INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. 34: 235–248 (2014) Published online 12 March 2013 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/joc.3683

Seasonal precipitation variability in regional climate simulations over Northern basins of Tunisia Zoubeida Bargaoui,a* Yves Tramblay,b Emmanuel A. Lawinc and Eric Servatb a

Ecole Nationale des Ing´enieurs de Tunis (ENIT), Universit´e Tunis El Manar, Tunisia b IRD, Hydrosciences-Montpellier, France c Laboratoire d’Hydrologie Appliqu´ee (LHA), University of Abomey-Calavi, Benin

ABSTRACT: Northern Tunisia is the rainiest part of the country where most of the water management structures (dams, reservoirs, etc) are located. Its strategic situation with respect to surface water resources encourages the investigation of the climate change impacts projected by climate models. The goal of this study is first to compare the observed precipitation with climate model outputs, and then to evaluate the future changes projected by different climate models. The study area is subdivided into four regions: the upstream and downstream transboundary Medjerda basin, the northern coastal basins and the eastern coastal basins. A database provided by the Tunisian hydrological service includes 388 stations with complete monthly precipitation data over the period 1961–2000. An ensemble of Regional Climate Models (RCM) simulations provided by the European Union-funded project ENSEMBLES are used. Six RCM model runs (CNR-A, DMI-A, DMI-B, ICT-E, SMH-B and SMH-E) are analysed, for the control period 1961–2000 and two projection periods, 2011–2050 and 2051–2090. The models efficiency in reproducing seasonal precipitation amounts and variability is evaluated. A 1km monthly precipitation reference grid is computed through the interpolation of rainfall observations during the period 1961–2000 with kriging techniques. Monthly precipitation series averaged over the four basins are built for comparison during the control period. The RCM outputs are evaluated with respect to the annual precipitation cycle and rainfall frequency distribution using robust statistics. For the control period, features of the seasonal regimes are well reproduced by all models. It is found that models underestimate seasonal precipitation on average by 20%. The discrepancy between model outputs and observations depends on the season. For the future, in summer and autumn the different models do not project major changes in the seasonal distributions. However, for winter and spring, all the models project a significant decrease of precipitations. KEY WORDS

Tunisia; precipitation; climate change; regional climate models

Received 13 February 2012; Revised 7 November 2012; Accepted 11 January 2013

1. Introduction The Northern part of Tunisia is the rainiest region of the country where, according to Gaussen and Vernet (1951), the average annual precipitation ranges from 806 mm in the northern coastal part to 489 mm in eastern coastal part. The main surface water management works are located in that area. They include several dams together with a concrete channel of 120-km length (Medjerda – Cap Bon canal). They help satisfying the water demand (mainly domestic water and irrigation) for the whole country. The strategic situation of the study area with respect to surface water management encourages the investigation of the climate change impacts as projected by climate models. The main goal of this article is first to evaluate the potential of climate model simulations to reproduce past observations of precipitation and the statistical characteristics of future climate projections. * Correspondence to: Z. Bargaoui, Ecole Nationale des Ing´enieurs de Tunis (ENIT), Universit´e Tunis El Manar, Tunis, Tunisia. E-mail: [email protected]

 2013 Royal Meteorological Society

For the future, climate change projections indicate a possible precipitation decrease that could exceed −25% in the southern part of the Mediterranean basin, including Tunisia (Giorgi and Lionello, 2008; Paeth et al., 2009; Patricola and Cook, 2010; Garc´ıa-Ruiz et al., 2011). In a recent study, Nasr et al. (2008) adopted the HadCM3 Global Climate Model (GCM) with the A2 greenhouse effect scenario to analyse the future dry projections in Tunisia by comparison to the period 1961–1990. They suggest a fall of 5–10% in precipitation amounts for the 2020–2050 period. However, they did not investigate the bias of model outputs by comparing them to observations during a control period. Recent advances in dynamical downscaling have provided climate simulations at a resolution of 25 km, which are better at reproducing regional climate features such as orographic precipitation, while the topographic complexity is very poorly resolved at GCM resolution (Fowler et al., 2007; Paeth et al., 2009). Dynamical downscaling refers to the use of Regional Climate Models (RCM) with large-scale and lateral boundary conditions from GCMs to produce higher resolution outputs. The temperature and precipitation biases in RCM

