 10 days); significant increases are projected for Tmax90 in all seasons, most intense in summer and autumn, when the mean change over all stations is predicted to be around 4 C and around 2 C, respectively. During spring the increase will be around 1 C, while during winter it will be less than 1 C (mean over the region), when the mean value of Tmin is greater over the whole region; significant increases are projected to occur for the HWD index over the whole region during spring, with values up to 10 days. With regards to other seasons, HWD shows significant increases, especially during summer, but the SD model does not have significant skill and this particular projection must be regarded with caution. Similar signals of increases in mean and extreme temperature are projected to occur under B2 scenario conditions, but be less intense than under the A2 scenario. These results are consistent with those obtained within the STARDEX project for other European regions, such as Greece, the German Rhine, and the Iberian Peninsula using other statistical downscaling techniques (http:==www.cru.uea.ac.uk=cru= projects=stardex=reports=STARDEX_FINAL_ REPORT.pdf) or within the MICE project (www.cru.uea.ac.uk=cru=projects=mice) using dynamical downscaling techniques. In these studies, more intense decreases are projected to occur in winter Tnfd across Europe than in EmiliaRomagna, with values between 60 to 120 days per year, and the greatest reductions in Northern Europe. In addition, the annual number of ‘very hot’ days (index similar to HWD) is projected to R. Tomozeiu et al. increase by several weeks across the UK and Scandinavia, while increases up to 120 days are projected to take place in other parts of Europe. Finally, a set of PRUDENCE RCM experiments reveals that HWD will increase over Europe and the intensity of extreme temperatures will increase more rapidly than mean temperature by the end of 21st century (Beniston et al., 2005). Acknowledgments This work was supported by the European Commission as part of the STARDEX (STAtistical and Regional dynamical Downscaling of EXtremes for European regions) contract EVK2-CT-2001-00115. The HadAM3P climate model output were provided by the Hadley Centre for Climate Prediction and Research, UK. The comments of two anonymous reviewers were helpful in improving the quality of the paper. References Akima H (1984) On estimating partial derivatives for bivariate interpolation of scattered data. Rocky Mt J Math 14: 1 Barnett T, Preisendorfer R (1987) Multi-field analogue prediction of short-term climate fluctuations using a climate state vector. J Atmos Sci 35: 1771–1787 Benestad RE (2001) A comparison between two empirical downscaling strategies. Int J Climatol 21: 1645–1668 DOI: 10.1002=joc.703 Beniston M, Stephenson DB, Christensen OB, Ferro CAT, Frei C, Goyette S, Halsnaes K, Holt T, Jylha K, Koffi B, Palutikof J, Scholl R, Semmler T, Woth K (2005) Future extreme events in European climate: an exploration of regional climate model projections. ClimaticChange, PRUDENCE special issue (submitted) Buishand TA, Shabalova MV, Brandsma T (2004) On the choice of the temporal aggregation level for statistical downscaling of precipitation. J Climate 17: 1816–1827 Busuioc A, Chen D, Hellstr€ om C (2001) Performance of statistical downscaling models in GCM validation and regional climate change estimates: application for Swedish precipitation. Int J Climatol 21: 557–578 Busuioc A, von Storch H (2003) Conditional stochastic model for generating daily precipitation time series. Climate Res 24: 181–195 Busuioc A, Tomozeiu R, Cacciamani C (2005) Statistical downscaling model for winter extreme precipitation events in Emilia-Romagna region. Int J Climatol (submitted) Cacciamani C, Nanni S, Tibaldi S (1994) Mesoclimatology of winter temperature and precipitation in the Po Valley of Northern Italy. Int J Climatol 14: 777–814 Cassou C, Terray L, Phillips AS (2005) Tropical Atlantic influence on European heat waves. J Climate 18: 2805–2811 Cavazos T, Comre AC, Liverm DM (2002) Inter-seasonal variability associated with wet monsoons in southeast Arizona. J Climate 15: 2477–2490 Christensen JH, Christensen OB (2003) Climate modelling: severe summertime flooding in Europe. Nature 42: 805–806 Colacino M, Conte M (1995) Heat waves in the Central Mediterranean. A synoptic climatology. Nuovo Cimento 18: 295–405 Fuentes U, Heimann D (2000) An improved statistical– dynamical downscaling scheme and its application to Alpine precipitation climatology. Theor Appl Climatol 65: 119–135 Giorgi F, Bi X, Pal JS (2004a) Mean, interannual variability and trends in a regional climate change experiment over Europe. I. Present-day