This study is a part of the K-EGOKITZEN project (http://info.labein.es/k-egokitzen/descripcion-del-proyecto/; http://188.8.131.52/secciones/cambio_climatico/pdf/ad_hoc_resumen.pdf), which aims to coordinate efforts of local research groups from the Basque Country (Spain) focusing on the impact and adaptation to climate change at the regional scale. The scientific platform of the project has a multidisciplinary background and covers a large sample of subjects from biology, hydrology, coastal oceanography, territorial management or urban climate among others. In this context, information on climate change with high spatial resolution is needed to feed impact, vulnerability and adaptation models. The most up-to-date projections of potential impacts of climate change over the region were published by Abanades et al. (2007). The report analyses the climate change impact of climate scenarios extracted from IPCC at the regional scale for Spain from the EU-FP5 project PRUDENCE (Déqué, 2007) at 50 × 50 km of horizontal resolution. Due to advances in computing capability, a recently finished European project, the EU-FP6 ENSEMBLES (Hewitt and Griggs, 2004), performed ensemble simulations for Europe at 25 × 25 km horizontal resolution. The higher resolution of these new available scenarios should improve the analysis on the changes of temperature and precipitation, key parameters for the analysis on the impact, vulnerability and adaptation to climate change. Particularly, the interest of the local policy-makers and resource managers are zones with a characterized irregular orography in the Basque Region.
The article aims to explore the expected changes over the Basque Country area in the tails of the distribution of temperature by means of climate change indices derived from daily and seasonal near-surface temperature (a complementary study focusing on changes in precipitation over the Basque Country area has been developed by Moncho (2009)). The second objective is to contribute to the discussion with respect to the latest tendencies in the fundamental aspects of model validation and current bias correction methods.
2. Regional setting
The Basque Country is located in the Northern coast of the Iberian Peninsula with complex terrain, covering a surface of 7235 km2 with 4.6% of Spanish population, about 2.17 million people (INE, 01 January 2009). It is located in the transition strip of Atlantic and Mediterranean climates and it is dominated by the Westerlies and the Polar front. Its climate is influenced by zonal mountain ranges that cross the region from North to South and rise 1500-m high. The inferring Siberian anticyclone causes numerous frost episodes and when föhn wind situations appear temperatures can reach upto 35 °C also due to the emerging of the Azores anticyclone (Saenz et al., 2001). The meso-scale climate is controlled by the irregular terrain and its proximity to the Cantabric Sea (Figure 1).
3. Data and methods
Global climate models (GCM) constitute a fundamental tool for simulating the global climate system behaviour. They represent the processes responsible for maintaining the general atmospheric/oceanic circulations and their variability. However, given their coarse resolution, downscaling techniques are applied for impact assessment studies in order to provide an accurate regional and local climate detail. Recent European projects, such as EU-FP5 PRUDENCE (Déqué, 2007), EU-FP6 ENSEMBLES (Hewitt and Griggs, 2004) and Mistra-SWECIA Project (http://www.mistra-swecia.se/; Kjellstrom, 2007) performed ensemble simulations to assess uncertainties in future climate change projections for Europe. The reason for developing regionalization techniques is to capture the effect of fine scale forcing in areas characterized by fine spatial variability of features such as topography and land surface conditions. However, this increase of spatial resolution adds an extra factor to the uncertainty cascade. In addition to the natural variability of the atmosphere, the emissions scenarios, and the GCM uncertainties, the uncertainties linked to the regional climate model (RCM) should also be taken into account. Climate model outputs are only a synthetic reality and should be compared with the observed reality of a reference period. The goodness of fit (bias) that summarize the discrepancy between the observations and the model outputs should be evaluated before analysing future climatic scenarios from such models. In the case of RCM, accuracy assessment is evaluated by means of the statistical climate representation. The interpretation of the produced climatic series is carried out by means of climatic indices based on statistics for extreme climatic events.
3.1. Models and observational data
For this study, 6 RCMs from 13 included in the EU-FP6 ENSEMBLES project were selected: CLM (Steppeler et al., 2003), HIRHAM (Christensen et al., 1996), RACMO (Lenderik et al., 2003), REMO (Jacob, 2001), PROMES (Castro et al., 1993) and ALADIN (Spiridonov et al., 2005). These RCMs were driven by three different GCM: ARPEGE (Gibelin and Déqué, 2003), HadCM3 (Rowell, 2005) and ECHAM5 (Roeckner et al., 1999). Each pair of RCMs was driven by the same GCM allowing an assessment of the global model influence on the future projections (Table I).