236

Z. BARGAOUI et al.

outputs have been thoroughly investigated on various spatial and temporal scales over Europe (Schmidli et al., 2007; Herrera et al. 2010; Kjellstr¨om et al., 2010). In the particular case of the Mediterranean region (Southern Europe), Jacob et al. (2007) observed a bias of −0.5 K for temperatures and −1 mm d –1 for precipitation in winter. Also, in the winter season, it is found that inter-annual variability is reduced by comparison with the observations. In the summer season, models are generally too warm and dry while the modelled temperature variability is larger than for observed values. More recently, Heinrich and Gobiet (2011) delineated the range of deviations based on daily gridded precipitation and surface air temperature interpolated from observed data for the period 1961–1990 (Haylock et al., 2008). Using eight different RCMs, the resulting deviations were ranging from −12 to +10.9% for monthly precipitation data and from −0.5 to 0.9 K for monthly surface air temperature. In Tunisia, no studies to our knowledge have evaluated the ability of climate models to reproduce the monthly or seasonal patterns of precipitation. This is mainly due to the absence of reference gridded data sets such as the E-OBS data set available for Europe (Haylock et al., 2008). Consequently, there is a need for the study region to validate the RCM outputs with respect to observed precipitation, in order to make future projections in the light of the model performances in present climate. This is a key question since, as outlined by Teutschbein and Seibert (2010) or B´ardossy and Pegram (2011), biases in RCM outputs have enormous impacts on the calculation of the hydrological cycle. Correction procedures such as Quantile–Quantile transforms are thus performed to obtain unbiased RCM outputs to be used with hydrological models (Piani et al., 2010). B´ardossy and Pegram (2011) proposed to correct the RCM outputs using a frequency distribution of daily rainfall series dependent on synoptic circulation patterns. This methodology was applied to Rhine basin. In this study, we did not implement a bias correction procedure because it is a first step to evaluate the model outputs. However, for hydrological impact studies using climate model outputs such correction would be required. This study intends to compare various RCM outputs with ground precipitation in present climate, and the convergence of the different model projections in future climate. Several studies came to the conclusion that both the choice of the GCMs and RCMs are major sources of uncertainty (D´equ´e et al., 2007; Tebaldi and Knutti, 2007; D´equ´e et al., 2012). Indeed, the analysis of the convergence in the projections given by different models may help to evaluate the uncertainties on rainfall projections. To avoid under-sampling of uncertainty, several authors recommend using at least two or more RCMs that are forced by at least two GCMs (van der Linden and Mitchell, 2009; Heinrich and Gobiet, 2011). Surface observation data and model simulations are presented in the Section 2. In Section 3, the interpolation methodology is presented with the statistical framework  2013 Royal Meteorological Society

of model-surface data and model–model comparison. Results are presented in Section 4.

2. Data 2.1. Surface precipitation data The study area (Figure 1) lies between 36◦ N and 38◦ N and 7◦ E and 8◦ E. It is subdivided into four subregions: (1) the Upper Transboundary Medjerda river basin, upstream the Sidi Salem dam (focusing on the Tunisian part), denoted as Medjerda U ; (2) the river Medjerda valley (downstream Sidi Salem dam), denoted as Medjerda D; (3) the northern coastal catchments (Zouara, Sidi El Barrack, Lake Ichkeul basins), denoted as North and (4) the eastern coastal catchments (Cap-Bon region and wadi Meliane basins), denoted as Cap Bon. A 40-year period (1961–2000) is considered to compare historical rainfall data with simulated RCM data. Rainfall data are collected in the study area since the late 19th century. Precipitation database is provided by the Tunisian hydrological service (DGRE) including 388 stations in operation with complete monthly data over the period 1961–2000. However, only 356 stations have well-identified geographical locations (latitude and longitude) with 152 stations in the Medjerda river basin, 71 stations in the northern coastal catchments and 133 stations in the eastern coastal catchments (Figure 1). The number of available stations by year and by subregion is reported in Figure 2 which outlines the growth of rainfall networks through time. The Medjerda river basin covers 15 789 km2 in its Tunisian part. The northern coastal catchments cover about 4791 km2 . The eastern coastal catchments represent 7309 km2 . Finally, it is important to outline the importance of network coverage (a density of 0.012 stations per km2 is achieved). 2.2. Data from regional climate model simulations An ensemble of different regional climate model (RCM) simulations is considered. The RCM data is provided by the FP6-ENSEMBLES European project (http://www.ensembles-eu.org/), a set of high-resolution (25 km) RCM simulations driven by different general circulation models (GCM) from 1950 to 2100. Beyond 2000, they use the A1B scenario for greenhouse gas concentration. A summary of grid configurations, parameterizations and boundary conditions that have been carried out in several RCMs is presented in Jacob et al. (2007). A detailed description of the ENSEMBLES models and further references can be found in D´equ´e et al. (2012). More generally, the link http://ensembleseu.metoffice.com/papers.html provides published papers related to ENSEMBLES. This set of simulations has been used in several studies to evaluate the adequacy of the model simulations with past observations of precipitation and the future projections for Europe (Halenka et al., 2006; May, 2007; D´equ´e et al., 2007; Herrera et al., 2010; Heinrich and Gobiet, 2011). In this study, a subset of simulations from the ENSEMBLES Int. J. Climatol. 34: 235–248 (2014)

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

237

Figure 1. Study area and available precipitation stations (dots).

Figure 2. Number of available stations by year and by subregion.

data set is considered, including the RCMs of the French meteorological service (CNRM, Radu et al. 2008), the Danish Meteorological Institute (DMI, Christensen et al. 1996), the International Center for Theoretical Physics in Trieste, Italy (ICTP, Giorgi and Mearns 1999) and the Swedish Meteorological and Hydrological Institute (SMHI, Kjellstr¨om et al. 2005). The driving GCMs for these RCM simulations are the Max Planck Institute (ECHAM5) model, the Bergen Climate Model (BCM) and the ARPEGE model developed in M´et´eo-France. Table 1 reports the different simulations characteristics of the multi-model ensemble. The rainfall simulated by the RCMs was averaged over the four basins  2013 Royal Meteorological Society

considered in the study at the monthly and seasonal timescales.

3.