climate (1961–1990). Clim Dyn 22(6=7): 733–756 Giorgi F, Bi X, Pal JS (2004b) Mean, interannual variability and trends in a regional climate change experiment over Europe. II. Climate change scenarios (2071–2100). Clim Dyn 23: 839–858. DOI 10.1007=S00382-004-0467-0 Hanssen-Bauer I, Førland EJ, Haugen JE, Tveito OE (2003) Temperature and precipitation scenarios for Norway: comparison of results from dynamical and empirical downscaling. Climate Res 25: 15–27 Hanssen-Bauer I, Achberger C, Benestad RE, Chen D, Førland EJ (2005) Statistical downscaling of climate scenarios over Scandinavia. Climate Res 29: 255–268 Hellstr€ om C, Chen D (2003) Statistical downscaling based on dynamical downscaled predictors: application to monthly precipitation in Sweden. Adv Atmosp Sci 20: 951–958 Huth R (2002) Statistical downscaling of the daily temperature in Central Europe. J Climate 15: 1731–1742 Huth R (2004) Sensitivity of local daily temperature change estimates to the selection of downscaling models and predictors. J Climate 17: 640–652 Huth R, Kysely J, Dubrovsky M (2001) Time structure of observed, GCM-simulated, downscaled, and stochastically generated daily temperature series. J Climate 14: 4047–4061 Johns TC, Gregory JM, Ingram WJ, Johnson CE, Jones A, Lowe JA, Mitchell JFB, Roberts DL, Sexton DMH, Stevenson DS, Tett SFB, Woodage MJ (2003) Anthropogenic climate change for 1860 to 2100 simulated with the HadCM3 model under updated emission scenarios. Climate Dyn 20: 583–612 Kalnay E et al. (1996) The NCEP=NCAR 40-year – reanalysis Project. Bull Amer Meteor Soc 77: 437–472 Katz R, Parlange MB, Tebaldi C (2003) Stochastic modelling of the effects of large-scale circulation on daily weather in the south-eastern U.S. Climate Change 60: 189–216 Murphy JM (2000) Prediction of Climate change over Europe using statistical and dynamical downscaling techniques. Int J Climatol 20: 489–501 Palutikof JP, Goodess CM, Watkins SJ, Holt T (2002) Generating rainfall and temperature scenarios at multiple sites: examples from the Mediterranean. J Climate 15: 3529–3548 Pavan V, Tomozeiu R, Selvini A, Marchesi S, Marsigli C (2003) Controllo di qualita dei dati giornalieri di temperatura minima e massima e di precipitazione. Quaderno Tecnico ARPA-SIM 15: 66 pp Climate change scenarios for surface temperature in Emilia-Romagna Pavan V, Marchesi S, Morgillo A, Cacciamani C, DoblasReyes FJ (2005) Downscaling of DEMETER winter seasonal hindcasts over Northern Italy. Tellus 57A: 424–434 Pope VD, Gallani ML, Rowntree PR, Stratton RA (2000) The impact of new parameterizations in the Hadley Centre model: HadAM3. Climate Dyn 16: 123–146 Räisänen J, Hansson U, Ullerstig A, D€ oscher R, Graham LP, Jones C, Meier HEM, Samuelsson P, Willen U (2004) European climate in the late twenty-first century:regional simulations with two driving global models and two forcing scenarios. Climate Dynamics 22: 13–31 Schmidli J, Frei C, Vidale PL (2006) Downscaling from GCM precipitation: a benchmark for dynamical and statistical downscaling. Int J Climatol (in press) Schubert S (1998) Downscaling local extreme temperature changes in south-eastern Australia from CSIRO Mark2 GCM. Int J Climatol 18: 1419–1438 Tomozeiu R, Pavan V, Cacciamani C, Amici M (2005) Observed temperature changes in Emilia-Romagna: mean values and extremes. Climate Res (in press) Trigo RM, Palutikof JP (2001) Precipitation scenarios over Iberia: a comparison between direct GCM output and different downscaling techniques. J Climate 14: 4422–4446 Von Storch H (1995) Spatial Patterns: EOFs and CCA. In: von Storch H, Navarra A (eds) Analysis of climate variability. Application of statistical techniques. Springer pp 227–258 Von Storch H, Zorita E, Cubasch U (1993) Downscaling of climate change estimates to regional scales: an application to the Iberian winter time. J Climate 6: 1161–1171 Widmann M, Bretherton CS, Salathe EP Jr (2003) Statistical precipitation downscaling over the Northern United States using numerically simulated precipitation as a predictor. J Climate 16: 799–816 Zorita E, von Storch H (1999) The analogue method as a simple statistical downscaling technique: comparison with more complicated methods. J Climate 12: 2474–2489 Authors’ addresses: Rodica Tomozeiu (e-mail: rtomozeiu@ arpa.emr.it), Carlo Cacciamani (e-mail: ccacciamani@ arpa.emr.it), Valentina Pavan (e-mail: [email protected]), Antonella Morgillo (e-mail: [email protected]), ARPAServizio IdroMeteorologico Regionale,Viale Silvani 6, 40122 Bologna, Italy; Aristita Busuioc (e-mail: busuioc@ meteo.inmh.ro), National Meteorological Administration, Sos. Bucuresti-Ploiesti 97, Sector 1, Bucharest, Romania.
© Copyright 2026 Paperzz