Table I. RCMs included in the EU-FP6 ENSEMBLES project selected for this study, their driven GCM and the institution that generated the outputs
Danish Meteorological Institute
Centre National de Recherche Météorologiques
Castilla La Mancha University
Swiss Institute of Technology
Royal Netherlands Meteorological Institute
Max Plank Institute for Meteorology
In the ENSEMBLES project, simulations run under the IPCC A1B climate scenario (Nakicenovic et al., 2000) was the only available. During the project, two types of experiments were defined: first, driven RCMs by the ERA—40 reanalysis of the ECMW from 1961 to 2000 and second, transient experiments (1951–2050) have driven by different GCM that were coupled to an ocean model. Even if previous validation work of the first had been done, the results shown here correspond to the transient experiments (http://ensemblesrt3.dmi.dk/).
The analysed variables were the daily maximum (T2max), daily minimum (T2min), the daily mean (T2mean) air temperatures and relative humidity (RH) at 2 m. The periods selected were summer [June, July and August (JJA)] and winter [December, January and February (DJF)] seasons during the years 1961–2100. The bias between observations and model outputs were evaluated for a ‘reference or control period’. IPCC defines a control period as representative of the present-day or recent average climate in the study region that does not exhibits anthropogenic climate changes. The current WMO normal period is 1961–1990, which provides a standard reference for many impact studies. In this study, the control period used is 1978–2000 as a consequence of a lack of data.
Model data were compared with observed data provided by the Spanish National Meteorological Agency (AEMET, http://www.aemet.es/es/portada). Four observatories covering the Basque Region were selected fulfilling the criteria: observatories must be placed in rural areas giving good coverage to the study area; they do not contain missing values over the threshold (0.5% for JJA and 0.8% for DJF) and provided data during the control period. A reconstruction of these observed series were carried out for the control period (1978–2000); the missing values were filled following Beckers and Rixen (2003): a multivariate linear regression and empirical orthogonal function analysis Preisendorfer (1988). The description of each station with the characteristics of its close environment is shown in Table II. The geographical position of the observatories and the grid resolution of the models are portrayed in Figure 1.
Table II. Characteristics of the meteorological stations References: The Basque Meteorological Agency (EUSKALMET)
Continental station, located near an airport, protected against the landing strips.
Continental station in a rural area, clear surroundings without vegetation.
Coastal-hill station placed in the summit of a mountain. Located in an observatory with no vegetation.
Coastal station placed in a valley. Located inside an airport close to landing strips.
From the single local series of the observatories, a regional series was computed for each season through a principal component analysis as explained in Saenz et al. (2001). Finally, seven regional series were obtained per variable. One for the observational data for the period 1978–2000 and six series from models corresponding to the (1961–2100) or (1961–2050) period, depending on the RCM model.
3.2. Climatic indices
Impact studies often focus on extreme events because changes in such quantities are likely to have greater societal impact than changes in the mean of the distribution (Beniston et al., 2007, Energy policy, Kjellström et al., 2007). The option chosen here was to interpret the temperature projections for the present and future by means of standardized and widely used indices. The STARDEX indices defined during the EU-FP5 STARDEX project (http://www.cru.uea.ac.uk/projects/stardex) are focused on relatively moderate extremes and many of them are based on thresholds defined using percentile values. The indices selected reflect events occurring during the two seasons studied (JJA, DJF) and they are classified into different frequencies; that is, the analysis is made by means of indices which assess the seasonal or daily impact:
Hot day thresholds, defined as 90 percentile of maximum temperatures per summer (Tmax90p) and cold night thresholds, defined as 10 percentile of minimum temperatures per winter (Tmin10p).
Number of frost days defined as days where minimum temperature is less than zero degrees (125Fd); the percentage of days where maximum temperature is greater than the 90 percentile (192Tx90) and the percentage of days where minimum temperature is less than the 10 percentile (193Tn10).
Indices related with heat and cold waves duration and their frequency (txhw90, 145CWDI) defined hot/cold wave as six consecutive days having temperatures higher/lower than the seasonal temperature for the control period plus/minus 5 °C
For summer seasons, the discomfort index or temperature humidity index (THI) which includes parameters such as the specific humidity, the changes in the surface pressure and the behaviour of the saturation vapour pressure (Wexler, 1976; Bolton, 1980; Buck, 1981). It is a human bio-meteorological index derived from the human energy balance (Mayer and Höppe, 1987 and Matzarakis and Mayer (1990)).