Methodology

To reflect the spatial variability of monthly surface rainfall patterns, the first step was to provide an interpolated regular grid. This interpolation issue is an important step as mentioned by Haylock et al. (2008) who performed such work for the European region at the daily time scale for the period 1950–2006 using various kriging methods. In this study, monthly precipitation was interpolated Int. J. Climatol. 34: 235–248 (2014)

238

Z. BARGAOUI et al.

Table 1. RCM simulations. Institute

Scenario

Driving GCM

RCM model

Resolution (km)

Acronym

CNRM DMI

A1B A1B A1B A1B A1B A1B

ARPEGE ARPEGE BCM BCM ECHAM5-r3 ECHAM5-r3

Aladin DMI-HIRHAM5 DMI-HIRHAM5 RCA RCA RegCM

25 25 25 25 25 25

CNR-A DMI-A DMI-B SMH-B SMH-E ICT-E

SMHI ICTP

using point kriging at every grid node with 1-km resolution covering the study region. The Medjerda basin (Medjerda U and Medjerda D) was considered as a single domain in the kriging study because it is a consistent physical unit. Krigged monthly rainfall values during the period 1961–2000 were then aggregated for each subregion (1)–(4) by averaging monthly rainfall. Seasonal totals were subsequently calculated adopting four seasons: Autumn SON (September, October, November), Winter DJF (December, January, February), Spring MAM (March, April, May) and Summer JJA (June, July, August). Annual totals were also computed. 3.1.

Kriging monthly precipitation data

Let ξ (x ) the monthly rainfall at the location x for a given month of a given year; h is the inter-distance between locations x and (x + h). The semivariogram function γ (h) which is a second-order statistic reflects the variability for a stationary covariance structure (which is assumed here).  1
γ (h) = E (ξ (x + h) − ξ (x ))2 (1) 2 In Equation (1), E(X ) holds for the expectation (mean) of the random variable X . For every month, the sample semivariogram γ exp (h) is estimated using pairs of stations: 
 1 γexp (h) = Nh−1 (2) (ξ (x ) − ξ (x + h))2 2 where Nh is the number of pairs of stations at the inter-distance h. A step of 3.5 km was assumed to build inter-distance classes. The median inter-distance of each class was adopted for h. Ordinary kriging (OK) assumes unknown mean as well as second-order stationary process. At each grid point i , the ordinary kriged value ξi∗ is obtained by averaging the observations at N neighbouring stations ξ j (N is delineated by trials and errors). N  λj ξj (3a) ξi∗ = j =1

Weights λj are obtained by solving the kriging system (Equation (3b)). The kriging system is obtained through the minimization   of the variance of the unbiased kriging error ξi∗ − ξi . The minimization of the error variance at a given point j under the non-bias condition involves the computation of a Lagrange parameter µ and the estimation of the  2013 Royal Meteorological Society

model semivariogram γ j 0 for the corresponding interdistance between the point j and any observed point ξ 0 . N 

λi γij + µ = γj 0 , j = 1, N

(3b)

j =1

The condition

N 

λj = 1 corresponds to non-bias

j =1

condition. To evaluate whether the estimations are sufficiently accurate, a cross validation (Jackknife) is performed: observations are dropped one by one and then kriging is performed at the dropped location using the data from the remaining stations. Subsequently, the bias and standardized residuals are computed. Standardized residuals are obtained by dividing the residuals between the observed and krigged values by the standard deviation of the kriging error (Glatzer and Muller 2004). If the absolute standardized error is higher than two, outliers are suspected. 3.2. Variogram assessment We computed the root mean square error (RMSE) between the experimental and the modelled semivariograms to evaluate the quality of the model:   NC  γi sample − γi mod 2 (4) RMSE =  γi mod i =1

where NC the total number of inter-distance classes, γ i sample the sample semivariogram, γ i mod the model semivariogram in the inter-distance class i . The spherical model was adopted to adjust the sample semivariogram. Model parameters are the range parameter AA, the sill parameter OM and the nugget effect parameter C .   h 3 h γmod (h) = OM 1.5 AA +C − 0.5 AA for h = 0; h < AA γmod (h) = OM + C γmod (h) = C

for h > AA for h = 0

(5)

The spherical model is a widely used semivariogram model, characterized by a linear behaviour at the origin (Goovaerts, 2000). In addition, the spherical model is a Int. J. Climatol. 34: 235–248 (2014)

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

convenient tool for precipitation kriging because it provides a value of the decorrelation distance – given by the value of the range – (Lebel et al., 1987; Lebel and Laborde, 1988). Therefore, it is most often used with precipitation data (Ly et al., 2011; Verworn and Haberlandt, 2011). It must be noted that the exact matching of the experimental and theoretical semivariogram is considered more important than the choice of semivariogram model itself (Wackernagel, 1995). To adjust the spherical model parameters, the minimization of RMSE was performed using simulated annealing (SA) which is a stochastic optimization algorithm introduced by Kirkpatrick et al. (1983) and inspired from the thermodynamics of metal annealing. This method was recently adopted as optimization method in the automatically search of the best locations to implement new rain gauges (Chebbi et al., 2011) in the study area. The calibration was performed under two assumptions: C = 0 and C = 0. Consequently, monthly rainfall semivariograms were automatically calibrated for every month of every year. In the case where the nugget effect was assumed equal to zero, monthly climatological semivariograms were obtained after averaging the 40year adjusted AA and OM parameters for every month. Finally, OK using climatological as well as year-to-year semivariograms were performed at every grid node. For comparison purposes an interpolation by Thiessen polygons (nearest neighbor) was also carried out. 3.3. Evaluation of RCM model outputs of precipitation using robust statistics RCM models efficiency in reproducing precipitation amounts over the study domain is addressed through the diagnosis of the ability of RCMs to reflect historical precipitation variability. First, RCM outputs are evaluated with respect to the annual precipitation cycle and spatial variability during the control run 1961–2000. Then, the degree of matching between RCM and ground data estimates is evaluated for the control run by comparing the seasonal distributions with the Kolmogorov–Smirnov (KS) test (Darling, 1957). The KS test statistic is computed from the maximum difference between two empirical cumulative distribution functions (CDF): KS = max |CDFdata (x ) − CDFmod (x )|

(6)

The null hypothesis (H 0 ) is that the two samples are from the same continuous distribution. The alternative hypothesis is that they are from different continuous distributions. The critical values for the KS statistic are obtained by a bootstrap procedure. To evaluate the future changes projected by the climate models, the changes between the control period and two different projection periods are analysed with robust statistics. Starting from the estimation of quantiles X p and assuming independent data, the evaluation of the differences was based on order statistics according to Ferro et al. (2005) and Hannachi (2006) who adopted nonparametric estimation methods  2013 Royal Meteorological Society

239

to assess the changes in probability distribution functions of climate time series. This framework was adopted to evaluate both the statistical significance and the magnitude of the differences. Comparisons of location measures (the median X 0.5 ), spread measures (interquartile range, IQR = X 0.75 − X 0.25 ) as well as lower and upper quantiles (X 0.1 and X 0.9 ) were performed in this work. For x 1 , . . . xn and y 1 , . . . ym , two samples of the same variable for two different periods of a model run, the changes in the pth quantile can be estimated from: mp = yp − xp

(7)

Similarly, changes in the IQR can be estimated with: IQR = IQRy − IQRx

(8)

The statistical significance of the differences can be evaluated by computing their standard errors. In the present study, the 5% confidence level is considered. The full equations for this computation can be found in Hannachi (2006). 3.4.

Model ranking

An attempt of model sorting relatively to their efficiency in reproducing the precipitation distributions was achieved. The KS test statistic was recently considered by Giraldo Osorio and Garcia Galiano (2013) to obtain reliability factors for different RCMs over Spain. A similar measure, provided by the Cram´er–von Mises test statistic was used by Tramblay et al. (2012) to weight RCM outputs over Morocco. Another equivalent metric frequently used for climate model evaluation is the Perkins score (Perkins et al., 2007). This metric calculates the cumulative minimum value of two distributions, thereby measuring the common area between two probability distribution functions (Perkins et al., 2007). Most often, weights are derived from a statistic based on the mean square difference between observed and modelled climate variables (equivalent to the Cram´er–von Mises statistic). The use of nonparametric measures of distances allows a flexible use with different types of distributions. Two different methods were adopted as measures of model efficiency in reproducing the seasonal precipitation distributions. The first method is based on the KS statistic, which quantify the maximum distance between two CDF. The quantity (1−KS) is assumed as a measure of model performance. Thus, for every model i (i = 1, 6) a score W i ,KS was attributed according to KS seasonal values (by giving all seasons the same weight). Wi ,KS =

   1
 1 − KSi ,SON + 1 − KSi ,DJF 4     + 1 − KSi ,MAM + 1 − KSi ,JJA

(9)

In order to classify the models, the scores W i ,KS were normalized using the ratio: Wi ,KS Win,KS =  Wi ,KS

(10)

Int. J. Climatol. 34: 235–248 (2014)

240

Z. BARGAOUI et al.

Figure 3. Time series of the adjusted parameters AA and OM for the month of December.

Figure 4. Comparison of annual and climatological variograms for December 1967.

The second method is based on the IQR differences. Thus, the difference in the spread of the distributions between historical data and the models are computed for every season to build a score: Wi ,IQR

1
IQRi ,SON + IQRi ,DJF = 4  +IQRi ,MAM + IQRi ,JJA

4. Results 4.1. Analysis of historical data (11)

The normalized score using the change in IQR is: 

Win,IQR

−1 Wi ,IQR  =  Wi ,IQR

 2013 Royal Meteorological Society

For the two scores considered, higher values indicate a better model agreement with the observed distributions.

(12)

Figure 3 reports the time series of the adjusted parameters AA and OM corresponding to the month of December as example. These time series are obtained with spherical semivariograms without nugget effects. Year-to-year variability is noticed for both parameters. In Figure 4, the adjusted semivariogram for December 1967 is compared with the climatological semivariogram. Despite the Int. J. Climatol. 34: 235–248 (2014)

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

241

Figure 5. Monthly climatological variograms.

Figure 6. Kriged precipitation estimates versus observed monthly precipitation for the Medjerda basin in January.

differences in sill and ranges estimates under the two assumptions (annual or climatological semivariograms), only small differences are found between interpolation methods results (OK with year-to-year semivariograms, OK with climatological semivariograms, Thiessen interpolation – results not shown). This indicates the robustness of the spatial interpolation due to the high-station density and the good coverage available. Figure 5 reports the plots of monthly climatological semivariograms. The ranges are little varying depending on the season, from 10–20 km, but the sills show larger variations with higher values for the wet months (September to January). Indeed the sills are estimators of the spatial variance of the rainfall field and their temporal variability suggests a seasonal pattern. The quality of OK is reflected in cross validation (Jackknife) results. As an example, Figure 6 reports krigged versus observed monthly rainfalls for basin 3 in  2013 Royal Meteorological Society

January. The mean standardized error is equal to 0.015. The quality of fitting may be qualified as very good since 3. On the other hand, 9% of the krigged values have |u| > 2. Finally, we adopted OK of monthly rainfall with climatological semivariograms in further investigations. Time series of resulting annual cumuli are reported in Figure 7. They display important year to year variability. Indeed, the range of the averaged grid values varies from 200 to 700 mm in Cap Bon; from 300 to 850 mm in Medjerda U, from 250 to 750 mm in Medjerda D and from 450 to 1200 mm in the North. They are found in general good agreement with the Gaussen and Vernet (1951) map. The inter-annual variability of monthly rainfall totals is reported using box plots (Figure 8). The hypotheses of independence and stationarity have been tested on the Int. J. Climatol. 34: 235–248 (2014)

242

Z. BARGAOUI et al.

Figure 7. Time series of annual precipitation over the four basins .