3.3. Bias correction methodology
Several methodologies and discretions exist to quantify the goodness of fit or bias but could be grouped in two main ways: the first approach is devoted to a statistical methodology, which identifies correlations between observable and predicted quantities. It is based on the conforming of the observed data distribution to a probability density function (PDF) and they assumed to be approximated by a gamma distribution or by a gauss distribution. A correction is obtained for daily, monthly or seasonally variables by the transfer cumulative distribution function (Alexander et al., 2005; Kjellstrom et al., 2007; Chauvin and Denvil, 2007; Planton et al., 2008; Piani et al., 2009).
The second technique calculates the bias based on empirical distributions in terms of a delta change temperature calculated from monthly temperature means over the entire year (Van Roossmalen, 2009) or seasonal temperature means (Ferro et al., 2005; Christensen and Christensen, 2007; Déqué, 2007; Jacob et al., 2007; Kjellstrom et al., 2007). Also, it is possible to combine the delta temperature method together with a calibration curve that provides the bias, daily, and monthly or seasonally, based on percentile corrections (Christensen and Christensen, 2007; Déqué, 2007; Kjellstrom et al., 2007).
The methodology used here follows the second approach based on the delta temperature change method combined with the calibration curve. One calibration curve is considered for each model and for each season (maximum daily temperatures for summer and minimum daily temperatures for winter). Twenty one percentiles are calculated for the control period series (1978–2000). The delta temperature (ΔT) is calculated by means of the subtraction of the corresponding percentiles temperatures (i.e. ΔT = Tmodelpercentile − Tobservedpercentile and so on).
The calibration curve provides the ΔT between models and observations stated for the control period. It is considered that the systematic bias during the control period corresponding to one percentile in the calibration curve will be stationary in the future, and future temperature projections are corrected on the basis of this curve. A linear interpolation method is applied for modelled temperatures, which value falls in the middle of two percentiles. For temperatures out of the curve, an extrapolation is performed depending on the adjustment of the fitting curve of each RCM. Different linear, quadratic or exponential extrapolations are applied for each model. The results are corrected series with modelled pdfs approximated to the observed pdf. The difference between them is due to the slight error introduced when applying the linear interpolation between two percentiles.
The calibration curves (Figure 2) showed that the six models are closer to the observations in the median range, while the bias in the extremes temperatures are further from the observations as already pointed out by other authors (Jacob et al., 2007; Kjellström et al., 2007; Christensen et al., 2008). It could be seen that the calibration curve shape is highly dependent on the GCM (Figure 2). For the Basque Country, extreme minimum temperatures are overestimated by all models, while the correction applied to extreme maximum temperatures varies in a wider range. Modelled values overestimate (in particular for HIRHAM–ARPEGE) or underestimate the observations. Table III shows, for the reference period, the interval range for extreme temperatures (10th and 90th percentile for wintertime minimum and summertime maximum temperatures, respectively) and their associated bias correction. This allows visualize in the calibration curve the range of extremes temperatures for summer or winter.
Table III. For the reference period, interval for the modelled wintertime 10th percentile of the minimum temperature and summertime 90th percentile of the maximum temperature associated to their bias correction
Temperature ( °C) not corrected 10p
ΔT ( °C) 10p
Temperature ( °C) not corrected 90p
ΔT ( °C) 90p
4. Projected temperature changes
4.1. Variability and goodness of fit between RCM and observations
A statistical analysis was performed to assess how the main systematic bias varies across the models and to evaluate the bias correction methodology results. Figure 3 shows the box plots for the original and corrected daily maximum temperatures for summertime and minimum temperatures for wintertime versus the equivalent observed ones. As all models underestimates wintertime minimum temperatures, before the correction, the modelled series contain more extreme cases (circles and asterisks in the lower part of the figure). Although the interquartile range (IQR) for modelled minimum temperatures is smaller than the observations, it falls inside of the observation range. The median is overestimated by HIRHAM, ALADIN, REMO, RACMO models and underestimated by CLM and PROMES models. For the maximum temperatures, series present higher variability so the IQR and the box plots are larger than for the minimum temperatures. The median fluctuates overestimating and underestimating the observed median, in accordance with Table III. Even an identical distribution of probability could be expected for corrected series and the observed ones after using the bias method, differences are introduced (as explained in section 3.3) during the interpolation process. It is for this reason that after the correction, the size of the boxes and the IQRs for models are near identical to the observations with the quantiles slightly shifted (<1 °C).