Figure 8. Box-plots of the inter-annual variability of monthly rainfall totals. The boxes have lines at the lower quartile, median and upper quartile values; the whiskers extend from each end to the most extreme values.

resulting seasonal totals of precipitation for each basin. The Wald and Wolfowitz (1943) test was applied to assess the independence of series of seasonal totals. For all basins and seasons, at the 5% significance level, the hypothesis of independence is accepted, except for the Cap Bon in summer (JJA). Similarly, the Mann–Kendall test (Mann, 1945) was adopted to verify that observed seasonal totals are stationary. No trends were detected for all seasons and basins.  2013 Royal Meteorological Society

4.2. Analysis of the model control runs Comparison of the spatial variability of annual totals is reported in Figure 9 using the 25-km-grid mesh. A relative bias is computed between the annual precipitation from the RCMs and from the krigged observations. All models tends to underestimate the total precipitation, except for the southwestern region where precipitation totals seems to be overestimated over the upstream of the Medjerda basin. The model describing the best the Int. J. Climatol. 34: 235–248 (2014)

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

243

Figure 9. Observed mean annual precipitation (left) on a 25-km grid and relative bias in the RCM simulations for the same variable (right).

Figure 10. Monthly precipitation regime observed (thick blue line) and simulated by the different RCMs in each basin.

spatial distribution of annual precipitation is ICT-E (with a correlation coefficient r = 0.87 between the simulated and observed precipitation). At a monthly resolution, the comparison of spatial mean value for control and simulated rainfall in each basin suggested that the features of seasonal regimes are verified by all models (Figure 10). The seasonal patterns are well reproduced with dry summer and wet winter and the existence of two interseasons. Figure 11 which compares CDFs of control data and RCM data by region and by season outlined the differences either between models or between models and control data. It is found that model underestimation is in average equal to 20%. It is worth noting that differences between models depend on the season. This suggests that both location measures and spread ranges may differ notably between the models and between models and observed data. The differences between the observed and modelled distributions are evaluated with the KS test.  2013 Royal Meteorological Society

The test scores are presented in Figure 12. Results are mainly differing from season to season and do not depend on the basin (geographical location). The greatest number of H 0 acceptations (when the KS scores are below the critical value at the 5% level, the horizontal lines in Figure 12) is found for SMH-E (14 times) then ICT-E (9 times) followed by CNR-A (7 times). This finding shows the better ability of the model SMH driven by the Echam GCM, to reproduce with accuracy the seasonal distributions of precipitation in Northern Tunisia. Model skill scores averaged for all the basins according to KS statistic as a measure of performance (Equation (10)) gave the highest score to SMH-E and the second highest score to ICT-E (Figure 13(a)). Similarly, model ranking using the spread of the distribution criteria (Equation (12)) suggested that the highest score corresponds to ICT-E followed by CNR-A and then by SMH-E (Figure 13(b)). Thus, it is encouraging to find Int. J. Climatol. 34: 235–248 (2014)

244

Z. BARGAOUI et al.

Figure 11. Empirical cumulative distribution functions of seasonal precipitation observed and simulated by the different RCMS for each basin.

Figure 12. Kolmogorov–Smirnov test statistic between the observed and simulated seasonal precipitation CDFs during the control period 1961–2000. The horizontal solid lines are critical values corresponding to the 5% significance level.

that the selected three best models are similar using the two criteria which is a guaranty of robustness. 4.3.

Analysis of future model projections

The differences in the 10th, 50th, 90th percentiles and the IQR of the distributions between the control period 1961–2000 and two projection periods, 2011–2050 and 2051–2090, are tested using robust statistics. Tables 2 and 3 report the significant relative changes at the  2013 Royal Meteorological Society

5% confidence level by season and by model for the two projection scenarios 2011–2050 and 2051–2090. The mean relative changes from all the six models (ENSEMBLE, in Tables 2 and 3) and the mean of ICT-E and SMH-E (ICT+SMH, in Tables 2 and 3) are also presented. For summer and autumn seasons, the models do not project major changes in the seasonal distributions, except the RCMs driven by BCM projecting a decrease in summer. For winter and spring seasons, all Int. J. Climatol. 34: 235–248 (2014)

245

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

Figure 13. Model skill scores averaged for the four basins, computed from the KS statistic (Equation (10) – left) and from the IQR differences (Equation (12) – right).

Table 2. Significant relative differences (at the 5% level) in the 10th, 50th and 90th percentiles (X10, X50, X90) and the IQR between the reference period 1961–2000 and the projection 2011–2050. Cap Bon Season

Models

Medjerda U

Q10 Q50 Q90 IQR Q10

CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH DJF CNR-A DMI-A DMI-B −0.44 SMH-B ICT-E SMH-E ENSEMBLE −0.23 ICT+SMH MAM CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH −0.22 JJA CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH

Medjerda D

Q50 Q90 IQR Q10

North

Q50 Q90 IQR

Q10

Q50

Q90 IQR

SON

−0.25

0.02 1.07 −0.36

−0.34

−0.48 −0.30 −0.21 −0.32 −0.32 −0.08

−0.22 0.14 −0.32 −0.15

−0.20 −0.25

−0.45 −0.41 −0.25 0.75 0.90

models project a decrease in the median and the 10th and 90th percentiles, in particular for the period 2051–2090. These projected changes are on average of −23% in DJF and −35% in MAM for 2051–2090. For 2011–2050, it is mainly in the Medjerda and North basins that model  2013 Royal Meteorological Society

−0.51 −0.26 −0.17

−0.33 0.21 −0.45 −0.38 −0.15 −0.37 −0.24 −0.25

−0.11

0.16

−0.12 −0.41 −0.44

0.90 0.78

projects a decrease, in particular for the 10th percentile and the median. For 2051–2090, all the distribution is shifted towards lower values. Therefore, the changes are expected to amplify during the period 2051–2090 by comparison to 2011–2050. Int. J. Climatol. 34: 235–248 (2014)

246

Z. BARGAOUI et al.

Table 3. Significant relative differences (at the 5% level) in the 10th, 50th and 90th percentiles (X10, X50, X90) and the IQR between the reference period 1961–2000 and the projection 2051–2090.

Season

Models

CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH DJF CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH MAM CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH JJA CNR-A DMI-A DMI-B SMH-B ICT-E SMH-E ENSEMBLE ICT+SMH

Q10

Cap Bon

Medjerda U

Q50 Q90 IQR Q10

Q50

Q90

Medjerda D IQR

Q10

Q50

Q90

North IQR

Q10

SON

Q90 IQR

−0.28 −0.29 −0.12

−0.09

−0.42 −0.22 −0.34 −0.31 −0.27 −0.16 −0.22 −0.63 −0.39 −0.38 −0.27 −0.47 −0.36 −0.21 −0.39

−0.47 0.79 −0.35 −0.31 −0.26 −0.31 −0.29

−0.31 −0.28 −0.21 −0.30 −0.23 −0.25

−0.30

−0.70 −0.36 −0.37 −0.29 −0.29 −0.27 −0.37 −0.27 −0.49 −0.28 −0.37 −0.27 −0.16 −0.43 −0.28 −0.41 −0.33 −0.33 −0.34

−0.18

Conclusions

This study provided the first attempt to validate the precipitation simulated by regional climate models over Tunisia. The methodology to validate the RCM outputs with respect to observed precipitation data is first based on the generation of an interpolated grid of monthly rainfall time series obtained by ordinary kriging. The krigged values where computed using a dense rainfall network of monthly precipitation data during a 40-year  2013 Royal Meteorological Society

−0.15

−0.22 −0.31 −0.18 0.18 −0.27 −0.23 −0.25 −0.30 −0.22 −0.22 −0.20 −0.26 −0.24

On the opposite, very little changes are found in the spread of the distributions, measured here by the IQR. However, all the significant changes detected for the IQR are towards an increase, indicating a greater seasonal variability for the future. It must be noted also that the ensemble average of model outputs sometimes yields significant changes even if no significant changes are detected in the individual model projections. Therefore, ensemble of climate projections needs to be evaluated with care. On the opposite, the ensemble mean changes computed with the six models or only with the ICT-E and SMH-E models yields very similar results. This finding is similar to the results obtained by D´equ´e and Somot (2010). 5.

Q50

−0.07 −0.30 −0.44 −0.27 −0.46 0.41 −0.40 −0.26 −0.40 −0.28 −0.19 −0.27 −0.32

−0.74 −0.39 −0.41 −0.24 −0.24 −0.35 −0.19 −0.34 −0.22 −0.09 −0.19 −0.31

−0.16

−0.17

−0.27 −0.26 −0.24 −0.29 −0.24 −0.26 −0.27 −0.72 −0.42 −0.43 −0.33

−0.27 −0.23 −0.18 −0.27 −0.22 −0.22

−0.43 −0.32 −0.23 −0.10 −0.35 −0.28 −0.29 −0.27 −0.48 −0.56

control period (1961–2000). The validation itself was based on the comparison of cumulative distributions of the control series and the RCM simulated series at the seasonal scale. Six RCM simulations from the ENSEMBLES data set were considered. The KS test was implemented to compare the observed and simulated empirical distributions. A model skill score based on seasonal distributions was proposed and the three best models with respect to this criterion were identified (SMH-E, ICT-E and SMH-B). On the other hand, robust statistics were used to compare the quantiles and the IQRs of the observed and simulated distributions. Another model skill score has been introduced, based on the difference in IQRs between the control and simulated series. The three best models according to this criterion (ICT-E, CNR-A and SMH-E) were found analogous to those identified by the KS criterion; but, they were ranked in a different order. Consequently, the choice of a skill score metric can lead to different interpretation of the results, they need to be chosen with care and in relation with the objective pursued. Overall, the RCMs and in particular those driven by ECHAM are able to reproduce the spatial distribution, the inter-annual variability and the seasonal patterns of precipitation over Tunisia Northern basins. Int. J. Climatol. 34: 235–248 (2014)