4.2. Climatic indices
After correction, during winter seasons, for the reference period (1978–2000), the 10th percentile of the minimum temperature (Tmin10p) is − 3.7 °C (with a standard deviation of 2.42 °C) and − 4.4 °C (with a standard deviation of 0.37 °C) for the model ensemble daily average. The ensemble future projection (2001–2100) shows an expected increase in the temperature upto 3 °C until the end of the XXIst century (Figure 4). Table IV summarizes the statistics of the tendency for two 30-year-window periods extracted from the future projections. The number of days where the temperature is lower than the 10th minimum temperature percentile (193Tn10) shows the same upwards trend in temperature: in general, the models indicate 30% of decrease along the future projection (2001–2100). For the number of the frost days, the spread in the goodness of fit with the observations is ± − 3 days in the control period (Figure 5). For the future projection, the decrease in this climatic index is 70% until the end of the XXI century with respect to the control period (Table IV).
Table IV. Ensemble average and ensemble standard deviation, the wintertime 10th percentile of minimum temperature and the number of frost days and in summer, the 90th percentile of maximum temperature
Ensemble Standard Deviation
Tmin10p ( °C)
Number of frost days
Tmax90p ( °C)
In the reference period, the ensemble models showed that, on average, the duration range of a cold-wave episode (145CWDI) varied from 7 to 20 days. The maximum number of days involved in cold waves for a single winter was of 60% and happened in 1996. In the period 2020–2050, the duration of each cold wave, on average, is reduced with a range varying between 6 and 10 days. The maximum number of days involved in cold waves for a single winter will decrease in of 20%. During the last 30 years of the century CLM, PROMES and ALADIN do not simulate any cold waves, while HIRHAM, RACMO and REMO models indicate that the duration of the cold waves varied from 6 to 10 days, with a maximum of 20% of the days. The occurrence and the total number of cold days involved in a cold wave will decrease but the ensemble modelled temperature involved will remain almost constant (−6.67 °C for the control period, − 6 °C for the 2020–2050 and − 5.8 °C for the period 2070–2100).
During summer, for the reference period, the 90th percentile of daily temperature (Tmax90) for each model shows a good fit with the observations (Figure 6): the observed temperature on average is 34.4 °C in this period (with a standard deviation of 1.9 °C) and 33.9 °C (with a standard deviation of 0.4 °C) on average for the ensemble models. The future projection (2000–2100) shows an expected increase in temperature upto 3.5 °C on average for the ensemble model. Table IV summarizes the trend for two 30-year-window periods extracted from the projections.
All the models show an increase in the total number of days involved in a heat wave (txhw90): while in the reference period (1978–2000), on ensemble model average, the duration of each heat wave is 15 days, at the end of the century the ensemble average is 24 days. Hence, the total number of a heat wave and the occurrence of the waves, on average, would increase by 30%. However, the temperature involved will remain almost constant: 31.7 °C for the control period, the average of 32.4 °C for the 2020–2050 and 32.4 °C for the period 2070–2100. This is the same tendency found in other European studies such as Schär et al. (2004) and Beniston and Diaz (2004).
The Basque Country region, due to the characteristics of its climate, presents a great level in the comfort of the population comparing to other parts of the Iberian Peninsula. However, due to the results obtained in the indices based on the maximum temperatures, considerations should be given to the trend of the comfort of the population for future scenarios. There are several methodologies to analyse the thermal comfort of human in a thermo-physiologically point of view using meteorological data. The predicted mean vote (Mayer, 1990), the physiologically equivalent temperature (Mayer and Höppe, 1987), the discomfort index (Thom, 1959), heat stress (Terjung, 1967), the THI, etc are considered human bio-meteorological indices derived from the human energy balance. Here, the THI index was chosen because, it classifies the thermal environment for the population using only the temperature, RH, vapour pressure and the dew point temperature that are fields available in the ENSEMBLES dataset. The THI expresses the discomfort of the population by means of several values: between 70 and 75 from 10 to 50% of the population feels uncomfortable, and when the THI index goes upto 80, between 90 and 100% of the population feels uncomfortable (Gates, 1972).
Figure 7 shows that during the control period, from the observations, between 10 and 25% of the population they do not feel very comfortable. This interval corresponds to the most sensitive population such as elderly people and pregnant women. During the entire period, the spread of the models varies between 4 and 5 units of THI and there is no clear increase or decrease in the ensemble models, it can be due to the fact the models simulate a decrease in the RH. The CLM–HadCM3 is the only model indicating a sign of increase: from 67 in the reference period and THI equal to 75 by the end of the century. The REMO–ECHAM5 model does not fit well with the observations: it shows a high internal variability in the THI along the entire period and in the last decade of the century the THI increases from 70 to 75 units.
When comparing the RCMs with their driving GCM, a great influence on the pairs of RCMs driven by the same GCM is found, as it is shown in the indices analysed and depicted in Figures 4–6. The pairs of RCM follow a similar behaviour along the time and the correction of the systematic bias does not affect the trend of the models.
The objective of the study is to analyse the changes over the Basque Country region in the tails of the temperature distribution for application on the impact and adaptation to climate change studies coordinated in the frame of the multidisciplinary project K-EGOKITZEN. An ensemble of outputs from six RCMs from the EU-FP6 ENSEMBLES project at 25 × 25 km horizontal resolution is used to analyse daily maximum and minimum near-surface air temperature and RH over land. Extreme events are assessed using climate change indices derived from daily and seasonal temperature. These indices are defined by the STARDEX methodology for summer and winter seasons for the control period (1978–2000) and future projections (2001–2100). The approach taken here for the correction of the bias between models' outputs and observations is based on 21 temperature percentiles which provide a calibration curve. The calibration curve is calculated by means of the daily temperatures during the control period, the minimum temperatures are used for winter season and the maximum temperatures are utilized for summer seasons.
It is found that the six models present the largest systematic bias for the extreme minimum or maximum temperatures in accordance with other studies (Moberg and Jones, 2004; Kjellstrom et al., 2007). When comparing the RCMs with their driving GCM, a great influence is found on the pairs of RCMs driven by the same GCM, the pairs of RCM follow a similar behaviour along the time and the correction of the systematic bias do not affect the trend of the models.
The climatic indices show both for maximum (90th percentile) and minimum (10th percentile) temperatures an increase upto 3 °C for winter seasons and 3.5 °C for summer seasons during the XXIst century (2000–2100). The number of frost days, the days with temperature less than 0 °C, indicates a decrease of 70% in the total days in the winter from the control period to the last decade of the century. The ensemble models in the reference period showed that, on average, the duration of each cold-wave episode varied from 7 to 20 days. The maximum number of days involved in the cold waves for one winter was 60%. During the period 2020–2050, the duration of each cold wave, on average, is reduced from 6 to 10 days, with a maximum of 40% of the days involved. During the last 30 years of the century, CLM, PROMES and ALADIN do not simulate any cold waves, while HIRHAM, RACMO and REMO models indicate that the duration of the cold waves varied from 6 to 10 days, with a maximum of 20% of the days involved in the cold waves. The ensemble modelled temperature involved in the cold waves will remain almost constant (−6 °C). The heat wave duration index (txhw90) indicates that all the models show an increase in the total number of days involved in each heat wave: while in the reference period, on ensemble model average, the duration of each heat wave is 15 days, at the end of the century is 24 days. Hence, the total number of a heat wave and the occurrence of the waves, on average, would increase by 30%. However, the temperature involved in the heat waves will remain constant (32 °C ± 1) during the XXth century (2000–2100).
The THI, which expresses the discomfort of the population, indicates that during the control period from observations, between 10 and 25% of the population feel not very comfortable corresponding to the most sensitive population. During the future period (2001–2100), there is no clear increase or decrease in the ensemble models. The CLM–HadCM3 is the only model indicating a sign of an increase from 10% of the population in discomfort during the reference period, to 50% for the population in discomfort by the end of the century.
Financial support for this work is gratefully acknowledged to the K-EGOKITZEN project and, LABEIN-Tecnalia and—Iñaki Goenaga Technological Foundations. Also, thanks to Dr Jon Saenz, Dr Michel Deque and Dr Hamish Mair for constructive discussions. The ENSEMBLES data used in this work were made available through funding from the EU FP6 Integrated Project ENSEMBLES (Contract number 505539) whose support is gratefully acknowledged.