REGIONAL CLIMATE SIMULATIONS OF PRECIPITATION OVER NORTH TUNISIA

Future projections for the periods 2011–2050 and 2051–2090 were compared with the simulations during the control period 1961–2000. The median, 10th and 90th percentiles values as well as the IQR were considered to carry out the comparison. A significant decrease in the median and the 10th and 90th percentiles in winter and spring seasons was observed for all models and basins. This global signal is intensified for the projection period 2051–2090 by comparison to 2011–2050. This indicate that climate models under the scenario A1B projects a shift towards lower precipitation for the central part of the distributions as well as the extremes. On the opposite, fewer significant changes are detected in summer and autumn. These findings appear critical for the future water resources management in Tunisia. In the light of these projections made from recent climate models, less surface water would require adaptation measures to mitigate the future impact on water supplies. The methodology developed here can be useful to build robust future scenarios for water resources in Northern Tunisia. The next step would be to provide bias-corrected precipitation inputs for hydrological models to quantify the consequences of climate change on surface water resources.

Acknowledgements Thanks are due to the DGRE of Tunisia for providing monthly precipitation data. Also, this research project has been supported by a travel fund from the Institut de Recherche pour le D´eveloppement (IRD, France) and a travel fund from the NPT 153 - B´enin project. The financial supports provided are gratefully acknowledged.

References B´ardossy A, Pegram G. 2011. Downscaling precipitation using regional climate models and circulation patterns toward hydrology. Water Resources Research 47: W04505. DOI: 10.1029/2010WR009689. Chebbi A, Kebaili Bargaoui Z, Da Conceic¸ao CM. 2011. Optimal extension of rain gauge monitoring network for rainfall intensity and erosivity index interpolation. Journal of Hydrologic Engineering 16(8): 665–676. Christensen JH, Christensen OB, Lopez P, van Meijgaard E, Botzet M. 1996. The HIRHAM4 regional atmospheric climate model. Scientific Report DMI, Copenhagen, Report 96-4. Available at DMI, Lyngbyvej 100, DK-2100 Copenhagen Ø, Denmark. Darling DA. 1957. The Kolmogorov–Smirnov, Cram´er–von Mises tests. Annals of Mathematical Statistics 28: 823–838. D´equ´e M, Somot S. 2010. Weighted frequency distributions express modelling uncertainties in the ENSEMBLES regional climate experiments. Climate Research 44: 195–209. D´equ´e M, Rowell DP, L¨uthi D, Giorgi F, Christensen JH, Rockel B, Jacob D, Kjellstr¨om E, de Castro M, van den Hurk B. 2007. An intercomparison of regional climate simulations for Europe: assessing uncertainties in model projections. Climatic Change 81: 53–70. DOI: 10.1007/s10584-006-9228-x. D´equ´e M, Somot S, Sanchez-Gomez E, Goodess CM, Jacob D, Lenderink G, Christensen OB. 2012. The spread amongst ENSEMBLES regional scenarios: regional climate models, driving general circulation Models and interannual variability. Climate Dynamics 38: 951–964. DOI: 10.1007/s00382-011-1053-x. Ferro CAT, Hannachi A, Stephenson DB. 2005. Simple nonparametric techniques for exploring changing probability distributions of weather. Journal of Climate 18: 4344–4354.  2013 Royal Meteorological Society

247

Fowler HJ, Blenkinsop S, Tebaldi C. 2007. Linking climate change modelling to impacts studies: recent advances in downscaling techniques for hydrological modelling. International Journal of Climatology 27: 1547–1578. Garc´ıa-Ruiz JM, L´opez-Moreno I, Vicente-Serrano SM, Lasanta–Mart´ınez T, Beguer´ıa S. 2011. Mediterranean water resources in a global change scenario. Earth-Science Reviews 105(3–4): 121–139. Gaussen H, Vernet A. 1951. Carte des pr´ecipitations de Tunisie, feuille 2. Cartes en couleur a` 1/500000 . Facult´e des Sciences, Service carte internationale: Toulouse. Giorgi F, Lionello P. 2008. Climate change projections for the Mediterranean region. Global and Planetary Change 63: 90–104. Giorgi F, Mearns LO. 1999. Introduction to special section: regional climate modeling revisited. Journal of Geophysical Research: Atmospheres 104: 6335–6352. Giraldo Osorio JD, Garcia Galiano SG. 2013. Assessing uncertainties in the building of ensemble RCMs over Spain based on dry spell lengths probability density functions. Climate Dynamics 40: 1271–1290. DOI: 10.1007/s00382-012-1381-5. Glatzer E, Muller WG. 2004. Residual diagnostics for variogram fitting. Computers and Geosciences 30: 859–866. Goovaerts P. 2000. Geostatistical approaches for incorporating elevation into the spatial interpolation of rainfall. Journal of Hydrology 228: 113–129. Halenka T, Kalvova J, Chladova Z, Demeterova A, Zemankova K, Belda M. 2006. On the capability of RegCM to capture extremes in longterm regional climate simulation - comparison with the observations for Czech Republic. Theoretical and Applied Climatology 86: 125–145. Hannachi A. 2006. Quantifying changes and their uncertainties in probability distribution of climate variables using robust statistics. Climate Dynamics 27: 301–317. Haylock MR, Hofstra N, Klein Tank AMG, Klok EJ, Jones PD, New M. 2008. A European daily high resolution gridded data set of surface temperature and precipitation for 1950–2006. Journal of Geophysical Research 113. DOI: 10.1029/2008JD010201. Heinrich G, Gobiet A. 2011. The future of dry and wet spells in Europe: a comprehensive study based on the ENSEMBLES regional climate models. International Journal of Climatology. DOI: 10.1002/joc.2421. Herrera S, Fita L, Fern´andez J, Guti´errez JM. 2010. Evaluation of the mean and extreme precipitation regimes from the ENSEMBLES regional climate multimodel simulations over Spain. Journal of Geophysical Research 115. DOI: 10.1029/2010JD013936. Jacob D, Barring L, Christensen OB, Christensen JH, de Castro M, Deque M, Giorgi F, Hagemann S, Hirschi M, Jones R, Kjellstrom E, Lenderink G, Rockel B, van Ulden AP, van den Hurk BJJM. 2007. An inter-comparison of regional climate models for Europe: model performance in present-day climate. Climatic Change 81: 31–52. DOI: 10.1007/s10584-006-9213-4. Kirkpatrick S Jr, Gelatt CD, Vecchi M. 1983. Optimization by simulated annealing. Science 220: 671–680. Kjellstr¨om E, B¨arring L, Gollvik S, Hansson U, Jones C, Samuelsson P, Rummukainen M, Ullersig A, Willen U, Wyser K. 2005. A 140year simulation of European climate with the new version of the Rossby Centre regional atmospheric climate model (RCA3). Reports meteorology and climatology, 108, SMHI, SE-60176 Norrkoping, Sweden, pp. 54. Kjellstr¨om E, Boberg F, Castro M, Christensen HJ, Nikulin G, S´anchez E. 2010. Daily and monthly temperature and precipitation statistics as performance indicators for regional climate models. Climate Research 44: 135–150. Lebel T, Laborde JP. 1988. A geostatistical approach for areal rainfall statistics. Stochastic Hydrology and Hydraulics 2: 245–261. Lebel T, Bastin G, Obled C, Creutin JD. 1987. On the accuracy of areal rainfall estimation. Water Resources Research 23: 2123–2134. van der Linden P, Mitchell JFB. 2009. ENSEMBLES: Climate Change and Its Impacts: Summary of Research and Results from the ENSEMBLES Project. Exeter: Met Office Hadley Centre. Ly S, Charles C, Degr´e A. 2011. Geostatistical interpolation of daily rainfall at catchment scale: the use of several variogram models in the Ourthe and Ambleve catchments, Belgium. Hydrology and Earth System Sciences 15: 2259–2274. Mann HB. 1945. Nonparametric tests against trend. Econometrica 3: 245–259. May W. 2007. The simulation of the variability and extremes of daily precipitation over Europe by the HIRHAM regional climate model. Global and Planetary Change 57: 59–82. Int. J. Climatol. 34: 235–248 (2014)

248

Z. BARGAOUI et al.

Nasr Z, Almohammed H, Gafrej Lahache R, Maag C, King L. 2008. Drought Modelling under climate change in Tunisia during the 2020 and 2050 periods. Option M´editerran´eennes, S´eries A 80: 365–369. Paeth H, Born K, Girmes R, Podzun R, Jacob D. 2009. Regional climate change in tropical and Northern Africa due to greenhouse forcing and land use changes. Journal of Climate 22: 114–132. Patricola CM, Cook KH. 2010. Northern African climate at the end of the twenty-first century: an integrated application of regional and global climate models. Climate Dynamics 35: 193–212. Perkins SE, Pitman AJ, Holbrook NJ, McAneney J. 2007. Evaluation of the AR4 climate models’ simulated daily maximum temperature, minimum temperature, and precipitation over Australia using probability density functions. Journal of Climate 20: 4356–4376. Piani C, Haerter JO, Coppola E. 2010. Statistical bias correction for daily precipitation in regional climate models over Europe. Theoretical and Applied Climatology 99: 187–192. Radu R, D´equ´e M, Somot S. 2008. Spectral nudging in a spectral regional climate model. Tellus A 60: 898–910. Schmidli J, Goodess CM, Frei C, Haylock MR, Hundecha Y, Ribalaygua J, Schmith T. 2007. Statistical and dynamical downscaling

 2013 Royal Meteorological Society

of precipitation: an evaluation and comparison of scenarios for the European Alps. Journal of Geophysical Research 112: D04105. DOI: 10.1029/2005JD007026. Tebaldi C, Knutti R. 2007. The use of the multi-model ensemble in probabilistic climate projections. Philosophical Transactions of the Royal Society A 365: 2053–2075. Teutschbein C, Seibert J. 2010. Regional climate models for hydrological impact studies at the catchment scale : A review of recent modeling strategies. Geography Compass 4/7: 834–860. DOI: 10.111/j.1749-8198.2010.0037.x. Tramblay Y, Badi W, Driouech F, El AS, Neppel L, Servat E. 2012. Climate change impacts on extreme precipitation in Morocco. Global and Planetary Change 82–83: 104–114. Verworn A, Haberlandt U. 2011. Spatial interpolation of hourly rainfall effect of additional information, variogram inference and storm properties. Hydrology and Earth System Sciences 15: 569–584. DOI: 10.5194/hess-15-569-2011. Wackernagel H. 1995. Multivariate Geostatistics. Springer-Verlag: Berlin, Heidelberg, Germany. Wald A, Wolfowitz J. 1943. An exact test for randomness in serial correlation. The Annals of Mathematical Statistics 14: 378–388.

Int. J. Climatol. 34: 235–248 (2014)